A text-mining approach for classification of genomic fragments

Genome identification is an emerging area of interest due to the study of environmental DNA samples. We show that performance approaches 50% for classifying 500 bp fragments when using 12 mer features, but more importantly, the performance linearly increases for large N. Secondly, we determine that an inverted TF-IDF measure performs 16% better when only using 80% of the words, as opposed to taking the fullset (100%). This increase implies that while too sparse of a feature subset does not produce good results, a carefully selected set has the potential to improve genome classification over a random feature set. Computing even 80% of all possible features can result in a significant savings in computation. The Euclidean classifier and TF-IDF measures will pave the way for more discriminative classification techniques.