Weighted Neural Collaborative Filtering: Deep Implicit Recommendation with Weighted Positive and Negative Feedback

Being able to generate personalized recommendations is a widespread objective in (online) retail. The focus of this research is to estimate the relevance of user-item combinations based on previous interactions using implicit feedback. We do this in a situation where interactions are often repeated, focusing on new ones. We bring two weighting schemes of positive and negative implicit feedback together into a single Weighted Matrix Factorization (WMF) model to handle the uncertainty associated with implicit preference information. Next, we bring the concept of these weighting schemes to a Deep Learning framework by introducing a Neural Weighted Matrix Factorization model (NeuWMF). We experiment with different weights, loss functions, and regularization terms, and evaluate both models using purchase data from an online supermarket. Our WMF model with both weighted positive and negative feedbacks gives superior performance in terms of NDCG and HR over regular WMF models. Even better results are obtained by our NeuWMF model, which is better capable of capturing the complex patterns behind item preferences. Especially the weighting of positive terms gives an extra boost compared to the state-of-the-art NeuMF model. We confirm the practical use of our model results in an experiment on real customer interactions.