Hi Pedro, I used Gaussian multivariate model for anomaly detection. Basically, I had 24 models for the interested usage pattern, one for each hour during the day. Every model will learn mean and standard distribution for that hour. When new data comes in, rank the anomaly score based how far it deviates from the mean.
It is a simple model, and can scale to tens of thousands gyms (or servers). By no means I was implying it’s “correct” or optimal. But it does capture abnormal incidents in both high and low volume hours. It’s based on the assumption that the distribution is normal, so when actual data is highly skewed, I had to adjust the scoring function. Marek O. gave me some pointers how to tune CLA. I’m going to try it out and report the results. Regards, Tom _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
