Data-driven approaches for discovery and prediction of user-preferred picture settings on smart TVs

Abstract

We discover user-preferred picture settings on smart TVs and investigate whether it is possible to predict the users' picture setting preferences through machine learning methods. We first perform K-means clustering on large-scale smart TV usage log data to understand how users fine-tune the factory default picture settings. Clustering results recognize 3-4 user groups who have reasonably different preferences toward the default settings. By characterizing these user preferences, we come up with new user-preferred picture settings. We perform an in-depth analysis of the newly discovered picture settings regarding diverse TV device characteristics. We also perform lab experiments to demonstrate how these new settings deliver different picture quality than the default. Next, we construct a deep learning-based classifier that learns and predicts the picture setting preferences of the users. The final trained model shows 86% accuracy in predicting users' decisions to choose a specific picture setting out of four available options.