Final Year Project Showcase

Under the COVID-19 pandemic, it is common for one to wear mask in the daily life. In this project, we will develop a smart face mask using off-the-shelf sensors and soft circuit. We will use this prototype of smart mask to capture the signals induced by different human facial expression, and develop and experiment different machine-learning methods to classify different facial expression. The work of this project include: developing the hardware prototype, collecting the dataset, and implementing the classification algorithms.

In a metaverse, 3D objects and environments are essential to the senses of the metaverse. The quality and quantity of the construction of 3D objects can be time and labor-consuming, as the traditional method is to create the object one by one. As a result, developers often reuse the same object in the virtual world which made many items look and feel the same which reduces the sense of reality in such a virtual world. Especially for a large project like building a metaverse, creating all 3D objects manually is not feasible. This project aims to recreate the CityU in the most effective ways, like via phone image or LiDi scan. And use post-processing for background and object to create the virtual world. The methodology of the creation of 3D objects and environments via camera and LiDar will be investigated. And enhance the scan-produced object quality and cohesiveness with the metaverse with texture mapping.

Although end-to-end encryption has dramatically improved communication content privacy, communication metadata can still reveal much private information about the users, such as who communicates with whom and when and how much they communicate. Existing metadata hiding systems face the challenge that limited user volume will explicitly reveal the user group, which makes metadata hiding nearly meaningless. To tackle this problem, we propose a browser extension based front-end client compatible with existing metadata hiding systems, which can greatly increase the effective users to the back-end metadata hiding system.

The metaverse has become a trending topic recently. Metaverse is a multiuser post-reality universe that users can experience and interact with digital artifacts in real time. According to Koc-Januchta, MM(2019), visual cognitive learning style correlates with a better learning outcome. As metaverse allows users to be immersed in a virtual world filled with a wide range of multimedia, it could allow for more effective learning. As Schlemmer, E., & Backes, L. (2014) mentioned, metaverse breaks new ground for learning, as being inside a carefully curated virtual world provokes the thought process of how to learn in a new context. In an age where digital learning is an inevitable part of education, this project intends to discover how the metaverse impacts education and teaching, and what metaverse means for the future of education.

Handwriting skills of a person usually determined by it's speed and accuracy. In order to determine these factor, a device must be used to collect a person's handwriting into digital information. Using typical tablet with pen is not ideal as using digital pen to write on a tablet screen has a totally different handwriting sensation compared with writing on paper. Hence, using typical tablet is different from writing on paper. To address this issue, a tablet that could clipped with paper can be used to capture a person's handwriting while keeping the handwriting sensation. These paper clipping tablet allows user to write on paper and capture the handwriting into digital information simultaneously. This project aims to develop a application for a paper clipping tablet to determine a user's speed and accuracy for English handwriting and Chinese handwriting, where the accuracy is determined by character/alphabet recognition and speed is determined by the calculation of word per minute from the handwriting capturing.

Recent developments in the areas of data science and computing have allowed us to create increasingly complex and advanced ML models, enabling us to tackle important challenges in different areas of science. A prominent area of development, in the past few years, has been biochemistry, with models for molecular graph generation enabling easier evaluation and examination of potential drugs. One of the most important problems in drug discovery is determining the molecular structure and arrangement, as the structure defines the chemical properties of drug molecules, determining which other compounds it can interact with, and in what way. Development in this area has huge potential to help in development of new medication, and improve the lives of many by improving the accuracy and speed of drug examination, making drugs cheaper and more effective. Inspired by the recent developments in flow-based generative models, an architecture based on the principles of generative adversarial network (GAN) combined with a flow-based approach would be proposed, to possibly allow for both an adversarial and an MLE-based approach to training. It would be used to predict the molecular graph, and determine its configuration.

Blood types are known for its correlations to host personalities. The personality correlations can impact on society structures and its developments. We conjecture that country-wide blood type distributions are correlated to different country society development measures such as GDP and divorce rates. In this project, we wish to conduct a data science project with a focus on country-wide blood typing.

