Concept drift detection based on Fisher's Exact test

Concept drift detectors are software that usually attempt to estimate the positions of concept drifts in large data streams in order to replace the base learner after changes in the data distribution and thus improve accuracy. Statistical Test of Equal Proportions (STEPD) is a simple, efficient, and well-known method which detects concept drifts based on a hypothesis test between two proportions. However, statistically, this test is not recommended when sample sizes are small or data are sparse and/or imbalanced. This article proposes an ingeniously efficient implementation of the statistically preferred but computationally expensive Fisher's Exact test and examines three slightly different applications of this test for concept drift detection, proposing FPDD, FSDD, and FTDD. Experiments run using four artificial dataset generators, with both abrupt and gradual drift versions, as well as three real-world datasets, suggest that the new methods improve the accuracy results and the detections of STEPD and other well-known and/or recent concept drift detectors in many scenarios, with little impact on memory and run-time usage.