A Weight-Sharing Gaussian Process Model Using Web-Based Information for Audience Rating Prediction - Technologies and Applications of Artificial Intelligence - page 197

A Weight-Sharing Gaussian Process Model Using
Web-Based Information for Audience Rating Prediction
Yu-Yang Huang 1 , Yu-An Yen 1 , Ting-Wei Ku 1 , Shou-De Lin 1 ,
Wen-Tai Hsieh 2 , and Tsun Ku 2
1 Dept. of Computer Science and Information Engineering,
National Taiwan University, Taipei, Taiwan
2 Institute for Information Industry, Taipei, Taiwan
{myutwo150,lovelove6402,martin79831}@gmail.com,
[e-mail],
{wentai,cujing}@iii.org.tw
Abstract. In this paper, we describe a novel Gaussian process model for TV
audience rating prediction. A weight-sharing covariance function well-suited
for this problem is introduced. We extract several types of features from Google
Trends and Facebook, and demonstrate that they can be useful in predicting the
TV audience ratings. Experiments on a dataset consisting of daily dramas show
that the proposed model outperforms the other conventional models given the
same feature set.
Keywords: Time Series, Gaussian Process, Audience Rating Prediction.
1 Introduction
Time series analysis is an active research area with many real-world applications
including price forecasting [12] and sales prediction [9]. A typical method relies on
historical data sequences to predict upcoming data points. In this paper, we focus on
television audience rating prediction. The goal is to accurately predict the rating of
an upcoming TV episode, and to analyze the most crucial factors that cause the
audience ratings to fluctuate.
Gaussian Process Regression (GPR) [16] is used to predict the audience ratings.
We analyze variants of GPR models, and propose a weight-sharing kernel to deal with
the overfitting issue caused by the increasing number of hyperparameters. The expe-
riments show that, with the weight-sharing technique applied, our GPR model outper-
forms the other competitors in predicting the audience ratings.
Furthermore, we propose three novel types of features to boost the prediction per-
formance. Different from previous works, our model relies not only on classic time
series features such as historical ratings, but also on features extracted from social
networks and search engines. For example, trend information from Google Trends,
opinion polarity and popularity information from Facebook are used in the prediction.
We show how such information can be adapted in the proposed GPR model.