Projects
Projects
2025/26
Research Grants Council – Faculty Development Scheme
Project Title:
Design of A Carbon-Neutral Cold Chain E-Fulfilment Model
Principal Investigator:
Dr. WU Chun-ho Jack
Co-Investigator (HSUHK):
Dr. LAM Hoi-yan Cathy
Project Period: 2026-01-01 to 2027-12-31
Funding Amount: $385,189
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/B14/25
Abstract:
This research explores balancing cost, productivity, product quality, and carbon emissions in cold chain e-fulfilment, addressing the conflict between maintaining quality and achieving carbon neutrality amid growing e-commerce demands. In this research, a carbon-neutralised e-fulfilment model for the cold chain is proposed with the following components. Firstly, the temporal inventory and pick face replenishment decisions for managing cold storage facilities are modelled using deep actor-critic reinforcement learning to eliminate temperature fluctuations caused by frequent facility access, resulting in energy-efficient warehousing operations. Secondly, a novel IoT-based hybrid thermal packaging design for multi-temperature perishable products is designed and validated through full factorial experiments, where biodegradable and recyclable materials, such as expanded polypropylene (EPP), and thermoelectric cooling technologies are considered. Subsequently, the relationship between specific thermal packaging design and maximal quality assurance time for transportation can be determined. Thirdly, a combinatorial optimisation model is formulated to determine the optimal order of packing materials and delivery routes, considering green aspects, thereby ensuring the wise utilisation of packaging materials, fuel, and electricity. Lastly, the carbon footprint of refined e-fulfilment processes is calculated and standardised for benchmarking, while its carbon emissions can be analysed towards carbon neutrality.
Project Title:
A Multi-Agent Reinforcement Learning Framework for Optimizing E-Fulfillment Processes
Principal Investigator:
Dr HO To-sum George
Project Period: 2026-01-01 to 2027-12-31
Funding Amount: $629,035
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/B10/25
Abstract:
E-commerce is experiencing substantial growth worldwide, yet most businesses in Hong Kong are small and medium enterprises (SMEs) and rely on rented storage spaces, making large investments in automation challenging. Artificial intelligence (AI) provides a practical alternative by optimizing resources through data-driven solutions. Currently, inefficient coordination among key e-fulfillment processes—stock receipt, replenishment, and order picking—often leads to delays and inflexibility, particularly when urgent orders emerge. Conventional optimization techniques are limited in adapting to these dynamic conditions, frequently requiring costly or ad-hoc adjustments.
This research proposes the use of a multi-agent reinforcement learning (MARL) model, which autonomously coordinates these fulfillment stages, identifies high-demand products, generates optimized picking routes, and allocates tasks to workers based on real-time priorities. The model’s flexibility allows it to be applied across various warehouse settings, making it particularly accessible for SMEs. By offering real-time, data-driven decision support, the proposed MARL model is positioned to enhance fulfillment efficiency and responsiveness in competitive logistics environments.
Project Title:
Towards Robust and Cost-Efficient Vertical Federated Learning
Principal Investigator:
Dr. ZHANG Chen
Co-Investigator (HSUHK):
Prof LIU Hai
Project Period: 2026-01-01 to 2028-12-31
Funding Amount: $1,101,800
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/E02/25
Abstract:
Vertical federated learning (VFL) is a privacy-preserving machine learning framework that enables multiple parties with different feature spaces to collaboratively train a model without sharing their raw data. For instance, an insurance company could collaborate with several banks to develop a fraud detection model. This collaboration enables the model to benefit from the comprehensive knowledge of all parties involved while keeping each party’s data localized.
To achieve efficient VFL, the initial task is private entity alignment, which requires pinpointing the common sample ID intersection among all parties without revealing any additional information. A practical private entity alignment protocol should be both robust and cost-efficient. However, existing protocols struggle with significant overhead challenges, especially when the number of participants is large. Moreover, current designs do not adequately accommodate participant dropouts, an issue frequently encountered in VFL due to geographical diversity and network heterogeneity among participants. Thus, it is crucial to bridge the gap by designing a cost-efficient multiple-party private entity alignment protocol for VFL systems that is tolerant of participant dropout during the alignment process.
In addition, the distributed learning architecture of VFL presents significant challenges to its robustness. If attackers control some participants and cause them to upload poisoned model updates during training, the performance of the collaboratively trained model can significantly deteriorate. Unlike horizontal federated learning, where participants transmit model updates, VFL involves sharing local model outputs. These outputs are abstract and reveal little about the specifics of the local models, making it difficult to detect malicious activity. Existing defenses are designed to increase tolerance to abnormal data and protect against data poisoning attacks. However, these designs are inadequate against sophisticated attackers who can directly manipulate the local model training process, known as model poisoning attacks, instead of merely altering the training datasets. There is an urgent need to design more robust defense schemes that can effectively detect malicious participants and defend against untargeted model poisoning attacks.
The principal goal of this project is to fill the above-mentioned research gaps and design a robust and cost-efficient VFL framework. There are three main tasks in this project: 1) Design a robust and cost-efficient multi-party private entity alignment protocol that enables multiple parties to efficiently compute data intersections without revealing dividual samples, while also accommodating participant dropouts during the alignment process. 2) Design defense schemes to protect the proposed VFL framework against untargeted model poisoning attacks. The proposed design should be able to efficiently detect malicious parties during the training process. 3) Develop a prototype system of the proposed robust and cost-efficient VFL framework to facilitate performance evaluations. Evaluate the proposed privacy-preserving entity alignment protocol across diverse datasets and party scales, and assess the robustness of the proposed defense mechanisms against various poisoning attacks in different network environments.
Project Title:
Graph Neural Networks Assisted Dynamic Task Offloading in Internet of Vehicles
Principal Investigator:
Dr. HOU Yun
Project Period: 2026-01-01 to 2028-12-31
Funding Amount: $1,489,982
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/E07/25
Abstract:
The Internet of Vehicles (IoV) is transforming transportation by enabling vehicles to communicate with each other and with nearby roadside units (RSUs). This communication can enhance safety and efficiency on the road. However, because vehicles are constantly moving and their connections are always changing, deciding how to best share computing tasks among them is quite challenging. This project stands out by exploring the unique potential of sharing computational power among vehicles, rather than solely offloading tasks to roadside infrastructures. This approach makes decision-making even more challenging, as the contact window between two vehicles is often unpredictable due to their high mobility. Our main goals include creating a system that can predict how long two vehicles will stay connected. This information is crucial for helping vehicles decide when and where to offload tasks. We will also use Graph Neural Networks (GNNs) to analyse the network of vehicles, which will help us understand how to best allocate communication resources and timings for tasks that need to be shared. The insights gained from these analyses will improve the decision-making process for the Deep Reinforcement Learning (DRL) based task offloading schemes.
The project is organized into three key objectives. First, we will develop a prediction framework for estimating how long vehicles will be in contact with RSUs and other vehicles, which is essential for making informed offloading decisions. Second, we will create a GNN model to learn the relationships between network entities and efficiently manages communication resources between them. Finally, we will build a flexible DRL framework that can adapt to the ever-changing network conditions, allowing vehicles to share tasks not only with RSUs but also with each other.
Research Grants Council – Inter-institutional Development Scheme
Project Title:
Data Science Frontier: Expanding Horizons through Mathematical and Statistical Innovations
Principal Investigator:
Dr LAM Shu-yan
Co-Investigators:
Prof CHOY Siu-kai, Dr SIU Chi-chung, Dr CHEN Yongzhao,
Funding Scheme/Source: Research Grants Council – Inter-institutional Development Scheme
Project Period: 2026-01-01 to 2026-12-31 (On-going)
Funding Amount: $678,600
Other Collaborating Parties: University of Hertfordshire
Project Reference No.: UGC/IIDS14/P01/25
Abstract:
The field of data science is undergoing rapid expansion, profoundly reshaping various aspects of our society. Technological advancements have granted us access to vast amounts of data in diverse formats, presenting us with both challenges and opportunities. One domain where data science has made a significant impact is e-commerce, where the analysis of streaming data provides real-time insights into customer behavior, enabling personalized recommendations and enhancing shopping experiences. Similarly, in the healthcare sector, data analysis of medical imaging and sensor readings yields valuable information for researchers, leading to innovative approaches in the treatment of chronic diseases.
