Fuzzy Logic Insight Analysis
In this article
Fuzzy Binary Insights
After a fuzzy logic model has been built and deployed, you may view the Explainable AI(XAI) analysis for all the records.
Drivers are the specific aspects of a feature that points towards or is associated with a given outcome. An example of a driver would be "Income is High" if "Income" is a feature for the model. Rules are combinations of drivers, individually which constitute an antecedent in a fuzzy logic rule. An example would be "This instance is likely to be Good because Income is High, Other_expenses is Low, and Past_delinquencies is None".
Training and testing samples are presented immediately after deployment, while live instances are added as and when they are performed.
This page will show you the inference ID, the date and time it was performed, the predicted class, the prediction score, the correct class, and thereafter all the features and values of the instance itself.
Clicking on the green link of inference ID will take you to a fuzzy insights page, where the predicted class and predicted score are shown, as well as the fuzzy logic reasoning behind why that decision was made. Underneath the class and score, the default view is the drivers pointing towards each output class, ranked by importance as indicated by the solid bar underneath each driver.
Clicking on the drop-down arrow on each driver will bring up a list of each of the rules that contributed to it, along with the proportional percentage.
Drivers that are included in rules associated with both output classes are marked with an icon of two blue and red horizontal arrows pointing in opposite directions. The driver is only displayed with the class it is most associated with and its importance, or relative driver weight, is determined by summing over all rules in its drop-down: $driver\ weight = \sum{\frac{rule\ score}{number\ of\ antecedents\ in\ rule}}$
The normalised driver weights are then obtained by dividing each driver weight by the value of the maximum driver weight so that the normalised values range from 0 to 1.
Clicking on any feature/driver, or any antecedent in the drop-down rules will bring up a display showing the boundaries of the low, medium, and high fuzzy sets. If the feature is a categorical, then it will not be displayed.
Moving on to the rules view under the Rules tab, the fuzzy rules belonging to each of the classes are shown adjacently. Rules below the 1% contribution threshold are hidden automatically. The display shows the rule in full, along with its dominance, and percentage contribution to deciding the class.
There are a set of options that allow you to customise how the fuzzy insights page is displayed. You can choose between collapsed, interleaved, or column display.
- Collapsed: Where the drivers are on two sides and are ranked only according to the relevance of its own class.
- Interleaved: Where the drivers are on two sides but ranked according to absolute importance.
- Column: Where the drivers of both classes are in one column, ranked according to absolute importance.
You can set the number of drivers to display before being automatically hidden, and the result chart style in the views. you can also hide the decision on inference page by checking the box in the screenshot below.
If the model is a regression model, you can select the "Show Output As Integer" checkbox to display the predicted values as integers in the fuzzy instance view page.
You can also select the checkbox to display the feature values in "Scientific Notation" and set the Negative Exponent Limit and Positive Exponent Limit for the representation as shown below.
Here is the display with collapsed drivers (ranked only amongst their own class rather than on absolute importance), and doughnut view.
Fuzzy Estimator Insights
The default view when clicking on a specific instance from the list will take you to a page similar to the binary intelligence task.
Hovering over the fuzzy sets show their boundaries.
The black double arrows indicate that the particular driver is used in more than one output label. As seven labels are too many to display side by side, it is just associated with the one that has the most influence.
Switching from the Drivers to the Rules tab will show a display of the amount of contribution the rules pointing to each output label put forth towards the outcome. The display is defaulted to the majority/largest fuzzy set, clicking on each of the others will display on the right-hand side the rules that point to that outcome. Once again due to space constraints, the different sets cannot be displayed adjacently.
The options for the fuzzy estimator insights view are more restrictive compared to the binary display. The rules and drivers pertaining to different sets cannot be shown adjacently, so all those options are removed. The thermometer view is also redundant, and the doughnut is shown by default in the rules pane.
Fuzzy Multiclass Insights
The default view when clicking on a specific instance from the list will take you to a page similar to the regression intelligence task.
The black double arrows indicate that the particular driver is used in more than one output label. As more than two labels are difficult to display side by side, it is just associated with the one that has the most influence.
The Rules tab is also similar to that of the regression model as shown below:
The options for the fuzzy estimator insights view are more same a that of the binary display:
Records tab
