How to Learn Item Representation for Cold-Start Multimedia Recommendation?

The ability of recommending cold items (that have no behavior history) is a core strength of multimedia recommendation compared with behavior-only collaborative filtering. To learn effective item representation, a key challenge lies in the discrepancy between training and testing, since the cold items only exist in the testing data. This means that the signal used to represent an item varies during training and testing --- in the training stage, we can represent an item with both collaborative embedding and content embedding; whereas in the testing stage, we represent a cold item with content embedding only. Nevertheless, existing learning frameworks omit this critical discrepancy, resulting in suboptimal item representation for multimedia recommendation. In this work, we pay special attention to cold items in multimedia recommender training. To address the discrepancy, we first represent an item with dual representation, i.e., two vectors where one follows the traditional way that combines collaborative embedding and content embedding, and the other assumes that the item is cold by replacing the collaborative embedding with zero vector. We then propose a Multi-Task Pairwise Ranking (MTPR) framework for model training, which enforces the observed interactions ranking higher than the unobserved ones even if the item is assumed to be cold. As a general learning framework, Our MTPR is agnostic to the choice of the collaborative (and/or content) encoder. We demonstrate it on VBPR, a representative multimedia recommendation model based on matrix factorization. Extensive experiments on three datasets of diverse domains validate MTPR, which leads to better representation for both cold and non-cold items in the testing stage, thus improving the overall performance of multimedia recommendation.