10. Case Study of Network-Based Semi-Supervised Learning: Stochastic Competitive-Cooperative Learning in Networks

Information reaches us at a remarkable speed and the amount of data it brings is unprecedented. In many situations, only a small subset of data items can be effectively labeled. This is because the labeling process is often expensive, time consuming, and requires intensive human involvement. As a result, partially labeled data sets are more frequently encountered. In order to get a better characterization of partially labeled data sets, semi-supervised classifiers are designed to learn from both labeled and unlabeled data. It has turned out to be a new topic of machine learning research that has received increasing attention in the past years. In this chapter, the semi-supervised classification with focus on methods based on complex networks is explored. In special, the particle competition model that we have introduced in the previous chapter is adapted to this new learning paradigm. Specifically, this enhancement is achieved by introducing the idea of cooperation among the particles and by changing the inner mechanisms of the original algorithm so as to fit it into a semi-supervised environment. In contrast to the unsupervised learning model, where the particles are randomly spawned in the network because no prior analysis of the groups is available, the semi-supervised learning version does have some external knowledge by definition. This knowledge is represented by the labeled data items, usually offered as a small fraction of the entire data set. In this scenario, the objective is to propagate the labels from the labeled set to the unlabeled set. Likewise the previous chapter, a mathematical formalization of the model, as well as a theoretical analysis, is also provided. A great portion of this analysis is based on the model that we have studied in the last chapter. A validation is also presented linking the numerical and theoretical results. An application in imperfect data learning is also presented, where the particle competition model is employed to detect and prevent error propagation in the learning process due to noisy or wrongly labeled data .