AC-Stream: Associative classification over data streams using multiple class association rules

摘要

Data stream classification is one of the most interesting problems in the data mining community. Recently, the idea of associative classification was introduced to handle data streams. However, single rule classification over data streams like AC-DS implicitly has two flaws. Firstly, it tends to produce a large bias on simple rules. Secondly, it is not appropriate for data streams that are slowly changed from time to time. To overcome this problem, we propose an algorithm, namely AC-Stream, for classifying a data stream using multiple rules. AC-Stream is able to find k-rules for predicting unseen data. An interval estimated Hoeffding-bound is used as a gain to approximate the best number of rules, k. Compared to AC-DS and other traditional associative classifiers on large number of UCI datasets, AC-Stream is more effective in terms of average accuracy and F1 measurement.