At the heart of data science lie mathematics and statistics, serving as foundational pillars for the development of powerful tools such as deep learning models, a widely employed form of artificial intelligence. Essential to solving data science problems is the transformation of real-world challenges into mathematical and statistical formulations. A comprehensive understanding of the underlying mathematics and statistics behind these tools, along with their adaptation to specific scenarios, applications, and problems, not only facilitates the discovery of improved solutions but also empowers researchers to develop novel data science tools. However, the current focus in data science education and practice often leans toward computer programming and software utilization, with insufficient attention given to the critical role of mathematics and statistics.
Therefore, the primary objective of this project is to organize a symposium that places a strong emphasis on the application of mathematics and statistics in the development of advanced data science tools for tackling complex problems. The symposium will bring together distinguished speakers from local organizations and renowned universities worldwide, who will share the latest trends and developments in four key thematic areas:
1. Machine Learning and AI: Exploring how these technologies can enhance operational efficiency and productivity, ultimately enabling automation and liberating human workers.
2. Economics and Finance: Investigating how data science empowers policymakers to make well-informed decisions, leading to more effective risk management, investment decisions, as well as fiscal and monetary policies.
3. Biostatistics and Health: Delving into the ways data science can uncover disease patterns, identify innovative treatments, and contribute to advancements in medical knowledge.
4. Industrial Data Science: Highlighting the pivotal role of data science in optimizing production processes, enhancing quality control, and improving resource allocation within industrial settings.
Each thematic session will conclude with an open discussion aimed at fostering meaningful interaction between speakers and participants, with a focus on effective research strategies. Many of our speakers are also associate editors or editors of academic journals, offering invaluable insights into identifying promising research directions, building networks, and publishing impactful work. This creates an important platform for all participants to exchange ideas and share their unique perspectives, enriching the academic and industry connections essential for the development of young researchers.
The symposium will be open to the public, including students and faculty members from local universities funded by the University Grants Committee (UGC), as well as self-financing degree-awarding institutions. By ensuring accessibility, the symposium strives to promote data science research to prospective students and expand the research interests of local scholars.
To summarize, this project aims to address the existing gap in highlighting the critical role of mathematics and statistics in data science. Through the symposium, it seeks to facilitate knowledge sharing, foster collaboration, and promote the development of advanced data science tools, ultimately benefiting academia and society.
2024/25
Research Grants Council – Faculty Development Scheme
Project Title:
Robust Federated Learning Empowered Microservice Migration in Mobile Edge Computing
Principal Investigator:
Dr ZHANG Chen
Co-Investigator (HSUHK):
Prof LIU Hai
Project Period: 2025-01-01 to 2027-12-31
Funding Amount: $1,054,650
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/E03/24
Abstract:
With the rapid development of 5G network technology, mobile edge computing (MEC) has emerged as a promising paradigm to support latency-sensitive applications. Service migration is critical in MEC to ensure the seamless provision of services as users move. In many applications such as autonomous driving, user location often changes during service provision, which may cause service interruption if the user is far away from the server running the requested service. Ensuring that such services can follow users’ movements is crucial for maintaining low response latency and providing a seamless user experience. While service migration has the potential to bring significant benefits, it also introduces additional overheads. An inappropriate service migration scheme may even increase service latency.
Reinforcement learning is a promising method for making service migration decisions owing to its high adaptability to dynamic environments. Most reinforcement learning-based designs focus on the migration of independent services. However, in real-world scenarios, the task requested by a user usually contains multiple subtasks with dependencies, and these subtasks follow a clear directed acyclic graph (DAG) structure. Therefore, dependencies between microservices need to be considered in migration decisions. Existing DAG-based service migration designs assume that all microservices corresponding to a DAG task will be migrated to a single edge server, which may lead to high service latency when the target edge server lacks sufficient resources to accommodate these microservices. It is important to allow microservices to be migrated to different edge servers to distribute workloads.
Additionally, the data used to train the service migration model contains sensitive user information (e.g., user trajectory data). Adopting a centralized training architecture will result in high transmission overheads and risks of privacy leakage. Federated learning enables edge servers to collaboratively train the service migration model without sharing local data. However, the distributed learning architecture of federated learning poses challenges to its robustness. If some edge servers are controlled by attackers and upload poisoned model updates during training, the performance of the service migration model will be significantly degraded. Although some defense schemes have been proposed to sift through model updates to identify outliers, they fail to consider the weight importance factor, resulting in significant degradation in defense performance under model poisoning attacks enhanced by weight importance.
The principal goal of this project is to fill the above-mentioned research gaps and design a robust federated learning based microservice migration in MEC. There are three main tasks in this project: 1) Design a federated reinforcement learning-based microservice migration framework for MEC to minimize the service latency of user-offloaded DAG tasks. The framework shall support the migration of microservices of a DAG task to different edge servers. 2) Design defense schemes to protect the federated learning-based microservice migration against model poisoning attacks. The proposed schemes should have strong defense under model poisoning attacks enhanced by weight importance. 3) Develop a prototype system of robust federated learning-based microservice migration for performance evaluation and fine tuning of the proposed designs under different datasets and network environments.
2023/24
Research Grants Council – Faculty Development Scheme
Project Title:
Probabilistic Classification Error Modeling for Distance Metric Learning and Deep Metric Learning Theory and Applications
Principal Investigator:
Dr. LAM Benson Shu-yan
Co-Investigator (HSUHK):
Prof CHOY Siu-kai, Dr Carisa YU
Project Period: 2024-01-01 to 2026-12-31
Funding Amount: $828,859
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/P04/23
Abstract:
Feature extraction plays a crucial role in classification problems in the field of machine learning, especially big data problems. Its purpose is to extract discriminant information from data to enhance the accuracy of classification methods. It has numerous applications in various fields, including language detection, sentiment analysis, fraud detection, image recognition, disease classification, etc. One example is image recognition in computer vision. The problem is to recognize the types of objects, such as basketball, on the basis of a set of given images. This is a challenging problem because images of an object can be taken from different views. Furthermore, the images may have noise and be taken against various backgrounds. However, the noise and the backgrounds do not carry much discriminant information and can confuse many classification methods.
To cope with the noise and background problems, metric learning methods have been developed. The purpose of metric learning is to extract discriminant features from data by optimizing a mathematical objective function. The function converts the data features to a feature space so that samples from the same classes move closer to each other while samples from different classes move further away. Metric learning methods can be broadly categorized into distance metric learning and deep metric learning. Distance metric learning extracts the features from the data using the linear projection method, while deep metric learning extracts the features using multi-layer neural networks, such as convolutional neural networks. However, many objective functions were designed by intuition. It is hard to find the connections between these objective functions and the theoretical error rate. The theoretical error rate is a fundamental concept in machine learning that quantifies the best possible accuracy any classifier can achieve on a fixed probability distribution. Theoretically, extracting the features that can minimize the theoretical error achieves good performance. Moreover, the designs of intra-class and inter-class information are based on some heuristic rules. Intra-class and inter-class information refers to information for the samples drawn from the same classes and from different classes, respectively. However, some designs only include one type (such as intra-class) of information and ignore the other type (such as inter-class).
In this project, we aim to develop a new type of metric learning method based on the theoretical error rate. This new method can be applied to both distance metric learning and deep metric learning problems. The class information can be modelled by probability density functions. Based on the nature of probabilities, both intra-class and inter-class information are implicitly formulated. No heuristic rules are applied. We will also develop new optimization strategies. For distance metric learning, the linear projection can be found by a two-stage optimization method. The solutions obtained are unique. In other words, unlike the current body of work, which uses optimization methods that are sensitive to initial guesses, the methods we propose are not sensitive to the initial guess set. Different initial guesses lead to different solutions. For deep metric learning, we aim to develop a fast approach to find the deep learning parameters. Currently, stochastic gradient descent methods are mainly used to find a solution for deep learning models. The essential idea is to select a min-batch of samples from the data and update the parameters of the layers using gradient descent methods. However, if the min-batch of samples is drawn uniformly from the data, the convergence speed can be slow (take several days) because not all samples carry the same discriminant information. Given a big data problem, the sample size can be over a million. It is hard to draw a min-batch with lots of discriminant information. To alleviate this problem, we aim to develop a new sampling strategy based on probability simulation, namely importance sampling. The idea of importance sampling is to select samples on the basis of a proposal distribution that is like the target distribution. In this project, we will develop a new probability function that has a similar form to the deep learning function. With this new function, we will be able to draw more effective samples and enhance the convergence speed in the training of deep learning models.
