Local-aggregate modeling for big data via distributed optimization: Applications to neuroimaging.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Technological advancements have generated large, matrix-valued datasets, particularly in multi-subject neuroimaging.
• Analyzing this tensor data (brain locations by time points) with existing supervised models is computationally challenging due to its ultra-high dimensionality.

Purpose of the Study:

• To develop a novel modeling and algorithmic strategy for applying generalized linear models (GLMs) to massive tensor neuroimaging data.
• To address the computational burden and enhance predictive modeling for complex, high-dimensional datasets.

Main Methods:

• Proposed a Local-Aggregate Model fitting GLMs to each location separately, then blending information via an aggregating penalty.
• Employed an Alternating Direction Method of Multipliers (ADMM) for distributed computation, significantly reducing the computational load.
• Introduced a novel sequence of algorithmic solutions, akin to regularization paths, for efficient model selection.

Main Results:

• Demonstrated significant computational advantages through distributed fitting using ADMM.
• Showcased improved predictive modeling capabilities compared to existing methods.
• Validated the method's effectiveness via simulations and an electroencephalogram (EEG) classification task.

Conclusions:

• The Local-Aggregate Model provides an efficient and scalable solution for analyzing ultra-high-dimensional tensor neuroimaging data.
• The proposed ADMM strategy and regularization path approach offer substantial computational and predictive benefits.
• This method advances the field of machine learning for complex, big data in neuroscience.