On the Effectiveness of Convolutional Autoencoders on Image-Based Personalized Recommender Systems

Over the years, the success of recommender systems has become remarkable. Due to the massive arrival of options that a consumer can have at his/her reach, a collaborative environment was generated, where users from all over the world seek and share their opinions based on all types of products. Specifically, millions of images tagged with users’ tastes are available on the web. Therefore, the application of deep learning techniques to solve these types of tasks has become a key issue, and there is a growing interest in the use of images to solve them, particularly through feature extraction. This work explores the potential of using only images as sources of information for modeling users’ tastes and proposes a method to provide gastronomic recommendations based on them. To achieve this, we focus on the pre-processing and encoding of the images, proposing the use of a pre-trained convolutional autoencoder as feature extractor. We compare our method with the standard approach of using convolutional neural networks and study the effect of applying transfer learning, reflecting how it is better to use only the specific knowledge of the target domain in this case, even if fewer examples are available.