Current browse context:
stat.ML
Change to browse by:
References & Citations
Statistics > Machine Learning
Title: Local Optimality of User Choices and Collaborative Competitive Filtering
(Submitted on 4 Oct 2010 (this version), latest version 25 Feb 2011 (v2))
Abstract: We describe a novel framework for learning recommender models for recommendation systems, which views user-system-item interactions as an opportunity give-and-take process, and encodes both "collaboration" and "competition" mechanisms underlying the interaction. The proposed framework leverages the latent factor models of collaborative filtering to encode "collaboration" (via factor sharing); and in the meanwhile, it utilizes a type of objectives that implies local optimality of user choices to encode "competition". Specifically, it takes into account both the revenue and the opportunity cost of each user decision; and, by optimizing a new objective that are analogous to the economic profit, it encourages that every opportunity being taken by a user be locally the best among the opportunities being offered to him/her. Such competition among candidates opportunities imposes stronger supervision and in turn leads to better generalization to unseen interactions. Empirical results indicates that the collaborative-competitive filtering (CCF) approaches improve dramatically recommendation performance compared with traditional collaborative filtering models (e.g., nDCG score is boosted from 0.14 to 0.71 on Yahoo! Pulse data, which is a huge 400% improvement).
We also discuss a Bayesian Generative model that enables joint learning of explore-exploit strategy and factorization recommender models. We call for attentions from recommendation industry to test this model on real recommendation system.
Submission history
From: Shuanghong Yang [view email][v1] Mon, 4 Oct 2010 15:29:33 GMT (21kb)
[v2] Fri, 25 Feb 2011 21:37:16 GMT (32kb)
Link back to: arXiv, form interface, contact.