A Note on the Compaction of long Training Sequences for Universal Classification -a Non-Probabilistic Approach

TL;DR

This paper proposes a non-probabilistic method for compacting long training sequences into a suffix-tree with linear leaves, reducing storage needs while maintaining classification accuracy in sequence analysis.

Contribution

It introduces a universal data compaction technique for feature-based classifiers that does not rely on probabilistic models, applicable to biological sequence classification.

Findings

Suffix-tree with O(N) leaves for long training sequences

Minimal increase in classification error rate

Applicable to biological data without probabilistic assumptions

Abstract

One of the central problems in the classification of individual test sequences (e.g. genetic analysis), is that of checking for the similarity of sample test sequences as compared with a set of much longer training sequences. This is done by a set of classifiers for test sequences of length N, where each of the classifiers is trained by the training sequences so as to minimize the classification error rate when fed with each of the training sequences. It should be noted that the storage of long training sequences is considered to be a serious bottleneck in the next generation sequencing for Genome analysis Some popular classification algorithms adopt a probabilistic approach, by assuming that the sequences are realizations of some variable-length Markov process or a hidden Markov process (HMM), thus enabling the imbeding of the training data onto a variable-length Suffix-tree, the size of which is usually linear in $N$, the length of the test sequence. Despite of the fact that it is not assumed here that the sequences are realizations of probabilistic processes (an assumption that does not seem to be fully justified when dealing with biological data), it is demonstrated that "feature-based" classifiers, where particular substrings (called "features" or markers) are sought in a set of "big data" training sequences may be based on a universal compaction of the training data that is contained in a set of $t$ (long) individual training sequences, onto a suffix-tree with no more than O(N) leaves, regardless of how long the training sequence is, at only a vanishing increase in the classification error rate.