Memory-efficient Recommendation

The School of ITEE is hosting the following PhD progress review 1 seminar

Memory-efficient Recommendation

Speaker: Yunke Qu
Host: Dr Rocky Chen

Abstract: Recommender systems predict the preference that users have for the items based on their preferences, browsing history, and past interactions. With the increasing sophistication of machine learning algorithms and the availability of vast amounts of user data, recommender systems have become indispensable tools, empowering individuals to discover new content, make informed decisions, and optimise their digital interactions.

However, the performance of recommender systems is often bottlenecked by the limited memory resources available. This can be attributed to the embedding table of recommender systems, where fixed and uniform embedding sizes are assigned to each user and item. This approach leads to memory inefficiency as it requires allocating a substantial amount of memory to accommodate all users and items, regardless of their varying characteristics and importance. In response to this challenge, researchers have proposed numerous approached. In this report, we identify four major shortcomings of the existing approaches. We further propose two innovative frameworks that dynamically assign flexible embedding sizes to each user and item to address them. We verify the superior performance of the proposed approaches through extensive experiments on several benchmark datasets.

Bio: Yunke Qu completed his bachelor’s degree in information technology in 2021. He is currently pursuing a doctoral degree in computer science. His research focuses on memory efficient recommender systems.