Beyond Local Power: Functional Connectivity Analysis for Subject-Independent Learning Style Recognition

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.

The paper evaluates PLV against traditional localized features (PSD, entropy, and statistical measures) for FSLSM classification. Using Support Vector Machines (SVM), the authors classify the Active–Reflective and Verbal–Visual dimensions. To ensure robust generalization, a dual-validation strategy is employed: a standard 70:30 train-test split and a rigorous Leave-One-Subject-Out Cross-Validation (LOSO-CV) to assess subject-independent performance.

The contributions of this paper are threefold. First, it provides a complete feature evaluation across the frequency, time, complexity, and connection domains to discover highly accurate EEG biomarkers for the Felder-Silverman Learning Style Model (FSLSM). Second, to address the underlying neuronal heterogeneity in EEG data, this study establishes a subject-independent generalization framework applying a rigorous Leave-One-Subject-Out Cross-Validation (LOSO-CV) methodology. Finally, a full neurophysiological characterisation is conducted by directly mapping the FSLSM dimensions to phase-locking brain networks to understand the neurophysiological foundation underpinning individual learning styles.

Twenty-eight healthy college students (aged 18–20) with no history of neurological or psychiatric disorders participated in this study. To ensure robust labeling and minimize “neural inversion” artifacts, only participants with a score of ≥ 5 (moderate-to-strong preference) on the Index of Learning Styles (ILS) survey were included; balanced learners (scores 1–3) were excluded. Following a balanced-class approach, 18 participants were selected for the Active–Reflective (AR) dimension (9 Active, 9 Reflective) and 10 for the Verbal–Visual (VV) dimension (5 Verbal, 5 Visual). This 50:50 distribution establishes a strict theoretical chance-level accuracy of 50.00%, eliminating majority class bias.

Continuous EEG signals were recorded at a 250 Hz sampling rate using an OpenBCI Cyton board from eight electrodes (Fp1, Fp2, F3, F4, C3, C4, O1, and O2) with a left earlobe (A1) reference. Cognitive stimulation was induced via 20 independent, 15-second trials of Raven’s Advanced Progressive Matrices (RAPM) Set II. RAPM serves as a language-free, high-load cognitive catalyst that forces style-specific processing strategies under the FSLSM framework. The total recording time was approximately five minutes per participant.

Raw EEG data were band-pass filtered between 8–30 Hz using a fourth-order, zero-phase Butterworth filter. The 15-second trials were segmented into 1-second non-overlapping windows. To ensure reliable functional connectivity estimation, features were calculated at the whole-trial level by combining these windowed segments. In line with previous protocols, no additional artifact rejection or re-referencing was performed to maintain methodological consistency.

Four feature sets were extracted across the time, frequency, complexity, and connectivity domains: Power Spectral Density (PSD) via Welch’s method for Alpha (8–13 Hz) and Beta (13–30 Hz) bands across 8 channels (16 features); five statistical measures (mean, standard deviation, skewness, kurtosis, and RMS) per channel (40 features); Sample Entropy (SE) per channel to index cognitive load (8 features); and Phase Locking Value (PLV) to quantify phase synchronization consistency between channel pairs. For the 8-channel system, 28 unique pairs were computed per window, representing the global synchronization of the brain network.

To address inter-subject heterogeneity, individual Z-score normalization was applied to each participant’s data. Classification was performed using a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel. Subject-level predictions were determined using a hierarchical two-stage majority voting system across windows and trials. Model generalization was evaluated using two distinct validation schemes: an epoch-wise 70:30 split and Leave-One-Subject-Out Cross-Validation (LOSO-CV). The epoch-wise 70:30 split was conducted as an internal validation baseline to confirm model learnability and intra-subject feature consistency. Meanwhile, LOSO-CV was implemented to rigorously evaluate subject-independent performance while completely eliminating data leakage risks; notably, all core findings and subject-independent claims in this study rely exclusively on the LOSO-CV scheme.

Functional connectivity was estimated by calculating the Phase Locking Value (PLV) between pairs of electrodes among a set of eight electrodes located at frontal (Fp1, Fp2, F3, F4), central (C3, C4), and occipital (O1, O2) sites. Class-specific connectivity patterns were obtained by retaining only the top 25% strongest synchronizations for each class. To detect features that best discriminate between classes, a difference connectivity matrix (∆PLV) was derived through a double threshold technique. Edges were displayed only when they satisfied two requirements: (1) statistical significance by means of an independent t-test (p 80th percentile).