Project Title:
Optimizing Model Freshness for Federated Transfer Learning with Extreme Cases in Autonomous Driving Networks
Principal Investigator:
Dr HOU Aileen Yun
Project Period: 2024-01-01 to 2026-12-31
Funding Amount: $1,384,538
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: University of Manchester, EduHK, Wuyi University
Project Reference No.: UGC/FDS14/E02/23
Abstract:
This project aims to tackle the training data shortage problem and optimize the freshness of the machine learning models for extreme cases in the C-V2X enabled autonomous driving networks. Autonomous vehicles rely on artificial intelligence (AI), visual computing, radar, monitoring equipment, and positioning system to work together so that the system can automatically and safely operate the automobile without active operation. However, an unsafe ‘decision’ of the autonomous driving system is likely to endanger human life and cause huge economic losses. To maintain the high safety requirement of autonomous driving, the accuracy and freshness of the AI networks used for perception and decision-making play a decisive role. However, decision reinforcement learning requires a lot of interaction with the real world to learn strategies and deal with various situations. This makes it difficult for reinforcement learning to learn scenarios rarely seen in reality, e.g., extreme cases. Therefore, in this project, we aim to build a Federated Transfer Learning Framework at roadside units with extreme cases generated and pre-trained in a Digital Twin with optimization for model freshness. To achieve this goal, we will (1) develop a digital twin for extreme case generation in autonomous driving; (2) build a federated transfer learning framework for autonomous driving to improve road safety using the extreme cases emulated in the digital twin; and (3) optimize the federated transfer learning framework in C-V2X networks depending on the real-time network status to dynamically decide how broad the model should be federated and how mature the model should be trained at edge nodes towards the optimal freshness of transferred models.
The research outcomes from this project will lay the foundation for the future optimizations of collaborative deep learning in vehicular edge networks toward the new metric “model freshness”. Given the rapid evolvement in high-level autonomous driving, the research outcomes are expected to open a new dimension in providing all-round cooperative sensing and decision planning to assist vehicles safely navigate through complex and extreme scenarios while ensuring safety in all aspects.
Project Title:
Sustainable and Resilient Automated E-Fulfilment Operations in the Era of Industrial 5.0
Principal Investigator:
Dr. HO George To-sum
Co-Investigator (HSUHK):
Dr. Cathy LAM
Project Period: 2024-01-01 to 2025-12-31 (Completed)
Funding Amount: $479,650
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: Cardiff University, University of Liverpool
Project Reference No.: UGC/FDS14/E03/23
Abstract:
The rapidly growing e-commerce sector has significantly transformed customer behaviour worldwide. In the European Union, the gross value of retail sales in April 2020 diminished by 17.9%, whereas sales via e-commerce orders increased by 30% (OECD, 2020). The burgeoning of e-commerce purchasing has highlighted the growing importance of e-commerce logistics. In the era of Industry 4.0, automated e-fulfilment centres are adopting technologies such as artificial intelligence (AI), Internet of things (IoT), and automated guided vehicles (AGVs) to enhance the capability and reliability of industries. In the modern business environment, apart from the use of technology, a more value-driven movement is a concern that drives the revolution of Industry 5.0.
Studies on Industry 5.0 have focused on sustainability and resilience (Ghobakhloo et al., 2022; Azadeh et al., 2019). Moreover, concerns about environmental, social, and governance (ESG) performance are growing across business research and industries around the world, including Hong Kong (Linnenluecke, M. K, 2022; The Standard, 2022). Considering the boost in the efforts to mitigate environmental pollution, China has committed to achieving carbon neutrality before 2060 and curbing carbon dioxide (CO2) emissions before 2030, as China is one of the leading contributors of CO2 emissions–accounting for 28% of global CO2 emissions in 2018. However, the lack of integral consideration of interactions between different operation processes in automated e-fulfilment centres would affect operational efficiency (Azadeh et al., 2019). Despite the growth of technology, which increases the operational efficiency of automated e-fulfilment centres, robustness and security remain important for the industry to cope with a changing environment within the supply chain and other factors that might affect everyday processes (van Geest, M., 2022). To meet the concept of Industry 5.0, automated e-fulfilment centres need to integrate additional facilities to reduce non-renewable energy consumption and develop a resilient operation model. However, the implementation of Industry 5.0 is challenging owing to the absence of a practical framework for Industry 5.0. As such, this research gap in the practical implementation framework is a major barrier to the development of Industry 5.0.
This project aims to design a digital twin (DT) to jointly optimize sustainability and resilience in automated e-fulfilment centres. Specifically, solar power is employed as a source of renewable energy for an automated e-fulfilment centre. The proposed model contributes to infrastructure design recommendations by developing a DT that transforms the actual automated e-fulfilment centres to editable virtual automated e-fulfilment centres. Based on the DT, a simulated study is performed to determine the optimal renewable energy model to achieve self-reliance. With the optimal renewable energy model, operational resilience can be achieved by considering the interactions between different operational processes and renewable energy consumption. From the perspective of automated e-fulfilment centres, the proposed model presents a framework to transition from non-renewable energy to renewable energy. In addition, automated e-fulfilment centres can handle the rapidly changing supply chain environment by accomplishing resilient operations. With the aid of the proposed model, sustainable and resilient automated e-fulfilment centres can be realized, resulting in better economic competitiveness and environmental benefits.
Project Title:
Generating novel customer needs for new product development
Principal Investigator:
Prof LIU Hai
Project Period: 2024-01-01 to 2026-06-30
Funding Amount: $1,128,250
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: International Digital Economy Academy
Project Reference No.:UGC/FDS14/E08/23
Abstract:
Feature extraction plays a crucial role in classification problems in the field of machine learning, especially big data problems. Its purpose is to extract discriminant information from data to enhance the accuracy of classification methods. It has numerous applications in various fields, including language detection, sentiment analysis, fraud detection, image recognition, disease classification, etc. One example is image recognition in computer vision. The problem is to recognize the types of objects, such as basketball, on the basis of a set of given images. This is a challenging problem because images of an object can be taken from different views. Furthermore, the images may have noise and be taken against various backgrounds. However, the noise and the backgrounds do not carry much discriminant information and can confuse many classification methods.
To cope with the noise and background problems, metric learning methods have been developed. The purpose of metric learning is to extract discriminant features from data by optimizing a mathematical objective function. The function converts the data features to a feature space so that samples from the same classes move closer to each other while samples from different classes move further away. Metric learning methods can be broadly categorized into distance metric learning and deep metric learning. Distance metric learning extracts the features from the data using the linear projection method, while deep metric learning extracts the features using multi-layer neural networks, such as convolutional neural networks. However, many objective functions were designed by intuition. It is hard to find the connections between these objective functions and the theoretical error rate. The theoretical error rate is a fundamental concept in machine learning that quantifies the best possible accuracy any classifier can achieve on a fixed probability distribution. Theoretically, extracting the features that can minimize the theoretical error achieves good performance. Moreover, the designs of intra-class and inter-class information are based on some heuristic rules. Intra-class and inter-class information refers to information for the samples drawn from the same classes and from different classes, respectively. However, some designs only include one type (such as intra-class) of information and ignore the other type (such as inter-class).
In this project, we aim to develop a new type of metric learning method based on the theoretical error rate. This new method can be applied to both distance metric learning and deep metric learning problems. The class information can be modelled by probability density functions. Based on the nature of probabilities, both intra-class and inter-class information are implicitly formulated. No heuristic rules are applied. We will also develop new optimization strategies. For distance metric learning, the linear projection can be found by a two-stage optimization method. The solutions obtained are unique. In other words, unlike the current body of work, which uses optimization methods that are sensitive to initial guesses, the methods we propose are not sensitive to the initial guess set. Different initial guesses lead to different solutions. For deep metric learning, we aim to develop a fast approach to find the deep learning parameters. Currently, stochastic gradient descent methods are mainly used to find a solution for deep learning models. The essential idea is to select a min-batch of samples from the data and update the parameters of the layers using gradient descent methods. However, if the min-batch of samples is drawn uniformly from the data, the convergence speed can be slow (take several days) because not all samples carry the same discriminant information. Given a big data problem, the sample size can be over a million. It is hard to draw a min-batch with lots of discriminant information. To alleviate this problem, we aim to develop a new sampling strategy based on probability simulation, namely importance sampling. The idea of importance sampling is to select samples on the basis of a proposal distribution that is like the target distribution. In this project, we will develop a new probability function that has a similar form to the deep learning function. With this new function, we will be able to draw more effective samples and enhance the convergence speed in the training of deep learning models.
