Group based Self Training for E-Commerce Product Record Linkage

In this paper, we study the task of product record linkage across multiple e-commerce websites. We solve this task via a semi-supervised approach and adopt the self-training algorithm for learning with little labeled data. In previous self-training algorithms, the learner tries to convert the most confidently predicted unlabeled examples of each class into labeled training examples. However, they evaluate the confidence of an instance only based on the individual evidence from the instance. The correlation among data instances is rarely considered. To address it, we develop a novel variant of the self-training algorithm by leveraging the data characteristics for the task of product record linkage. We joint consider a candidate linked pair and its corresponding correlated pairs as a group at the selection of pseudo labeled data. We propose a novel confidence evaluation method for a group of instances, and incorporate it as a re-ranking step in the self-training algorithm. We evaluate the novel self-training algorithm on two large datasets constructed based on real e-commerce Websites. We adopt several competitive methods as comparisons and perform extensive experiments. The results show that our method outperforms these baselines that do not consider data correlation.