Distributed Random Forests

• Activity Number: 581
• Type: Invited
• Date/Time: Wednesday, August 12, 2015 : 2:00 PM to 3:50 PM
• Sponsor: International Chinese Statistical Association
• Author(s): Adam Bloniarz* and Bin Yu and Ameet Talwalkar
• Companies: UC Berkeley and UC Berkeley and UCLA
• Keywords: random forests ; spark ; hadoop ; distributed algorithms ; classification ; regression

Abstract

Random forests are among the most successful general-purpose methods for classification and regression tasks. Thus, it is desirable that they be able to scale to large datasets and maintain statistical efficiency. We present a novel distributed algorithm for random forests with a small communication footprint. Instead of focusing on parallelized tree construction, we leverage the interpretation of random forests as a potential nearest-neighbor type algorithm to design a divide-and-conquer algorithm. We demonstrate the accuracy and scalability of our algorithm on multiple real-world applications using Spark, a widely used open-source platform for distributed computation. Spark is an open-source implementation of MapReduce, similar to Hadoop but with increased flexibility for in-memory caching of intermediate results.