One of successful predictive and descriptive approaches in machine learning is decision rule learning. Decision rules achieve reasonable classification accuracy and are interpretable, which is important to end users, who trust predictions more if they are supported with explanations. The challenge in mining decision rules is to find a short and comprehensible rule list with high classification accuracy. This led to many different types of classification rules like crisp rules, soft rules, probabilistic rules, etc.
We developed two new methods for mining classification rules based on ant colony optimization, which is a successful discrete optimization method. In the first part of the dissertation, we present a new method called nAnt-Miner, which can, contrary to most other ant colony based approaches, handle numeric attributes. This leads to an increased search space and affects the running time and use of memory. We showed that the nAnt-Miner method is comparable to other ant colony optimization based rule learning methods, but is worse than fuzzy rules based method FURIA. The advantage of the nAnt-Miner method is that it can detect strong dependencies between attributes.
In the second part of the dissertation we present the ProAnt-Miner method, which mines probabilistic rules. We introduce a novel interpretation of pheromone values for this approach. ProAnt-Miner is faster, achieves better prediction accuracy than nAnt-Miner, and uses less memory due to a different search graph. We showed that the ProAnt-Miner classification accuracy does not statistically differ from the state-of-the-art methods like FURIA and RIPPER. The ProAnt-Miner method has new rule form, which can give the user new insights. We evaluated both methods on real and artificial datasets.
|
