Confidence of SVM Predictions using a Strangeness Measure

Support vector machines is increasingly used for developing models for online process control. One limitation to its wide-spread use is the lack of information about the confidence in a prediction. Existing approaches to overcome this problem are not suitable for industrial applications due to limited prior information or problematic data sets. A new approach, called the strangeness measure, enables confidence limits for SVM models that are suitable for industrial applications. The advantages of the new measure over other measures are that it requires less a priori information, it takes the data density into account and it is less sensitive to noise and outliers.