김민정/정재헌/마수드’s paper has been accepted in

Title: PolarDR: Polarity-aware Preference Trajectory Modeling for Effective Dynamic Recommendation,
Author: Min-Jeong Kim, Jae-Heon Jeong, Masoud Reyhani Hamedani, and Sang-Wook Kim
Abstract
In the dynamic setting of real-world applications, a user’s preference can be divided into a pre-use preference, inferred from the occurrence of an interaction (e.g., a click or a purchase), and a post-use preference inferred from the polarity of her feedback (e.g., positive or negative), which influences her future interactions. However, existing dynamic recommender systems (DRSs) rely solely on the former, while completely ignore the latter; consequently, they may fail to accurately model how a user’s preference changes over time. To address this limitation, we propose PolarDR, a novel model-agnostic framework via Polarity-aware preference trajectory modeling for Dynamic Recommendation that simultaneously takes into account both pre-use and post-use preferences in the recommendation process. To this end, PolarDR employs the three following modules: (1) Pre-use Preference Update generates PRE embeddings based on pre-use preferences, (2) Post-use Preference Update generates P-POST and N-POST embeddings based on post-use preferences using only interactions from the corresponding polarity, positive and negative, respectively, and (3) Residual Polarity Disentanglement decouples P-POST and N-POST embeddings through a two-step mechanism via a projection-removal operator and a residual polarity orthogonalization (RPO) loss. The results of our extensive experiments with five real-world datasets and six state-of-the-art DRSs demonstrate that PolarDR consistently and significantly improves the recommendation accuracy of all augmented DRSs, with gains ranging from 3.33% on the Yelp dataset to 14.58% on the Instant Video dataset in terms of NDCG@10 for TraceRec, the most recent DRS. The source code is available via Anonymous GitHub.