Improving Cross-Day EEG-Based Emotion Classification Using Robust Principal Component Analysis

Yuan-Pin Lin, Ping-Keng Jao, Yi-Hsuan YangFrontiers Comput. Neurosci 2017 · OpenReview

TL;DR

This paper tackles the challenge of day-to-day variability in EEG signals, which hinders reliable emotion classification in real-world applications. They propose using Robust Principal Component Analysis (RPCA) to filter out this variability, showing it significantly improves cross-day emotion classification accuracy by isolating more consistent emotional response patterns.

What Problem Does This Paper Address?

Electroencephalogram (EEG) signals are a promising avenue for non-invasive emotion recognition, enabling more intuitive human-computer interactions. However, a major hurdle for real-world deployment is "day-to-day variability" – the EEG patterns recorded on different days for the same emotional state can differ significantly. This variability makes models trained on one day's data perform poorly on data from another day, rendering them unreliable for ecological applications. Existing research often focuses on single-day datasets, overlooking this critical challenge. This paper aims to bridge this gap by developing a method to mitigate inter-day variability and improve cross-day emotion classification.

Methods

The core of the proposed method is Robust Principal Component Analysis (RPCA). RPCA is a matrix factorization technique that decomposes an input data matrix into two components: a low-rank matrix (representing regular, background activity) and a sparse matrix (representing deviant, or in this case, emotion-specific activity). The authors hypothesize that the emotion-related EEG signals are the "deviant" or sparse components, while the day-to-day variability resides in the more regular, low-rank background activity.

They applied RPCA to EEG features extracted using the MESH method, which captures spectral oscillations, laterality, and caudality across electrodes. After RPCA decomposition, they focused on the sparse component (RPCA-S) for emotion classification, hypothesizing it would contain more consistent emotional signals.

For classification, they used a Gaussian Naïve Bayes (GNB) classifier. Crucially, they employed an "add-day-in" (ADI) validation framework. This realistic approach involves training the classifier with data from an increasing number of days (1 to 4) and then testing it on data from a completely unseen subsequent day. This simulates a real-world scenario where a model needs to adapt to new data over time. Feature selection was performed using the F-score method.

Datasets

The study utilized a five-day EEG dataset collected from 12 subjects. Participants engaged in a music-listening task designed to elicit emotions of happiness and sadness. Each day involved multiple sessions, and the data was collected using a 14-channel Emotiv EEG headset. The dataset is described as being used for both training and evaluation, and the ADI validation scheme effectively creates multiple cross-day evaluation scenarios from this single dataset. No new benchmark dataset was introduced.

Key Insights and Ideas

  • Emotion signals are "sparse" against background noise: The core idea is that the transient, emotion-specific EEG signals are akin to "deviant events" that can be isolated from the more stable, "regular" background EEG activity using RPCA.
  • Day-to-day variability is in the "low-rank" component: The paper suggests that the day-to-day fluctuations in EEG patterns are captured by the low-rank component of the RPCA decomposition, not the emotion-specific sparse component.
  • Focusing on the sparse component is key: By filtering out the low-rank component and focusing on the sparse component (RPCA-S), the researchers were able to isolate EEG patterns that were more consistent across different days, leading to better emotion classification.
  • Realistic validation is crucial: The "add-day-in" validation scheme is highlighted as essential for assessing the practical utility of the method in real-world, time-evolving scenarios, as opposed to simpler offline cross-validation.
  • Neurophysiological grounding: The study attempts to link the mathematical decomposition of RPCA to actual neurophysiological phenomena, suggesting the sparse component reflects emotion-related oscillations.

