Collaborative Data Mining on a BDI Multi-agent System over Vertically Partitioned Data
Jorge Melgoza-Gutierrez, Alejandro Guerra‐Hernández, Nicandro Cruz-Ramírez · 2014
This paper presents a collaborative learning protocol dealing with vertical partitions in training data, i.e., The attributes of the instances are distributed in different data sources. The protocol has been modeled and implemented following the Agents and Artifacts paradigm. The artifacts provide Weka based learning tools to induce and evaluate Decision Trees (a modified version of J48), While the agents manage the workflow of the learning process, using such tools. The proposed protocol, and slightly faster variation, are tested with some known training sets of the UCI repository, comparing the obtained accuracy against that obtained in a centralized scenario. Our collaborative learning protocol achieves equivalent accuracy to that obtained with centralized data, while preserving privacy.