Because we cannot get back to the past time, the memory of the past time becomes valuable. The recent development of VR and 3D representation techniques, we can have a chance of preserving the scene of what is happening now and making them as valuable historical data provided for future generations.We hope to obtain and input the 3D scene data that can be reached within a specific range in some way. Through these data, we can calculate and generate the corresponding object model that can run in a VR environment, and finally enable users to visit the scene in this environment. On this basis, we also pay attention to user experience, so we hope to add the interaction between people and objects, sound effects, light effects, and feeling in the scene to increase the effect of immersive experience based on the above description. Through the realization of the above operations, a scene collected in the corresponding historical period can be constructed. The preservation of the scene is conducive to promoting cultural exchange now and preserving the historical appearance in the future.

Modern feminists utilize social media for various activities, while the Internet incurs opposite views. Prior work has explored feminist practices in computer-mediated social interfaces and online activism for social justice. However, less investigated how the public, especially grassroots feminists, participates on social media in discussing gender-related social controversy in a non-Western context. I chose the dichotomized debate around the Tangshan beating incident in China as the case to explore the characteristics of user participation and the psychological status during online collective actions.

Quantitative analysis was used to approach the questions. I first fine-tuned a Bidirectional Encoder Representations from Transformers-based (BERT-based) text classifier to classify the posts into three groups: (1) Crime-related, (2) Gender-related, and (3) Irrelevant or Ineffective. Each of the groups represents the content as well as the perspective is taken when viewing the Tangshan incident. Then I examined users’ temporal participation and how they posted content in different categories. After describing the landscape, I moved to a comprehensive analysis of user psychological characteristics based on the posts. Lastly, I examined their correlations to post volume to understand better the key factors associated with the popularity of online discussions.

The classification results suggest that the two most common perspectives: the Crime-related and Gender-related were equal in size but smaller when compared to the Irrelevant and Ineffective posts. Examination of user participation indicates that for each day, new users contributed to a much larger portion of online discussion than continuing users who posted before. I found that Crime-related posts showed more analytic thinking than Gender-related posts, which were more authentic and used more personal pronouns and words of social processes. I also compared different temporal trends of these features across content categories. Last but not least, I demonstrated the considerable predictability of LIWC features, especially since words related to emotional expressions were highly associated with the volume of online discussion. The findings echo prior work on networked online collective actions where participants seek to connect and construct affiliation and that emotions rather than cognitions greatly intensify online activities.

By conducting this research on Chinese online activism, we better understand the characteristics of such collective actions. Thus, we can better guide and support future online civic participation in social justice issues to reach its maximal potential of creating an equal and equitable society.

The wearable device based sign language recognition system plays a crucial role in the daily communication between people with disabilities and other ordinary people. A product that can quickly configure and accurately translate words is the goal people pursue. However, the current products or scientific research projects in related fields have more or fewer shortcomings, either because the customer samples that need to be collected are too complicated and the configuration time is too long, or because they only focus on a specific step for improvement. This project aims to study the common problems of existing products and improve sign language recognition system performance on wearable devices through our analysis and improvement measures. Finally, a complete set of end-to-end sign language recognition systems will be developed on smartphones or related devices, hoping to provide greater convenience to the daily life of people with disabilities in the future.

Time-series data prediction is a challenging task. Time-series prediction on economic data is even harder due to the complicated relation of different parties and factors in reality. In this project, the author developed a method for oil price prediction using information from the news. The author assumes events, especially geopolitical events, that happened in the past cause impact on traders and other stakeholders of oil futures and change oil prices. In particular, sentiment and the topic of news are chosen as the perspectives of studying the problem with the above assumption. As the deliverable, a hybrid model of Linear Regression and Gate Recurrent Units with STL decomposition is proposed for the task.

