Siamese Neural Networks for Content-based Visual Art Recommendation

Abstract

The global art has experienced a steady growth to tens of billion dollars in annual sales. The huge profits behind art trades unfortunately have been largely overlooked and rarely been studied in most of the machine learning and recommendation system (RS) research. As a popular Deep Metric Learning (DML) model, the Siamese Neural Network (SNN) has been widely used in music and other e-commerce RS, but not been used in art recommendation tasks. In this paper we propose an art similarity metric with SNN, and based on which built a content-based art RS, followed by clustering for reducing comparison numbers. Performance evaluation of the proposed SNN-based art RS has been conducted, in comparison with our original, simpler model basing on cosine similarity. Results shows that the SNN-based visual art RS performs significantly better in every experiment subgroup, is more robust with strong resistance to overfitting and confusion. Additional experiments show that it is nontrivial to further improve these recommendation results. To the best of our knowledge, this is the first visually-aware RS that took advantage of both SNN and content-based recommendation framework in visual art recommendation. We believe that this work opens wide opportunities for applying machine-learning and deep-learning techniques in the exciting area of visual art recommendation.