Fuzzy Rules Analysis
In this article
Fuzzy rules
When viewing the model results of a fuzzy logic model, click on the Fuzzy Rules panel, the viewing panel will display. Here the fuzzy rules are shown in order of dominance. You can select whether to display instances from Training, Testing or Live inferences.
The statistics shown are:
- Dominance: How important a rule is overall considering all relevant metrics that go into deciding its rank, which are then normalised and bucketed into five ranges - the star rating. The dominance combines support (how frequently a rule applies), and confidence (how frequently the rule is correct).
- Frequency: The number of instances that the rule applies to.
- Winning: this metric counts the following: whenever a rule is fired and points to the right classification, if the rule had the highest contribution to that decision, then it counts as the "winner". This figure counts how many times the rule was the winner.
You can further filter based on antecedents, by clicking in the column title. Select the feature to filter by, and then the antecedent associated with it. All the rules in the universe that contain that feature-antecedent pair will be displayed in order of dominance. Rules can also be filtered by the Result, or consequent, of rules.
After applying a filter, you can then click the Matching Rules button, which will take you to a page containing all rules that satisfy the filter criteria. Then you can inspect those individual instances, or download them. The reverse is also true, where you can click on the Matching Instances button in the rules view, to see all the instances.
Clicking the subsection export under Fuzzy Rules in the Analysis tab on the sidebar takes you to a page where you can download the fuzzy rules data as a CSV using the Download Rules button. The membership functions data for the features are exportable separately to the rules themselves. You can select the partition to download from Training, Testing or Live.
Fuzzy coverage
The Coverage panel shows metrics of every rule included in the model.
RuleID: the identifier of the rule in the final rule-base. %Coverage: The proportion of the instances the rule fired for. Firing for class: The number of instances the rule fired for. %Crisp Confidence: The proportion of instances the rule correctly fired for.