High costs and the lack of available golf courses have been a barrier for entry for a lot of people willing to familiarize themselves into the world of golfing. As a result, people have resorted into indoor golf solutions and mini-golf courses, which have been increasing in popularity over the recent years. Current indoor golf training solutions in the market involve simulated golf environments with projectors detecting golf ball swings. However, in a professional setting, such system would not suffice as it would not allow players to be able to walk around and familiarize themselves with the characteristics of the golf course. This project aims to build a VR-based virtual golf course system for training to complement existing solutions in the market. By using 3D scans of actual golf courses, the system would be able to emulate the golf courses such that players are able to experience a lifelike golf environment.

Quantitative Trading is one of the most successful strategy for not only retail investors, but also institutional investors. With the help of mathematical and statistical models, finding out opportunities to maximize gain and minimize lost within a period of time. Although this might sound promising, and its relative risk factor is low comparing to blatantly purchasing stocks. Stocks price movement are random and unpredictable, even with the most sounding fundamental analysis and technical analysis from experts.

Since late 2016, the public launch of virtual influencers foreshadows sensation soon after its release, receiving critical acclaim worldwide. Recently, one of the remarkable VTuber, Kizuna AI was selected as one of Asia's top 60 influencers. This highlights the potentiality of this field with the advancements of skyrocketing technology. Influenced by the big market share of ACGN (anime, comic, game, novel) culture, Vtubers' behaviors are more acceptable than real-person YouTubers. Many audiences are motivated to engage with virtual Youtubers (VTubers) as they would like to interact with favorite anime characters. In consequence, VTubers can provide relaxation and entertainment for viewers. Yet, a higher sense of distance is found between viewers and virtual avatars contrary to the variety of video genres in traditional streamers. Whereas VTubing has limited genre and content diversity under its performing style, Nakanohitos can hardly build unintuitive content under identity management. These directly affect the performance and quality of VTubing. This bottleneck leaves unsolved consequences for supporting immersive viewer engagement with VTuber. This paper proposes a novel design in creating agents to share the workload of Nakanohitos and an innovative user interface in virtual reality. Significantly, possible measurements and functionalities in virtual live streaming are designed for assisting the management of characters' persona in order to match the viewers' expectations.

This Fintech project is an integrated smart investing platform to streamline and automate the business operation with an intelligent investment portfolio advisor for Financial Institutions and individual investors. The platform makes user behavior & preferences understandable and smart investing at very low cost, automated, secure and transparent. The aim is to make sophisticated financial advisory services.

Stock market forecasting is always a challenging problem in academia. Different research state that the stock-prediction deep learning models have a significantly improved accuracy compared with machine learning and traditional statistic models. To further evaluate the forecasting performance of the deep learning models, this project employs the state-of-the-art deep learning models in academia, including bi-directional Gated recurrent unit (BGRU) and Generative adversarial network (GAN) with Long short-term memory (LSTM) and Convolutional neural network (CNN), trained separately with Hong Kong stock data and the United States stock data to predict the future stock price. This project also conducts a comparative study to evaluate the respective performance and accuracy in forecasting the stock prices in Hong Kong and the United States market with orthodox metrics, including Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). The evaluation results reveal that the BGRU models outperform the GAN models, and the models trained with the United States stock data have higher generalizability than those trained with Hong Kong stock data, consequently implying that the models trained with US stock data have a higher ability to predicts the stock closing price in different regions.

The academic writing style taught in colleges and universities aims to assist scholars and students in communicating precisely. On the other hand, the formal writing style has wider application scenarios in business and industry. This project proposed a new task to rewrite informal sentences in a formal style with academic writing features. To finish this task, a new corpus GYAFC-academic dataset is generated and utilized in the training process. Through using transformer model and warm-starting mechanisms, the proposed models perform well in style transfer accuracy and outperform the benchmark models by a significant margin in terms of grammar accuracy.

