How are the anomaly scores computed for pca, knn, svm,

Davek

Hi Anomaly Detection Experts, I have carefully read the papers underlying these anomaly detection  algorithms, but getting surprising results. I am puzzled how is temporal aspects captured in these spatial algorithms. What have you done ? It is not  stated anywhere that I can find. I am pulling my hair out, trying to figure out if the per timeslot outlier scores are computed on a per time slot basis, meaning if I have 9 data time series, each timeslot has 9 data points, or are outlier scores computed in sliding window scheme , +1, -1 time slots etc ? How are you computing the per time slot outliers scores using these algorithms which themselves are well documented? What data points do you use to compute the per time slot outlier scores ? 
Thanks, great great product !

jacobcybulski

Hi @Davek , it is not clear what exactly is your data and what is your time series representation (is it a time point per example, are time point one per attribute, or perhaps it is stored as an object collection). Most importantly which of the anomaly detection operators do you use. In general, anomaly detection operators do not understand that your data is a time series. Thus anomaly scores are computed based on attributes passed to them. They do not worry what you do with the scores afterwards, i.e. if you want to use PCA, k-NN or SVM. Jacob