Deep Semi-Supervised Learning via Tensor Label Propagation for High-Dimension-Low-Sample-Size Data

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Semi-supervised learning (SSL) leverages limited labeled data with abundant unlabeled data for improved model performance.
• Label propagation is a common SSL technique using graph structures to spread label information.
• Existing methods struggle with high-dimension-low-sample-size (HDLSS) data due to reliance on pairwise similarities.

Purpose of the Study:

• To develop a novel deep SSL framework to address limitations of current label propagation methods on HDLSS data.
• To enhance label propagation by incorporating tensor-based similarity for modeling high-order relationships.
• To improve prediction performance using limited labeled data in HDLSS settings.

Main Methods:

• A deep SSL framework is proposed, integrating tensor-based similarity with label propagation.
• Feature extraction network pretraining is performed on labeled data.
• Iterative tensor label propagation and feature network fine-tuning are employed, utilizing high-order similarities to estimate pseudo-labels.

Main Results:

• The proposed method effectively models high-order relationships among samples in HDLSS data.
• Pseudo-labels are estimated more accurately, leading to robust feature representations.
• Significant enhancement in prediction performance was observed compared to existing baselines on multiple HDLSS datasets.

Conclusions:

• The novel deep SSL framework with tensor label propagation overcomes limitations of traditional methods on HDLSS data.
• High-order similarity modeling is crucial for effective label propagation in complex data settings.
• The approach demonstrates superior performance in leveraging limited labeled data for improved predictions.