Deep Semi-supervised Learning with Weight Map for Review Helpfulness Prediction

Hua Yin, Zhensheng Hu, Yahui Peng, Zhijian Wang, Guanglong Xu, Yanfang Xu

Helpful online product reviews, which include massive information, have large impacts on customers’ purchasing decisions. In most of e-commerce platforms, the helpfulness of reviews are decided by the votes from other customers. Making full use of these reviews with votes has enormous commercial value, especially in product recommendation. It drives researchers to study the technologies about how to evaluate the review helpfulness automatically. Although Deep Neural Network(DNN), learning from the historical reviews and labels, computed by the votes, has demonstrated effective results, it still has suffered insufficient labeled reviews problem. When the helpfulness of a large number of reviews is unknown for lack of votes, or some useful latest reviews with less votes are submerged by the past reviews, the accuracy of current DNN model decreases quickly. Therefore, we propose an end-to-end deep semi-supervised learning model with weight map, which makes full use of the unlabeled reviews. The training process in this model is divided into three stages:obtaining base classifier by less labeled reviews, iteratively applying weight map strategy on large unlabeled reviews to obtain pseudo-labeled reviews, training on above combined reviews to obtain the re-training classifier. Based on this novel model, we develop an algorithm and conduct a series of experiments, on Amazon Review Dataset, from the aspects of the baseline neural network selection and the strategies comparisons, including two labeling and three weighting strategies. The experimental results demonstrate the effectiveness of our method on utilizing the unlabeled data. And our findings show that the model adopted batch labeling strategy and non-linear weight mapping method has achieved the best performance.