Catlinks - a category clustering algorithm based on multi-class regression

This paper proposed CatLinks, a category clustering algorithm based on multi-class regression. In recommender systems for e-commerce web sites, users' experience of recommendations highly relies on the diversity of purchase suggestions. Taking inexpensive training data as products' literal information and their categories, CatLinks extracts latent features of categories and construct presentation of them as vectors. With vector presentation categories can be clustered by similarity measure and aggregation methods such as KNN or K-Means. Algorithm of CatLinks is based on training of a multi-class category predictor of products. After the predictor is trained, its weight matrix is taken as feature vectors of categories. With similarity of categories, recommender system can suggest users to purchase products from extended categories, when their interest on a certain category is discovered. Through our experiments on Alibaba's product and order dataset, CatLinks is proved a novel method to predict category co-occurrence of user's joint orders.