Debiasing Recommender Systems: Applying DANCER to Neural Collaborative Filtering Models

In recent years, the demand for improving the performance of recommender systems (RSs) has become increasingly important due to the exponential growth of big data. Many prediction models have been developed and further optimized, but improvements based on several traditional models are still needed because dynamic user preferences and selection bias are seldom considered. In this paper, we use a semi-synthetic dataset based on MovieLens-Latest-small and apply a deep-learning model — Neural Collaborative Filtering (NCF) — to the DANCER debiasing method. We then evaluate the performance of NCF models with different propensities and neural architecture. Although preliminary experimental results don’t exceed those obtained by the DANCER-TMF model, we focus on the sensitivity analysis of NCF and provide guidance for further tuning to investigate how well NCF pairs with DANCER.