Topic-based Relevance Modeling of Metadata to Transactions for Collaborative Filtering

概要

論文の詳細を見る
We propose a probabilistic topic model that infers the relevance of content-descriptive metadata to assist in the task of collaborative filtering. Metadata can be used to add functionality to purely user-based models, such as, in a purchase recommendation service, the ability to recommend items that have not yet been purchase by any user in the system. However, content-based methods are vulnerable overfitting on metadata that is not descriptive of the users interests, and thus detrimental to recommendation accuracy. We describe a model based GM-LDA proposed by Blei, et al. that incorporates a relevance-sensitive topic model for metadata to assist collaborative filtering methods. We seek to mitigate the potential negative impact of fitting to unhelpful metadata by assuming that metadata words are either relevant or irrelevant in terms of descriptive power, and allow our model to infer the relevance of each metadatum before using the information. We compare our model's normal and cold-start recommendation accuracy on data taken from MovieLens against that of conventional methods found in the literature.
2011-10-27