Source-linked AI summary
Beyond Local Power: Functional Connectivity Analysis for Subject-Independent Learning Style Recognition
Wiga Maulana Baihaqi, Indriana Hidayah, Sri Kusrohmaniah, Noor Akhmad Setiawan
TL;DR
The paper addresses the time demands of behavioral learning-style classification by evaluating EEG connectivity and localized features across two FSLSM dimensions. It uses PLV-based and localized EEG features with SVM classification and dual validation, finding stronger subject-level VV performance than AR generalization while revealing systematic individual reversals that limit rigid global models.
Problem
Behavioral learning-style methods can require prolonged interaction-log accumulation, motivating a rapid objective EEG alternative.
Method
The study compares PLV connectivity with localized EEG features for AR and VV classification using SVMs and subject-level validation.
Results
The VV dimension achieved 70.00% accuracy under LOSO-CV, while AR achieved 55.56% and showed high-confidence inverted predictions with 20–0 voting margins.
Takeaways & Limitations
Stable individual connectivity signatures can oppose population-averaged decision boundaries, limiting rigid one-size-fits-all classifier generalization.
Takeaways & Limitations
Cross-subject generalization is bounded by inter-subject biological diversity, motivating adaptive feature transformation and distribution-alignment strategies.
Abstract
from arXiv · showhide
Identifying individual learning styles optimizes pedagogical efficacy. While traditional questionnaires are structured, behavioral tracking methods require prolonged interaction log accumulation. To overcome these temporal constraints, this paper proposes an objective Electroencephalography (EEG) approach evaluating Phase Locking Value (PLV) connectivity against localized features across the Active-Reflective (AR) and Verbal-Visual (VV) Felder-Silverman dimensions. EEG signals were recorded from 28 participants during Raven's Advanced Progressive Matrices tasks. Support Vector Machine classification used Leave-One-Subject-Out Cross-Validation (LOSO-CV) alongside a 70:30 intra-subject split. The VV dimension achieved 70.00% subject-level accuracy driven by distinct fronto-occipital polarization. Conversely, the AR dimension yielded lower cross-subject generalizability (55.56%) due to overlapping executive networks and a "Systematic Neural Inversion" phenomenon, where stable individual connectivity signatures operated diametrically opposed to global boundaries (up to 20-0 voting margins). Ultimately, these outcomes demonstrate that rigid "one-size-fits-all" classifiers are bounded by biological diversity, emphasizing the need for future adaptive feature transformation techniques to bridge the cross-subject generalization gap.
I. INTRODUCTION
The paper motivates rapid, objective EEG-based learning-style recognition and evaluates connectivity alongside localized features with subject-independent validation.
- Behavioral learning-style classification can require multiple course sessions to accumulate sufficient interaction logs.
- EEG offers a rapid, direct alternative for identifying learning tendencies.
- Distributed neural networks support cognitive functions associated with the Active–Reflective and Verbal–Visual dimensions.
- The study compares PLV with PSD, entropy, and statistical features for Active–Reflective and Verbal–Visual classification.
- The methodology spans EEG recording and labeling, preprocessing, multi-domain feature extraction, machine-learning classification, and neurophysiological evaluation.
A. Participants
The study records EEG from selected young adult participants performing language-free cognitive tasks, then filters and segments the signals for analysis.
- Twenty-eight healthy college students aged 18–20 participated, with balanced learners excluded using an ILS score threshold.
- EEG was recorded at 250 Hz from eight electrodes during twenty independent 15-second Raven’s Advanced Progressive Matrices trials.
- Raven’s Advanced Progressive Matrices provided a language-free, high-load cognitive task intended to elicit style-specific processing strategies.
- Signals were filtered from 8–30 Hz with a fourth-order zero-phase Butterworth filter and segmented into 1-second nonoverlapping windows.
D. Feature Extraction
The feature-extraction stage derives time, frequency, complexity, and connectivity descriptors, with PLV representing phase synchronization between EEG channels.
- Four feature domains were extracted: time, frequency, complexity, and connectivity.
- PSD features quantified Alpha and Beta band power across eight channels, producing 16 features.
- Five statistical measures were computed per channel, producing 40 features.
- Sample Entropy was computed per channel to index cognitive load, producing eight features.
- PLV quantified phase-synchronization consistency between channel pairs, with 28 unique pairs computed for the eight-channel system.
E. Classification and Evaluation Framework
The framework normalizes each participant’s data, classifies features with an RBF-kernel SVM, aggregates predictions by voting, and evaluates intra- and inter-subject generalization.
- Individual Z-score normalization was applied to address inter-subject heterogeneity.
- Classification used a Support Vector Machine with a Radial Basis Function kernel.
- The SVM decision function combines support-vector coefficients, class labels, bias, and the RBF kernel.
- Subject-level predictions used hierarchical majority voting across windows and trials.
- Evaluation used an epoch-wise 70:30 split as an internal baseline and LOSO-CV for subject-independent performance without data leakage.
- Accuracy, Precision, Recall, and F1-Score quantified performance for each FSLSM dimension.
F. Neurophysiological Connectivity Analysis
The study estimates functional connectivity with PLV between eight frontal, central, and occipital electrodes, then isolates statistically significant, high-magnitude class differences. These differential networks provide neurophysiologically supported features for distinguishing learning-style classes.
- PLV measures connectivity between electrode pairs across frontal, central, and occipital sites.Class-specific patterns retain the 25% strongest synchronizations.
