|
Abstract:
|
Statistical inferences from multiple data sources can often be fused together to yield more effective inference than from individual source alone. Such fusion learning is of vital importance for big data where data are often assembled in various domains. This paper develops a fusion methodology called individualized fusion learning (iFusion), to enhancing inference for an individual via adaptive combination of confidence distributions obtained from its clique (i.e., peers of similar individuals). iFusion begins with obtaining inference for each individual, then adaptively forming a clique, and finally obtaining a combined inference from the clique. iFusion explores heterogeneity in the database to form a clique for each individual and, by drawing inference from the clique, it allows borrowing strength from similar peers to enhance the inference efficiency for each individual. Furthermore, iFusion can be performed without using the entire data simultaneously and thus allow split-&-conque to be implemented on individuals to substantially reduce the computational expense. We provide supporting theories for iFusion and also illustrate it using a real data example.
|
