Paint Price Prediction Using a Triplet Network-Multimodal Network-LSTM Combined Deep Learning Approach

Painting prices fluctuate over time, as does the performance of the art market. Obtaining an accurate art price prediction by exploring the dynamic change pattern of artwork painting prices remains a challenge. We propose a paint price prediction using a triplet network-multimodal network-LSTM combined deep learning approach for dynamically predicting paint prices. By using triplet sets of paintings classified by similar transaction prices, our painting price classification model can be adjusted to identify paintings with similar prices over time. After that, we use the multimodal painting price prediction model to predict the set of paintings with similar prices within specified time intervals and to determine the raw simulated painting transaction prices within those intervals. By time sorting the simulated prices of the paintings, the simulated time series data of the paintings can then be obtained, and from this, a model of long short-term memory networks is used to predict the dynamic predicted prices of the paintings. The results of this experiment demonstrate that the designed model had a 5.7% and 10.3% decrease in RMSE and MAE for 110 paintings compared to the prediction case using the model based on image and numerical information. Thus, the method is more effective at predicting painting prices and can reveal better dynamic patterns in painting prices.