Clustering Product Features of Online Reviews Based on Nonnegative Matrix Tri-factorizations
Wang Jiajia, Yezheng Liu, Jiang Yuanchun, Sun Chunhua, Sun Jianshan, Yanan Du · 2016
Clustering product features is the essential task to mine opinions from unstructured online reviews because different customers usually express the same feature with different words or phrases. Several supervised and unsupervised methods have been applied to accomplish this task. In this paper, we propose an orthogonal nonnegative matrix tri-factorizations model to solve the problem. We first construct the feature-opinion relation matrix and two constraint matrixes (i.e., cannot-link and must-link) based on three assumptions, and then integrate those matrixes to construct the orthogonal nonnegative matrix tri-factorizations model. The proposed model takes feature-opinion pairwises into consideration, caters to the principle of mutual reinforcement, and clusters product features by incorporating the cannot-link and must-link constraints. We develop an optimization algorithm to solve the matrix factorization, and prove the correctness and convergence. Experimental results on real datasets show that the proposed method is valid.