The Leave-One-Subject-Out (LOSO) evaluation results show that PSD (Alpha/Beta) features yielded the highest accuracy for the Active–Reflective (AR) dimension (61.11%), indicating that AR styles are better captured by rhythmic oscillations and spectral power distributions. Conversely, PLV achieved the highest accuracy for the Verbal–Visual (VV) dimension (70.00%). This demonstrates that VV processing is predominantly a large-scale cortical network phenomenon relying on inter-regional synchronization (e.g., occipital-temporal integration) rather than localized power alterations.

A substantial “generalization gap” was observed when comparing the subject-independent LOSO scheme against the intra-subject 70:30 train-test split baseline. In the 70:30 split, PLV achieved perfect accuracy (100.00%) for both dimensions. However, the sharp performance decline in LOSO validation highlights pronounced inter-subject variability. While the SVM model effectively learns individual neural “fingerprints” from known subjects, cross-subject generalization remains highly challenging due to the idiosyncratic nature of EEG signals.

The hierarchical majority voting mechanism at the trial and subject levels effectively filtered out transient, noisy EEG segments to stabilize classification. Aggregating window-level predictions into trial- and subject-level decisions consistently enhances classification stability. For instance, the AR subject-level accuracy (61.11%) outperforms its window-level baseline (55.89%). This indicates that while brief EEG epochs may be ambiguous, the cumulative majority vote over the entire session provides a more resilient reflection of a learner’s cognitive style.

Functional synchronization networks were visualized using PLV mapping to investigate the biological underpinnings of classification performance. For the Active–Reflective (AR) dimension, connectivity patterns reveal distinct synchronization densities. The Active style exhibits a streamlined network localized along the Frontal–Occipital axis (Fp2-O2, F4-O2). Conversely, the Reflective style displays a denser, more complex global integration across Frontal–Central–Occipital regions. The differential map is dominated by blue edges (Reflective > Active), showing that the Reflective style recruits significantly higher global synchronization driven by intensive mental simulation and executive control. Key discriminative features include dense inter-hemispheric coupling (Fp1-Fp2, C3-C4) and robust long-range prefrontal-to-occipital connectivity. This topological overlap and shared executive circuitry explain why the classifier struggles to distinguish baseline synchronization profiles across unseen subjects during LOSO validation.

For the Verbal–Visual (VV) dimension, neural signatures demonstrate clear spatial polarization. The Verbal style forms a dense hub centered in Left-Frontal and Prefrontal regions (Fp1, F3) associated with linguistic circuits and verbal rehearsal. Conversely, the Visual style activates localized Right-Hemisphere and posterior occipital pathways (Fp2-O2, F4-O2) tailored for spatial information processing. Significant differences confirm an overwhelming dominance of red edges (Verbal > Visual). This pronounced, distinct spatial divergence (Anterior vs. Posterior) directly accounts for the superior subject-level classification accuracy (70.00%) in the VV dimension compared to the overlapping functional networks of AR.

The Leave-One-Subject-Out (LOSO) evaluation reveals a prominent “Systematic Bias” rather than a stochastic tracking failure across subject-independent distributions. As demonstrated in the individual voting analysis, several subjects achieved a 0% classification accuracy despite maintaining highly decisive voting margins (e.g., 20–0). Within the Active–Reflective (AR) dimension, multiple Reflective participants (such as Subjects 11, 12, 13, and 18) were classified as Active with 100% certainty. This empirical outcome uncovers a distinct “Systematic Neural Inversion” effect, indicating that while functional connectivity signatures remain highly stable within individuals, their biological manifestation can be mathematically oriented in a direction diametrically opposed to the population-averaged global model boundary. Similarly, the Verbal–Visual (VV) dimension exhibits a clear binary “Correct or Inverted” split (e.g., Subject 8), confirming that while Phase Locking Value (PLV) acts as a highly structured biomarker, its interpretation remains strictly person-specific.

These pronounced 20–0 voting errors provide critical empirical insights into the explicit limitations of utilizing a rigid, “one-size-fits-all” global classifier on volatile biological data. A generalized framework inherently presumes a homogeneous distribution of feature representations across an entire population; however, human neurophysiology sharply violates this assumption due to high inter-subject heterogeneity. Because the absolute consistency of the misclassifications is not driven by random sensor noise, it confirms that while a population-averaged model can successfully map prominent spatially polarized networks—such as the anterior-posterior divergence during Verbal–Visual processing—its generalizability remains strictly bounded by individual biological diversity. This limitation is heavily reflected in the substantial “generalization gap” where perfect intra-subject learnability (100.00% in the 70:30 split) drops significantly under cross-subject validation, particularly in the AR dimension, where the cross-subject PLV accuracy is limited to 55.56% due to topological overlaps in shared executive circuits.

