DEEPSENSE: A Deep-Learning Predictive Tool for Evaluating Effectiveness of Video Commercials

Testing and validation of television commercials require sophisticated audience response collection schemes at various production and deployment stages of a commercial’s lifecycle including script writing, actor casting and footage shooting. Audience response at these stages allows interactive adjustment to the production and deployment processes which increases the effectiveness of the television commercials. Current testing advertising approaches include manual annotation, focus group studies and audience/customer opinion surveys. Besides their high operational cost, these approaches for collecting audience response data are usually time-consuming and lack immediacy.

We propose a computational “media cognition” tool that uses deep neural networks on video, audio and natural language content of commercials as training data and the audience response attributes as the training labels to form a predictive analysis framework for video commercials. Our data source is YouTube video commercials. The audience response data include the manual annotations as well as YouTube likes/dislikes and comments.

We find that the direct content modeling system performs better when we combine the video, audio, and text dimensions of the content than employing any of these dimensions alone. For predictive analysis based on a single dimension, the video stream has the highest predictive power. The central implication of our study is that firms can train deep learning algorithms directly on the content of existing commercials and use them to reliably predict audience response of unreleased media content.

Keywords: deep learning, advertising effectiveness, multimedia content analysis, long short-term memory, convolutional neural network