Abstract
In high-dimensional, low-sample size (HDLSS) regimes, determining whether two groups in a dataset are truly distinct is a fundamental challenge. Existing methods are either fully supervised, requiring expensive labeling for all samples, or entirely unsupervised, often lacking the power to detect subtle differences. We introduce SigPal (Significance Analysis for Partially Labeled Data), a semi-supervised framework that combines a small amount of labeled data with a larger pool of unlabeled data. SigPal provides significantly more power to detect weak signals than unsupervised methods while remaining statistically valid in high dimensions.
Introduction
In fields such as genomics, researchers frequently encounter datasets with thousands of features (e.g., genes) but only a few dozen samples. This "High-Dimensional, Low-Sample Size" (HDLSS) problem renders standard statistical tests, like Hotelling’s $T^2$, unusable because covariance matrices become non-invertible.
Currently, researchers must choose between two extremes:
- Supervised Methods (e.g., DiProPerm): These require every sample to be labeled, which is often cost-prohibitive.
- Unsupervised Methods (e.g., SigClust): These ignore labels to see if data naturally clusters, but they frequently miss subtle differences that even a few labels could help identify.
SigPal addresses this gap by providing a rigorous significance test for the semi-supervised middle ground.
Methodology
SigPal follows a three-step procedure to bridge classification and clustering:
- Initialization: The algorithm uses a semi-supervised learner (such as $S^3LDA$ or COP-KMEANS) to estimate labels for the unlabeled data based on the few known labels. It then calculates a Cluster Index (CI), measuring the variation explained by this split.
- Simulation: To establish a null distribution, the authors simulate a "null world" where no real difference exists. They estimate the data's covariance structure using soft-thresholding and generate new data from a single Gaussian distribution, randomly assigning labels to a small portion to mimic the original dataset's structure.
- Testing: The labeling and CI calculation are repeated on thousands of simulated datasets. The final $p$-value represents the proportion of simulated datasets that produced a CI stronger than the observed data.
Experimental Results
The framework was evaluated using synthetic Gaussian mixtures, lung carcinoma microarrays, and the Breast Cancer (BRCA) dataset from TCGA (348 samples, 4,000 genes).
Key Insights
- Labels as a Compass: Even a tiny fraction of labels (e.g., 20 out of 348) acts as a guide, allowing the algorithm to ignore noise and focus on dimensions where group differences exist.
- The Weak Signal Advantage: SigPal excels when group differences are subtle. In the BRCA dataset, SigPal identified significant differences between LumA and LumB subtypes ($p = 0.002$), whereas SigClust failed to distinguish them ($p = 0.298$).
- Asymptotic Consistency: The authors prove that as dimensionality ($d$) increases, the SigPal $p$-value converges to zero if a real difference exists, even with a small sample size ($n$).
Conclusion and Future Work
SigPal offers a powerful new tool for HDLSS data analysis. However, it is sensitive to covariance estimation; the authors recommend applying PCA to rotate data before testing to mitigate false positives from complex non-diagonal structures. Future work will focus on extending this framework from binary classification to multi-class significance testing and exploring the optimal choice of underlying semi-supervised learners for different data types.
Summary updated