Neuro-Symbolic Recommender Systems

Recommender systems technologies represent an essential component of commercial platforms and e-services (e.g., Amazon, Netflix, Spotify). Generally speaking, a recommendation system aims to provide suggestions for items based on user preferences. By suggesting novel items to users, recommender systems mitigate information overload. This problem occurs when a user is overwhelmed by the vast quantity of items the platform proposes and cannot efficiently decide which items to interact with. Despite being studied for over twenty years, recommender systems still suffer from important drawbacks that limit their applicability in real-world scenarios. These issues are data sparsity, cold-start, and explainability. Intuitively, these problems can be addressed by incorporating additional information, usually expressed in the form of content (e.g., movie genres, actors, directors) or contextual (e.g., location, user's mood) side information in the models. Despite the vast effort devoted to designing hybrid recommendation models that can integrate this information, current state-of-the-art approaches still struggle with these limitations. This is probably due to the advent of deep learning and the consequent vast use of neural network models to tackle the recommendation problem. In particular, it is well-known that neural networks require huge quantities of data for training and are not explainable due to black-box models. To this end, in this thesis, we propose to mitigate recommendation limitations using Neuro-Symbolic integration, which we believe represents an excellent candidate to incorporate knowledge into deep learning models. Specifically, Neuro-Symbolic computing is a new branch of artificial intelligence that aims to integrate (neural) learning and (symbolic) reasoning to obtain the best from both worlds. In particular, symbolic approaches can learn from a few examples (i.e., few-shot learning) and, in some cases, even in the absence of data (i.e., zero-shot learning). Then, symbolic approaches are explainable by design as they are usually implemented as a set of logical rules. Intuitively, when these capabilities are integrated with deep learning, it could be possible to obtain neural networks that can learn with fewer examples and whose predictions are less opaque. This motivates us to investigate the application of Neuro-Symbolic computing to recommendation systems. In particular, by leveraging the advantages of a Neuro-Symbolic architecture, it could be possible to obtain a recommendation system that can deal well with data sparsity, thanks to few-shot learning, mitigate cold-start cases, thanks to zero-shot learning, and eventually explain its predictions. In this thesis, we investigated the use of Logic Tensor Networks (LTNs) to implement our recommendation models. LTN is a Neuro-Symbolic framework that allows the satisfaction of a first-order logic knowledge base to be used as the loss function of a neural network. Motivated by the promises of this recent Neuro-Symbolic framework, we conducted preliminary experiments utilizing this framework to develop Neuro-Symbolic architectures tailored to the recommendation task. We specifically applied the logical reasoning capabilities of LTN to perform different kinds of logical regularization in recommendation systems. In particular, we investigated the application of Logic Tensor Networks to the hybrid recommendation task, knowledge transfer in recommender systems, and finally, the cross-domain recommendation task. The experiments we performed prove that our approaches have successfully exploited the advantages of a Neuro-Symbolic architecture to mitigate the aforementioned recommendation limitations.