Key Results

  • Improved Classification Accuracy: The RPCA-S method significantly improved cross-day binary emotion classification accuracy. When trained on data from one to four recording days and tested on an unseen day, accuracy increased from 58.31 ± 12.33% to 64.03 ± 8.40% for the music-listening task.
  • Increased Informative Features: The RPCA-S approach also led to a progressive increase in the number of informative features exploited, growing from 12.67 ± 5.99 to 20.83 ± 7.18 as more training days were included.
  • Original EEG Underperforms: The original EEG features (without RPCA processing) showed accuracies around chance level and did not exhibit the same improvement trend, highlighting the negative impact of inter-day variability.
  • RPCA-L (Low-Rank) Ineffective: The low-rank component (RPCA-L) did not improve classification accuracy and was often worse than the original EEG features, suggesting it captured the problematic background variability.
  • Baseline Performance: For the eye-closed resting baseline condition, neither RPCA nor original features showed significant improvements or distinct differences, indicating the method's effectiveness is tied to the presence of emotion-eliciting stimuli.

Limitations and Future Work

  • Plausible Causes of Perturbations: The paper acknowledges a limitation in fully elucidating the specific causes of background EEG perturbations. Future work could incorporate comprehensive behavioral and mental questionnaires (e.g., fatigue, attention, sleep quality) alongside emotional labels to better understand these influences.
  • Generalizability to Other Emotions/Tasks: While successful for binary classification (happiness vs. sadness) in a music-listening task, the generalizability of the RPCA approach to a wider range of emotions or different emotion-elicitation paradigms (e.g., image viewing, video watching) is not fully explored.
  • Comparison to Offline Validation: The authors note that their realistic ADI validation differs from some previous works that used offline validation (averaging across all day combinations). This difference makes direct comparison challenging but emphasizes the ecological relevance of their approach.
  • Further Exploration of RPCA-L: While RPCA-L was found to capture variability, its potential role in understanding baseline EEG dynamics or its interaction with RPCA-S could be further investigated.

Summary updated

Abstract

p>Constructing a robust emotion-aware analytical framework using non-invasively recorded electroencephalogram (EEG) signals has gained intensive attentions nowadays. However, as deploying a laboratory-oriented proof-of-concept study toward real-world applications, researchers are now facing an ecological challenge that the EEG patterns recorded in real life substantially change across days (i.e., day-to-day variability), arguably making the pre-defined predictive model vulnerable to the given EEG signals of a separate day. The present work addressed how to mitigate the inter-day EEG variability of emotional responses with an attempt to facilitate cross-day emotion classification, which was less concerned in the literature. This study proposed a robust principal component analysis (RPCA)-based signal filtering strategy and validated its neurophysiological validity and machine-learning practicability on a binary emotion classification task (happiness vs. sadness) using a five-day EEG dataset of 12 subjects when participated in a music-listening task. The empirical results showed that the RPCA-decomposed sparse signals (RPCA-S) enabled filtering off the background EEG activity that contributed more to the inter-day variability, and predominately captured the EEG oscillations of emotional responses that behaved relatively consistent along days. Through applying a realistic add-day-in classification validation scheme, the RPCA-S progressively exploited more informative features (from 12.67 ± 5.99 to 20.83 ± 7.18) and improved the cross-day binary emotion-classification accuracy (from 58.31 ± 12.33% to 64.03 ± 8.40%) as trained the EEG signals from one to four recording days and tested against one unseen subsequent day. The original EEG features (prior to RPCA processing) neither achieved the cross-day classification (the accuracy was around chance level) nor replicated the encouraging improvement due to the inter-day EEG variability. This result demonstrated the effectiveness of the proposed method and may shed some light on developing a realistic emotion-classification analytical framework alleviating day-to-day variability.</p>
BibTeX
@inproceedings{lin2017improving,
  author = {Yuan-Pin Lin and Ping-Keng Jao and Yi-Hsuan Yang},
  title = {{Improving Cross-Day EEG-Based Emotion Classification Using Robust Principal Component Analysis}},
  booktitle = {Frontiers Comput. Neurosci},
  year = {2017},
  url = {https://openreview.net/forum?id=F8zHVlwUad},
}
Figure