Cross-Domain Relationship Prediction by Efficient Block Matrix Completion for Social Media Applications

Abstract

The online social media has experienced vigorous evolution. Diversified needs of information acquisition and retrieval on social media platforms have been evoked by massive users. While all sorts of application demands meet with explosive data growth, the development of effective methodologies has become emergent. By taking full advantage of rich context, we propose a heterogeneous object relation matrix completion approach (EBMC) which jointly complements the relationship between the heterogeneous data objects. Specifically, we detect the Place-of-Interest (POI) with mean shift algorithm on the GPS information of the social image collection. Then, a batch matrix completion and learning method is developed by optimizing a unified objective function to learn the POI-specific user-image, image-tag and user-tag relationships. Finally, we decompose the whole learning problem into a set of POI-specific subtasks, which corresponding to the relation data blocks separated by the POI structure. Through experiments on tasks of image annotation and user retrieval based on image similarity of real-world social media datasets, we found that our proposed method achieved good performance.