ReliefF-Based EEG Sensor Selection Methods for Emotion Recognition

Jianhai Zhang, Ming Chen, Shaokai Zhao, Sanqing Hu, Zhiguo Shi, Yu CaoSensors 2016 · DOI

TL;DR

This paper introduces novel methods for selecting the most informative EEG channels for emotion recognition, significantly reducing the number of sensors needed without sacrificing much accuracy. This is crucial for developing practical, wearable emotion-aware systems by making them more computationally efficient and user-friendly.

What Problem Does This Paper Address?

Recognizing human emotions using Electroencephalogram (EEG) signals is a rapidly growing field, driven by the demand for more immersive human-computer interfaces (HCIs) and affective computing. While multi-channel EEG offers rich data, using a large number of sensors increases computational complexity and inconveniences users. This paper tackles the challenge of efficiently selecting a minimal yet effective subset of EEG channels for emotion recognition, aiming to balance accuracy with practicality.

Methods

The core of this research lies in applying and extending the ReliefF algorithm, a well-established feature selection technique, to the problem of EEG sensor channel selection. Instead of just selecting features, the authors adapt ReliefF to evaluate the importance of entire channels. They propose three main strategies:

  1. Feature-based selection: Selecting channels based on the top-ranked features identified by ReliefF.
  2. Mean-ReliefF-channel-selection (MRCS): This method treats each channel as a unit. It calculates a channel's weight by averaging the ReliefF weights of all features within that channel. Channels are then ranked and selected based on these aggregated weights.
  3. SVM-MRCS: This is an advancement of MRCS. It iteratively adjusts channel weights based on their actual contribution to classification accuracy when using a specific classifier (Support Vector Machine - SVM). This "classifier-dependent" approach aims to further optimize channel selection for a particular recognition task.

The authors also compare their ReliefF-based methods against a similar approach using the F-score as the feature weighting mechanism.

Datasets

The study utilizes the DEAP (Database for Emotion Analysis using Physiological Signals) dataset. This publicly available multimodal dataset contains EEG and other physiological signals from 32 subjects while they watched music videos designed to elicit various emotional states. For their experiments, the authors selected 16 subjects and divided their trials into two sets: one for channel selection and another for performance validation.

Key Insights and Ideas

  • Channels over Features: The paper highlights that directly selecting channels, rather than just features, is more critical for reducing computational complexity and improving practical usability in EEG-based systems.
  • Channel Aggregation is Key: Treating channels as unified units (MRCS and SVM-MRCS) proves more effective for channel reduction than simply picking channels based on individual top features.
  • Classifier Influence Matters: The SVM-MRCS method, which incorporates classifier feedback to refine channel weights, demonstrates improved performance in channel reduction, suggesting that classifier-specific optimization can be beneficial.
  • Diminishing Returns: For most methods, classification accuracy tends to plateau or even slightly decrease after a certain number of channels are selected, underscoring the redundancy in a full EEG setup.
  • Subject-Independent Channels: Identifying common, subject-independent channels is crucial for building robust, real-world systems. The study successfully identifies a core set of channels that perform well across individuals.

Key Results

  • Significant Channel Reduction: The MRCS and SVM-MRCS methods achieved substantial reductions in the number of channels required. For instance, SVM-MRCS reduced the channel count to eight with only a slight accuracy drop (58.51% ± 10.05%) compared to using 19 channels (59.13% ± 11.00%).
  • ReliefF vs. F-Score: ReliefF-based methods generally outperformed F-score-based methods across the three strategies, particularly in channel reduction.
  • Subject-Independent Channels: The study identified a set of top-performing subject-independent channels, primarily located in the frontal and parietal lobes, which were symmetrically distributed across the scalp. Using the top 12 subject-independent channels yielded an average accuracy of 57.67% ± 10.02%.
  • Accuracy Plateau: The average classification accuracy curves showed a tendency to stabilize, indicating that beyond a certain point, adding more channels provided little to no benefit.

Limitations and Future Work

  • Classifier Dependency: The SVM-MRCS method's reliance on a specific classifier (SVM) means its optimized channel set might not be optimal for other classifiers. The authors acknowledge this but argue it's useful when the classifier is predetermined.
  • Stimulus Type: The identified subject-independent channels were influenced by the DEAP dataset's stimuli (music videos). Future work could explore how different types of emotional stimuli (e.g., visual, auditory only) might affect the optimal channel selection.
  • Practical Trade-offs: While the paper identifies points of best accuracy, it suggests that for practical applications, a slightly lower accuracy might be acceptable for a more significant reduction in channels, pointing towards finding an optimal trade-off.

The study lays a strong foundation for developing more efficient and practical EEG-based emotion recognition systems, particularly for wearable and real-time applications.

Summary updated

Abstract

Electroencephalogram (EEG) signals recorded from sensor electrodes on the scalp can directly detect the brain dynamics in response to different emotional states. Emotion recognition from EEG signals has attracted broad attention, partly due to the rapid development of wearable computing and the needs of a more immersive human-computer interface (HCI) environment. To improve the recognition performance, multi-channel EEG signals are usually used. A large set of EEG sensor channels will add to the computational complexity and cause users inconvenience. ReliefF-based channel selection methods were systematically investigated for EEG-based emotion recognition on a database for emotion analysis using physiological signals (DEAP). Three strategies were employed to select the best channels in classifying four emotional states (joy, fear, sadness and relaxation). Furthermore, support vector machine (SVM) was used as a classifier to validate the performance of the channel selection results. The experimental results showed the effectiveness of our methods and the comparison with the similar strategies, based on the F-score, was given. Strategies to evaluate a channel as a unity gave better performance in channel reduction with an acceptable loss of accuracy. In the third strategy, after adjusting channels' weights according to their contribution to the classification accuracy, the number of channels was reduced to eight with a slight loss of accuracy (58.51% ± 10.05% versus the best classification accuracy 59.13% ± 11.00% using 19 channels). In addition, the study of selecting subject-independent channels, related to emotion processing, was also implemented. The sensors, selected subject-independently from frontal, parietal lobes, have been identified to provide more discriminative information associated with emotion processing, and are distributed symmetrically over the scalp, which is consistent with the existing literature. The results will make a contribution to the realization of a practical EEG-based emotion recognition system.
BibTeX
@inproceedings{zhang2016relieff,
  author = {Jianhai Zhang and Ming Chen and Shaokai Zhao and Sanqing Hu and Zhiguo Shi and Yu Cao},
  title = {{ReliefF-Based EEG Sensor Selection Methods for Emotion Recognition}},
  booktitle = {Sensors},
  year = {2016},
  doi = {10.3390/s16101558},
}

Loading related content...

Figure