Avoiding the Hay for the Needle in the Stack: Online Rule Pruning in Rare Events Detection

In certain machine learning application domains such as those that arise in Industry 4.0 settings that involve rare events detection and prediction, classifiers based on Rule Mining often have an advantage over other algorithms due to their design that allows them to deal with highly skewed class distributions. We describe R4RE, an algorithm that mines all quantitative association rules from large datasets that meet certain user-defined criteria of interestingness that is not limited to the classical support/confidence framework. When dealing with very small classes, the possibility arises that rules with very small support may be valid in the training set that do not hold in the test set. To avoid generating such rules, we implement the approach found in the IREP and other systems where rules may be pruned immediately after discovery. Results from real world factory data show the efficiency of our approach for predictive maintenance purposes.