Department of Computer Science Washington University St. Louis, MO 63130
PROJECT SUMMARY In standard supervised learning each example is given a label with the correct (or possibly noisy) classification. In unsupervised learning, all the individual examples are unlabeled with just a single overall label. This project is studying two learning models that fall between these two extremes. In the multiple-instance model the learner only receives labeled collections (or bags) of examples. A bag is classified as positive if and only if at least one of the examples in the bag is classified as positive by the target concept. Supervised and unsupervised learning can be thought of as two special cases of this model. In supervised learning, each example is in its own bag, and in unsupervised learning, all examples are together in one bag. The multiple-instance model was motivated by the drug activity prediction problem where each example is a possible shape for a molecule of interest and each bag contains all likely shapes for the molecule. By accurately predicting which molecules will bind to an unknown protein, one can accelerate the discovery process for new drugs, hence reducing cost. Existing multiple-instance learning algorithms use boolean labels for the bags. However, in the drug activity prediction problem, the true label is a real-valued affinity value measurement which gives the strength of the binding. This project is performing an in-depth study of learning in the multiple-instance model with real-valued labels including empirical studies using real drug binding data. Other applications areas will also be explored. This project is also studying learning when much of the available data is unlabeled. In many application areas (e.g. the classification of web pages as appropriate or inappropriate for minors, or medical applications) there is a small amount of labeled data along with a large pool of unlabeled data. This project is studying techniques to use the unlabeled data to improve the performance of standard supervised learning algorithms. In particular, a method of co-training is being studied in which there are two independent learning algorithms which are originally trained on the labeled data. Then using statistical techniques, each learner will repeatedly select some of the unlabeled data to labeled for the other learner. This project will perform empirical studies and also theoretical studies to understand the limitations of various approaches to develop better learning algorithms.
