We are pleased to announce a new release of Alibi Detect featuring new drift detection capabilities. The new 0.5.0 release includes three new drift detectors focusing on mixed-type and tabular data. The new Chi-Squared detector is designed for datasets with categorical features while the Tabular Drift detector can handle tabular datasets with a mix of numerical and categorical features by performing either Kolmogorov-Smirnov or a Chi-squared test depending on the type of feature. Finally, a new Classifier Drift detector can make use of a custom classifier that classifies batches of instances into normal or drifted data at test time. Currently custom TensorFlow models are supported as classifiers but in a future release this will be extended to other kinds of models such as scikit-learn based ones.
We have also released a new version of Alibi Explain featuring some enhancements to existing methods and various bug fixes. The Integrated Gradients explainer now works on multi-input models, returning feature attributions for each input. The ALE explainer now allows for selecting a subset of features to be explained and also handles the case when certain input features are constant.
The full changelogs are available here: