Recommendation Versus Regression Neural Collaborative Filtering

Abstract

Neural Collaborative Filtering recommendations are traditionally based on regression architectures (returning continuous predictions, e.g. 2.8 stars), such as DeepMF and NCF. However, there are advantages in the use of collaborative filtering classification models. This work tested both neuronal approaches using a set of representative open datasets, baselines, and quality measures. The results show the superiority of the regular regression model compared to the regular classification model (returning discrete predictions, e.g. 1–5 stars) and the binary classification model (returning binary predictions: recommended, non-recommended). Results also show a similar performance when comparing our proposed recommendation neural approach with the state-of-the-art neural regression baseline. The key issue is the additional information the recommendation approach provides compared to the regression model: While the regression baseline only returns the recommendation values, the proposed recommendation model returns $\langle$value, probability$\rangle$ pairs. Extra probability information can be used in the recommender systems area for different objectives: recommendation explanation, visualization of results, quality improvements, mitigate attack risks, etc.