Enhancing recommendation with causal embedding: Considering social network influence

In the realm of recommendation models, we consistently rely on observational interaction data. This data encompasses a variety of aspects, such as user conformity and genuine user interests. The key challenge for recommender systems is to extract a user's authentic interests from this interaction data in order to provide accurate recommendations. The current method, DICE, attempts to separate conformity and interest by assigning distinct embeddings for each to users and items. The method ensures that each embedding captures only one causal factor through training with specific causal data. In our research, we've enhanced this existing method by incorporating social networks into the disentanglement of conformity and real interest from observational interaction data. The results from our proposed method surpass those of the prevailing baseline, demonstrating significant improvements across various backbone models using a real dataset. Furthermore, we conducted a sensitivity analysis and provided recommendations for scenarios in which our new model would be most effective.