An Application of Learned Multi-modal Product Similarity to E-Commerce

Product similarity search is an important tool for e-commerce companies to manage their portfolios of products and to find competitive prices on large electronic market places. The specific requirements for this similarity search application are (i) the similar products should be competitive products with respect to a given query product, (ii) related and just generally similar products should be treated as not similar products. Thus, the similarity between products should be learned from data. We propose to use classification models for entity matching and image classification to learn a multi-modal model for similarity search. Further, we propose a way to construct a meaningful training data set to learn the relevant similarities between product pairs. Extensive experiments show that a transformer based language model combined with Siamese convolutional neural networks outperform competitive baseline models.