Research Matching Grant Scheme (RMGS)
Project Title:
AI-driven One-stop Platform for enhancing Inventory Management through Big Data Analysis
Principal Investigator:
Dr Valerie TANG
Principal Co-Investigators (HSUHK):
Dr George HO, CHOW Wing Ho, Lika LEE
Project Period: 2024-06-01 to 2027-05-31
Funding Amount: $1,500,000 (Brilliant Business Centre Limited; in-kind) and $750,000 (RMGS)
Funding Scheme/Source: Brilliant Business Centre Limited and Research Matching Grant Scheme (RMGS)
Abstract:
Research objectives:
1) To embed an AI-driven inventory management platform in an AI-based data analysis model to collect sales data for Big data analysis.
2) To implement association rules and clustering techniques to identify patterns and relationships within extensive datasets to help optimize inventory procedures and categorize products based on their shared attributes.
3) To provide real-time insights into inventory level through an AI-driven inventory management platform to enhance the efficiency of inventory management and meet customer fulfilment needs.
2022/23
Research Grants Council – Faculty Development Scheme
Project Title:
MetaConfigurator: A Resource-Effective Method to Develop Needs-Based Configurators for Product Customisation
Principal Investigator:
Prof MO Yiu-wing, Daniel
Project Period: 2023-01-01 to 2026-06-30
Funding Amount: $1,261,600
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: Southern University of Science and Technology, cyLEDGE Media Group
Project Reference No.:UGC/FDS14/E05/22
Abstract:
Product configurators are considered to be critical toolkits for customised product design and have been successfully implanted in various companies, such as Dell, BMW and Nike. Among the various versions of product configurators, needs-based systems are particularly useful to map customer needs in natural language directly to the targeted product configurations. Although needs-based configurators are more user-friendly and applicable in B2C environment, they are resource intensive to be implemented. A large amount of product relevant data such as customer needs are required to be collected, annotated, and processed to train the configuration model. Such approach is, moreover, limited to one product family and cannot be generalised to other products or even the updated product family. This restricts its wide application in practice.
This proposal will develop a MetaConfigurator framework to overcome these challenges. We will leverage natural language processing, deep learning and transfer learning techniques to interpret ambiguous and probably ill-defined customer needs sentences and map them to well-defined product configurations. Using the massive amount of product review data from e-commerce website, we firstly extract product related knowledge or features in a product category. Then, the generic knowledge is adapted to a specific product using a relatively small amount of customer needs data to build the product-specific needs-based configurator. Through this pretraining-then-finetuning process, a more user-friendly product configurator can be derived in a more efficient way.
Research Matching Grant Scheme (RMGS)
Project Title:
Integration of Artificial Intelligence and Metaverse in Luxury E-commerce Retail for Enhancing Customers’ Online Experience
Principal Investigator:
Dr. Cathy LAM
Principal Co-Investigators (HSUHK):
Prof CHOY Siu Kai, Dr George HO, Dr Valerie TANG
Project Period: 2023-09-01 to 2027-08-31
Funding Amount: $8,375,583.6 (Chef Digital Limited: in kind) and $4,187,791.80 (RMGS)
Funding Scheme/Source: Chef Digital Limited and Research Matching Grant Scheme (RMGS)
Abstract:
Luxury retail market is specializing in selling high-end goods that are characterized by high quality, high prices and a high degree of exclusivity such as women’s designer clothes and handbags. During the Covid-19 pandemic, the global luxury retail market has suffered a decline of 20% to 22% in sales of personal luxury goods in 2020 (D’Arpizio et al., 2021). In light of this, retailers have been pushed to conduct their e-commerce platforms for stepped up to the challenge of government-imposed restrictions and consumer anxiety over the potential health risk associated with in-store shopping. According to the market research conducted by Global Industry Analysts Inc, the global e-commerce market for luxury goods is projected to reach 369.8 billion United States dollar by 2030, growing at a compound annual growth rate of 5.4% over the analysis period 2022 to 2030. However, purchasing such premium priced luxury goods through e-commerce platforms implies a high level of risk for customers since customers cannot evaluate the products based on sensual touch and tangible experiences (Yu et al., 2018). Moreover, rather than mere e-commerce websites, customers wanted to engage through chats, fan groups and online forums in order to listen to other luxury customers regarding innovative ways they engaged with luxury brands amidst the pandemic (Kumar, 2023). To enhance the customer satisfaction, metaverse would be adopted to facilitates goods’ trials and enable consumer interaction through creating interactive virtual shopping environment.
Project Title:
Digital Transformation in Warehouse Management using MES, RPA and AI
Principal Investigator:
Dr. Valerie TANG
Co-Investigators:
Prof Daniel MO, Dr Cathy LAM
Project Period: 2022-09-01 to 2025-08-31
Funding Amount: $HKD 49,00,000 (Httpeace Company Limited: in-kind) and HKD 2,450,000 ( RMGS)
Funding Scheme/Source: Httpeace Company Limited and Research Matching Grant Scheme (RMGS)
Abstract:
In this project, an AI-based intelligent model is proposed to facilitate the warehouse operation and enhance its performance. Two modules namely digital workforce module (DWM) and AI-based analysis module (AIAM) are involved in the proposed model. In the first module (DWM), robotic process automation (RPA) and Manufacturing Execution System (MES) are integrated to automatically collect useful data in the manufacturing process. Then, the AIAM will utilize AI to extract the data collected from DWM as the input to perform analysis of the manufacturing process using fuzzy association rules mining (FARM) for predicting the possible demand of materials and possible outputs of finished goods. As a result, the proposed model would facilitate digital transformation in traditional warehouses enabling relevant data to be digitalized to perform further analysis and visualization. Moreover, digital workforce is adopted for reducing manual mistakes and increasing operational efficiency. The AI-based intelligent model provides a better level of decision support ability for improving the quality of overall warehouse management. By doing so, traditional warehouses can accomplish error free operation and efficiency enhancement under the e-commerce environment with its existing capacity.
Three objectives are defined in this research project which are:
• To study the existing warehouse operations and manufacturing process under the e-commerce environment.
• To integrate RPA with MES for digital transformation in warehouse operations and manufacturing
processes so as to reduce human error and resources wastes.
• To design and develop the architecture of the AI-based intelligent model for warehouse operation analysis
to achieve performance enhancement.
Project Title:
Intelligent Digitalization Platform for Enhancing Business Workflow Management in E-commerce Industry
Principal Investigator:
Dr. George HO
Principal Co-Investigators (HSUHK):
Prof CHOY Siu Kai, Dr Valerie TANG, Dr Benson LAM
Project Period: 2023-03-01 to 2026-02-28
Funding Amount: $3,800,000 (Software donation):1,900,000 (RMGS)
Funding Scheme/Source: Digital (HKSAR) Limited and Research Matching Grant Scheme (RMGS)
Abstract:
E-commerce has become an indispensable part of the global retail business, owing to the pandemic’s long-lasting impact on global consumer behaviours. In 2020, 17.8% of overall sales were e-commerce sales, the share is predicted to rise to 24.5% by 2025, marking a 37.6% increase in just five years (Rita & Ramos, 2022). In China, the national online retail sales reached 6,300.7 billion yuan during the first half of 2022, which is an increase of 3.1 percent year-on-year (National Bureau of Statistics of China, 2022). In the current highly competitive business environment, Logistics industry in a Supply Chain play a more critical role in gaining the competitive advantage (Andiyappillai, 2020). To grasp the opportunity of e-commerce, some of the traditional logistics industries have adopted digital transformation (DT), which performed a transition from brick-and-mortar retail models to online for improving competitiveness and profitability. DT is the integration of digital technologies and new business models into all possible areas that enables major business improvements such as value creation for customers and productivity enhancement (Albukhitan, 2020). In logistics industry, applying digital technology provides significant benefits to operations management such as high optimization potential through big data analytics, device and location independent information gathering through cloud computing, low management complexity through decentralized, and better automation through human-machine interaction (Kayikci, 2018). However, the typically widespread DT technologies are ineligible to deal with the challenges of rapid changing business environment such as increasing complexity of operation management (Sanchis et al., 2019). Hence, logistics industry requires an intelligent and flexible solution to enhance the resilience capacity for overcome the issues of dynamically changing environment.
Low-code development platform (LCDP), a trendy mechanism to facilitate the rapid development of software applications (apps), would be eligible to support and facilitate resilient digital transformation. LCDP are provided on the cloud that enabling the development of fully functional apps via advanced graphical user interfaces and visual abstractions, all with minimal or no procedural code (Sahay et al., 2020). By using LCDP, users with no particular programming background can build their apps for conducting the activities or tasks in business operation, without the help of several developers. Thus, developers can focus on valuable work such as improve business logic of the application rather than dealing with unnecessary details related to setting up of the needed infrastructures, managing data integrity across different environments, or enhancing the robustness of the system. Additionally, adoption of LCDP would enhance flexibility and agility, speed up development time, enable quick response to market demands, reduce bug-fixing, lower deployment effort, and make maintenance easier (Al Alamin et al., 2021), and hence improve resilience capacity of logistics industries.
