Distributed prediction from vertically partitioned data

We address the problem of prediction of data that is vertically partitioned, that is where local sites hold some of the attributes of all of the records. This situation is natural when data is collected by channels that are physically separated. For distributed prediction, we show that a technique called attribute ensembles is simple, predicts almost as well as a centralized predictor, reduces the amount of communication required, distributes computation and data access well, and allows each local site to keep its raw data private. We show how to extend attribute ensembles to data that is partitioned both horizontally and vertically.