Source-linked AI summary
Enabling Smart Data: Noise filtering in Big Data classification
Diego García-Gil, Julián Luengo, Salvador García, Francisco Herrera
TL;DR
Big Data classification requires scalable ways to address disruptive noise and produce higher-quality data. The paper proposes Spark-based homogeneous and heterogeneous ensemble filters, evaluates them on large noisy datasets, and reports that the homogeneous ensemble is most suitable in most cases.
Problem
Traditional noise-preprocessing methods struggle with the scale of Big Data, although noise—especially incorrect class labels—degrades classification data quality.
Method
The paper implements homogeneous and heterogeneous classifier-ensemble filters in Apache Spark, with ENN-BD as a similarity-based comparison method.
Results
The homogeneous ensemble was most suitable in most cases for accuracy improvement and running time, while also providing a balanced trade-off between removing and retaining instances.
Takeaways & Limitations
The framework provides a Big Data noise-filtering approach for transforming raw classification data into Smart Data.
Abstract
from arXiv · showhide
In any knowledge discovery process the value of extracted knowledge is directly related to the quality of the data used. Big Data problems, generated by massive growth in the scale of data observed in recent years, also follow the same dictate. A common problem affecting data quality is the presence of noise, particularly in classification problems, where label noise refers to the incorrect labeling of training instances, and is known to be a very disruptive feature of data. However, in this Big Data era, the massive growth in the scale of the data poses a challenge to traditional proposals created to tackle noise, as they have difficulties coping with such a large amount of data. New algorithms need to be proposed to treat the noise in Big Data problems, providing high quality and clean data, also known as Smart Data. In this paper, two Big Data preprocessing approaches to remove noisy examples are proposed: an homogeneous ensemble and an heterogeneous ensemble filter, with special emphasis in their scalability and performance traits. The obtained results show that these proposals enable the practitioner to efficiently obtain a Smart Dataset from any Big Data classification problem.
1. Introduction
Big Data classification is hindered by noisy, low-quality data and preprocessing methods that struggle with massive datasets. The paper proposes Spark-based filtering approaches and evaluates their effect on classification performance.
- Data quality is central to extracting useful knowledge from Big Data, whose scale has exceeded classical data-mining capabilities.
- Noise can corrupt data acquisition and labeling, producing overly complex models with deteriorated performance and greater computing times.
- Traditional preprocessing is often too time consuming and resource demanding for Big Data, motivating scalable Smart Data approaches.
- Class noise affects labels, whereas attribute noise affects input attributes; class noise is described as the more disruptive type.
- The paper proposes Apache Spark implementations of homogeneous and heterogeneous classifier ensembles, alongside ENN-BD, to remove noisy examples.
- The framework is evaluated on four large datasets under induced class noise using decision trees and KNN to assess classification-accuracy improvements.
2. Related work
Related work distinguishes noise types and filtering strategies while framing Spark as a scalable platform for Big Data preprocessing. The paper addresses the remaining gap by providing an automated class-noise filtering framework for Smart Data.
- Class noise vs. attribute noise: Classification noise includes label errors and corrupted attributes, with class noise generally considered more harmful to learning.
- Noise-handling approaches: Noise-handling research includes robust algorithms and data-level filters that cleanse or repair datasets before classifier fitting.
- Class noise vs. attribute noise: Big Data’s dimensionality and example volume accumulate noise, while many classic filters cannot process huge datasets.
- Big Data frameworks: MapReduce uses parallel distributed Map and Reduce phases to process large datasets through divide-and-conquer computation.
- Big Data frameworks: Spark supports in-memory computation and iterative machine-learning workflows through operations on partitioned Resilient Distributed Datasets.
- Big Data to Smart Data: The paper presents an automated Spark preprocessing framework for class noise, intended to support the transition from Big Data to Smart Data.
3. Towards Smart Data: Noise filtering for Big Data
The framework distributes noise filtering under Apache Spark using ensemble-based algorithms and a similarity baseline. HME-BD and HTE-BD use partitioned training, classifier predictions, and filtering to identify noisy instances.
- The Apache Spark framework performs noise filtering through distributed MapReduce processes over large datasets.Its implementation uses Spark primitives to support the proposed filtering algorithms.
- HME-BD: Unlike CVCF, HME-BD uses Random Forest and predicts only unseen partitions, improving the voting step and reducing computation.The method associates each instance’s noise decision with the Random Forest prediction, avoiding a separate voting strategy.
- HME-BD: HME-BD partitions training data, learns Random Forest models on complementary folds, and predicts each held-out partition to mark noise.Spark’s kFold creates train-test pairs, after which mismatched predictions identify noisy instances for removal.
- HTE-BD: HTE-BD applies Random Forest, Logistic Regression, and KNN to each train-test fold before comparing their predictions with original labels.The three predictions are stored together for each test instance, supporting heterogeneous ensemble filtering.
- ENN-BD: ENN-BD provides a similarity-based baseline derived from Edited Nearest Neighbor filtering.It is included for comparison with the ensemble-based methods.
4. Experimental Results
Experiments on four large datasets compare HME-BD, HTE-BD, and ENN-BD under induced class noise using KNN and decision-tree classifiers. HME-BD generally provides the strongest combination of accuracy, noise removal, and computing efficiency, while filter behavior depends on classifier and voting strategy.
- Any noise treatment improves KNN accuracy over the Original value at the same noise level, including at the highest tested noise levels.
- HME-BD enables KNN to obtain the highest accuracy values, while HTE-BD consensus voting produces nearly 2% lower accuracy than its alternatives.
- For decision trees, HTE-BD can slightly outperform HME-BD at low noise, but HME-BD outperforms HTE-BD from 10% noise onward.
- ENN-BD removes the most instances but averages only 50% noisy-instance removal, and its aggressive filtering hinders noise-tolerant decision trees.
- HME-BD is the most balanced filtering technique, removing around 20% of instances at 0% noise and an average 3% more at each noise increment.
- HME-BD removes about 65% of noisy instances and HTE-BD with majority voting about 80%, with each outperforming the other on two of four datasets.
- HME-BD is about ten times faster than HTE-BD and ENN-BD because the latter two use computationally demanding KNN distance calculations.
5. Conclusions
The paper presents Spark-based ensemble noise filters for Big Data classification, enabling distributed processing of huge datasets. Across evaluations, the homogeneous ensemble is generally the most suitable approach, balancing accuracy improvement, runtime, and instance retention.
- Spark-based ensemble filters distribute noise removal across map operations, allowing practitioners to process huge classification datasets.The framework implements filtering techniques in a Big Data environment using ensembles of classifiers executed in different maps.
- The proposed techniques were evaluated for accuracy improvement, running time, and data reduction across several datasets.
- The homogeneous ensemble also provided a balanced trade-off between removing noisy instances and retaining sufficient training data.
- The framework addresses challenges posed by high redundancy and high dimensionality that limit classic Big Data noise-preprocessing algorithms.