Label Noise Cleaning with an Adaptive Ensemble Method Based on Noise Detection Metric

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Real-world datasets frequently contain label noise, impacting model training and performance.
• Accurate data labeling is often costly or time-intensive.
• Label noise can lead to increased training set size, model complexity, overfitting, and reduced prediction accuracy.

Purpose of the Study:

• To propose a novel data cleaning technique to address label noise in datasets.
• To develop a method that identifies and corrects or removes mislabeled samples effectively.
• To enhance the performance and robustness of machine learning models trained on noisy data.

Main Methods:

• The proposed method, Ensemble Method based on the Noise Detection Metric (ENDM), utilizes ensemble classifiers.
• Four metrics are derived from a learned ensemble classifier to assess the probability of a sample being mislabeled.
• Thresholds are optimized on a validation dataset for three ensemble classifiers (Bagging, AdaBoost, k-NN) to identify noisy samples.

Main Results:

• The ENDM technique was evaluated on 15 public datasets, demonstrating its effectiveness in data cleaning.
• A comparative analysis showed ENDM's performance against established methods like majority vote and consensus vote.
• The method successfully identified and handled corrupted samples, leading to improved classification outcomes.

Conclusions:

• The Ensemble Method based on the Noise Detection Metric (ENDM) is an effective approach for handling label noise in datasets.
• ENDM offers a robust solution for improving machine learning model accuracy and reliability.
• The proposed technique provides a valuable tool for data preprocessing in various machine learning applications.