Improving the Detection of Noise Artifacts in Gravitational-Wave Data With a Classifier Graph

We propose a method for improving the classification performance of a classifier, termed the classifier graph, by embedding it in a graph of classifiers. Our graph-based method has the advantage of enabling delicate classification from different levels of interpretation and abstraction. For the problem that thresholds corresponding to different classifiers are correlated and thus have mutual effects on the final performance, we provide a generalization of the receiver operator characteristic curve that properly tunes them jointly to obtain optimal performance. This method is successfully applied to the detection of noise artifacts (glitches) in the gravitational-wave data. We thus obtain an improvement up to 10% on the classification performance compared with that of a single classifier. The methods of this paper provide an effective way to improve the classification performance with multiple classifiers.