Gravitational Clustering: A Novel Approach for Efficient Supervised Learning with Minimal DataG

Dinesh Kumar Koilada

Keywords: Gravitational Clustering, Supervised Learning, Few-Shot Learning, Machine Learning Algorithms, Data Efficiency, Overfitting Resilience, Phyiscs-Inspired Models

Abstract:

Traditional supervised learning algorithms, such as
neural networks and support vector machines, often struggle
when training data is limited or when dealing with multiclass
classification tasks. In response to these challenges, this
paper introduces Gravitational Clustering, a novel algorithm
that eliminates the need for predefined cluster numbers and
effectively learns from small datasets. Drawing inspiration from
gravitational physics, this method models each cluster as a
planet with mass, radius, and class, allowing for dynamic cluster
formation without the risk of overfitting. Key advantages include
the ability to weight feature vectors, handle minimal data samples,
and maintain resilience against overfitting. The algorithm
demonstrates competitive performance across multiple datasets,
achieving higher classification accuracy while maintaining lower
computational complexity compared to traditional methods such
as K-Means and support vector machines. This paper explores
the algorithm’s theoretical foundations, computational efficiency,
and empirical results, offering a robust solution for classification
tasks with limited data availability.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 20035