- Differential connectivity matrices retain edges that are statistically significant and among the top 20% of absolute PLV differences.Edges require an independent t-test result of p < 0.05 and ∆PLV above the 80th percentile.
- The resulting differential networks encode positive and negative class differences through red and blue edges, with thickness proportional to ∆PLV.
III. EXPERIMENTAL RESULTS
The experiments process EEG into filtered signals and PLV connectivity, then compare multiple feature domains under subject-independent evaluation. PLV performs best for VV, while PSD performs best for AR.
- The processing pipeline applies an 8–30 Hz Butterworth filter before deriving the PLV connectivity matrix.The filter produces stable, zero-mean signals while isolating Alpha and Beta oscillations.
- LOSO cross-validation evaluates whether the classifiers generalize to unseen subjects across PLV, PSD, Sample Entropy, and time-domain statistics.
- 70.00% subject-level accuracy was achieved by PLV for the Verbal–Visual dimension.The result is reported in the LOSO comparison across feature domains.
- 61.11% subject-level accuracy was achieved by PSD Alpha/Beta features for the Active–Reflective dimension.This indicates that AR classification was best captured by rhythmic oscillations and spectral power distributions.
C. Analysis of the Generalization Gap (LOSO vs. 70:30 Split)
Comparing intra-subject and subject-independent validation reveals a substantial generalization gap. Hierarchical voting improves stability by aggregating predictions across windows and trials.
- 100.00% PLV accuracy was achieved for both dimensions in the 70:30 split, whereas LOSO validation showed a sharp decline.The decline highlights pronounced inter-subject variability and challenging cross-subject generalization.
- Subject-independent LOSO performance is constrained by idiosyncratic EEG signals that differ across participants.
- Hierarchical majority voting across trials and subjects filters transient noisy segments and stabilizes classification.
- 61.11% AR subject-level accuracy exceeded the 55.89% window-level baseline after aggregating predictions.The cumulative majority vote over the session provides a more resilient reflection of cognitive style.
E. Neurophysiological Connectivity Analysis
PLV maps reveal different connectivity organizations across learning-style dimensions. AR profiles overlap across executive networks, whereas VV profiles show stronger spatial polarization associated with higher subject-level accuracy.
- Active–Reflective (AR) Connectivity Profiles: Reflective participants show denser global synchronization across frontal, central, and occipital regions than Active participants.Active connectivity is more streamlined along the frontal–occipital axis.
- Processing Pipeline: The EEG pipeline progresses from raw signals to 8–30 Hz filtered signals and a derived PLV matrix.The matrix is shown for a Verbal learner in the processing visualization.
- Active–Reflective (AR) Connectivity Profiles: Overlapping executive circuitry makes baseline synchronization profiles difficult to distinguish across unseen subjects during LOSO validation.
- Verbal–Visual (VV) Connectivity Profiles: 70.00% subject-level accuracy in VV coincides with distinct anterior–posterior spatial divergence between Verbal and Visual profiles.Verbal forms a left-frontal and prefrontal hub, whereas Visual emphasizes right-hemisphere and posterior occipital pathways.
IV. DISCUSSION
The discussion identifies systematic, person-specific connectivity inversions as a major barrier to subject-independent learning-style classification. It links stronger VV separability to spatial polarization, while AR performance is constrained by overlapping executive networks and limited sample size.
- Systematic Neural Inversion: 20–0 voting margins show that some subjects were classified with high confidence in the opposite style, indicating systematic neural inversion rather than random tracking failure.Several Reflective participants were classified as Active with 100% certainty, despite decisive voting margins.
- Systematic Neural Inversion: Stable individual connectivity signatures can point opposite to the population-averaged model boundary, making PLV interpretation person-specific.The same pattern appears as a correct-or-inverted split in the VV dimension.
- Methodological Scope: The binary paradigm isolates raw discriminative boundaries, while extending the analysis to continuous or multiclass structures would require a substantially larger cohort than N = 28.The study uses binary classification to limit the SVM parameter space and reduce overfitting risk.
- Connectivity Comparisons: The AR figure compares grand-average Active and Reflective connectivity with significant differences, supporting analysis of their distinct network organization.The comparison concerns connectivity patterns across the AR dimension.
- Connectivity Comparisons: The VV connectivity figure contrasts grand-average Verbal and Visual networks with their significant differences, highlighting spatially polarized connectivity patterns.The figure organizes the comparison into Verbal, Visual, and significant-difference panels.
- Future Directions: Future work targets distribution alignment and adaptive feature transformation to reduce the cross-subject generalization gap before covering additional FSLSM dimensions.The planned extensions include the Sensing–Intuitive and Sequential–Global dimensions.
V. CONCLUSION
The paper evaluates PLV as an EEG biomarker for recognizing FSLSM learning styles under subject-independent validation. VV classification performs better than AR, while systematic neural inversion shows that global models remain bounded by individual biological diversity.
- Conclusion: 70.00% VV accuracy under LOSO-CV reflects distinct fronto-occipital polarization, whereas AR reaches 55.56% because executive networks overlap topologically.These results compare subject-independent performance across the two evaluated FSLSM dimensions.
- Conclusion: 20–0 voting margins reveal stable individual connectivity signatures operating directly opposite to population-averaged decision boundaries.The finding is identified as the Systematic Neural Inversion phenomenon.
- Conclusion: Global models remain strictly bounded by individual biological diversity, motivating adaptive feature transformations to mitigate the cross-subject generalization gap.The proposed methodological shift addresses heterogeneous neural patterns across individuals.