Black-box explainer for time series forecasting
A method, system, and computer program product for an interpretable, feature-based post-hoc black box explainer for univariate time series forecasters are provided. The method receives a set of time series forecasting predictions. The set of time series forecasting predictions are generated from a set of black-box models trained with an initial data set. The method generates a set of features based on at least a portion of the initial data set. A set of surrogate models are trained based on the set of time series forecasting predictions and at least a portion of the set of features. A subset of surrogate models is selected. Based on the subset of surrogate models, the method generates one or more explanation outputs for time series forecasting predictions of the set of black-box models.
1 . A computer-implemented method, comprising:
receiving a set of time series forecasting predictions, the set of time series forecasting predictions generated from a set of black-box models trained with an initial data set;
generating a surrogate data set by backtesting one or more time series forecasting predictions of the set of time series forecasting predictions, the surrogate data set including a surrogate training set and a surrogate testing set;
generating a set of features based on at least a portion of the initial data set;
training a set of surrogate models based on the set of time series forecasting predictions and at least a portion of the set of features, wherein the at least a portion of the set of features comprises one or more features selected from the group consisting of lag features, seasonal lag features, rolling window features, expanding window features, date features, time features, encoding cyclical features, holiday features, and trend features;
selecting a subset of surrogate models; and
based on the subset of surrogate models, generating one or more explanation outputs for time series forecasting predictions of the set of black-box models.
2 . The method of claim 1 , wherein generating the set of features based on the portion of the initial data set further comprises:
identifying a global feature set for a training data set, the training data set being at least a portion of the initial data set subject to one or more data perturbation operations; and
selecting a subset of features from the global feature set as the set of features.
3 . The method of claim 1 , wherein training the set of surrogate models based on the set of time series forecasting predictions further comprises:
training the set of surrogate models with the surrogate training set; and
evaluating the set of surrogate models with the surrogate testing set.
4 . The method of claim 3 , wherein training the set of surrogate models further comprises:
fitting the set of surrogate models using a tree-based regressor based on a set of feature vectors associated with a set of time points associated with the surrogate training set.
5 . The method of claim 1 , wherein generating one or more explanation outputs further comprises:
aggregating explanation outputs of each surrogate model of the subset of surrogate models.
6 . The method of claim 1 , wherein the one or more explanation outputs explain a mean and a prediction interval for the set of time series forecasting predictions of the set of black-box models based on a relative contribution of each feature to a specified black-box model.
7 . A system, comprising:
one or more processors; and
a computer-readable storage medium, coupled to the one or more processors, storing program instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
receiving a set of time series forecasting predictions, the set of time series forecasting predictions generated from a set of black-box models trained with an initial data set;
generating a surrogate data set by backtesting one or more time series forecasting predictions of the set of time series forecasting predictions, the surrogate data set including a surrogate training set and a surrogate testing set;
generating a set of features based on at least a portion of the initial data set;
training a set of surrogate models based on the set of time series forecasting predictions and at least a portion of the set of features, wherein the at least a portion of the set of features comprises one or more features selected from the group consisting of lag features, seasonal lag features, rolling window features, expanding window features, date features, time features, encoding cyclical features, holiday features, and trend features;
selecting a subset of surrogate models; and
based on the subset of surrogate models, generating one or more explanation outputs for time series forecasting predictions of the set of black-box models.
8 . The system of claim 7 , wherein generating the set of features based on the portion of the initial data set further comprises:
identifying a global feature set for a training data set, the training data set being at least a portion of the initial data set subject to one or more data perturbation operations; and
selecting a subset of features from the global feature set as the set of features.
9 . The system of claim 7 , wherein training the set of surrogate models based on the set of time series forecasting predictions further comprises:
training the set of surrogate models with the surrogate training set; and
evaluating the set of surrogate models with the surrogate testing set.
10 . The system of claim 9 , wherein training the set of surrogate models further comprises:
fitting the set of surrogate models using a tree-based regressor based on a set of feature vectors associated with a set of time points associated with the surrogate training set.
11 . The system of claim 7 , wherein generating one or more explanation outputs further comprises:
aggregating explanation outputs of each surrogate model of the subset of surrogate models.
12 . The system of claim 7 , wherein the one or more explanation outputs explain a mean and a prediction interval for the set of time series forecasting predictions of the set of black-box models based on a relative contribution of each feature to a specified black-box model.
13 . A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions being executable by one or more processors to cause the one or more processors to perform operations comprising:
receiving a set of time series forecasting predictions, the set of time series forecasting predictions generated from a set of black-box models trained with an initial data set;
generating a surrogate data set by backtesting one or more time series forecasting predictions of the set of time series forecasting predictions, the surrogate data set including a surrogate training set and a surrogate testing set;
generating a set of features based on at least a portion of the initial data set;
training a set of surrogate models based on the set of time series forecasting predictions and at least a portion of the set of features, wherein the at least a portion of the set of features comprises one or more features selected from the group consisting of lag features, seasonal lag features, rolling window features, expanding window features, date features, time features, encoding cyclical features, holiday features, and trend features;
selecting a subset of surrogate models; and
based on the subset of surrogate models, generating one or more explanation outputs for time series forecasting predictions of the set of black-box models.
14 . The computer program product of claim 13 , wherein generating the set of features based on the portion of the initial data set further comprises:
identifying a global feature set for a training data set, the training data set being at least a portion of the initial data set subject to one or more data perturbation operations; and
selecting a subset of features from the global feature set as the set of features.
15 . The computer program product of claim 13 , wherein training the set of surrogate models based on the set of time series forecasting predictions further comprises:
training the set of surrogate models with the surrogate training set; and
evaluating the set of surrogate models with the surrogate testing set.
16 . The computer program product of claim 15 , wherein training the set of surrogate models further comprises:
fitting the set of surrogate models using a tree-based regressor based on a set of feature vectors associated with a set of time points associated with the surrogate training set.
17 . The computer program product of claim 13 , wherein the one or more explanation outputs explain a mean and a prediction interval for the set of time series forecasting predictions of the set of black-box models based on a relative contribution of each feature to a specified black-box model and generating the one or more explanation outputs further comprises:
aggregating explanation outputs of each surrogate model of the subset of surrogate models.