Deep Learning Embeddings for User and Item Profiling in Recommender Systems

In the presented work, we focus on the development of the recommender system for stylistically compatible products, such as furniture and decor. The proposed model consists of a neural network for extraction of the visual features and neural collaborative filtration model. The former part captures the stylistic properties of an object, while the latter one uses the obtained features to form the recommendations. The goal is to offer the user products that best complement the ones already purchased by him both in terms of functionality and style. We proposed a method based on Siamese neural networks for fine-tuning the pre-trained net and enforcing it to focus on the extraction of style-related features. This approach was compared with a standard fine-tuning in the classification task. The experiments were conducted on the "Home and Kitchen" subset of an open dataset "Amazon product data". For fine-tuning the feature extraction network, an additional IKEA product dataset containing information about the stylistic compatibility of interior items was also introduced. The quality of the recommendations obtained with the proposed stylistic-aware approach was quantitatively and qualitatively compared with the result of standard neural collaborative filtration, which takes into account only information on user-item interactions.