Evolving accurate and comprehensible classification rules

摘要: In this paper, Genetic Programming is used to evolve ordered rule sets (also called decision lists) for a number of benchmark classification problems, with evaluation both predictive performance and comprehensibility. The main purpose compare approach the standard list algorithm JRip also evaluate use different length penalties fitness functions evolving type model. results, using 25 data from UCI repository, show that genetic lists accuracy-based outperform regarding accuracy. Indeed, best setup was significantly better than JRip. JRip, however, held slight advantage over these models when evaluating AUC. Furthermore, all setups produced were more compact models, thus readily comprehensible. effect very clear; in essence, performed on criterion function, worsening other criteria. Brier score provided middle ground, acceptable accuracy conclusion programming solves task well, but have immediate effects results. Thus, parameters can be control trade-off between aspects