Anomaly detection based preprocessing for better classification tasks with noisy and imbalanced datasets
A computer trains, based on many timeseries, many anomaly detectors. Each anomaly detector is configured with a respective distinct contamination factor. Each timeseries is a temporal sequence of datapoints that characterize a device. Each datapoint in the many timeseries has a respective label that indicates whether the device failed when the datapoint occurred. Each anomaly detector detects: a set of anomalous datapoints, the size of which is proportional to the contamination factor of the anomaly detector, a healthy count of anomalous datapoints in timeseries of devices not failed, and an unhealthy count of anomalous datapoints in timeseries of failed devices. For a particular anomaly detector, the computer detects that the magnitude of the difference between the respective healthy count and the respective unhealthy count is less than a predefined threshold. Based on the contamination factor of the particular anomaly detector, anomalous datapoints are oversampled. Based on the oversampled anomalous datapoints, a classifier is trained.
1. A method comprising:
training, based on a plurality of timeseries, a plurality of anomaly detectors, wherein:
each anomaly detector in the plurality of anomaly detectors is configured with a respective distinct contamination factor,
each timeseries in the plurality of timeseries comprises a temporal sequence of datapoints that characterize a device, and
each datapoint in the plurality of timeseries comprises a respective label that indicates whether the device failed when the datapoint occurred;
detecting, by each anomaly detector of the plurality of anomaly detectors after said training:
a plurality of anomalous datapoints in the plurality of timeseries, wherein a size of the plurality of anomalous datapoints is proportional to said contamination factor of the anomaly detector,
a respective healthy count of the plurality of anomalous datapoints in timeseries not containing a datapoint whose label indicates the device failed, and
a respective unhealthy count of the plurality of anomalous datapoints in timeseries containing a datapoint whose label indicates the device failed;
detecting, for a particular anomaly detector of the plurality of anomaly detectors, that a magnitude of difference between the respective healthy count and the respective unhealthy count is less than a threshold;
oversampling, based on said contamination factor of the particular anomaly detector, an oversampled plurality of anomalous datapoints from the anomalous datapoints of the plurality of anomaly detectors; and
training, based on the oversampled plurality of anomalous datapoints, a classifier;
wherein the method is performed by one or more computers.
2. The method of claim 1 wherein for each anomaly detector of the plurality of anomaly detectors, the size of the plurality of anomalous datapoints is equal to said contamination factor of the anomaly detector.
3. The method of claim 1 wherein said training the classifier is further based on the plurality of timeseries not containing a datapoint whose label indicates the device failed.
4. The method of claim 1 wherein said oversampled plurality of anomalous datapoints contains datapoints from the plurality of timeseries not containing a datapoint whose label indicates the device failed.
5. The method of claim 1 wherein said oversampled plurality of anomalous datapoints contains datapoints from the plurality of timeseries containing a datapoint whose label indicates the device failed.
6. The method of claim 1 wherein said training the classifier is unsupervised.
7. The method of claim 1 further comprising for a particular timeseries in the plurality of timeseries containing a datapoint whose label indicates the device failed, detecting an oldest datapoint that the plurality of anomaly detectors agree is not anomalous.
8. The method of claim 1 wherein the plurality of anomaly detectors are identically configured except for the contamination factor.
9. The method of claim 1 wherein each timeseries in the plurality of timeseries represents a respective distinct disk drive.
10. The method of claim 1 wherein said oversampling said oversampled plurality of anomalous datapoints comprises sampling a particular datapoint at a frequency that is proportional to a count of the plurality of anomaly detectors that detected the particular datapoint as anomalous.
11. One or more computer-readable non-transitory media storing instructions that, when executed by one or more processors, cause:
training, based on a plurality of timeseries, a plurality of anomaly detectors, wherein:
each anomaly detector in the plurality of anomaly detectors is configured with a respective distinct contamination factor,
each timeseries in the plurality of timeseries comprises a temporal sequence of datapoints that characterize a device, and
each datapoint in the plurality of timeseries comprises a respective label that indicates whether the device failed when the datapoint occurred;
detecting, by each anomaly detector of the plurality of anomaly detectors after said training:
a plurality of anomalous datapoints in the plurality of timeseries, wherein a size of the plurality of anomalous datapoints is proportional to said contamination factor of the anomaly detector,
a respective healthy count of the plurality of anomalous datapoints in timeseries not containing a datapoint whose label indicates the device failed, and
a respective unhealthy count of the plurality of anomalous datapoints in timeseries containing a datapoint whose label indicates the device failed;
detecting, for a particular anomaly detector of the plurality of anomaly detectors, that a magnitude of difference between the respective healthy count and the respective unhealthy count is less than a threshold;
oversampling, based on said contamination factor of the particular anomaly detector, an oversampled plurality of anomalous datapoints from the anomalous datapoints of the plurality of anomaly detectors; and
training, based on the oversampled plurality of anomalous datapoints, a classifier.
12. The one or more computer-readable non-transitory media of claim 11 wherein for each anomaly detector of the plurality of anomaly detectors, the size of the plurality of anomalous datapoints is equal to said contamination factor of the anomaly detector.
13. The one or more computer-readable non-transitory media of claim 11 wherein said training the classifier is further based on the plurality of timeseries not containing a datapoint whose label indicates the device failed.
14. The one or more computer-readable non-transitory media of claim 11 wherein said oversampled plurality of anomalous datapoints contains datapoints from the plurality of timeseries not containing a datapoint whose label indicates the device failed.
15. The one or more computer-readable non-transitory media of claim 11 wherein said oversampled plurality of anomalous datapoints contains datapoints from the plurality of timeseries containing a datapoint whose label indicates the device failed.
16. The one or more computer-readable non-transitory media of claim 11 wherein said training the classifier is unsupervised.
17. The one or more computer-readable non-transitory media of claim 11 wherein the instructions further cause for a particular timeseries in the plurality of timeseries containing a datapoint whose label indicates the device failed, detecting an oldest datapoint that the plurality of anomaly detectors agree is not anomalous.
18. The one or more computer-readable non-transitory media of claim 11 wherein the plurality of anomaly detectors are identically configured except for the contamination factor.
19. The one or more computer-readable non-transitory media of claim 11 wherein each timeseries in the plurality of timeseries represents a respective distinct disk drive.
20. The one or more computer-readable non-transitory media of claim 11 wherein said oversampling said oversampled plurality of anomalous datapoints comprises sampling a particular datapoint at a frequency that is proportional to a count of the plurality of anomaly detectors that detected the particular datapoint as anomalous.