Enhanced 1-NN time series classification using badness of records

Abstract

Time series classification has recently been addressed by a wide range of researchers, for its application in variety of fields. Many algorithms have been proposed. Until now, 1-NN (one-nearestneighbor) using Dynamic Time Warping (DTW) for distance measurement has been proven to be the best technique to produce the maximum accurate result. In this paper we present the idea of bad records that have the tendency to misclassify other records. To evade the misclassification by these bad records, we propose an enhanced 1-NN algorithm for time series classification that properly handles the badness of records without escalating the time complexity of the technique. As a result the accuracy of 1-NN using DTW is further improved. Experimental results show that the error rate can be minimized up to 100% in some cases.

Multiobjective Time Series Matching for Audio Classification and Retrieval

“…This comparison is based on the same set of time series and mean features for all classification criterion. Even if these methods might seem easy to overpower, several published studies confirm that 1-NN selection is still by far the top performing classification scheme for time series data [30], [36], [60]. Some authors point out "while there have been attempts to classify time series with decision trees, neural networks, support vector machines etc., the best published results (by a large margin) come from simple 1-NN methods" [20].…”

Supporting Query by Content on ECG Data with User Defined Functions

“…Over the last years, experimental evidence has shown that nearest neighbors algorithms are effective for similarity search queries on time series data [25,26]. These algorithms require a distance function that can be used to esti-mate similarity between two sequences.…”