Cases¶
Case is the core element of XAIB – it provides an interface to conduct experiments.
It defines one abstract method which accepts name of an explainer, and an explainer to evaluate. It receives a model and a dataset on creation. This is done intentionally to insist on the particular usage workflow. User first defines all settings – the model and data, which should be constant for one set of experiments and then they can evaluate several explainer in the same settings using the same model and data, which is correct way to compare different methods.
Another important aspect of Cases is that each one of them represents one abstract property that is measured in it. These theoretically defined properties are complex and very difficult to define numerically in exact way. This means that any metric will by definition miss the property that we should measure. However, if several ways to measure one property will be found (and they were) one property can and should be represented using several exact metrics, which will save users from overfitting even on the scale of one property.
Each case corresponds to one of the Co-12 properties.
Correctness¶
xaib.cases.feature_importance.CorrectnessCase
xaib.cases.example_selection.CorrectnessCase
The property named Correctness means that explanations describe the model correctly and truthfully. The explanations may not always be reasonable for the user, but they should be true to the model to satisfy this criterion.
Metrics¶
Completeness¶
Nothing in the explanation is left out, but it should be balanced with compactness and correctness: “don’t overwhelm”. This property can be measured from different points of view. For example reasoning completeness – white or black box, can be measured only qualitatively and refers to the type of explanation. Output-completeness – how the explanation covers output of model. Can be measured using deletion check – delete or mask all important features and measure the drop in model’s performance.
A common definition of completeness revolves around showing the user as much relevant information as possible, trying not to omit details of little influence in pursuit of a pleasant explanation.
Consistency¶
Identical inputs have identical explanations. This is different from Continuity, because this metric is about implementation invariance. It can be from example measured using implementation invariance check – construct identical models that have different weights and check the closeness between explanations made on them. This is based on the assumption that models with different values of their internal state when trained on one task, will eventually have the same way of reasoning that should be captured by explainer.
Continuity¶
xaib.cases.feature_importance.ContinuityCase
xaib.cases.example_selection.ContinuityCase
How continuous explanation function is. Continuous functions are desirable, because they are more predictable and comprehensive.
Metrics¶
Contrastivity¶
xaib.cases.feature_importance.ContrastivityCase
xaib.cases.example_selection.ContrastivityCase
How discriminative the explanation is in relation to different targets. The contrast between different concepts is very important and explanation method should explain instances of different classes in different ways.
Metrics¶
Covariate complexity¶
xaib.cases.feature_importance.CovariateComplexityCase
xaib.cases.example_selection.CovariateComplexityCase
The features used in the explanation should be comprehensible. Also non- complex interactions between features are desired.
Metrics¶
Compactness¶
xaib.cases.feature_importance.CompactnessCase
Compactness measures the size of explanations. Explanations should be sparse, short and not redundant.
Compositionality¶
Compositionality considers the format of presentation of the explanation. Some formats are considered more interpretable than others. Here in some cases metrics can be defined using for example Perceptual Realism – Fréchet Inception Distance (FID) for example synthesis. The challenge is to define quantitative measure for the presentation format of the explanations that will not involve human experiments into computation.
Confidence¶
Confidence describes the presence and accuracy of probability information. It may be defined as some quality value for agreement between confidence and true labels.
Context¶
Context is about relevance for the users of different needs. May be computed using Simulated User Studies – for example the identification of a better model or bad features.
Coherence¶
xaib.cases.feature_importance.CoherenceCase
xaib.cases.example_selection.CoherenceCase
To what extent the explanation is consistent with relevant background knowledge, beliefs and general consensus. The agreement with domain-specific knowledge can be measured, but this is difficult to define and very task-dependent.
Metrics¶
Controllability¶
How interactive the explanation is for user. This can be measured using Human Feedback Impact – improvement after feedback and the metric called Concept-level feedback Satisfaction Ratio for example.