With LCDP, traditional logistics industries can convert the data storage model from handwriting documents to digital, which prevent significant data loss and create the opportunity of data analysis. To perform data analysis, Artificial intelligence (AI) would be feasible technology that help industries to make fast and smart decisions to reacting to external change, so as to enhance the resilience capacity. AI would be defined as the engineering of intelligent machines with a special focus on intelligent computer programs that can operate without human intervention (Woschank et al., 2020). By employing big data analytics, such as the fuzzy association rule mining (FARM) technique, AI can discover hidden relationships to make decisions, and to reach different conclusions based on the analysis of different situations. Consequently, logistics industry not only can make real-time adjustment of the business operation to ensure operational efficiency in a timely manner but also provides a stable mechanism for long-term quality enhancement.
Project Title:
Mass Customizing Artificial Intelligence-Based Inspection Systems
Principal Investigator:
Prof Daniel MO
Co-Investigators:
Prof CHOY Siu Kai, Dr George HO, Dr Jack WU
Project Period: 2023-09-01 to 2026-08-31
Funding Amount: $3,691,667.2 (SmartMore Corporation Limited: in-kind) and $1,845,833.60 (RMGS)
Funding Scheme/Source: SmartMore Corporation Limited and Research Matching Grant Scheme (RMGS)
Abstract:
In the past century, manufacturing and production industry has undergone digital transformation in operations to achieve efficiency, flexibility and quality. Quality management strategies such as Total Quality Management (TQM), Six Sigma and Lean Operations Management allow many companies to meet the product specifications through quality assurance (QA) and inspection methods with lowering cost. In recent years, digital transformation technologies such as Artificial Intelligence (AI), Cyber- Physical Systems and Digital Twins advance the shop floor production environment with robotics and automation. With the support of artificial intelligence techniques, machine vision-based inspection system was adopted in the QA of finished products, to automatically distinguish between good and defective products by machines and computer devices instead of human eyes and hands [1]. However, most of the works that applied AI to visual quality inspection focused on improving the model performance of inspection accuracy. Very few of them take into consideration the various factors involved in the visual inspection process [2]. In other words, most previous studies focused on the performance of artificial intelligence techniques for a specific application model but did not design quality inspection systems for mass customization. This research initiative aims to mass customize quality inspection systems with artificial intelligence technology.
Project Title:
Towards Adaptive Federated Learning Against Gradient Leakage Attacks
Principal Investigator:
Prof LIU Hai
Co-Investigators (HSUHK):
Dr ZHANG Chen
Project Period: 2023-09-01 to 2026-08-31
Funding Amount: $4,698,000 (LENA Network Limited : in-kind) and $2,349,000 (RMGS)
Funding Scheme/Source: LENA Network Limited and Research Matching Grant Scheme (RMGS)
Abstract:
Machine learning has sparked a revolution in the way we tackle complex problems, ushering in a new era of problem-solving. By training models on large datasets, machine learning algorithms can identify patterns in data and leverage the knowledge to drive innovation and solve intricate challenges in a wide range of fields, such as healthcare (Shailaja et al., 2018) and finance (Ghoddusi et al., 2019). However, in many cases, the data used to train the model includes sensitive information such as personal health records and financial transaction records. Directly collecting and processing such data would cause serious privacy breaches (Liu et al., 2021). Meanwhile, traditional machine learning relies on centralized data storage and processing, which can be costly in terms of both time and resources, especially when dealing with large datasets (Elgabli et al., 2020).
To address these challenges, federated learning has been proposed. It is an emerging distributed machine learning paradigm that has been adopted by many leading companies such as Google (Konečný et al., 2016) and Apple (Paulik et al., 2021). Federated learning enables many edge devices (called clients) to collaboratively train a model without sharing their local private data. In a typical federated learning setting, there is a central server that maintains a global model and coordinates clients for training. In each round of training, the server broadcasts the global model to its randomly selected subset of clients. The selected clients will use their local datasets to fine-tune the model and then return the local model updates to the server. The server aggregates them following an aggregation rule and then updates the global model with the aggregation result. This approach preserves data privacy while enabling the aggregation of knowledge from multiple clients to improve model accuracy and robustness (Bonawitz et al., 2019).
Despite that federated learning is designed to preserve privacy by enabling clients to hold their data locally and train a global model collaboratively by only sharing local model updates with the server, recent research has revealed that the default setting is insufficient to resist gradient leakage attacks (Jin et al., 2021). To launch such attack, the attacker intercepts the local model updates before they are sent to the server, and then uses the intercepted gradient information to reconstruct the training data of clients. This type of attack is particularly concerning in sensitive applications such as finance, where the data used for training may contain a lot of confidential personal information.
To protect federated learning against gradient leakage attacks, a series of defense schemes have been proposed. A line of research adopts the differential privacy technique to protect the privacy of sensitive client data (Wei et al., 2020). The main idea is to add a controlled amount of random noise to the model updates of clients before they are sent to the server for aggregation. In this way, it is difficult for attackers to use the intercepted information with noise to reconstruct the training data. When more noise is added to local model updates, the privacy strength of the system increases, and it is harder for attackers to extract sensitive information from the data. In contrast, it also greatly decreases the accuracy of the global model because the noise added in local model updates makes it harder for the server to aggregate. Therefore, there is a trade-off between data privacy and system efficiency in federated learning, and finding the right balance between the two factors is important. Meanwhile, there is a great need to design an adaptive federated learning framework that can dynamically adjust the defense strength to ensure that the defense strength can just meet the expected privacy requirements.
2021/22
Research Grants Council – Faculty Development Scheme
Project Title:
Riding to Success in Cold Chain Digitalization: A Digital Twin based Closed-Loop Logistics Decision Model
Principal Investigator:
Dr Cathy LAM
Project Period: 2022-01-01 to 2025-06-30
Funding Amount: $769,519
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: The Hong Kong Polytechnic University, Nottingham University Business School
Project Reference No.:UGC/FDS14/E04/21
Abstract:
Due to the rapidly growing demand for reliable and high-quality cold chain logistics following the global pandemic, increasing concern has led to the development of a robust and comprehensive cold chain logistics system, in order to meet designated handling requirements and specifications. Different from general logistics services, time-temperature-sensitive products, such as pharmaceuticals and life sciences products, need to be refrigerated at extremely low temperatures during transportation and distributed within a short time period within the cold chain. Concerning the strict handling requirement of such time-temperature-sensitive products, appropriate cold chain packaging methods, monitoring devices and shipment routes must be specially designed by the Cold Chain Logistics Service Providers. Currently, most of the passive packaging materials and monitoring devices are designed for one-time consumption, such that the cost of reverse logistics in the supply chain network, can be eliminated. However, after receiving the pharmaceuticals and life science products, the downstream supply chain partners would simply dispose of the packaging materials. This results in poor sustainable development regarding cold chain logistics. Consequently, a certain amount of solid waste is created each time goods are received, with a significant environmental impact on society. Hence, this research proposes a digital twin-based closed-loop logistics decision model for handling time-temperature-sensitive shipments. The result of this project will reshape the cold chain in the digital age, benefit society in terms of sustainability and environmental impact and hence contribute to cold chain logistics development in Hong Kong.
Project Title:
Dual Channel Logistics Strategy with the Integration of Crowdsourced Vehicles for Ad Hoc Demand
Principal Investigator:
Prof Daniel MO
Co-Investigator (HSUHK):
Dr George HO
Project Period: 2022-01-01 to 2025-06-30
Funding Amount: $493,450
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: The Chinese University of Hong Kong, University of Oxford
Project Reference No.: UGC/FDS14/E05/21
Abstract:
Managing the dual channel of logistics resources has become more critical than ever, not only to achieve cost savings via enhanced process efficiency in operations, but also to utilise idle resources within and outside operations for social sustainability. With the success of crowdsourcing logistics platforms in recent years, many companies have sought to outsource some of their logistics orders to crowd networks. However, when compared with internal logistics resources, resources in the crowd network involve higher uncertainty, which creates many challenges for companies in determining the allocation of logistics orders to the crowdsourced platform for the fulfilment of ad-hoc demand. There is a lack of a holistic approach for integrating internal logistics resources among various storage facilities with crowdsourced vehicles via decision intelligence systems.
