Collaborative Embedding Features and Diversified Ensemble for E-Commerce Repeat Buyer Prediction

In online advertisement, after sales promotion, it is important to predict which buyers will return and become loyal repeat buyers. Given the action logs of users, brand and category information of items, and user profiles, we study the problem of repeat buyer prediction on E-commerce data, which aims to predict whether a new buyer of a merchant is a one-time deal hunter or will become a loyal repeat customer. We develop a set of useful features to capture the underlying structure of the data, including features regarding users, merchants, categories, brands, items and their interaction. We also propose to learn collaborative features with embeddings, which represent users and merchants in a shared feature space. We use logistic regression, gradient boosting decision trees, and factorization machines as individual classification models. We develop a diversified ensemble model to combine different feature sets and models. Our solution obtained AUCs of 70.4762% in stage 1 and 71.0163% in stage 2, ranked 2nd and 3rd places respectively in IJCAI 2015 Repeat Buyer Prediction Competition.