Learning with Vertically-Partitioned Data, Binary Feedback, and Random Parameter Update

Machine learning models can deal with data samples scattered among distributed agents, each of which holds a nonoverlapping set of sample features. In this paper, we propose a training algorithm that does not require communication between these agents. A coordinator can access ground-truth labels and produces binary feedback to guide the optimization process towards optimal model parameters. We mimic the gradient descent technique with information observed locally at each agent. We experimented with the logistic regression model on multiple benchmark datasets and achieves promising results in terms of convergence rate and communication load.