In order to understand how ML algorithms arrive at a certain output value (forecast), it can be helpful to determine the importance of
features used by the algorithm. The here introduced method, called SHAP, to determine feature importance for
complex ML algorithms is based on mathematical game theory. The method is capable of providing the feature importance
for the prediction of a single data point (e.g. the next one to predict) at a time and is hence a local method. The importance
assigned to a single feature can vary from data point to data point (from one day to predict to another day to predict).
In plain words, the method tries to answer how much a prediction was driven by the fact that a feature has a
certain value, instead of some baseline value.
As an example, the table below lists all 15 features with their current values. In bold you can see for each feature how much
the current prediction was driven by the fact that this feature currently has the shown value.
|
Avg_total = 104.2
-0.6
|
Avg_month = 80.2
-2.8
|
1_day_ago = 104
0.1
|
4_days_ago = 108
0.3
|
1_week_ago = 106
1.0
|
|
Avg_year = 87.3
1.0
|
Avg_week = 79.6
-3.7
|
2_days_ago = 74
-1.3
|
5_days_ago = 79
0.4
|
1_month_ago = 100
0.5
|
|
Avg_quarter = 82.7
-0.6
|
Avg_weekday = 116.4
7.3
|
3_days_ago = 78
-0.1
|
6_days_ago = 133
0.5
|
1_year_ago = 90
-4.0
|