Domain-Specialized Artworks Auction Price Prediction via Model-Level Mixtures-of-Experts

Mixture-of-Experts (MoE) models applied to the artworks auction price regression task can provide a novel structured approach to combining specialized models, improving prediction performance over non-ensemble-based pipelines while offering greater interpretability over how models and human expert critics perform across distinct artistic domains. In this work, we investigate model-level mixtures of heterogeneous regressors and analyze their suitability for improving both performance and interpretability in artworks price prediction. We propose an MoE architecture that partitions training data into semantically meaningful overlapping domains, defined along axes such as year of creation, artwork surface area, medium (e.g., paintings), materials (e.g., acrylic) and the geographic continent of the auction house, and trains a dedicated expert per domain. Different strategies are tried for the final gating mechanism, to combines expert contributions weighted by their relative validation set performance. The experts themselves range from interpretable linear regressors to advanced neural networks trained on multimodal data, including tabular auction features, artwork images and textual descriptions, enabling the framework to capture both structured and perceptual signals. Beyond predictive accuracy, the domain-based decomposition yields a dual interpretability benefit: it not only reveals which types of artworks are harder to price algorithmically, but more importantly exposes the systematic blind spots of human art critics, whose estimates serve as the primary expert baseline in this work, clarifying in which artistic domains their and our final model’s valuations are most and least reliable.