Information Technology Reference
In-Depth Information
The main contributions of this work are summarized as follows:
•
We modify a standard kernel of GPR model to avoid overfitting, and make it more
suitable for the TV audience rating prediction problem.
•
We propose three novel types of web-based features: trend features, social network
features, and opinion features for better performance.
•
We conduct experiments to verify the validity of the model and features.
2
Related Work
Audience rating prediction is treated as a time series forecasting problem in the field of
statistics and data analysis. Well-known models such as autoregressive model, moving
average model, or the hybrid (ARIMA), or the more advanced ones such as generalized
autoregressive conditional heteroskedasticity (GARCH) [2] and the nonlinear extension
of it (NGARCH) [12] are all plausible models that can be applied. However, those mod-
els may not be the best choice for audience rating prediction because they do not consider
specific characteristic of the ratings. Researchers have shown that using time-based and
program-based covariates provides a more effective way to forecast the audience ratings
[5, 6] than general time series models, as these models consider correlations between
rating, genre, show duration, live status, etc.
Another serious drawback of general time series models is that they cannot consid-
er external features such as information from social media and search engine. Such
external and social information has been shown effective in forecasting. In [1], a work
using chatters from Twitter to predict future revenue of movies is proposed. The
works in [4, 8] propose to perform audience rating prediction utilizing the count of
posts and comments from social media. Our work further extends the idea to exploit
opinion mining and search engine such as Google Trends to enhance the prediction
performance.
3
Framework and Features
In this paper, the TV audience rating prediction problem is modeled as a supervised
learning task. To forecast near-future ratings, historical data together with the follow-
ing listed features are used as the training input for the models.
1.
Basic Time Series Features
Similar to other basic time series forecasting models, the ratings of the previous epi-
sodes, the rating of the first episode, and binary indicator variables corresponding to
weekdays are used as features.
2.
Social Network Features
Nowadays, TV companies often host “Fan Pages” on social networking sites such as
Facebook. On these pages, companies run promotional campaigns, face-to-face
events, polls, and provide previews of the next episodes. Also, it provides a platform
for the fans to interact with each other and show their support or oppose to the show.
Search WWH ::
Custom Search