Meta-learning for recommender systems

Abstract

The widespread of mobile devices has enabled people to use various online services, such as video streaming, shopping, and news in everyday life. Because users seek only a few items out of a myriad of items in such services, the role of a recommender system is to quickly find them out to enhance user satisfaction and help the growth of the service. However, there are three challenges in recommender systems due to the vulnerability of the standard learning scheme: low-quality training data, online update delay, and unfair recommendations after learning.This dissertation aims to resolve the three challenges of recommender systems via meta-learning, which is also known as “learning to learn.” Meta-learning can deal with the vulnerability of recommender learning by adaptively guiding its learning process after observing how the recommender learns. In this regard, this dissertation suggests meta-learning-based approaches to overcome the three challenges.The first study proposes PREMERE that performs training data reweighting to avoid learning from low-quality data. PREMERE adaptively provides a high weight on useful but sparse data, and a low weight on missing data to induce the recommender to learn only from qualified data. Through extensive experiments on three real-world benchmark recommender datasets, PREMERE proved its effectiveness, improving performance by up to 26.9% compared with state-of-the-art algorithms. The second study proposes MeLON that online updates recommender systems with up-to-date information. While current user interests change constantly, their sparse signal in the user-item domain makes recommenders difficult to catch up due to the fixed scheme of standard learning. To this end, MeLON adaptively generates learning rates between each data and model parameter considering theirmutual relationship. Extensive empirical evaluations on three real-world online service datasets validated that MeLON can continuously maintain the high performance of a recommender system during online service by virtue of its high flexibility for updates.The third study first performs data analysis to demonstrate that recommender systems are prone to over-recommend popular items especially to new users, which is undesirable for both users and item providers. Because it is fair to recommend popular items to a user according to their ratio in the user’s history, to cope with this challenge, the third study proposes ColA that expedites learning of tail-itemsadapted to each user and recommender model learning state. Experimental results on a real-world recommender system dataset confirm that ColA can achieve higher performance and improved fairness compared with state-of-the-art popularity debiasing algorithms.This dissertation is expected to enhance user satisfaction, the ultimate goal of recommender systems by effectively addressing the challenges caused by learning vulnerability.