Classification Based On Association Rules: Complexity and Interestingness Guided Algorithm

Most of the classification methods proposed so far are based on the use of heuristics, greedy techniques or associations. Among them the latter ones have stand out due to their capability to reduce the noise easily and supply higher precision in real situations. However, even the most precise rule based classification methods consume a high execution time, since they work with the whole set of rules. To identify interesting but infrequent relationships, these methods have to establish very low threshold for support, so the number of frequent itemsets and rules increases even more. Here we propose an iterative association rules based classification technique, that reduces the complexity of the result by stopping the construction of the classifier through the identification of the conditions under which there is no improvement in the classification. To identify infrequent relationships, the methods use both a support schema (for frequent itemsets) and the use of an interest measure so these relationships can be identified without decreasing the support threshold. This way the size and number of rules decreases in the resultant set, increasing the interpretability of the final set. We test the proposal using a COVID19 patient database.