Brand recommendation leveraging heterogeneous implicit feedbacks

Collaborative Filtering (CF) is one of the most commonly used and successful recommendation techniques. The performance of traditional collaborative filtering is limited by the sparsity of data, and relies on explicit user feedbacks (ratings). However, the growing number of products and users will leave the user-item matrix sparser, and high-quality explicit feedbacks are not always available. In this paper, we focus on brand recommendation. A brand represents a set of products, which can ease the sparsity problem to some extent. We propose a learning-based recommendation approach, leveraging heterogeneous implicit feedbacks. First, we describe relations between brands by using the sequence of different types of behaviors, and extract path-based features to connect users to their unknown brands. Then we apply machine-learning approaches to predict the purchase probability using these features. Finally we recommend top-n ranked brands to users according to the predicted probability. Extensive experiments on large-scale real-world e-commerce data provided by Tmall.com demonstrate the effectiveness of our approach.