Latent Pattern Identification Using Orthogonal-Constraint Coupled Nonnegative Matrix Factorization

Anandkumar Balasubramaniam, Thirunavukarasu Balasubramaniam, Anand Paul, Richi Nayak

Abstract: The recent advancements and developments in Intelligent Transportation Systems (ITS) lead to the generation of abundant spatio-temporal traffic data. Identifying or understanding the latent patterns present in these spatio-temporal traffic data is very much essential and also challenging due to the fact that there is a chance of obtaining duplicate or similar patterns during the process of common pattern identification. This paper proposes an Orthogonal-Constraint Coupled Nonnegative Matrix Factorization (OC-CNMF) method and studies how to effectively identify the common as well as distinctive patterns that are hidden in the spatio-temporal traffic-related datasets. The distinctiveness of the patterns is achieved by the imposition of the orthogonality constraint in CNMF during the process of factorization. The imposition of the orthogonal constraint helps to ignore similar/duplicate patterns among the identification of the common patterns. We have shown that imposing orthogonality constraint in CNMF improves the convergence performance of the model and is able to identify common as well as distinctive patterns. Also, the performance of the OC-CNMF model is evaluated by comparing it with various performance evaluation measures.