SMU Office of Research – The National Research Foundation (NRF) Fellowship supports outstanding, early-stage career researchers to lead impactful research in Singapore. It is a globally competitive programme and is open to all areas of science and technology. The fellowship offers a five-year research grant of up to S$3 million to support projects that exhibit high likelihood of research breakthrough.
The Office of Research is pleased to announce that the research proposal from Assistant Professor Hady Wirawan Lauw was awarded with the NRF Fellowship grant. Below is the summary of his proposal, and we wish him the best in his research!
Project Title:
Dimensionality Reduction for Recommender Systems: Unified Latent Co-representation of Multi-Modal Preference Signals
Duration:
5 years
Overview:
With the increasing digitisation of our everyday lives, two trends are emerging. One is the proliferation of choices in our consumption of both physical and digital goods and services. The other is the "datafication" of our behaviours, whereby users’ preferences increasingly manifest themselves through multiple forms of preference signals, including consumption, ratings, text reviews, social connections, images, etc. The first trend points to the broad challenge of helping users to navigate the ever-expanding universe of options. The second points to the deep opportunity of leveraging on Big Data to model user preferences.
To support personalisation and recommendation, current approaches are based on mining historical preferences of users. However, for most users, the available information is too sparse. We propose to use additional preference signals mined from various sources to supplement the sparse data on historical preferences. Our approach encapsulates three main components. The first is gathering wide information on preference signals for a large number of items in specific domains. The second is to develop an abstract representation of this multi-format data in a shared representation space. Finally, this representation will be used to supplement historical preferences data in applications such as recommendation and visualisation of user preferences.
The research will have a significant role in advancing the state-of-the-art in dimensionality reduction and recommendation in a synergistic way. Through the integration of preference information of various modalities, as well as supplementation of proprietary datasets, the constructed knowledge bases could improve recommendation accuracy. The software tools developed in the project will also support computationally more efficient recommendation queries. The low-dimensional representations of data also naturally enable greater intuitiveness and usability through visual interfaces.
