An Adaptive Subspace Self-Organizing Map (ASSOM) Imbalanced Learning and Its Applications in EEG

Chin-Teng Lin, Yu-Ting Liu, Chun-Hsiang Chuang, Nikhil R. Pal, Yang-Yin Lin, Tsung-Yu Hsieh, Chieh-Ning Fang, Zehong Cao

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 application to three Brain Computer Interface (BCI) applications. 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 the three BCI based applications using electroencephalogram (EEG) datasets. These tasks are classification of motor imagery, drivers' fatigue states, and phases of migraine. Our results demonstrate the effectiveness of the ASSOM-based method in dealing with imbalance classification problem.

Introduction

Class imbalance is a pervasive challenge in machine learning, particularly in high-stakes domains like medical diagnosis, fraud detection, and Brain-Computer Interfaces (BCI). Standard classifiers are typically designed to minimize overall error, which often leads them to neglect the minority class—the very class that is often of primary interest. While existing solutions like the Synthetic Minority Over-sampling Technique (SMOTE) generate new samples through interpolation, they can be "blind" to the underlying structural invariants of the data. This research introduces a more sophisticated oversampling technique based on Kohonen’s Adaptive Subspace Self-Organizing Map (ASSOM) to capture the invariant characteristics of the minority class and generate more realistic synthetic data.

Methodology: ASSOM-based Oversampling

The core innovation of this approach lies in the use of subspaces rather than discrete prototype points. Unlike a standard Self-Organizing Map (SOM), ASSOM identifies sets of basis vectors that define linear subspaces, each representing specific invariant features (such as rotation or scale) of the data.

The oversampling workflow proceeds in four stages:

  1. Training: ASSOM modules are trained on minority class data to learn representative linear subspaces.
  2. Projection: Existing minority samples are projected into these learned subspaces.
  3. Reconstruction: These projections are inversely transformed back into the original data space, creating synthetic instances that occupy the learned "invariant" regions.
  4. Classification: The balanced dataset is then used to train standard classifiers like Artificial Neural Networks (ANN) or Support Vector Machines (SVM).
The ASSOM-based oversampling architecture showing multiple modules generating synthetic instances from learned subspaces.

Experimental Evaluation

The researchers evaluated the ASSOM-based method using two distinct categories of data:

  • Benchmark Datasets: Eight datasets from the UCI and KEEL repositories, including Abalone, Breast Cancer, E. coli, Glass, Pima, Vehicle, Yeast, and Ozone, were selected for their varying imbalance ratios.
  • EEG Driving Task: A real-world BCI application involving 33 subjects in a driving simulator. The objective was to classify fatigue states (low, medium, high) based on electroencephalogram (EEG) signals—a task characterized by significant noise and class imbalance.

Results and Discussion

The ASSOM-based method was compared against several state-of-the-art techniques, including SMOTE, ADASYN, MWMOTE, and ESPO.

Benchmark Performance

In terms of "Average Overall Rank" across metrics such as Recall, Precision, F1-value, and G-mean, the ASSOM method achieved the top rank for both ANN (4.59) and SVM (5.66) classifiers. The method showed particular robustness in extremely imbalanced scenarios, such as the Ozone dataset, where it nearly doubled the F1-value compared to the original imbalanced set.

BCI Application: Driver Fatigue

The driver fatigue study highlighted the practical utility of the method. The "High Fatigue" state, which is often the rarest and most critical to detect, saw a dramatic improvement in classification accuracy. Without oversampling, accuracy for high fatigue was only 45.0%; with ASSOM-based oversampling, this figure rose to 69.6%.

Comparison of classification accuracy for EEG-based fatigue states with and without ASSOM oversampling, showing significant gains in the high fatigue state.

Conclusion

This investigation demonstrates that leveraging the invariant subspaces of data through ASSOM provides a superior mechanism for oversampling in imbalanced learning. By generating synthetic samples that respect the underlying structure of the minority class, the method significantly enhances performance in both benchmark and real-world BCI applications. Future work will explore the adaptation of this approach for raw time-series generation and direct comparisons with modern generative models like GANs.

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 application to three Brain Computer Interface (BCI) applications. 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 the three BCI based applications using electroencephalogram (EEG) datasets. These tasks are classification of motor imagery , drivers' fatigue states, and phases of migraine. Our results demonstrate the effectiveness of the ASSOM-based meth od in dealing with imbalance classification problem.
BibTeX
@inproceedings{lin2020adaptive,
  author = {Chin-Teng Lin and Yu-Ting Liu and Chun-Hsiang Chuang and Nikhil R. Pal and Yang-Yin Lin and Tsung-Yu Hsieh and Chieh-Ning Fang and Zehong Cao},
  title = {{An Adaptive Subspace Self-Organizing Map (ASSOM) Imbalanced Learning and Its Applications in EEG}},
  booktitle = {Ifuzzy},
  year = {2020},
  doi = {10.1109/ifuzzy50310.2020.9297814},
}
Figure