Federated learning is a promising framework to collect the dispersed data and train a collaborative machine learning model. Incentive mechanisms are thus introduced to motivate clients to contribute data in the context of federated learning. To facilitate these mechanisms, data valuation is a state-of-the-art solution to measure clients' data quality for the payoff fairly. However, it suffers from high overheads of computation and communication. In this project, a round-based data valuation (RDV) approach is proposed to estimate data quality with efficiency. Besides, it helps to train better-performing models.

A sliding puzzle, or sliding block game, is a very interesting and challenging game. It requires players to move one block horizontally or vertically without overlapping or crossing the game board in each step. The main target is to use as few steps as possible to reach an end configuration. However, little work pays attention to developing an intelligent and compatible system that can automatically generate a series of sliding puzzles. In this project, a novel game generator for sliding puzzles is designed and implemented, producing different types of complex sliding puzzles with optimal solutions. Overall, plenty of methods and techniques are utilized, including the multi-source Breadth-First Search, fast hash operations, and image processing. As a result, a powerful game generator is achieved for generating three kinds of complex sliding puzzles with optimal solutions, i.e., Kltoski, 15-puzzle, and Sokoban. Besides, a complete search on Klotski is successfully carried out, generating the most complicated Klotski puzzle games. It allows users to design and produce diverse sliding puzzles by processing and reading information from pictorial or text inputs. It is also scalable to create many other categories of sliding puzzles automatically.

Spatial Transcriptomics is a series of novel methods that enable transcriptomes' quantitative spatial analyses in individual tissue sections. Although there exist several tools and packages now for spatial transcriptomics data, a platform that can have functionalities of better flexibility is in demand to satisfy the analytic needs of biological research. This project develops a novel online tool to display and examine spatial transcriptomic data. It creates comprehensive modules for the interactions and customizations of spatial transcriptomic data visualization. The visualization modules built include a correlation plot, 2-D and 3-D embedding maps, a U-map, a correlation plot, a violin plot, and a deconvolution plot. the actual visualization power of these modules in existing transcriptomics research projects. It also features visualizations of ten datasets as an output of this project, including tissue slices from distinctive organs such as the human brain or the mouse kidney.

App stores allow users to download and buy software apps, and share feedback on installed apps with star ratings and text comments. Based on app reviews, App developers can improve or maintain the apps by bug fixing, feature enhancement, and adding new functions. However, the vast number of user reviews with diversifying quality, and mixed sentiments in a review significantly affect the progress of the software maintenance and evolution done by developers. In this project, an automated approach is proposed to classify app reviews into four pre-defined categories, which helps developers maintain and evolve apps. Different machine learning algorithms are trained using different features from three extraction techniques: Text Analysis, Natural Language Processing, and Sentiment Analysis. After comparisons among all ML algorithms, it shows that the combined use of the feature extraction techniques achieves the most outstanding results (precision of 74% and a recall of 72%) with the Logistic Regression.

Recently, people pay more attention to weight and sub-optimal health issues. To address these issues, sports are effective methods. Due to the limited home space, Rope skipping that requires a small space and can be done individually is a good choice for Hong Kong people to play at home. Although there are many existing tools for evaluating sports performance like pedometer applications for monitoring running activities, there is no comprehensive tool for monitoring rope skipping training in the market. This project aims to develop a mobile application to provide auto Rope skipping counting and sports data recording functions for normal people, sports players, sports trainers, and judges. Also, to promote this sport and encourage people to build up regular sport behavior, this application provides a social media sharing function for users to share their training records and motive each other. The project introduces a multi-tracking point, markerless, single camera, mobile application for capturing the jumping action and calculating Jumping and Tripping.

Artificial Intelligence could bring music composition to another level with limitless possibilities as an assistant for human musicians or an AI musician itself. Living in a digital era, classical music plays a dominant role in commercial films, movie trailers, game soundtracks, and more. However, there are no existing works that generate classical music in different eras. To fill this gap, this project proposes an AI music generator for classical music. So that it would be possible for composers, musicians, or even non-specialists without prior knowledge of classical music theories and backgrounds could quickly compose classical music according to their favorite musical eras for many practical purposes. It uses generative models, i.e., Bi-LSTM and CNNGAN to compose classical music for some particular classical music genres and evaluate their performance respectively and collectively.