In this research project, we aim to design an integrated decision framework for managing the dual channel of logistics resources through the adoption of decision intelligence systems. With the support of decision intelligence systems, including systems simulation, data-driven models, and geospatial data analytics, the integration of internal and crowd logistics resources is expected to lead logistics operations to the next stage of operations management. The main contributions of this study are therefore focused on the management theory of dual channel logistics resource management. Apart from the management theory, we will collaborate with a company in this project. The collaborated case study would also serve as a guideline for practitioners.
Project Title:
Dynamic Pick Face Replenishment & Pallet Consolidation Model for Landing in the Next E-Fulfilment Normal
Principal Investigator:
Dr George HO
Project Period: 2022-01-01 to 2024-06-30
Funding Amount: $669,975
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: Cardiff University, The Hong Kong Polytechnic University, University of Liverpool
Project Reference No.: UGC/FDS14/E06/21
Abstract:
Due to the outbreak of COVID-19, customer behaviour around the world looks completely different today than it did even one year ago. For example, retail sales via e-commerce channels in both the United States and European Union recorded rapid growth (i.e., 15% and 30% respectively) in 2020, while the gross value of retail sales was in decline (OECD, 2020). The same trend could also be observed in Hong Kong. The changes in customer behaviour and the burgeoning of e-commerce purchasing indicated shrinking sales at physical stores and the emergence of the ‘next normal’: B2C e-commerce business. Amidst these changes, the value chain of the retail industry may be reconfigured and Logistics Service Providers (LSPs) are urged to transform their routine operations (i.e., orders placed by wholesalers or retailers) into a sound e-fulfilment process (i.e., orders placed by individual customers via e-commerce) with effective strategies.
Considering the needs of next-day or even same-day deliveries as an e-fulfilment process, ensuring a fast and efficient retrieval of Stock Keeping Units (SKUs) from shelves, has become crucial for today’s LSPs. To meet the trends of the ‘next’ e-fulfilment ‘normal’, LSPs need to be transformed with additional capabilities for handling discrete and fluctuating e-order demands. However, most LSPs in Hong Kong, especially small and medium (SME) -type LSPs, use rented warehouses to provide their services. They are unable to afford the large investments that would be entailed in adopting an automated storage and retrieval system and a sophisticated order picking system in rented warehouses; this limits their competencies and capabilities in handling e-orders.
This project aims to design and develop a e-fulfilment decision model for overcoming the new challenges presented to the logistics industry by today’s B2C e-commerce business in the wake of the COVID-19 pandemic. The proposed model allows LSPs to generate the optimal pick face replenishment strategy and fully utilize resources for handling the fluctuating demands of e-orders without needing to re-construct the whole premises and infrastructure. Considering the limited datasets obtained by SME-type LSPs, this project also contributes to establish an industry-wide solution for estimating quantity per SKU to be held in the pick face area. With the aid of the proposed decision model, the capabilities of the e-fulfilment process are enabled for LSPs, resulting in better competitiveness and service coverage when the ‘next normal’ emerges in the B2C e-commerce market.
Project Title:
AutoQFD: A Smart Quality Function Deployment Method for Product Development
Principal Investigator:
Dr Jack WU
Co-Investigator (HSUHK):
Prof LIU Hai
Project Period: 2022-01-01 to 2025-06-30
Funding Amount: $1,209,850
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: The Hong Kong University of Science and Technology
Project Reference No.: UGC/FDS14/E08/21
Abstract:
Quality Function Deployment (QFD) is a widely used toolkit to develop or improve products to better satisfy customer needs. It has also been widely used in other industries, such as service, healthcare, and software. While being proud of the history, QFD faces some difficult problems in today’s business. A QFD process is usually complicated, labour-intensive and time consuming. A smart and automatic QFD platform is proposed in this project to meet these challenges. The proposed methodology can automate the QFD process in a speedy manner and requires much less resources of expertise from companies. We will leverage the massive amount of product review text to inference the knowledge between customer requirements and product engineering characteristics. Natural language processing, deep learning and transfer learning techniques will be deployed to train a general QFD model for a product category, such as consumer electronics. Users just need to fine-tune the generic model using a small amount of product-specific data in the category, such as mobile phone, to get a product-specific QFD. In summary, AutoQFD enables users to train their customised QFD models with limited time, data and domain knowledge.
Project Title:
Enhancing Wireless Information Freshness via Physical-Layer Network Coding and Non-Orthogonal Multiple Access
Principal Investigator:
Prof LIU Hai
Project Period: 2022-01-01 to 2025-06-30
Funding Amount: $1,143,229
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: The Chinese University of Hong Kong, The Education University of Hong Kong
Project Reference No.: UGC/FDS14/E02/21
Abstract:
The Internet of Things (IoT) is an emerging wireless communications and networking technology that can be utilized to connect billions of devices and establish a close connection between our physical world and computer networks. Many time-critical applications, such as autonomous vehicles and industrial control, require the support of ultra-reliable low-latency communications (URLLC) to convey fresh information updates. However, information freshness cannot be accurately quantified by traditional metrics such as throughput and delay. Therefore, the age of information (AoI) metric has recently received extensive attention from researchers. AoI is defined as the elapsed time since the most recently received packet was generated. Literature shows that replacing traditional performance metrics with AoI may lead to fundamental changes in the communication system designs.
Most AoI research has focused on the upper layers of communication networks. Lower-layer solutions, such as multiple access schemes for the medium access control (MAC) layer and multi-user interference cancellation schemes for the physical (PHY) layer, have not been thoroughly studied for their impact on information freshness. Existing lower-layer designs cannot guarantee good information freshness when a large number of users access complicated and unreliable wireless channels. This problem seriously hinders the development of time-critical IoT applications. Moreover, information update packets in the IoT networks are usually very short. Shannon’s channel capacity formula in information theory assumes an infinite blocklength and is therefore not suitable for characterizing the performance of short-packet communications.
The purpose of this project is to fill the above-mentioned research gaps. To begin with, we would like to develop a theoretical framework for AoI analyses in various error-prone short-packet wireless communication models. Based on the developed framework, we then design lower-layer algorithms to enhance information freshness by physical-layer network coding (PNC) and non-orthogonal multiple access (NOMA). PNC alleviates the multi-user interference problem by utilizing the network-coded packets decoded from superimposed signals. NOMA improves spectral efficiency by serving multiple users at the same time and frequency. Our preliminary simulations show that PNC and NOMA can significantly improve the AoI performance of many channel models. To the end, we would investigate the combination of PNC and NOMA to improve the AoI performance further. If this research achieves favorable outcomes, it will be a solid step in the theory and practice of enhancing information freshness in the next-generation IoT networks.
Research Matching Grant Scheme (RMGS)
Project Title:
Blockchain-based Insurance and Financial Products Recommendation: Affordances and Actualisations
Principal Investigator:
Dr. Jack WU
Co-Principal Investigators:
Prof Daniel MO
Co-Investigators (HSUHK):
Dr Derrick FUNG, Dr Cathy LAM
Project Period: 2022-03-01 to 2024-12-31
Funding Amount: $4,236,970 (in-kind) and $4,236,970 (RMGS)
Funding Scheme/Source: Jedies Technology and Consultancy Limited and Research Matching Grant Scheme (RMGS)
Abstract:
This research investigates the affordances and actualisations of recommendation systems based on data and blockchain technology for insurance and financial products. To analyse the blockchain-based recommendation systems, our research objectives include:
1) To identify the critical moment(s) of health-conscious people purchasing an insurance and/or financial product
2) To characterise the insurance and/or financial products to be recommended
3) To extract association rules to enhance the recommendation process
The outcomes of this project may demonstrate the positive contributions insurance and financial companies can fine-tune their product configuration and combination based on a practical data-driven approach.
2020/21
Research Grants Council – Faculty Development Scheme
Project Title:
Variable Selection Methods for Complex Data Analysis
Principal Investigator:
Dr George HO
Project Period: 2021-01-01 to 2024-12-31
Funding Amount: $1,453,650
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/P05/20
Abstract:
Variable selection procedures aim to identify the correct covariates, which have a significant influence on the outcome variable and could provide robust model prediction. Traditional variable selection procedures such as forward selection procedure, backward elimination procedure, stepwise selection procedure, or model comparison via Bayes factor or some information criterion such as the Akaike information criterion may not be desirable for models with large number of covariates or complex structures. In this project, we particularly develop variable selection procedures for complex data modelling such as high dimensional additive model with interactions under marginality principle and composite quantile regression for ordinal longitudinal data.
