Convolutional Neural Network-Based Discriminator for Outlier Detection.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Training data noise and outliers degrade machine learning model performance by increasing overfitting.
• Identifying and removing outliers is crucial for robust big data analysis.
• Existing methods for outlier detection in training data have limitations.

Purpose of the Study:

• To propose a novel discriminator model for effective identification of outliers in training data.
• To develop a systematic approach for training this discriminator using limited trusted data.
• To enhance the reliability and performance of machine learning models by cleaning training datasets.

Main Methods:

• A convolutional neural network (CNN) was designed as a noise discriminator.
• Training datasets were systematically created using a small set of genuine instances (trusted data).
• The discriminator's performance was evaluated on benchmark datasets with varying noise ratios, including experiments with and without data augmentation.

Main Results:

• The proposed noise discriminator model demonstrated competitive performance against seven existing literature methods.
• The method showed superior performance in identifying and mitigating pair noise within datasets.
• Empirical results confirmed the effectiveness of the CNN discriminator in cleaning noisy training data.

Conclusions:

• The novel CNN-based discriminator offers a robust and effective solution for outlier detection in training data.
• The proposed systematic training approach enables high performance even with limited trusted data.
• This method significantly contributes to improving the accuracy and reliability of machine learning models by addressing data noise.