A robust approach to sequence classification
Ming Li, Ronan Sleep · 2005
We report results for classification of representations of music, spoken words, and text documents. Experimental comparisons with other state-of-the-art algorithms yield improved results for all three examples. We use a support vector machine (SVM) as our classifier in all experiments. This is driven by a kernel matrix of similarity measures between the sequences. Our similarity measure is based on n-grams of varying length (multi-grams), weighted to reflect discrimination ability. To alleviate the problem of the exponential growth of feature size with n, we use a modified LZ78 algorithm (Z. Jacob and L. Abraham, 1978) to guide feature selection. Our method exhibits good performance over the three widely distinct tasks reported here, and is very computationally efficient and may therefore be useful in real time applications.