Automatic Product Classification and Clustering Solutions in a Retail Context

In this report we propose a methodology to automatically classify products and cluster similar products together. This enhances user interaction and performance metrics for e-commerce (product shopping) websites, with applications in product search, site navigation, product comparison, etc. For this project, a variant of the multinomial Naive Bayes algorithm was used for classification. We present our findings of key aspects that demonstrate its accuracy. A clustering methodology was developed to handle large data-sets, overcoming the N 2 complexity, in a two-step approach. The first step of clustering utilizes Locality Sensitive Hashing, the second step uses a Shingles matching method. We describe a methodology that allows trading cluster-size versus similarity of items within a cluster and propose a way to use clustering to increment classification accuracy.