Explainability and Interpretability of Neural Network and Factorization Models of Recommender Systems

The thesis explores various methods for constructing recommender systems for a task with implicit feedback and reducible to it with explicit feedback and possible methods for their interpretation. In addition to classical factorization machines, neural network methods are considered —— Neural Collaborative Filtering (NCF) and Proto-MF. In particular, for each model, the representation of users and items is studied in the form of latent factors obtained when constructing the model. A way to compare the degree of interpretability of models is proposed by assessing how well the hidden parameters of the model are predicted using models based on interpretable features. An algorithm is also proposed for identifying prototypes of users and items based on embeddings obtained with the NCF model as one of the interpretation methods.