A survey of parallel algorithms for classification

Classification of objects based on their features into pre-defined categories is a widely studied problem with applications in fraud detection, artificial intelligence and many other fields. Parallel formulations of classification methods are desirable to improve training times for problems with large training sets as well as to exploit existing computing hardware. I survey the literature on parallel formulations of building decision trees from large training sets. Most approaches exploit data parallelism by distributing the work of evaluating splitting points at each node of the tree and building the tree in a breadth-first, synchronous manner. The alternative is to exploit task parallelism by assigning a processor or a group of processors the task of building a specific subtree of the global decision tree.