Neuro-Symbolic Cross-Domain Recommendation

Recommender Systems have nowadays become ubiquitous but continue to face challenges such as the cold start problem and lack of explainability. This work tries to address both issues by proposing a Neuro-Symbolic, cross-domain recommendation model. Specifically, we designed a two-stage architecture which trains a Matrix Factorization model through a Neuro-Symbolic AI framework called Logic Tensor Networks. By defining a set of logic rules in Real Logic, a fuzzy, differentiable first-order logic language, the framework enables training neural networks in an explainable manner, using gradient-based optimization to maximize the satisfaction of the rules. Additionally, we extracted from a Knowledge Graph a set of semantic connections between recommendation domains, allowing the model to transfer users’ preferences from the source domain to the target domain with the goal of mitigating the cold start issue. Such connections could then be used to provide explanations to the users about the recommendations produced by the model. However, our models underperformed compared to the baseline BPR-MF models, likely due to the simplicity of the adopted knowledge transfer method. Future work could improve performance by employing more advanced techniques, such as Knowledge Graph Embedding models for richer representations or integrating Large Language Models to enhance the semantic extraction and entity linking processes.