Learning to Predict in Complex Biological Domains

摘要

Protein secondary structure prediction and high-throughput drug screen data mining are two important applications in bioinformatics. The data is represented in sparse feature spaces and can be unrepresentative of future data. There is certainly some noise in the data and there may be significant noise. Supervised learners in this context will display their inherent bias toward certain solutions, generally solutions that fit the training set well. In this paper, we first describe an ensemble approach using subsampling that scales well with dataset size. A sufficient number of ensemble members using subsamples of the data can yield a more accurate classifier than a single classifier using the entire dataset. Experiments on several datasets demonstrate the effectiveness of the approach. We report results from the KDD Cup 2001 drug discovery dataset in which our approach yields a higher weighted accuracy than the winning entry. We then ex-tend our ensemble approach to create an over-generalized classifier for prediction by reducing the individual subsample size. The ensemble strategy using small subsamples has the effect of averaging over a wider range of hypotheses. We show that both protein secondary structure prediction and drug discovery prediction can be improved by the use of over-generalization, specifically through the use of ensembles of small subsamples.