Adaptive Subspace Sampling for Class Imbalance Processing-Some clarifications, algorithm, and further investigation including applications to Brain Computer Interface

Chin-Teng Lin, Kuan-Chih Huang, Nikhil R. Pal, Zehong Cao, Yu-Ting Liu, Chieh-Ning Fang, Tsung-Yu Hsieh, Yang-Yin Lin, Shang-Lin Wu

TL;DR

This paper introduces an improved method for handling imbalanced datasets, a common problem in machine learning where one class has significantly fewer samples than others. They adapt Kohonen's Adaptive Subspace Self-Organizing Map (ASSOM) to generate synthetic minority class samples, demonstrating its effectiveness on benchmark datasets and a real-world Brain-Computer Interface (BCI) application for fatigue detection.

What Problem Does This Paper Address?

Many real-world problems, such as medical diagnosis or fraud detection, suffer from class imbalance. This means that the data used to train machine learning models is heavily skewed towards a majority class, making it difficult for models to accurately identify the minority class. Conventional algorithms tend to favor the majority class, leading to biased predictions. While oversampling techniques exist to address this, generating synthetic samples that truly represent the minority class's underlying distribution remains a challenge. This paper aims to clarify and improve upon a previously proposed ASSOM-based oversampling method to better handle this class imbalance problem.

Methods

The core of this paper is the application and refinement of Kohonen's Adaptive Subspace Self-Organizing Map (ASSOM) for oversampling imbalanced datasets. ASSOM is a type of self-organizing map that learns invariant subspaces (filters) from data, essentially capturing underlying patterns. The key idea is that by generating synthetic samples within these learned subspaces, they can better mimic the distribution of the original minority class.

The paper clarifies the ASSOM architecture and its learning scheme, which involves competitive learning among modules to find the best subspace representation for reconstructing input data. A crucial aspect is the use of a specific kernel function in the ASSOM's objective function, tailored for good reconstruction, which is essential for generating realistic synthetic samples. The authors provide a clear algorithm for generating these synthetic samples:

  1. Train an ASSOM with N quadratic modules.
  2. Select a minority class data point.
  3. Obtain its subspace representations from each of the N modules.
  4. Inversely transform these representations back to the original data space to create N synthetic samples.

The paper also discusses the choice of N (the number of synthetic samples to generate per original minority sample), suggesting a formula based on the imbalance ratio.

Datasets

The paper evaluates its method on eight benchmark datasets from the UCI Machine Learning Repository and KEEL datasets. These datasets were transformed into two-class problems for simplicity. The specific datasets used are:

  • Abalone
  • Breast Cancer
  • E. coli
  • Glass
  • Pima
  • Vehicle
  • Yeast
  • Ozone

Additionally, the paper applies the ASSOM-based technique to a real-world Brain-Computer Interface (BCI) application involving electroencephalogram (EEG) data from a driving task to identify fatigue states.

Key Insights and Ideas

  • Subspace Invariance for Realistic Sampling: The ASSOM's ability to learn subspaces invariant to rotation and translation is key. This means the generated synthetic samples are not just random interpolations but are guided by the fundamental patterns of the minority class, making them more representative.
  • Clarifying Previous Work: The paper explicitly addresses and clarifies ambiguities in their prior work [20] regarding the use of ASSOM for oversampling, providing a more detailed and implementable algorithm.
  • Tailored Kernel Function: The use of a specific kernel function in the ASSOM's objective function, focused on good reconstruction, is highlighted as a critical component for effective synthetic sample generation.
  • EEG Application Demonstrates Practicality: Applying the method to EEG data for fatigue detection showcases its potential in domains where data collection is costly and challenging, and where accurate minority class identification (e.g., detecting fatigue) is critical.

Key Results

The paper presents extensive experimental results comparing the ASSOM-based oversampling method against several state-of-the-art techniques (SMOTE, ADASYN, ESPO, MWMOTE, SVM-light, SVM-balanced) using Multilayer Perceptron (MLP) and Support Vector Machines (SVM) classifiers.

  • Benchmark Datasets: Across the eight benchmark datasets, the ASSOM-based method consistently outperformed other oversampling techniques, particularly when evaluated using metrics like recall, precision, G-mean, and F1-value. The average overall rank for ASSOM was higher than other methods for both ANN and SVM.
  • EEG Fatigue Detection: In the BCI application, the ASSOM-based method significantly improved classification accuracy for identifying fatigue states. For instance, without oversampling, the classification accuracy for high fatigue was 45.0±2.2%, which improved to 55.9±1.0% with ASSOM oversampling. Similar improvements were observed for medium fatigue states.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, the training of ASSOM and the generation of synthetic samples can be computationally intensive, especially for very large datasets.
  • Generalizability to Multi-Class: The paper primarily focuses on two-class imbalanced problems, although some benchmark datasets were transformed from multi-class to two-class. Further investigation into its performance on inherently multi-class imbalanced problems could be beneficial.
  • EEG Data Specifics: The EEG experiment focused on theta power and coherence. Future work could explore other EEG features or more complex BCI tasks.
  • Future Directions: The authors suggest that the ASSOM's ability to capture invariant subspaces makes it suitable for other neuroscientific and neurological data analysis tasks where imbalance is common. They also mention the potential for generating artificial EEG signals, hinting at future exploration of generative models.

Summary updated

Abstract

Kohonen's Adaptive Subspace Self-Organizing Map (ASSOM) learns several subspaces of the data where each subspace represents some invariant characteristics of the data. To deal with the imbalance classification problem, earlier we have proposed a method for oversampling the minority class using Kohonen's ASSOM. This investigation extends that study, clarifies some issues related to our earlier work, provides the algorithm for generation of the oversamples, applies the method on several benchmark data sets, and makes an application to a Brain Computer Interface (BCI) problem. First we compare the performance of our method using some benchmark data sets with several state-of-the-art methods. Finally, we apply the ASSOM-based technique to analyze a BCI based application using electroencephalogram (EEG) datasets. Our results demonstrate the effectiveness of the ASSOM-based method in dealing with imbalance classification problem.
BibTeX
@inproceedings{lin2020adaptive,
  author = {Chin-Teng Lin and Kuan-Chih Huang and Nikhil R. Pal and Zehong Cao and Yu-Ting Liu and Chieh-Ning Fang and Tsung-Yu Hsieh and Yang-Yin Lin and Shang-Lin Wu},
  title = {{Adaptive Subspace Sampling for Class Imbalance Processing-Some clarifications, algorithm, and further investigation including applications to Brain Computer Interface}},
  booktitle = {Ifuzzy},
  year = {2020},
  doi = {10.1109/ifuzzy50310.2020.9297814},
}
Figure