MERCS: Efficient modeling of big data with multidirectional ensembles of decision trees

Analysis of digital data is needed everywhere. There is an increasing interest in analyzing so-called "big data", data that come in high volumes, are produced at high speed, and have high diversity. Methods for analyzing big data must be able to build models from terabytes of data, in an efficient manner, and ensuring that the models can also be used efficiently for inference of new information. In this project, we intend to develop and study a new method for building such models, one that departs completely from the probabilistic or compression-based methods that currently exist (and which have high inherent computational complexity). The method will be based on the construction of sets of ensembles of decision trees. It will make it possible to construct versatile models (useful for many kinds of inference) very fast, from very large datasets, and in a fully automatic manner (without any need for parameter tuning). The resulting models will allow highly efficient predictive inference, using a very simple, query-based, interface.