Project Title:
A Blockchain-enabled IoT System for Pallet-pooling Management
Principal Investigator:
Dr Jack WU
Project Period: 2021-01-01 to 2023-06-30
Funding Amount: $595,800
Other Collaborating Parties: The Hong Kong Polytechnic University, The University of Hong Kong
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.:UGC/FDS14/E06/20
Abstract:
Most recent studies on logistics and supply chain management have focused on improvements to operational efficiency, information management, and network. Pallet management is a crucial yet less-researched aspect of the logistics industry. Currently, the closed-loop network for pallet management is preferred in the logistics industry, as positive environmental and economic impacts can be obtained. However, logistics networks are relatively complex and difficult to manage, due to the presence of the reverse logistics process. Therefore, a blockchain-enabled IoT system for pallet-pooling management is proposed in this project. This system integrates the development of blockchain and IoT technologies to identify, control, and monitor pallets in a closed-loop logistics network. Consequently, pallet standardisation can be established in the Guangdong-Hong Kong-Macao greater bay area, while the efficiency of logistics operations can be further enhanced.
Research Matching Grant Scheme (RMGS)
Project Title:
Research and Development of Contactless Parking System for Hourly and Monthly Parking
Principal Investigator:
Dr. George HO
Co-Investigators (HSUHK):
Prof TANG Man Lai, Prof Daniel MO, Dr Jack WU, Dr Cathy LAM
Project Period: 2021-01-01 to 2021-09-30
Funding Amount: $4,236,970 (in-kind) and $4,236,970 (RMGS)
Funding Scheme/Source: Sino Parking Service Limited and Research Matching Grant Scheme (RMGS)
Abstract:
Sino Parking Services Limited intends to support the smart city related research and development project and offering management support and sites for conducting the site surveys, site testing and pilot implementation. Car parks managed and operated by the Sino Parking Services Limited cover small, medium and large sizes, hourly, daily and monthly based. Parking spaces for both standard cars, lorry and electric vehicles, are available in these car parks, providing sufficient coverage for development considerations in different situations. With the sponsor from the Sino Parking Services Limited, the research team, formed by Dr George T.S. Ho, Prof. Man-Lai Tang, Dr Daniel Mo, Dr Jack Wu and Dr Cathy Lam will work closely with the sponsor’s staff for conducting the research and development of contactless parking systems with the state-of-art of big data analytics and artificial intelligence techniques. This project is expected to last for five months and the activities involve: (i) to design and develop the contactless parking system architecture, and (ii) to implement the proposed system in one of sponsor’s car parks.
Project Title:
Applying Innovation Resistance Theory to Understanding the Adoption of Delivery Apps in the Wine Industry
Principal Investigator:
Dr Jack WU
Co-Principal Investigators:
Dr Cathy LAM
Co-Investigators:
Prof TANG Man Lai, Dr Stephen NG
Project Period: 2021-07-02 to 2023-12-31
Funding Amount: $1,500,000 & $150,000 (in-kind) and $1,425,000 (RMGS)
Funding Scheme/Source: Chinese Development International Ltd. and Research Matching Grant Scheme (RMGS)
Abstract:
As no prior study has examined consumer resistance towards Wine Delivery Applications (WDAs), the proposed research project will be the first to investigate different consumer barriers that result in resistance toward WDAs. In this empirical study, innovation resistance theory would be appropriately utilised to study and evaluate the association between consumer barriers, intention to use, and WOM toward WDAs. A mixed-method approach involving an open-ended essay (with a wide variety of scenarios created) and a cross-sectional survey for potential users will be conducted.
The study will identify factors in customer behaviour related to the use of Wine Delivery Applications and related technologies; refine our conceptual model and constructing demo unit(s) for experiencing various wineTech applications; collect data by a mixed-method research approach that consists of qualitative essays, cross-sectional survey and case-control studies with potential users; analyse the relationships among different factors identified and interview with
practitioners in the local wine industry.
It will also try to understand local customer behaviour related to the use of WDAs, examine the association between different consumer barriers and intention to use new technologies in the wine delivery, and finally explore managerial and theoretical implications to the wine industry by seminars and/or sharing sections.
Our proposed study will open new avenues of research on wine delivery and its extended technologies.
Project Title:
Design of an AI-based Intelligent Model for Enhancing Business Performance
Principal Investigator:
Dr Stephen NG
Co-Investigators:
Prof CHOY Siu Kai, Dr George HO, Dr Cathy LAM
Project Period: 2021-07-01 to 2025-08-31
Funding Amount: $1,545,719 (in-kind) and $1,545,719 (RMGS)
Funding Scheme/Source: Sage Software Asia Pte Ltd. and Research Matching Grant Scheme (RMGS)
Abstract:
The sales volume of e-commerce experienced a rapid growth after the outbreak of COVID-19. Considering that the demands are unpredictable and most of orders are small-sized, this increases the challenges for small and medium (SME) -type companies in handling e-orders in terms of order management, data analysis, demand forecast, and inventory optimization. Digital transformation could be the strategy adopted by companies to transfer their traditional warehouse under the e-commerce new normal. However, according to a survey, about half of SMEs in Hong Kong did not understand how to adopt digital technology and hesitated to perform digital transformation as they believed that could be complex and expensive. Therefore, this research proposes to design a new AI-based intelligent model by integrating the digital technologies and artificial intelligent-based predictive analytics for companies to achieve performance enhancement. With the proposed model, many routine processes could be performed automatically, and human staff would be freed from repetitive tasks to focus on more innovative, value-added, and serviced related jobs. Also, error free operation and efficiency enhancement could be achieved in traditional warehouses which can ultimately facilitates digital transformation.
Innovation and Technology Fund
Project Title:
Smart Robotic Workforce: A Digital Twin-based Solution for Process Re-Engineering in Small and Medium Enterprises
Principal Coordinator:
Dr George HO
Team Memebers:
Prof TANG Man Lai, Prof Daniel MO, Dr Stephen NG, Dr Jack WU, Dr Cathy LAM
Project Period: 2021-03-01 to 2023-02-28
Funding Amount: $864,911.50 (ITF-PRP) and $890,000 (Unlimic) = $1,754,911.50
Funding Scheme/Source: Innovation and Technology Fund – Partnership Research Programme (PRP) and Unlimic Limited
Project Reference No.: PRP/034/20FX
Abstract:
Sino Parking Services Limited intends to support the smart city related research and development project and offering management support and sites for conducting the site surveys, site testing and pilot implementation. Car parks managed and operated by the Sino Parking Services Limited cover small, medium and large sizes, hourly, daily and monthly based. Parking spaces for both standard cars, lorry and electric vehicles, are available in these car parks, providing sufficient coverage for development considerations in different situations. With the sponsor from the Sino Parking Services Limited, the research team, formed by Dr George T.S. Ho, Prof. Man-Lai Tang, Dr Daniel Mo, Dr Jack Wu and Dr Cathy Lam will work closely with the sponsor’s staff for conducting the research and development of contactless parking systems with the state-of-art of big data analytics and artificial intelligence techniques. This project is expected to last for five months and the activities involve: (i) to design and develop the contactless parking system architecture, and (ii) to implement the proposed system in one of sponsor’s car parks.
2019/20
Research Grants Council – Faculty Development Scheme
Project Title:
Blockchain-based E-Commerce Analytics Model for Facilitating Trusted Data Exchange and Digital Supply Chain Integration
Principal Investigator:
Dr George HO
Project Period: 2020-01-01 to 2023-06-30
Funding Amount: $997,840
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Other Collaborating Parties: The Hong Kong Polytechnic University, The University of York
Project Reference No.: UGC/FDS14/E06/19
Abstract:
The blooming of e-commerce in the past decade has not only brought significant economic growth to the e-retailers, but also new opportunities and challenges to the logistics industry. To seize the opportunities arising from the emerging e-commerce logistics in Hong Kong, logistics service providers (LSPs) are forced to take on new roles and adjust their operations to fulfill the dynamic customer demand. This research aims to develops a Blockchain-based E-Commerce Analytics Model, integrating blockchain technology and the machine learning algorithm for managing data across the supply chains and predicting dynamic e-commerce order demand.
This research enables industry practitioners, especially LSPs and e-retailers, to plan ahead for the subsequent e-commerce operations. From the perspective of LSPs, the prediction model allows the firm to realize the e-commerce order arrival patterns, enabling flexible re-allocation of the right amount of resources in real time to deal with the hour-to-hour fluctuating arrival of orders in distribution centers. From the perspective of a retailer, the generic prediction model allows the firm to predict, for example, the sales volume among various e-commerce sales channels, the sales volume from different customer segments, and the e-commerce sales performance of different product categories. By tackling the unpredictability of demand in the e-commerce business environment, this research contributes to an effective decision support strategy for logistics `operations planning, hence, enhancing e-commerce logistics competence in Hong Kong.
