DeepAiNet: Deep NLP-based Representations for a Generalizable Anime Recommender System

Traditionally, recommendation systems require a long history of user-item interactions in the form of a large preference matrix to perform well, making themimpractical without large datasets. We aim to build a successful content-driven recommendation system that takes a hybrid ground between collaborative filtering (CF) approaches based off of a preference matrix, and nearest neighbor approaches based off of self-supervised embeddings. Specifically, we develop a deep learning, NLP-based anime recommender system named DeepAniNet on top of representations of anime shows called anime2vec. We explicitly train our model to reconstructuser-anime relevance scores, for shows with few or zero interactions. Our goal is to demonstrate that deep NLP approaches can extract rich content features to improve both a recommender system's performance and ability to generalize to new users and anime.