Sparse multi-way DMDC for longitudinal classification in high dimension low sample size data

Area of Science:

• Statistics
• Machine Learning
• Bioinformatics

Background:

• High-dimensional, low sample size (HDLSS) data presents challenges for traditional methods.
• Neurodevelopmental disorder classification is a key application area for HDLSS data analysis.
• The Data Maximum Dispersion Classifier (DMDC) shows promise for HDLSS data.

Purpose of the Study:

• Introduce Sparse DMDC (SDMDC) and Longitudinal Sparse DMDC (LSDMDC) for HDLSS data.
• Evaluate SDMDC performance against other sparse models.
• Assess LSDMDC's effectiveness in longitudinal and complex data structures.

Main Methods:

• Generated simulated HDLSS datasets from multivariate normal distributions with sparse structures.
• Evaluated SDMDC with varying class imbalance, predictor counts, and correlation structures.
• Assessed performance using average F-measure and signal-to-noise ratio (SNR).
• Utilized multi-way data frameworks for LSDMDC in longitudinal simulations.

Main Results:

• SDMDC outperformed other sparse models in classification accuracy and predictor selection.
• Increased predictor variable correlation improved SDMDC performance.
• LSDMDC maintained performance with increasing predictors and handled rank-1 and rank-2 data.
• LSDMDC achieved optimal results with autoregressive correlation structures.

Conclusions:

• SDMDC offers significant advantages over existing sparse classifiers for HDLSS data.
• LSDMDC is highly effective for analyzing complex longitudinal data structures.
• These advanced classifiers enhance capabilities in fields like neurodevelopmental disorder classification.