Classification of examples by multiple agents with private features

We consider classification tasks where relevant features are distributed among a set of agents and cannot be cen- tralized, for example due to privacy restrictions. We are motivated by a key classification task that arises in a cal- endar management domain where software assistants clas- sify new meetings as likely to be difficult to schedule. Ac- curate prediction of the output class is difficult for an iso- lated single agent because the target concept may involve features to which the agent does not have access, for exam- ple each attendee's willingness to attend the meeting. To in - crease prediction accuracy, novel learning algorithms are required in which agents collaborate to classify new exam- ples while maintaining the privacy of features. We introduc e a novel distributed asynchronous decision-tree inspired a l- gorithm for such tasks named DDT. DDT differs from pre- vious approaches in that it applies to vertically partition ed data with categorical multi-valued features, it requires n o explicit hypothesis generation, and there is no a priori re- striction on number of agents. We present empirical results in our meeting scheduling domain and show that DDT out- performs a single agent learner and performs as well as a centralized learner with hypothetical access to all the fea - tures.