2D similarity kernels and representations for sequence data

Analysis of large-scale sequential data has become an important task in machine learning and pattern recognition, inspired in part by numerous scientific and technological applications such as the document and text classification or the analysis of music data, or biological sequences. In this work, we consider general, simple 2D matrix representations of sequences, and introduce a class of 2D similarity kernels that allows efficient inexact matching, comparison and classification of sequence inputs in the form of sequences of Rdim. feature vectors. The developed approach is applicable to a wide range of sequence domains, both discreteand continuousvalued, such as music, images, or biological sequences. Experiments using the new 2D representations and kernels on music genre and artist recognition show excellent predictive performance with significant 25%-40% improvements over the existing state-of-the-art sequence classification methods. Background. A number of state-of-the-art approaches to classification of sequences over finite alphabet Σ rely on measuring sequence similarity using fixed-length representations Φ(X) of sequences as the spectra (|Σ|-dimensional histogram) of counts of short substrings (k-mers), contained, possibly with up tommismatches, in a sequence, c.f., spectrum/mismatch methods [3, 4]. This essentially amounts to analysis of 1D sequences over finite alphabets Σ with 1D k-mers as basic sequence features. However, original input sequences are often in the form of sequences of feature vectors, i.e. each input sequence X is a sequence of R-dim. feature vectors which could be considered as R× |X| feature matrix. Examples of these include