A Hybrid Model of Tensor Factorization and Sentiment Utility Logistic Model for Trip Recommendation

Abstract

This paper proposes a hybrid model of aspect-oriented sentiment prediction which integrates tensor factorization (TF) and sentiment utility logistic model (SULM). First, using sentiment dictionary words as seeds, the aspect or opinion words can be extended iteratively through double propagation. Accordingly, the users’ reviews could be represented as the features in user-item-aspect space, in which prediction model could be built. Various combinations of the hybrid model were proposed and evaluated on the Chinese reviews on places of interest from Trip Advisor. Experimental results show that the hybrid model can achieve better prediction performance than TF or SULM. The hybrid model also outperforms either TF or SULM while handling cold-start problem.