Social choice in distributed classification tasks: Dealing with vertically partitioned data

In many situations, a centralized, conventional classification task can not be performed because the data is not available in a central facility. In such cases, we are dealing with distributed data mining problems, in which local models must be individually built and later combined into a consensus, global model. In this paper, we are particularly interested in distributed classification tasks with vertically partitioned data, i.e., when features are distributed among several sources. This restriction implies a challenging scenario given that the development of an accurate model usually requires access to all the features that are relevant for classification. To deal with such a situation, we propose an agent-based classification system, in which the preference orderings of each agent regarding the probability of an instance to belong to the target class are aggregated by means of social choice functions. We employ this method to classify microRNA target genes, an important bioinformatics problem, showing that the predictions derived from the social choice tend to outperform local models in this application. This performance gain is accompanied by other interesting advantages: the aggregation methods herein proposed are extremely simple, do not require transfer of large volumes of data, do not assume an offline training process or parameters setup, and preserve data privacy.