Information Technology Reference
In-Depth Information
each drama, a general scenario is considered, where all available features are used in
the prediction. With the capability to estimate the relative importance of different
groups of features and to avoid overfitting, the proposed weight-sharing method out-
performs the other standard GP-based models. Results are shown in Table 4.
Table 4.
Average MAPE using all features
Drama
D1
D2
D3
D4
Avg.
Model
GP_ard
0.1015
0.0887
0.1001
0.1184
0.1022
GP_iso
0.1018
0.0833
0.0929
0.0973
0.0938
Our Model
0.0991
0.0794
0.0911
0.0869
0.0891
5.3
Feature Analysis
In this section, we study the usefulness of the features based on our proposed model.
As previously mentioned, the features are categorized into six types, as shown in the
first column of Table 5. Holding the rest of the features identical, we compare the
performance with and without a certain type of features. If the resulting error is lower
when a certain type of features is used, we define it as a “win”. Conversely, if the
error is higher, then we define it as a “lose”. For instance, with all other conditions the
same, if removing Facebook features results in a higher MAPE for D1, then a “lose”
is assigned. Since there are total 63 different combinations of features, 31 compari-
sons are made for each drama. The winning percentages
100%
for each
type of the features are shown in Table 5. The higher the winning percentage, the
more useful it is. We can observe that the previous ratings and weekday information
are overall the most important features, while most of the features except opinion
feature generally improves the performance.
Table 5.
Winning percentage with or without a certain type of features
Winning Percentage (%)
D1
D2
D3
D4
TOTAL
Opinion
26
58
42
55
45
Google Trends
58
35
74
77
61
Facebook
29
71
87
55
60
Ratings of previous three episodes
100
100
100
100
100
Rating of the first episode
71
61
97
42
68
Weekday
81
100
100
84
91
Search WWH ::
Custom Search