InterpretML by Microsoft is designed with the aim of expanding **interpretability **of machine learning models. In other words, make those models easier to understand and ultimately facilitate human interpretation.
Microsoft’s Interpret-Community is an extension of this repository, which includes additional interpretability techniques.
In particular, one useful feature is what is called the **MimicExplainer. **This is a type of global surrogate model that allows for interpretability of any black box model.
Background
In this example, the MimicExplainer is used in interpreting regression models built using SVM (support vector machines) and XGBRegressor (XGBoost for regression problems).
Specifically, these two models are used as follows:
- SVM is used for predicting the **average daily rate **of a customer using specified features, such as their country of origin, market segment, among others. Original findings are available here.
- XGBRegressor is used as a time series regression model to predict the number of weekly cancellations by regressing a lagged series against the actual, i.e. 5 lagged series with sequential lags of up to t-5 are used as features in the model to predict the cancellation value at time t. Original findings are available here.
The original data is available from Antonio, Almeida and Nunes (2019): Hotel booking demand datasets.
For the purposes of demonstrating how MimicExplainer works, the original models and results are illustrated — with further information on how MimicExplainer can make such results more interpretable.
#data-science #machine-learning #regression #time-series-analysis
