Constraint-based MDL principle for Semi-Supervised Classification of Time Series

We propose a constraint-based method for the self-training process in semi-supervised classification of time series. Our constraint uses the Minimum Description Length principle to decide whether the instance should be added into the positive set or not. If the Description Length decreases when adding the new instance, we accept to add it; otherwise, we reject it. After the constraint-based self-training process, we continue to select more positive instances in the boundary of the positive set and the negative set. For the second step, we define a safe distance which is the sum of mean and standard deviation of the distances between pairs of nearest instances in the positive set. We select more instances to add into the positive set if its distance to the positive set is less than or equal to the safe distance. Experimental results show that our novel method can provide more accurate semi-supervised classifiers of time series.