Personalised Visual Art Recommendation by Learning Latent Semantic Representations

TL;DR

This paper presents a personalized visual art recommendation system that learns interpretable semantic representations from textual descriptions, outperforming visual feature-based methods.

Contribution

The authors introduce a latent semantic learning approach using LDA on textual descriptions for explainable and effective visual art recommendations.

Findings

LDA-based semantic representations outperform DNN visual features in recommendations.

The approach uncovers non-obvious semantic relationships between paintings.

The method provides more interpretable recommendations.

Abstract

In Recommender systems, data representation techniques play a great role as they have the power to entangle, hide and reveal explanatory factors embedded within datasets. Hence, they influence the quality of recommendations. Specifically, in Visual Art (VA) recommendations the complexity of the concepts embodied within paintings, makes the task of capturing semantics by machines far from trivial. In VA recommendation, prominent works commonly use manually curated metadata to drive recommendations. Recent works in this domain aim at leveraging visual features extracted using Deep Neural Networks (DNN). However, such data representation approaches are resource demanding and do not have a direct interpretation, hindering user acceptance. To address these limitations, we introduce an approach for Personalised Recommendation of Visual arts based on learning latent semantic representation of paintings. Specifically, we trained a Latent Dirichlet Allocation (LDA) model on textual descriptions of paintings. Our LDA model manages to successfully uncover non-obvious semantic relationships between paintings whilst being able to offer explainable recommendations. Experimental evaluations demonstrate that our method tends to perform better than exploiting visual features extracted using pre-trained Deep Neural Networks.