Program Popularity Prediction Approach for Internet TV Based on Trend Detecting

Predicting program popularity precisely and timely is of great value for content providers, advertisers, as well as Internet TV operators. Existing prediction methods usually need large quantity of samples and long training time, while the prediction accuracy is poor for the programs that experience a high peak or sharp decrease in popularity. This paper presents our improved prediction approach based on trend detecting. First, we apply a dynamic time warping (DTW) distance based k-medoids algorithm to group programs popularity evolution into 4 trends. Then, 4 trend-specific prediction models are built separately using random forests (RF) regression. According to the features extracted from electronic program guide (EPG) and early view records, newly published programs are classified into the 4 trends by a gradient boosting decision tree. Finally, combining forecasting values from the trend-specific models and classification probability, our proposed approach achieves better prediction results. The experimental results show that, compared to the existing prediction models, the prediction accuracy can increase more than 20%, and the forecasting period can be effectively shortened.