Deep semi-supervised learning with weight map for review helpfulness prediction

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