A Cognitively Confidence-Debiased Adversarial Fuzzy Apriori Method

By discretizing continuous attributes of data with fuzzy rather than crisp sets and then generating fuzzy association rules from data to summarize the relationship between the attributes and class labels, fuzzy Apriori method (FAM) features in both its promising data mining performance and its strong uncertainty-handling capability. However, FAM successively expands shorter rules into longer ones and simultaneously discards short rules that perform poorly on the training data, which inevitably results in highly correlated rules and hence deteriorates its generalization capability. By designing a novel cognitively confidence-debiased adversarial attack on fuzzy association rules, an adversarial fuzzy Apriori method (FA
²
M) is proposed in this study to ensure enhanced generalization capability of FAM. FA
²
M has three distinct merits: 1)
Reliable analysis
about why adversarial attacks should be directly exerted on confidence and support's values of fuzzy association rules instead of inputs and/or outputs. 2)
Cognitively behavioral inspiration
by actively debiasing a small amount of cognitive base-rate biases in a disturbed way during the generation of FAM's rules while such a bias means that humans tend to ignore the base rates of fuzzy association rules during their plausibility evaluation. In other words, the active usage of the proposed cognitively confidence-debiased adversarial attack may be beneficial for FA
²
M to obtain higher generalization capabilities. 3)
Theoretical guarantee
about FA
²
M's enhanced generalization and overfitting-avoidance capabilities. Extensive experimental results show that FA
²
M attains satisfactory classification performance and enhanced generalization capability while maintaining the interpretability of fuzzy association rules therein.