Project Title:
Unsupervised Fuzzy Superpixel-based Image Segmentation
Principal Investigator:
Prof CHOY Siu Kai
Co-Investigator (HSUHK):
Dr Carisa YU
Project Period: 2020-01-01 to 2023-06-30
Funding Amount: $985,144
Funding Scheme/Source: Research Grants Council – Faculty Development Scheme
Project Reference No.: UGC/FDS14/P02/19
Abstract:
Image segmentation is a challenging problem in computer vision and has a wide variety of applications in various fields such as pattern recognition and medical imaging. One of the main approaches to this problem is to perform superpixel segmentation followed by a graph-based methodology to achieve image segmentation. Crucial to the successful image segmentation using this method is the superpixel generation algorithm and superpixel partitioning algorithm. Existing superpixel generation algorithms have various priorities and place emphasis on boundary adherence, superpixel regularity, computational complexity, etc, but normally do not perform well in all of the above simultaneously. Superpixel partitioning algorithms are typically based on graph-based approaches and could have high computational costs, which makes them inefficient in practical contexts. In the proposed project, we will investigate a fast and effective unsupervised fuzzy superpixel-based image segmentation algorithm to remedy the aforementioned difficulties for a wide range of applications. In particular, we will study the combined use of a novel fuzzy clustering-based superpixel generation technique and fuzzy graph-theoretic superpixel partitioning approach for image segmentation applications. The proposed segmentation method will be assessed by extensive comparative experiments using complex natural and textural images.
Research Matching Grant Scheme (RMGS)
Project Title:
Applications of SAS Viya in Big Data Analytics
Principal Investigator:
Prof CHOY Siu Kai
Co-Investigator:
Dr Benson LAM
Project Period: 2020-04-01 to 2026-12-31
Funding Amount: $7,008,344 (in-kind) and 7,008,344 (RMGS)
Funding Scheme/Source:SAS Institute Ltd (Hong Kong) and Research Matching Grant Scheme (RMGS)
Abstract:
SAS Viya is a powerful Big-data analytics tools that run in the cloud. It provides rather comprehensive range of analytical tools (e.g., the newest image processing, text analytics and machine learning algorithms) that help solve Big-data related problems in business, healthcare, engineering and education. The toolsets available covers the entire analytics life cycle, from data to discovery and deployment. With the sponsored Viya from SAS Institute Ltd (Hong Kong), the research team, formed by Man-Lai Tang (PI), S. K. Choy (Co-PI), S. Y. Lam (Co-I) and Ricky S. K. Wong (Co-I), will address the outstanding issues that are related to Big Data. The research team will explore the applications of SAS Viya in four different areas, namely (I) variable selection methods in regression; (II) social media research; (III) image segmentation and (IV) business negotiation. These research projects are expected to last for three years. The activities involve (i) the exploration of the primary Big-data analytics tools available in SAS Viya; (ii) the applications of appropriate Viya analytics tools to the aforementioned Big-data related research areas; (iii) the comparisons between the methodologies developed by the research team and those available from SAS Viya; (vi) the submission of the results for possible publications in international journals and preparation of cases for SAS demonstration; and (v) the presentation of the results in international conferences/seminars to disseminate the research findings.
Project Title:
Study on Digital Standard for Assessing Students’ Learning Performance with Data Analytics
Principal Investigator:
Prof Daniel MO
Co-Principal Investigator:
Dr George HO
Co-Investigators:
Dr Stephen NG, Dr Jack WU
Project Period: 2019-09-20 to 2023-01-01
Funding Amount: $5,400,000 (in-kind) and $5,400,000 (RMGS)
Funding Scheme/Source: Iron Mountain and Research Matching Grant Scheme (RMGS)
Abstract:
This research initiative aims to evaluate the effectiveness of establishing a digital platform that enables the use of data analytics for assessing students’ learning performance and facilitates the standardization of the assessment process in higher education sector. While different pedagogical assessment methods (e.g., assignments and examinations) are selected to achieve the learning outcomes, their effects on the outcomes, however, have not been systematically estimated and validated yet, much less identifying the contingencies that can strengthen such expected relationships. Teachers and module coordinators are still in a quandary what the most effective assessment methods are and whether the existing design of the assessment methods could reflect the module learning objectives. In addition, the existing learning outcomes are defined qualitatively lacking rigorously validated operationalization.
In our study, we will (1) use an established e-platform to digitize the students’ academic outputs such as assignments, project reports, mid-term test papers and final examination papers, (2) develop a standardized assessment process based on the e-platform to establish a digitized database capturing students’ academic outcomes obtained from various pedagogic assessment methods and to measure their learning outcomes via the platform for teachers’ data analysis, and (3) evaluate the effectiveness of such doing in improving the validity of teachers’ pedagogic assessment methods.
All in all, establishing a digitized platform guided by a standardized protocol to enable data analytics is a significant step toward success, though the associated effort may be huge. By adopting an established e-platform provided by Iron Mountain, we shall be able to define the assessment process and identify the practices of teachers, external examiners, and module coordinators in the course of performance assessment via the digital platform. Second, the platform facilitates teachers/module coordinators to consolidate digitized data collected from different stages of the assessment process so that detailed analytics can be undertaken according to the need of teachers/module coordinators. Having said that, the cost of the platform development is huge. According to Iron Mountain, the market price amounts to HKD 5.4M. The principal investigator of this research, Daniel Y. Mo, conducted studies on systems design (Mo et al., 2009; Mo et al., 2014) and process analytics (Mo et al., 2019) for various supply chain management systems, and he aimed at exploring process analytics for various supply chain management systems, and he aimed at exploring process analytics for e-leaming purpose in this research. The research objectives of this project are as follows:
• To design a digital platform for standardizing the digitalization process of assessment in higher education sector;
• To investigate the decision rules mapping between various student assessment results and learning outcomes; and
• To share the research results in the forms of workshop and case studies.
Project Title:
Estimating Blockchain IoT Project Completion Times – Simulation and Analytic Approach
Principal Investigator:
Dr Jack WU
Co-Principal Investigator:
Dr Cathy LAM
Co-Investigators:
Dr George HO, Dr Stephen NG
Project Period: 2019-08-01 to 2022-12-31
Funding Amount: $2,800,000 (in-kind) and $933,333 (RMGS)
Funding Scheme/Source: Simio LLC and Research Matching Grant Scheme (RMGS)
Abstract:
For any business and organizations in adopting Blockchain and IoT, they need new project management techniques rather than the traditional project management practices. This project aims to investigate the benefits of a proposed simulation and analytics approach when compared with the traditional approaches such as Program/Project Evaluation and Review Technique (PERT) and critical path method (CPM) used in IT project management. In order to analyse how new factors in Blockchain IoT projects will affect project time management, a simulation approach will be adopted to facilitate researchers and project students in collecting and consolidating data from participants in the form of batch testing (participants with and without project experience) of the Blockchain IoT project at different stages, so that project/program analysis can be conducted based on the decisions made by each individual in the simulation.
Project Title:
Design of a Blockchain-based Intelligent Model for Enhancing Supply Chain Reliability
Principal Investigator:
Dr Cathy LAM
Co-Principal Investigator:
Dr Jack WU
Co-Investigators:
Prof Daniel MO, Dr Stephen NG
Project Period: 2020-03-01 to 2024-11-30
Funding Amount: $594,740 (in-kind) and $594,740 (RMGS)
Funding Scheme/Source:Edge Tech Consultancy Company Ltd. and Research Matching Grant Scheme (RMGS)
Abstract:
The growing trend of e-commerce has put high pressure on supply chain parties, especially for distribution centres. By shifting the traditional business model to e-commerce, the order fulfilment cycle become shorter as customers always expect to receive the goods as soon as possible. In addition, small order size with high product varieties, and, increasing demand for customized value-added services are received, which definitely increase the complexity in managing the operation process. Any mistake or wrong information in the operation process would pose great challenges to deliver e-commerce orders on time and in good quality. In order to fulfill e-commerce orders efficiently, all supply chain partners have to seamlessly collaborate with each other to increase the information visibility and operations efficiency. However, the uncertainty of the information provided by different supply chain parties is still a challenge of the e-commerce supply chain reliability. The variances during the operation process will affect the quality of goods and hence make on-time delivery impossible. This research focuses on a new research direction in designing an intelligent model by integrating Blockchain mechanism with AI supported data analytics for investigating the possible cause of delay delivery and product recall based on trustworthy process data. The result of which will contribute to enhancement in supply chain reliability under e-commerce environment.