How to validate my result of anomaly detection using k means clustering??
3 views (last 30 days)
I had a synthetic data set with an artificially injected anomaly at some point. I wanted to detect that anomaly using unsupervised learning techniques. So, I ran some of the processing on the data and in the end finally used k means clustering to detect it. I used 3 clusters (coz that was the best solution possible) and finally got the locations of the added anomaly (using the distance metric from centroid) with some results for the other clusters as well.
Now how to critically evaluate the performance of my approach??
Image Analyst on 18 Aug 2017
How about just computing the percentage of time it correctly detected the anomaly?
If you wanted to go further you could very the distance from known centroids and compute the ROC curve.