An Adaptive Ensemble Based Filtering Approach for Noise Removal in Online Data Streams

Data preprocessing is essential for improving the performance of machine learning models, especially in streaming environments, where training data is continuously collected and updated. Traditional approaches to filtering noisy training data either evaluate each instance only once, increasing the risk of misclassifications, or check instances against all available classifiers, leading to high computational costs. In this paper, we propose an ensemble-based adaptive filtering approach for cleaning training data in streaming settings, ensuring that models are trained on high-quality datasets. Our method partitions incoming training data into, let us say n number of streaming segments, on these segments we incrementally train n Random Forest classifiers, and evaluates each instance using a dynamically selected subset of three classifiers. If all three classifiers agree with the actual class label, the instance is retained in the training dataset. Otherwise, it undergoes a majority voting mechanism to determine its correctness. Misclassified instances are filtered out, ensuring that only reliable and high-quality data contributes to model training. This approach optimally balances computational efficiency and accuracy, preventing unnecessary recompilations while maintaining robust classification. Experimental results on benchmark streaming datasets demonstrate that our method effectively removes noisy instances, leading to cleaner training data and better-performing machine learning models compared to conventional training data preprocessing techniques. We are able to achieve significant performance without compromising much on accuracy aspect.