Binary classification with pFDR-pFNR losses.

Area of Science:

• Statistical Learning
• Machine Learning
• Biomedical Data Analysis

Background:

• Binary classification is crucial for analyzing complex datasets.
• Existing methods often assume data independence, limiting their application to autocorrelated data.
• Autocorrelation in data, common in time series, poses challenges for traditional classification.

Purpose of the Study:

• To develop a false discovery rate (FDR)-based classification approach for two-class mixture models with autocorrelated data.
• To generalize existing classification methods to higher dimensions (Rd) and incorporate autocorrelation.
• To provide flexible classification procedures that allow for prior knowledge and asymmetric misclassification costs.

Main Methods:

• Derivation of an FDR-based classification strategy for mixture models.
• Utilizing multivariate estimation methods for probability density functions and density ratios.
• Development of two distinct classification algorithms with options for prior probabilities and misclassification weighting.

Main Results:

• The proposed methods successfully handle autocorrelated time series data with moving average structures.
• Computer simulations demonstrate the practical applicability and effectiveness of the algorithms.
• The approach generalizes previous findings for independent data in lower dimensions.

Conclusions:

• The developed FDR-based classification approach is effective for autocorrelated data in mixture models.
• The methods offer flexibility in incorporating prior information and handling asymmetric misclassification costs.
• The approach shows promise for applications in brain-computer interfacing and electroencephalography data processing.