Personalized Adaptive Meta Learning for Cold-start User Preference Prediction

A common challenge in personalized user preference prediction is the cold-start problem. Due to lack of user-item interactions, directly learning from new users' log data causes serious over-fitting Recently, many existing studies regard as a few-shot problem, where each task and recommended items are classes, gradient-based meta method (MAML) leveraged address this challenge. However, real-world application, users not uniformly distributed (i.e., different may have browsing history, items, profiles. We define major groups with large numbers sharing similar information, other minor users), MAML approaches tend fit ignore users. To task-overfitting we propose novel adaptive approach consider both three key contributions: 1) first present rate meta-learning improve performance by focusing on 2) provide better rates for user, introduce similarity-based find reference tree-based store features fast search. 3) reduce memory usage, design agnostic regularizer further space complexity constant while maintain performance. Experiments MovieLens, BookCrossing, production datasets reveal that our outperforms state-of-the-art methods dramatically