Object imaging by utilizing different kinds of signals is already not a recent topic, including optical signals (visible light), radio frequency signals [11, 12], and acoustic signals (human audible [3] and inaudible). In this project, a solution in the form of mobile application on iOS using 20 Hz-20 kHz acoustic signals will be designed, with machine learning and ultrasonic signal extension in the late stage. For the deliverable of this project, apart from the imaging system, an attack model and a gesture classification system are designed.

Nowadays more and more people use and hold smart devices such as smart phones and these devices have become a part of people. However, in addition to using voice to control these devices, the more situations are interacting with them directly through touching. Therefore, in some cases, people find it more difficult to use these smart devices. For example, when the phone is occupied by putting it in the pocket, the user cannot touch the mobile phone directly, which means he cannot do the interaction with the phone such as answering the call or adjusting the volume. In this project, by tracking the user's finger motion to control the smart device, it can provide another interface to send the command as input to the computer.

Route planning problem has been a classic problem in the automatic driving area. With the development of computer vision and sensing techniques, automatic vehicles have gained the capability of capturing rich environment information for drive. However, how to utilize both internal and external information to plan an optimistic route is still not that satisfied. To produce an optimistic path, we are supposed to take user-experience, efficiency, accuracy and safety into consideration. Current state-of-the-art algorithms basically pay more attention to efficiency and accuracy, but they tend to ignore the importance of user-experience and safety. For instance, it is not desired that the vehicle drives at high speed in urban areas and with high angular velocity when road condition is bad (e.g. rainy days). Besides, the points used to construct the road are randomly sampled in the current algorithm. It may be time-consuming when some noisy points are selected and used to extend the path. We truly believe that with a learning algorithm involved in the sampling process, we are able to produce more high quality sampling points and thus eliminate the bad effect of useless points. Thus, I would like to focus on an automatic vehicle driving system that can take all these factors into account to provide an optimistic route planning algorithm.

Investors are optimizing their algorithm and model on predicting stock movement since it is hard to estimate the future market dynamic, which is affected by different factors. Some examples are the foreign market news, the effects of correlated stocks, and government politics. Therefore, investors are now using different approaches, like fundamental analysis and technical analysis, with various sources of data. Therefore, this paper attempts to use another approach to predicting the stock price movement. Instead of telling how exactly the stock price increases or decreases, this paper aims to provide a probabilistic measure on whether the next day stock price increases or decreases by comparing with today's closing price. Investors can make a better buy/sell decision based on the score according to the risk that they can bear or the risk diversification strategy on their financial portfolio. Five stocks in the Hong Kong sector are selected to be the target stocks of the prediction.

Football has become one of the popular Sports in the World. Nowadays, this sports game has further developed rapidly with over billions of fans or audiences in the world. The big five football leagues in the world - Premier League, LaLiga, Bundesliga, Serie A and Ligue1 have many football fans concerned about how their supporting teams performed in the world. In this project, a predictive model is built to predict the ranking of different teams in the mentioned league of the coming season for the five leagues mentioned.

Recommender systems become a major component in almost every Internet system nowadays, such as Taobao, eBay, Amazon, TikTok. In this project, we will study recent advances in recommender systems. The project will focus on contextual recommendations. Here contextual information refers to various situational information, such as time, location, browsing history, that can influence user preference for items. The insurance recommendation would be mainly focused. Based on the user's personal information such as age, height, weight, BMI, habit, digital footprint, income and so on to recommend the most suitable insurance to the user.

Although nowadays many frameworks do the data visualization, there is no framework for people who want to customize their data into the tree structure. Therefore, in this project, I am going to design the generic, compatible and optimized algorithms in order to represent the data in different tree structures.

All cancers are caused by somatic mutations. A software tool will be developed to explain what processes may cause the somatic mutations. That will help to find potential therapy for cancer patients.