GCN-based Interest and Popularity Disentangled Representations Framework for Recommendation

Recommender systems (RecSys) guide users toward items using observational interaction data. Present approaches often mitigate popularity bias, seeking to reveal pure user interest while down-playing the importance of item popularity. Yet, actions like clicks and purchases may frequently be motivated more by an item’s popularity than by user interest. Overcorrecting for popularity bias can cause RecSys models to overemphasize user interest while overlooking item popularity. Our paper introduces a unique method to disentangle user and item attributes into two components: interest and popularity, utilizing distinct GCNs. Based on the volume of interactions, we adaptively merge interest and popularity to form a more holistic embedding vector representation of users and items. Furthermore, in the context of disentangling, we introduce an innovative fusion method: Adaptive Merging. Our experimental validation substantiates the soundness of our method’s disentangling and merging, while benchmark tests indicate that our methodology achieves new state-of-the-art results. Data and codes are available at: <uri xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="https://github.com/lijialeplus/CIPS">https://github.com/lijialeplus/CIPS</uri>.