Abstract:

This paper formulates the protein function prediction into a typical LPU. Aiming at imbalance or overfitting from LPU with few positive examples, it proposes a method creating synthetic examples to enlarge the set of positive examples based on the nearest neighbor and convex combination, and meanwhile modifies the procedure learning optimal classifier for the classic LPU algorithm by using oneclass SVM(support vector machine) to identify the most probable negative examples, running iteratively SVM to move the classification hyperplane to a suitable place and obtaining representative negative examples through cross validation. For the yeast genomic data, the experiments show that our algorithm outperforms several classic prediction methods, particularly, for function classes with few positive examples.

Key words: protein function prediction;SVM;LPU