Visual product recommendation using neural aggregation network and context gating

In this paper we focus on the problem of user interests' classification in visual product recommender systems. We propose the two-stage procedure. At first, the visual features are learned by fine-tuning the convolutional neural network, e.g., MobileNet. At the second stage, we use such learnable pooling techniques as neural aggregation network and context gating in order to compute a weighted average of image features. As a result we can capture the relationships between the products images purchased by the same user. We provide an experimental study with the Amazon product dataset. It was shown that our approach achieves a F1-score of 0.90 for 15 recommendations, which is much higher when compared to 0.66 F1-measure classification of traditional averaging of the feature vector.