Survival Prediction for Ad Creative Discontinuation

Abstract

In this study, we develop a prediction framework for ad creative discontinuation. Our framework consists of deep neural networks (DNN) take the text, category, image and numerical information of the targeted ad. It estimates the appropriate discontinuation time with a survival prediction strategy. Here, we propose two simple but extremely effective techniques to enhance the prediction performance; (1) sales-based loss function, and (2) two-period estimation. The former considers the importance of ads by weighting the loss function according to the click-through rate (CTR). The latter separately estimates ``short-term censoring'' as a short-term discontinuation and ``long-term wear-out'' as a long-term discontinuation, taking into account these different properties. We evaluated our framework using the real-world 1,000,000 ad creatives provided Gunosy Inc. with a concordance index (CI). Our proposal multi-modal DNN-based framework performed better than the conventional method. Our two-period estimation largely improved the prediction performance by approximately 20 point on both short-term and long-term discontinuation. The introduction of sales-based loss further improves performance by an average of approximately 3 point in CI.