Contrastive Learning for Topic-Dependent Image Ranking

In e-commerce, users' feedback may vary depending on how the information they encounter is structured. Recently, ranking approaches based on deep learning successfully provided good content to users. In this line of work, we propose a novel method for selecting the best from multiple images considering a topic. For a given product, we can commonly imagine selecting the representative from several images describing the product to sell it with intuitive visual information. In this case, we should consider two factors: (1) how attractive each image is to users and (2) how well each image fits the given product concept (i.e., topic). Even though it seems that existing ranking approaches can solve the problem, we experimentally observed that they do not consider the factor (2) correctly. In this paper, we propose CLIK (Contrastive Learning for topic-dependent Image ranKing) that effectively solves the problem by considering both factors simultaneously. Our model performs two novel training tasks. At first, in topic matching, our model learns the semantic relationship between various images and topics based on contrastive learning. Secondly, in image ranking, our model ranks given images considering a given topic leveraging knowledge learned from topicmatching using contrastive loss. Both training tasks are done simultaneously by integrated modules with shared weights. Our method showed significant offline evaluation results and had more positive feedback from users in online A/B testing compared to existing methods.