Eager, Lazy and Hybrid Algorithms for Multi-Criteria Associative Classification

Classification aims to map a data instance to its appropriate class (or label). In associative classification the mapping is done through an association rule with the consequent restricted to the class attribute. Eager associative classification algorithms build a single rule set during the training phase, and this rule set is used to classify all test instances. Lazy algorithms, however, do not build a rule set during the training phase, the rule set generation is delayed until a test instance is given. The choice between eager and lazy algorithms is not simple. Using a single rule set to perform all the predictions may not take advantage of specific characteristics of a test instance. On the other hand, building a specific rule set for each test instance may incur in excessive computational efforts. In this paper we propose new eager and lazy algorithms for associative classification. Also, we present heuristics and caching mechanisms to alleviate computational costs during lazy classification, and a new hybrid algorithm which combines characteristics of eager and lazy approaches. Finally, we also propose a multi-criteria rule selection technique which is motivated by the intuition that a single criterion cannot give a global picture of the rule (i.e., rules with high confidence but low support). We performed a systematic evaluation of the proposed approaches using real data from an important application: spam detection.