Undergraduate Research (UResearch) in Computing Programme

A smart contract is an executable code that runs on the blockchain to facilitate, execute and enforce the terms of an agreement. A smart contract can execute the terms in a contract or agreement automatically when the specific conditions are met. Hence, resulting in a lower transaction fee when compared to the traditional method of executing a contract or terms of agreement (Alharby and Moorsel, 2017). Smart contracts are believed to revolutionize many industries in the near future and Ethereum is the most popular blockchain implementation that supports smart contracts.

In contrast to traditional programs, a deployed smart contract cannot be patched easily. If a bug is found in the smart contract or its underlying semantics provided by the blockchain implementation, it will result in the system being vulnerable to attacks. Thus, it is extremely important to ensure that smart contracts are executed correctly. This requires both the smart contract and the blockchain implementation to be bug-free. To achieve that, this study will be testing smart contract transactions on both the Conflux and Ethereum blockchain and analyse the possible differences in the test results.

Stack Overflow (SO) is the most popular Q&A website where developers can share their expertise by being the answerer and/or the questioner of a technical question. While receiving answers from the community helps to boost productivity and reduce the time-to-market for applications, the answers might not be provided by a relevant expert and could in turn promote security issues for real-life applications. In a recent study, researchers have identified security vulnerabilities in the suggested code of SO answers, which eventually found their way into applications installed by millions every day.

To mitigate the identified problem, SO uses the "Reputation" mechanism that measures the amount of trust the community has in a user. Users are able to earn reputation points by having his answer up-voted and/or having his answer accepted by the thread starter.However, a limitation to a user's reputation is that it is an aggregated across all domains and could hence be unrepresentative of the user's domain expertise for a particular question. As a result, there will be discrepancies between a user's reputation and the quality of their answers when answering questions out of their expertise.

In this work, we are proposing a new trust mechanism, “Contextual Expertise". Contextual Expertise measures the trustworthiness of a user's answer by evaluating his domain expertise using his activities across multiple platforms (i.e. SO and GitHub). We achieved a better accuracy in predicting whether the user's answer contains a security vulnerability.

The purpose of this project is to research, design and prototype microservices-based architecture for financial organizations with large transaction volumes and complex internal operations.

The presentation will first discuss about the current industry trend of IT systems development and deployment with microservices-based architecture, with further elaborations on how it can be potentially implemented for IT systems in financial organizations. Comparisons will be drawn between a monolithic system and microservices in terms of architecture designs and performance considerations. Lastly, there will be discussions on concrete steps in decomposing a monolithic application into microservices-based application.

A prototype of microservice, based on the Master Data Management (MDM) Service of SMU tBank, will be demonstrated in the presentation and performance measures will be used to compare the characteristics of different architectural approaches, such as CPU, memory and disk usages.

Classical permissioned distributed systems achieve consensus slowly. This is due in part to the quadratic messaging system utilised for its 3-phase agreement protocol, which achieves consensus in O(n2) time-complexity (Castro & Liskov, 1999). Current proposed quantum byzantine agreements present consensus resolution of up to O(1) time-complexity, with varying pros and cons (Ben-Or & Hassidim, 2005). The paper describes how we could utilise various aspects of quantum computing to instantly determine agreement, or disagreement, of proposed values between participants of a distributed network. We also explore the concept of private verifiability – how the network can come into consensus despite the presence of private, obfuscated data – without relying on zero knowledge proofs, which can be computationally expensive.

In this paper, we propose a novel consensus mechanism utilising quantum entanglement of photonic qubits to enhance the 3-phase agreement protocol by removing the need for multicast responses but instead, exploit the property of quantum entanglement to ensure that in each phase, only one multicast is required. This is made feasible with the recent revelation in which single-photon qubits can achieve a coherence time of >100ms, enough for data travel to most locations (K¨orber, et al., 2017). We also leverage the E91 quantum key distribution protocol, which requires photonic entanglement, to securely transmit values and preventing a man-in-the-middle attacks or system disturbance (Ekert, 1991). Finally, we also leverage the quantum entanglement of photonic qubits to achieve a verifiable “bridge” for private transactions between a subset private network and it’s corresponding consortium network

The ageing population is rapidly increasing, both in Singapore and worldwide. Due to the shortage of healthcare professionals and institutionalized care, there is a pertinent need for seniors to age-in-place - safely and in the familiarity of their neighborhoods. In addition, changing family structures and heavy work commitments of family members, coupled with the desire for more personal space and independence, have resulted in a significant proportion of seniors who live alone at home.

