Range Based Confusion Matrix for Imbalanced Time Series Classification

The recent flood of machine generated data allows for the detection of anomalous behaviors in the corresponding systems, something previously impossible. Consequently, the anomaly detection problem has grown in importance in industrial settings. This data often has a time-series nature and when a machine learning approach is used the labels are unbalanced. Classical point-to-point confusion matrices have been found misleading for scoring model performance. We propose an enhanced approach, based on the well-known confusion matrix, to evaluate binary classification on imbalanced time series datasets. This approach is utilized in a project which seeks to predict application failures in servers that provide web services to real customers and it results in improved estimates of classification models