Home

The COSNet R package

Prev. : Software installation
Next : A step-by-step application of COSNet  to protein function prediction

Main functionalities of COSNet

COSNet is a package which implements the homonym algorithm to infer node labels in biological networks [reference 1, 2 of the on-line section Overview of the COSNet R Package].

COSNet is specifically designed for problems characterized by unbalanced labellings, that is when negative examples largely outnumber those positive.

Its setting is semi-supervised: it requires positive and negative instances for the class being predicted (training set), and the unlabelled instances (test set), for which the algorithm infers a label (positive or negative) with regard to the current class.

The main steps of COSNet are summarized in the following.

  • Data processing. The package provides functions to partition input data in folds (find.division.strat and find.division.not.strat), and to generate temporary labels for the unlabelled instances (generate_labels), to be used in the learning phase. See Section 5.1.1 of  reference [1] for details about this step.

  • Learning of parameters. This part of the package realizes the learning procedure of the COSNet algorithm (see Section 5.2 of [1]). The function generate_points projects nodes in the training set into labelled points in the plane (Section 5.2.1), whereas functions optimizep and optimize_pos_above learn the optimal straight line (Section 5.2.2).

  • Network dynamics and regularization. Finally, the package provides the function runSubnet to realize the dynamics (with the learned parameters) of the sub-network restricted to unlabelled nodes (Section 5.3 of [1]). Moreover, the function reg_data allows to simulate a regularized dynamics (Section 5.6 of paper [1]).

The function COSNet executes all these steps in the mentioned order. 

Finally, the function cosnet.cross.validation allows to perform an entire cross validation procedure, with a desired number of folds, to estimate to generalization capabilities of the method.
For details about the single functions of the package, please refer to the Reference Manual.