Research

Vision: Empowering Users through Web Intelligence

Missions:
1) Achieve more accurate and complete discovery of knowledge.
2) Excel in research on video and multimedia.
3) Transform design, development and deployment of applications;
4) Provide alternative paradigm in engaging users.
5) Nurture top quality research, new ideas and innovative technologies.

Projects > Recommendation and Influence

Recommendation and Influence

25 January, 2018 by Tek Min Tan

Recommendation is an essential function of any user-oriented online services to improve user engagements and encourage active usage. Traditional recommendation techniques based on Matrix Factorization (MF) have been very successful in many recommendation tasks. However, MF is a linear model and is hard to leverage other available resources such as item and user attributes, and hence it has limited expressiveness and performance. The era of big data brings in large-scale rich data with strong computational power, making it possible to employ more complex and powerful deep learning models to enhance the recommendation performance.

This project aims to advance the state-of-the-art recommendation technology by exploring more data resources and deep learning technology. We will carry out research in three directions. First we will develop nonlinear (deep) models that are able to learn second and higher order correlations among data features (users and items), along with rich user/item attributes as well as auxiliary information such as Web and other social media information. Our initial research focus on enhancing FM (Factorization Machine), a nonlinear model, by combining it with deep learning framework into NeuraFM, for complex recommendation tasks. The work is preliminary.

Second, we will explore explainable framework for recommendation. One approach is to combine tree-based methods, such as Random Forrest, with deep learning framework to derive recommendations that are both accurate and explainable. Another approach involves the use of knowledge in deep learning framework. We also need to examine the basis and quality of explanation in this research.

The third direction is to explore recommendation technology that are incremental and able to handle cold start problems effectively. This is important in line with recent legislation in Europe on privacy in which users own their interaction data and have the right to exclude their personal data for use in tasks like recommendation. We will focus on the applications of the domains of: a) E-commerce, b) healthcare, c) FinTech (e.g., portfolio recommendation); and d) online news.