Explainable Digital Creatives Performance Monitoring using Deep Feature Attribution

A key challenge in marketing and advertising research is understanding when and why digital assets such as promotional content perform well during a marketing push. By leveraging raw image feature vectors extracted from large datasets, we can train performance prediction models using online social signals such as likes or views. While the resulting models make accurate predictions, they are opaque and rely on abstract features within the model, making attribution almost impossible. This paper demonstrates an approach to performance prediction modelling for image based digital creative assets. Utilising a combination of pre-trained vision model embeddings with a pipeline of generative Artificial Intelligence (AI) for image synthesis and manipulation, we establish a means of determining the performance of explainable components. This enables flexible performance prediction, even with smaller datasets, with high degree of explainability through the attribution of image features correlating with high or low performance.