Towards Privacy-Preserving Model Selection

Model selection is an important problem in statistics, machine learning, and data mining. In this paper, we investigate the problem of enabling multiple parties to perform model selection on their distributed data in a privacy-preserving fashion without revealing their data to each other. We specifically study cross validation, a standard method of model selection, in the setting in which two parties hold a vertically partitioned database. For a specific kind of vertical partitioning, we show how the participants can carry out privacy-preserving cross validation in order to select among a number of candidate models without revealing their data to each other.