Parallel collective factorization for modeling large heterogeneous networks

Relational learning methods for heterogeneous network data are becoming increasingly important for many real-world applications. However, existing relational learning approaches are sequential, inefficient, unable to scale to large heterogeneous networks, as well as many other limitations related to convergence, parameter tuning, etc. In this paper, we propose Parallel Collective Matrix Factorization (PCMF) that serves as a fast and flexible framework for joint modeling of a variety of heterogeneous network data. The PCMF learning algorithm solves for a single parameter given the others, leading to a parallel scheme that is fast, flexible, and general for a variety of relational learning tasks and heterogeneous data types. The proposed approach is carefully designed to be (1) efficient for large heterogeneous networks (linear in the total number of observations from the set of input matrices), (2) flexible as many components are interchangeable and easily adaptable, and (3) effective for a variety of applications as well as for different types of data. The experiments demonstrate the scalability, flexibility, and effectiveness of PCMF for a variety of relational modeling tasks. In particular, PCMF outperforms a recent state-of-the-art approach in runtime, scalability, and prediction quality. Finally, we also investigate variants of PCMF for serving predictions in a real-time streaming fashion.