Discovering novelty in spatio/temporal data using one-class support vector machines

Abstract

Novelty or anomaly detection in spatio/temporal data refers to the automatic identification of novel or abnormal events embedded in data that occur at a specific location/time. Traditional techniques used in process control to identify novelties are not robust for noise in the data set. We present an algorithm based on the support vector machine approach for domain description. This technique is intrinsicly robust for outliers in the data set but to make it work, several extensions are needed which form the contribution of this work: an extended representation of the spatio/temporal data, a tensor product kernel to separately deal with the distinct features of time and measurements, and a voting function which identifies novelties based on different representations of the time series in a robust way. Experimental results on both artificial and real data demonstrate that our algorithm performs significantly better than other standard techniques used in process control.