JAM: Joint Action Matrix Factorization for Summarizing a Temporal Heterogeneous Social Network

This paper presents JAM (Joint Action Matrix Factorization), a novel framework to summarize social activity from rich media social networks. Summarizing social network activities requires an understanding of the relationships among concepts, users, and the context in which the concepts are used. Our work has three contributions: First, we propose a novel summarization method which extracts the co-evolution on multiple facets of social activity – who (users), what (concepts), how (actions) and when (time), and constructs a context rich summary called "activity theme". Second, we provide an efficient algorithm for mining activity themes over time. The algorithm extracts representative elements in each facet based on their co-occurrences with other facets through specific actions. Third, we propose new metrics for evaluating the summarization results based on the temporal and topological relationship among activity themes. Extensive experiments on real-world Flickr datasets demonstrate that our technique significantly outperforms several baseline algorithms. The results explore nontrivial evolution in Flickr photo-sharing communities.