Sparsity Constraint Nonnegative Tensor Factorization for Mobility Pattern Mining

摘要

Despite the capability of modeling multi-dimensional (such as spatio-temporal) data, tensor modeling and factorization methods such as Nonnegative Tensor Factorization (NTF) is in infancy for automatically learning mobility patterns of people. The quality of patterns generated by these methods gets affected by the sparsity of the data. This paper introduces a Sparsity constraint Nonnegative 'tensor Factorization (SNTF) method and studies how to effectively generate mobility patterns from the Location Based Social Networks (LBSNs) usage data. The factorization process is optimized using the element selection based factorization algorithm, Greedy Coordinate Descent algorithm. Empirical analysis with real-world datasets shows the significance of SNTF in automatically learning accurate mobility patterns. We empirically show that the sparsity constraint in NTF improves the accuracy of patterns for highly sparse datasets and is able to identify distinctive patterns.