Dynamic Noise-Aware Preprocessing for Data Streams: Joint Treatment of Feature Drift, Label Errors, and Instance Redundancy

Main Article Content

Vranda Jajoo
Sanjay Tanwani

Abstract

The increasing ubiquity of real-time analytics across domains such as Internet of Things (IoT), healthcare, finance, and cybersecurity has amplified the challenges associated with mining highvelocity data streams. A key obstacle in this setting is ensuring data quality, as streaming environments are prone to noise, feature drift, label errors, and redundant instances, each of which can significantly degrade learning performance. Traditional preprocessing methods, although effective in static datasets, often fail to cope with the temporal evolution and transient nature of streaming data. This paper proposes the concept of dynamic noise-aware preprocessing, a joint strategy that simultaneously addresses feature drift, label errors, and instance redundancy under streaming conditions. By integrating adaptive feature selection, probabilistic label correction, and redundancy-aware instance reduction, the proposed approach enhances resilience and stability of stream mining algorithms. The paper also emphasizes the importance of lightweight, incremental mechanisms to ensure computational feasibility in resource-constrained environments. Through a critical synthesis of recent advances and empirical trends, this work positions dynamic noise-aware preprocessing as a pivotal component for next-generation real-time data mining systems.

Article Details

Section
Articles