Watch a fuzzy rule-based classifier learned with ex-fuzzy decide, one step at a time: fuzzify the measurements, fire the rules, weight them, aggregate per class, and pick the winner. Drag any value and every step updates.
Each input variable is covered by overlapping fuzzy sets with everyday names. A crisp value belongs to each set to a degree μ between 0 and 1, read off the membership function. ex-fuzzy places these trapezoids at quantiles of the training data (ticks along the bottom), so neighbouring sets overlap and, at any value, their degrees add up to 1. Drag inside a chart to move the value.
A rule is true to the degree that all of its conditions hold. ex-fuzzy combines the conditions with the product t-norm, so the firing strength is the product of their μ values. A single condition at 0 silences the whole rule; partly true conditions give a partly firing rule.
Not all rules deserve the same trust. During training ex-fuzzy measures, for each rule, its support (how much of the training data it covers with the right class) and its confidence (how often the class is right when the rule fires). Their product, the dominance score DS, scales the firing strength into the association degree.
Bars are scaled to the strongest association for this specimen.
Several rules can vote for the same class. Their votes are joined with a fuzzy OR, the
maximum: a class is as supported as its strongest rule. Normalising the class degrees to sum to 1 gives the
probabilities that predict_proba returns.
The class of the single rule with the highest association degree is the prediction (winner takes all). Because one rule decides, every prediction comes with a readable reason.
These are all the rules the genetic algorithm kept. Switch rules off to see what each one contributes: the accuracy and the confusion matrix are recomputed on the spot, and the steps above follow.
Every dot is a sample, coloured by its true class; a dark ring marks a sample the rules get wrong. The shaded background shows what the rules would predict across the two chosen variables, holding the other variables at the specimen's values, so it changes as you move the sliders. Click a dot to study it.
The rules above were found by ex-fuzzy's genetic algorithm. Change the budget and run it again: fewer rules or shorter rules are easier to read, more search may be more accurate. A new seed gives a different search.