Fig. 2

Benchmarking the R packages for rule-based machine learning applied onto the autism-control dataset. The packages were evaluated with various methods for default and tuned parameters. For the R.ROSETTA package, Johnson reducer (Johnson) and Genetic reducer (Genetic) were used. For the package C50, the C5.0 method (C50) was used. For the package RoughSets, Learning from Examples Module (LEM2), CN2 rule induction (CN2) and Quasi-optimal covering Algorithm (AQ) were used. For the package RWeka Repeated Incremental Pruning to Produce Error Reduction (JRip), 1-rule classifier (OneR) and partial decision trees-based (PART). Several methods were tuned for the number of boosting iterations (trials), times covering objects by rules (tc), algorithm complexity (K) and number of optimizations (O). Other methods were evaluated with default parameters. The results of benchmarking are presented for a accuracy distribution of classifiers, b ROC AUC distribution of classifiers, c number of estimated rules (logarithmic scale) and d average runtime of the algorithms (logarithmic scale). Two standard deviations were marked above each bar. The time was measured from inputting a decision table to receiving a model