PET-TURTLE: Deep Unsupervised Support Vector Machines for Imbalanced Data Clusters

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Foundation models in vision, audio, and language enable zero-shot task performance.
• Unsupervised learning for discovering data group structure is a growing area in deep learning.
• The TURTLE algorithm is a state-of-the-art deep clustering method using alternating label and hyperplane updates.

Purpose of the Study:

• To address the limitations of the TURTLE deep clustering algorithm with imbalanced data.
• To propose an improved algorithm, PET-TURTLE, that handles imbalanced data distributions effectively.
• To enhance clustering accuracy and performance for both imbalanced and balanced datasets.

Main Methods:

• Generalizing the cost function of TURTLE using a power law prior to accommodate imbalanced data.
• Introducing sparse logits in the labeling process to simplify the search space.
• Evaluating PET-TURTLE on synthetic and real-world imbalanced and balanced datasets.

Main Results:

• PET-TURTLE significantly improves clustering accuracy on imbalanced data sources.
• The proposed method effectively prevents the over-prediction of minority clusters.
• Enhanced overall clustering performance is observed for both imbalanced and balanced datasets.

Conclusions:

• PET-TURTLE offers a robust solution for deep clustering with imbalanced data.
• The algorithm generalizes existing methods, improving accuracy and reliability.
• PET-TURTLE represents a significant advancement in unsupervised learning for data clustering.