Building a 1500-Class Listing Categorizer from Implicit User Feedback

In peer-to-peer used goods marketplaces like letgo, accurate fine-grained item categorization is a necessary step for ensuring a smooth buyer and seller experience. The cost of building such a system is typically dominated by the cost of obtaining the millions of annotated examples necessary to train the Machine Learning models.

In this talk we present our approach for building an item categorizer using implicit feedback from buyers performing searches in our marketplace, as well as some tricks we learned in the process. This approach allows us to (1) train our models with millions of examples at a fraction of the cost we would incur were we relying on manual annotations and (2) build a taxonomy of classes that is relevant to our users.