RecTen: A Recursive Hierarchical Low Rank Tensor Factorization Method to Discover Hierarchical Patterns from Multi-modal Data

Abstract

How can we expand the tensor decomposition to reveal a hierarchical structure of the multi-modal data in a self-adaptive way? Current tensor decomposition provides only a single layer of clusters. We argue that with the abundance of multimodal data and time-evolving networks nowadays, the ability to identify emerging hierarchies is important. To this effect, we propose RecTen, a multi-modal hierarchical clustering approach based on tensor decomposition. Our approach enables us to: (a) recursively decompose clusters identified in the previous step, and (b) identify the right conditions for terminating this process. In the absence of a well-established benchmark, we evaluate our approach with synthetic and five real datasets. First, we test the sensitivity of the performance to different scenarios and parameters. Second, we apply RecTen on four online forums and a dataset that represents user interaction on GitHub. This analysis identifies meaningful and interesting behaviors, which further increases our confidence in the usefulness of our approach. For example, we identify some real events like ransomware outbreaks (55 users, 86 threads, December 2015, February 2016), the emergence of a black-market of decryption tools (34 users, 12 threads, February 2016), and romance scamming (82 users, 172 threads, March 2018). To maximize the impact of our work, we intend to: (a) develop a usable tool, (b) make the tool and our datasets publicly available. However, RecTen is a hierarchical clustering approach that can be used to take the pulse of large multimodal data and let the data reveal its own hidden structures.