This drastically speeds and scales up the detectors to large dataset sizes, with dataset sizes in the order of 100,000’s easily achievable on a single consumer grade GPU.

The metric/loss can be customized. The plot below shows the feature importance wrt to accuracy and F1 metrics of a random forest predicting whether employees are likely to leave a company. The feature "satisfaction_level" is the most important one regardless of the metric.

Furthermore, 2-way PD plots show the combined effect of a pair of features. For example, the temperature and windspeed together show a non-trivial interaction on the output whilst weather situation and month of the year shows that the situation is the main driver of predictions.

For example, training a bike rental prediction model on weather data, including temperature and season, uncovers that predicted rentals increase with temperature up to some point where it flattens, whilst there is a downward effect on model predictions is outside of summer.

Monitoring for drift can be a complex endeavour, involving multiple ML components. To make life easier, we are pleased to announce the release of Alibi Detect v0.10.0!

Moreover, as illustrated before, one can understand the importance of a prototype in a 2D image scatter plot. The following figure illustrates the prototypes' importance for the CIFAR10 dataset - the larger images, the more important it is.

Inner class diversity is an important characteristic of a dataset. One can gauge the diversity of each class by the number of prototypes returned.

For example, for the Adult Census dataset, besides identifying more obvious features such as 'Capital Gain' and 'Capital Loss', one can observe more interesting statistical associations within the 'Education' and 'Marital Status' features.

We are happy to introduce a new XAI tool in Alibi Explain 0.7.0 arsenal. ProtoSelect - Bien and Tibshirani (2012) arxiv.org/abs/1202.5933 - is a prototype selection method to construct a condensed view of a dataset and an interpretable classifier: bit.ly/3MpXdMt

An important feature of the cosine distance is the normalization coefficient which makes the method insensitive to outliers (i.e. instances with large norms) - something that even other SoTA methods can struggle with.