User-specific Adaptive Fine-tuning for Cross-domain Recommendations

Making accurate recommendations for cold-start users has been a longstanding and critical challenge recommender systems (RS). Cross-domain (CDR) offer solution to tackle such problem when there is no sufficient data the who have rarely used system. An effective approach in CDR leverage knowledge (e.g., user representations) learned from related but different domain transfer it target domain. Fine-tuning works as an learning technique this objective, which adapts parameters of pre-trained model source However, current methods are mainly based on global fine-tuning strategy: decision layers freeze or fine-tune taken all In paper, we argue that RS personalized should their own policies better preference learning. As such, propose novel User-specific Adaptive method (UAF), selecting network fine-tune, per-user basis. Specifically, devise policy with three alternative strategies automatically decide be fine-tuned frozen each user. Extensive experiments show proposed UAF exhibits significantly more robust performance recommendation.