Explainability

Data generators generate geodata from a preference model an/or a particular geometry computation. They are able to decompose a specific result and explain you where it comes from.

This information is retrieved for you and available in the explanation tab of the data analysis menu, whenever you are checking the result of a DataGenerator in a project. Then you click on a geometry feature on the map to see its explanations.

Note

If this tab is available for all project data generators, only the preference model itself is explainable.

Note

Only uptodate data have explanation, outdated data must be reprocessed in order to view its explanation.

There are two types of processing models:

  • preference models: computes result based on prefernces modelling

  • geometric models: modifies geometries/features of the data, potentially aggregating multiple input features together

Preference Models

Categorical rules

The “categorical rules” preference model displays the following information in its explanation tab:

  • Category: the category resulting from the computation (i.e. the output)

  • Rule: the category rule that got activated and thus explains the result (the first activated one)

  • Data tab: shows the input value for this feature used to compute its category

Below a screenshot showing an example:

../_images/explainability_categorical_rules.png

MR-Sort

MR-Sort displays the following information in its explanation tab:

  • Data comparison with its category bounding profiles: a plot showing the values of the feature for each input attribute, compared to the bounding profiles (note: higher is better, missing values for the data are not represented)

  • Profile cumulated weights comparison: a plot showing for each profile, the stacked attribute weights (added whenever the feature input is better or equal to the profile on this attribute), and the threshold used to rank the feature as better than the profile

  • Data tab: shows the input value for this feature used to compute its category

Please find in the following 2 screenshots an example, where the feature compares thusly with the profiles:

../_images/explainability_mr_sort_1.png ../_images/explainability_mr_sort_2.png

  • Between the lower (Very Bad - Bad) and upper (bad - Medium) profiles on the Current.current_criterion and Human_activities..human_activities_criterion attributes: therefore contributing their weights for the comparison with the lower profile

  • Better than both profiles on the Seabed.seabed_criterion and Depth.depth_criterion attributes: therefore contributing their weights for both profile comparisons

  • Missing value for Distance.distance_criterion attribute: thus considered worst than both and not adding its weight

The cumulated weights (0.9) for the comparison with the lower (Very Bad - Bad) profile is higher than the model threshold (0.55), thus the feature is better than this profile. The cumulated weights fell short (0.3) for the comparison with the upper profile (Bad - Medium), thus the feature is worst than it. This explains the final category (Bad) returned by the MR-Sort model.

Weighted Sum

The weighted sum shows the following information in its explanation tab:

  • Raw input values of the feature

  • Weights per attributes

  • Weighted input values for the feature (value multiplied by the corresponding weight for each attribute)

  • Data tab: shows the input value for this feature used to compute its result value

Please find the screenshots showing an example of such explanation:

../_images/explainability_weighted_sum_1.png ../_images/explainability_weighted_sum_2.png

Python

The python model does not have a computed explanation, as it uses user-supplied arbitrary code. It does show the feature input values though.

../_images/explainability_python.png

Geometric Models

Buffer

The buffer shows the following information in its exlanation tab:

  • Buffer width: indicates the width of the buffer applied to the input layer

  • Data tab: shows the values for this feature, directly taken from its input

Below a screenshot showing an example:

../_images/explainability_buffer.png

Merge Adjacent

This “Merge Adjacent” model squashes multiple jointed geometries sharing the same values. Therefore it does not have a particular explanation (result values is the shared input values).

Below a screenshot showing an example:

../_images/explainability_merge_adjacent.png

Merge Overlap

This “Merge Overlap” model aggregates multiple overlapping geometries into one. It shows the folowing explanation:

  • Aggregation method: method used to aggregate every value (per attribute) of its input geometries

  • Data tab: shows for each attribute, the values of each overlapping input features (i.e the values aggregated by this model)

../_images/explainability_merge_overlap.png