- Bettering Content material Advice: Data Graph-Based mostly Semantic Contrastive Studying for Variety and Chilly-Begin Customers
Authors: Yejin Kim, Scott Rome, Kevin Foley, Mayur Nankani, Rimon Melamed, Javier Morales, Abhay Yadav, Maria Peifer, Sardar Hamidian, H. Howie Huang
Summary: Addressing the challenges associated to information sparsity, cold-start issues, and variety in suggestion programs is each essential and demanding. Many present options leverage information graphs to deal with these points by combining each item-based and user-item collaborative indicators. A standard development in these approaches focuses on bettering rating efficiency at the price of escalating mannequin complexity, decreasing variety, and complicating the duty. It’s important to supply suggestions which can be each personalised and various, somewhat than solely counting on attaining excessive rank-based efficiency, equivalent to Click on-through Charge, Recall, and so on. On this paper, we suggest a hybrid multi-task studying strategy, coaching on user-item and item-item interactions. We apply item-based contrastive studying on descriptive textual content, sampling optimistic and unfavorable pairs based mostly on merchandise metadata. Our strategy permits the mannequin to raised perceive the relationships between entities inside the information graph by using semantic data from textual content. It results in extra correct, related, and various person suggestions and a profit that extends even to cold-start customers who’ve few interactions with gadgets. We carry out intensive experiments on two broadly used datasets to validate the effectiveness of our strategy. Our findings display that collectively coaching user-item interactions and item-based indicators utilizing synopsis textual content is very efficient. Moreover, our outcomes present proof that item-based contrastive studying enhances the standard of entity embeddings, as indicated by metrics equivalent to uniformity and alignment.
