PREDICTING EXHIBITION SUCCESS WITH NEURAL NETWORKS

Abstract [English]

The forecast of art and cultural exhibition success has become more of a concern to galleries, museums, and creative institutions aiming at data-driven methods to increase audience attendance, utilize resources to capacity and improve financial performance. Conventional methods of statistics do not provide much potential to model multifaceted, nonlinear relationships between the various factors, including visitor behaviour, marketing impact, social sentiments, and exhibit properties. In this paper, a neural-network-based predictive model is suggested to incorporate heterogeneous data (logs of ticketing, demographic factors, social media analytics, the performance of promotional channels and post-event survey) and predict the most important success metrics (volume of attendance, revenue, media exposure, and satisfaction of the visitors). The conceptual model relates the measurable exhibition features to the latent patterns, which are learned using multilayer perceptrons (MLP), convolutional neural networks (CNN), and long short-term memory (LSTM) models, making it the possible learning of structured variables in addition to the temporal or textual ones. The vast use of feature engineering is aimed at deriving meaningful indicators out of unstructured text, engagement metrics, and history. Adam, SGD and RMSprop are all optimized models and systematic tuning of hyperparameters and k-fold validation are all exploited. The findings indicate that neural networks are superior to the classic regression models especially in modeling nonlinear behavioral dynamics, the effects of social influences and temporal effects in visitor interest.