Multimodal Approach for Painting Price Prediction

The art market's subjective and unpredictable nature has led to a growing interest in predicting the prices of paintings using machine learning models. This study aims to explore a multimodal approach combining CNN, BERT, and regular NN to predict painting prices based on diverse inputs, including image, text, and categorical data. The dataset was collected from Sotheby's art marketplace website, consisting of various features such as image, dimensions, movement, period, material, condition, and price. Results from the model evaluation show that the multimodal approach outperformed single-modal models, achieving a MAPE loss of 30.66% and an MAE loss of 3348.73. The integration of numerical, categorical, image, and text data enhanced the model's ability to capture the complexities involved in painting valuation. The positive impact of the multimodal approach lies in its ability to enhance prediction accuracy and capture the relation of many factors affecting painting prices, ultimately providing valuable insights and tools for the art market.