In this project, we study two programmes that are established by a social enterprise to improve the wellbeing of community-dwelling seniors. We analyze seven months of weekly survey data from the community care programme and befriending programme, comprising a total of 191 clients and 39 volunteers in the south-westernmost fringe of the Central Region of Singapore. Preliminary results show that: (i) there is an unbalanced distribution of clients, whereby 25% volunteers care for more than 67% of the clients; (ii) the average client duration in the community care programme is 18 months; and (iii) befrienders with less clients appear to be paired with clients who will require more emotional needs.

For wheelchair users, social participation and physical mobility play a significant part in determining their quality of life outcomes. However, little is known about how wheelchair users move about and engage in social interactions in their life-spaces. In this project, we investigate the social participation performance of the wheelchair users based on a combination of geolocational and lifestyle survey data collected over a period of 3 months.

This paper adopts a multi-variate approach combining geolocational travel patterns and various factors (independence, willingness, self-perception, etc.), deviating from the univariate approach taken by previous studies. We provide profiles of wheelchair users by combining these factors in an empirical study-based analysis. With the geolocational data of our volunteers, we can demonstrate the influences of other attributes that affect the social participation performance of wheelchair users with regards to life-space mobility.

The increasing popularity of social media platforms in the last decade or so has changed the way people communicate and interact with one another. With over 900 active social media platforms globally, millions of everyday social media users are contributing to content creation for various purposes. Especially in recent years, Reddit community has grown significantly around the world. In Singapore, Reddit has grown to be a go-to platform for sharing information and voicing opinions. This study focuses on the phenomenon of suicide and depression in Singapore. Recent studies have shown that Twitter, for example, can be used to prevent online suicide. Such studies have identified terms that are used by suicidal individuals. In our study, we further this investigation with a focus on Singapore society. We built a custom Singlish corpus and performed natural language processing to analyze text indicative of depression and suicide. With this study, we hope to help make a happier society where early detection of depression and suicidal behavior can be made and proper interventions employed in a timely manner.

Pinterest is a social media platform where users save images as pins and categorise them into boards. Each board groups pins that a user deems similar. Our problem of interest is recommending relevant pins to a user so as to increase the engagement level. Traditional approaches presume a single standard of image similarity. Hypothetically, on Pinterest each user creates boards based on her own unique notion of similarity, as similar pins can be grouped differently by two different users. We investigate this hypothesis by comparing neural networks that condition pin similarity on the context of the user with those that learn a generic notion of similarity.

Hospitals have been trying to improve the utilization of operating theatres (OTs) as they affect patient satisfaction, surgery throughput, revenues and costs. An existing surgical prediction model used post-surgery data showed promising results but contained key predictors which were absent during surgical listing. This study aims to provide an implementable surgical duration prediction model which bridges the “research-to-practice” gap. To build an implementable model, we seek to understand the listing process where patient is scheduled for a surgery to gain insights on domain knowledge and to isolate unavailable data. We used these insights found for feature engineering of key predictors and proposed new implementable prediction models. We tested our proposed models and the experimental results showed that our proposed models fared similarly to existing model and it can be implemented in practice as the predictors used are now available during listing.

In today’s day and age information and communications technology has permeated into almost every activity of our daily life. This extends to education delivery assisted by these devices. The main challenge with in-class discussions is that they are conducted orally, giving rise to difficulty in retaining uncovered knowledge. The most common way to ensure that this knowledge is not lost is to record the entire discussion, transcribe and subsequently analyse it. Using speech recognition, this effort of transcribing can be reduced and finally eliminated. Interaction analysis is the area of research that analyses the discussions to discover the behaviour of participants and topics of interests. Using a Teaching Assistant for the capture of such knowledge is painstaking, tedious, and mundane for the TA, while not providing many insights to the faculty. Therefore, automated interaction analysis can be used to facilitate the faculty’s understanding of classroom interactions along three dimensions namely; participation of students, the evolution of topics and peer dynamics. For our study, we used classroom discussions which are converted into text by Liveclass App developed internally. We applied Natural Language Processing, and text mining techniques to process the discussions and analyse interactions in the three dimensions. Lastly, we evaluated the topics with human gold truth and generated visual reports for faculty in the previously mentioned aspects.

