Sample Weighting Methods for Compensating Class Imbalance in Elephant Flow Classification

Abstract

Accurately identifying and classifying elephant flows is crucial in many network traffic management applications. However, the inherent class imbalance between elephant and mouse flows presents a challenge for machine learning models, often leading to poor classification accuracy. This paper compares various sample weighting techniques to compensate this imbalance during model training. We evaluate recommended approaches, as well as propose novel methods based on roots, powers, and logarithms of flow size. Analysis reveals that one of our proposed methods based on square root weighting significantly outperforms standard class balancing, offering up to 72% gains in flow operations reduction metric across multiple algorithms. These findings provide valuable insights for researchers and practitioners working on flow classification problem, contributing to more efficient network traffic management systems.