A Joint Representation Learning Approach for Social Media Tag Recommendation.

In this paper, we analyze the mutual relationship between social media text and tags and explore how to integrate sematic information for tag recommendation. Our key motivation is to jointly map all words, tags and posts to vectors in a same hidden sematic space by modeling the syntagmatic and paradigmatic information simultaneously. We propose two novel distributed representation learning models for tagged documents: Tag Representation Learning (TRL) and Tag and Word Representation Learning (TWRL). The first models the immediate relationship between tags and words. The second one adds a skip-gram output layer to the first model, in order to enhance the semantic relationship among words. Extensive experiments are conducted on large scale datasets crawled from Twitter and Sina Weibo. By simulating two typical recommendation tasks, we discover that both models mentioned above outperform other competitive baselines remarkably.