Product Image Representation Learning on Large Scale Noisy Datasets

Learning product similarity using distance metric learning from real world catalog needs to take care of large number of product categories and noisy labels. On one hand, large number of product categories makes online hard mining (OHM) less effective as hard triplets become sparse and thus difficult to find. On the other hand, the validity of the hard-triplets themselves is less certain in the case of noisy labelled training data. In this paper, we address the problem of large-scale product representation learning in the presence of noisy training data. To address these challenges, we propose a novel co-teaching based label correction scheme for distance metric learning, that is motivated by the inconsistencies of variations relationships in the product catalog. To validate our approach, we conducted experiments on 20 different product categories, where we achieve up to 4% improvement in PR-AUC compared to the SOTA baseline and conclude by discussing the durable learnings we gained from these experiments and directions for future research.