Distributed Representations for Content-Based and Personalized Tag Recommendation

We consider the problem of learning distributed representations for documents from their content and associated tags, and of distributed representations of users from documents and tags provided by users. The documents, words, and tags are represented as low-dimensional vectors and are jointly learned with a multi-layered neural language model. We propose a two stage method where in the first stage which consists of two layers, we exploit the corpus wide topic-level information contained in tags to model one layer of neural language model and use document level words sequence information to model other layer of the proposed architecture. In the second stage, we use thus obtained document and tags representations to learn user representations. We utilize these jointly trained vector representations for personalized tag recommendation tasks. Our experiments on two widely used bookmarking datasets show a significant improvements for quality of recommendations. These continuous vector representations has the added advantages of conceptually meaningful which we show by our qualitative analysis on tag suggestion tasks.