In Singapore, 788,500 foreign unskilled and semiskilled workers, also known as Work Permit (WP) Holders, are hired on a 2-year contractual basis to alleviate the manual labour shortage in industries like construction and marine. While the WP Holders benefit relatively high-income jobs offered in Singapore compared to their home countries (mainly Bangladesh, Myanmar, India, China and Sri Lanka), they have lower salaries compared to local workers, fewer rights, and typically work in places with the highest risk of injuries. These issues seem to remain to despite of the recent efforts by Ministry of Manpower (MOM) and Non-Governmental Organisations (NGO). In this research, we seek to understand how computer-mediated support can improve accessibility and empower migrant communities. We conduct a formative semi-structured interview study with 2 local NGO representatives and a survey study with 22 WP Holders, of which 6 experience employment-related issues. The results of the study have shown that the abovementioned problems remain for the WP Holders despite of both technical and non-technical support provided to the WP Holders. Based on the insights gained through the study, we list design implications for future technologies to mitigate the challenges that WP Holders face.

The amount of structured knowledge on the web is growing and this has led to the creation of large knowledge bases (KB) such as Freebase. This has sparked the interest in developing methods to allow users to access these resources. Knowledge Based Question Answering systems provide users an intuitive access to information via natural language by encapsulating the technical complexities of data modelling, vocabularies and query languages.

Representations of the KBs are often learned only from the QA training data, leading to the following problems. First, the global information of the KB is insufficient. This means that for a question answer pair in the training data, there may be a more suitable answer from the global KB information. However, current KB-QA systems may not be able to learn this. Second, there is the problem of out-of-vocabulary (OOV) as the testing data contains answer entities that was not observed in the training data, thus the performance of the KB-QA system would be negatively affected.

This research aims to successfully learn the semantic vectors that model the entities and relationships of the KB to mitigate the effects of the above-mentioned problems.

In an increasing digitised world, an event registration chatbot (“chatbot”) serves as a beneficial use case to provide users with a means of automated querying and receiving of information related to events. A key technology enabling such a chatbot is Natural Language Processing, specifically in the realm of sequence classification and labelling. This research project studies the use of LSTM in sequence classification and labelling, with a direct application in TensorFlow to provide the chatbot with the ability to (1) process Natural Language input, (2) retrieve the desired information, and (3) return the information in a structured format.

In this study, we dive into an understanding of what the FinTech Payments industry is going through. These companies are becoming prevalent in the current world. In the last decade, there has been a transformational change in the way we deal with payments. Innovative FinTech Payment Companies are up and about providing alternative ways that we have not imagined payment services could provide. In this paper, we take a look into what these companies are from three different categories, studying how these companies work and function to understand the impacts they would have on the payment industry. Next, we take a look at a response regulatory bodies have implemented on the changing payment industry. The paper focuses on the new Payment Services Directive 2 (PSD2) implemented by European Union in 2018 which promotes the concept of “Open Banking” where banks are pressured to share their information through services for other companies to utilize as well as the impacts it has on the payment industry. In view of the rapid changes happening, what can banks do towards this? We propose in the paper as well, a recommended approach that banks should take in order to keep competing in this changing industry.

An everyday dilemma close to many Singaporeans' heart is deciding where and what to eat for a particular meal. Being a food haven, Singapore offers a great diversity of options. FoodRecce is an Android application that allows users to survey their immediate locality for desired eateries, by recommending food places and exhibiting their established information and photos obtained from various online sources. The personalized recommendation is powered by a machine learning algorithm, which is dynamically updated based on implicit feedback from users’ actions through an intuitive swipe-based interface. The underlying principle is to correlate the preferences of users who are similar and likely to love eateries enjoyed by other similar users. With a platform to run controlled experiments, the recommendations are continually evaluated and refined, so as to improve the performance and user experience over time.