Handling missing features with boosting algorithms for protein-protein interaction prediction

Combining information from multiple heterogeneous data sources can aid prediction of protein-protein interaction. This information can be arranged into a feature vector for classification. However, missing values in the data can impact on the prediction accuracy. Boosting has emerged as a powerful tool for feature selection and classification. Bayesian methods have traditionally been used to cope with missing data, with boosting being applied to the output of Bayesian classifiers. We explore a variation of Adaboost that deals with the missing values at the level of the boosting algorithm itself, without the need for any density estimation step. Experiments on a publicly available PPI dataset suggest this overall simpler and mathematically coherent approach may be more accurate.