Combination fairness with scores in outlier detection ensembles

We revisit score combinations for outlier detection ensembles (ODEs). Different detectors' scores vary in ranges and scales. Normalization transforms these into a common range. We assert that ODE score combinations may remain biased despite normalization. Determining a suitable normalization for an outlier model is challenging, which captures original score statistics apt for different parameters across different datasets. Thus, treating higher normalized scores of one detector as superior outlying probabilities to other detectors may be inappropriate. With differences between top rankers' scores, detectors with higher normalized scores dominate during combination. This limits the effective utilization of diversity, undesired for ensembles. We propose a framework FairComb for providing unbiased combination treatment to outputs of all detectors. It stands for combination fairness in ODEs. We define combination parity, which quantifies differences between detectors' participation in ensemble results. Higher combination parity indicates higher combination unfairness. This fairness work differs from others, which aim for fair treatment of protected data groups. The proposed FairComb checks for significant score variations across an ensemble. A reduction approach is used to mitigate these variations. We use a baseline approach WtSR to prove that addressing this issue promotes fairer combinations. Experiments on benchmark datasets demonstrate the effectiveness of FairComb.

Automated summary

We discuss the issue of score combinations in outlier detection ensembles (ODEs). Despite normalization, ODE score combinations may still be biased. Determining suitable normalization and avoiding dominance of specific detectors is challenging. We propose a framework called FairComb to address this issue and promote fairer combinations in ODEs.