Selecting Creators to Sign on a Content-Sharing Platform: A Deep-DiD Approach

Abstract

A large video-sharing platform introduced a "Creator Signing Program'" aimed at signing creators and motivating them to generate more high-quality video content on the platform. Leveraging a matched dataset from the platform, we employ Difference-in-Difference (DiD) analyses to demonstrate the significant positive impact of the signing program on signed creators' performance, measured by the number of uploaded videos as well as the total user time and user engagement contributed by the creators' videos. More importantly, we propose a novel Deep-DiD model that combines deep neural networks with DiD to estimate the individual-level heterogeneous treatment effects of the signing program. Based on the estimated individual-level treatment effects as a function of creators' pre-treatment characteristics, the platform can optimize creator selection by selecting creators with the highest estimated treatment effects. Comparing creators selected using our Deep-DiD model to those selected by the platform, we show that the former have significantly higher estimated treatment effects and experience substantially larger actual performance jumps. Lastly, we demonstrate the importance of incorporating unstructured data (visual and audio features) in the model.