Nonparametric Distributed Learning Architecture for Big Data: Algorithm and Applications

Dramatic increases in the size and complexity of modern datasets have made traditional “centralized” statistical inference prohibitive. In addition to computational challenges associated with big data learning, the presence of numerous data types (e.g., discrete, continuous, categorical, etc.) makes automation and scalability difficult. A question of immediate concern is how to design a data-intensive statistical inference architecture without changing the basic statistical modeling principles developed for “small” data over the last century. To address this problem, we present MetaLP, a flexible, distributed statistical modeling framework suitable for large-scale data analysis, where statistical inference meets big data computing. This framework consists of three key components that work together to provide a holistic solution for big data learning: (i) partitioning massive data into smaller datasets for parallel processing and efficient computation, (ii) modern nonparametric learning based on a specially designed, orthonormal data transformation leading to mixed data algorithms, and finally (iii) combining heterogeneous “local” inferences from partitioned data using meta-analysis techniques to arrive at the “global” inference for the original big data. We present an application of this general theory in the context of a nonparametric two-sample inference algorithm for Expedia personalized hotel recommendations based on 10 million search result records.