Optimal Filtering for Time Series Classification

The application of a (smoothing) filter is common practice in applications where time series are involved. The literature on time series similarity measures, however, seems to completely ignore the possibility of applying a filter first. In this paper, we investigate to what extent the benefit obtained by more complex distance measures may be achieved by simply applying a filter to the original series (while sticking to Euclidean distance). We propose two ways of deriving an optimized filter from classified time series to adopt the similarity measure to a given application. The empirical evaluation shows not only that in many cases a substantial fraction of the performance improvement can also be achieved by filtering, but also that for certain types of time series this simple approach outperforms more complex measures.