Large-scale product classification via text and image-based signals using a fusion of discriminative and deep learning-based classifiers

Typically, 810% of product URLs in ecommerce sites are misclassified. Sreeni Iyer and Anurag Bhardwaj discuss a machine-learning-based solution that relies on an innovative fusion of classifiers that are both text- and image-based, along with human touch to handle edge cases, to automatically classify product URLs according to a canonical taxonomic organization with a high F-score.

UX on ecommerce sites typically requires users to navigate the taxonomy and arrive at the leaf-node in the taxonomy (i.e., the shelf) and then onto product URLs within that shelf. Studies have shown that typically 8–10% of product URLs available at the shelf are misclassified. Given the lack of standardization across ecommerce merchants in terms of taxonomic morphology, what if the problem is about classifying into a canonical taxonomy? Can we use machines to solve these problems? Sreeni Iyer and Anurag Bhardwaj explain how Quad Analytix has effectively solved this business-critical problem at scale, without breaking the bank. Sreeni and Anurag share Quad Analytix’s real-world approach, which uses machine-learning discriminative techniques (SVM modeling) to highlight signals buried within titles, breadcrumbs, and other text, deep learning techniques (ConvNet) to highlight signals in the product image associated with the URL, and innovative classifier fusion to deliver the F-score that the SLA demands. Sreeni and Anurag then discuss results observed in the real world, along with the physical infrastructure (Amazon AWS) this translates to, and also outline where human touch is needed, including the art of arriving at an appropriate “canonical” taxonomy and using the crowd to handle long tail data. This includes edge cases where the machine fails to get it right—where the response is considered a likely misclassification or a failure to classify (i.e., responses from the system have a score below an acceptable threshold).