From a methodological perspective, choosing a binary classification paradigm in this preliminary study serves as a critical first step to mathematically isolate these raw discriminative boundaries without over-parameterizing the model. Attempting to map high-dimensional phase-synchronization features to a continuous psychometric scale or a multi-class structure requires a substantially larger cohort to prevent severe overfitting. Given the current sample size (N = 28), the binary paradigm effectively restricts the parameter space of the SVM classifier, ensuring that the documented neural inversions are genuine representations of neurophysiological variance rather than statistical artifacts of overfitting. Furthermore, these findings underscore that standard subject-wise Z-score normalization remains insufficient to fully eliminate individual neural fingerprints. Consequently, subsequent stages of this research framework will focus on structural methodological advancements specifically geared toward mitigating the cross-subject generalization gap. Future work will explore more robust distribution alignment strategies and adaptive feature transformation techniques to better harmonize the heterogeneous neural patterns across different individuals before expanding the scope to systematically incorporate the remaining Sensing–Intuitive and Sequential–Global dimensions of the FSLSM.

This paper demonstrates that PLV serves as an effective quantitative biomarker for FSLSM recognition. Under LOSO-CV, the VV dimension achieved 70.00% accuracy due to distinct fronto-occipital polarization, whereas the AR dimension dropped to 55.56% due to topologically overlapping executive networks. A major contribution of this foundational study is the empirical evaluation of a rigid “one-size-fits-all” model, highlighted by the identification of the “Systematic Neural Inversion” phenomenon. The presence of high-confidence yet completely inverted predictions (20–0 voting margins) reveals that stable, individual functional connectivity signatures can operate in direct opposition to population-averaged boundaries. Ultimately, these outcomes establish a critical neurophysiological baseline, demonstrating that global models remain strictly bounded by individual biological diversity. These findings strongly motivate future methodological shifts toward adaptive feature transformation techniques specifically designed to mitigate the cross-subject generalization gap.

Improvements to AI Systems:

1. Adaptive Feature Transformation via Subject-Specific Normalization

• Implement a meta-learning layer that learns per-subject feature transformations (e.g., learnable affine mappings or adversarial domain adaptation) to align individual neural signatures to a shared latent space before classification.

• The improved AI system can dynamically reorient inverted connectivity patterns (e.g., flipping subject-specific PLV distributions) to match global boundaries, reducing Systematic Neural Inversion errors.

2. Confidence-Aware Rejection and Personalization

• Add a confidence thresholding mechanism that flags high-margin but low-agreement predictions (e.g., 20–0 votes) as uncertain rather than forcing a binary label.

• The system can then trigger a personalized calibration step—e.g., fine-tuning on a small set of subject-specific trials—to correct inverted mappings, improving cross-subject robustness.

3. Multi-View Ensemble with Domain-Invariant Feature Selection

• Combine PLV connectivity features with localized features (PSD, entropy) using an ensemble that weights views based on their cross-subject stability (e.g., via domain-adversarial training).

• The improved system can automatically select features that generalize across subjects (e.g., fronto-occipital polarization for VV) while suppressing subject-idiosyncratic ones, boosting AR dimension accuracy from 55.56% toward VV-level performance.

4. Hierarchical Voting with Subject-Level Uncertainty Propagation

• Replace hard majority voting with a probabilistic aggregation that models subject-level noise and inversion probability (e.g., Bayesian hierarchical model or graph neural network over trial-level predictions).

• The system can output calibrated probabilities per subject, distinguishing true classifications from inverted ones, and provide interpretable confidence scores for each prediction.

5. Generative Data Augmentation for Inter-Subject Variability

• Train a conditional variational autoencoder (CVAE) on PLV matrices to generate synthetic subject-specific connectivity patterns, augmenting the training set with diverse neural topologies.

• The improved system can learn more robust decision boundaries by exposing the classifier to a wider range of biological variance, reducing overfitting to population-averaged patterns and improving LOSO-CV accuracy.

6. Adaptive Thresholding for Connectivity-Based Biomarkers

• Replace fixed top-25% and 80th-percentile thresholds with per-subject adaptive thresholds learned via reinforcement learning or gradient-based optimization.

• The system can dynamically adjust connectivity significance levels to each individual’s baseline synchronization, mitigating the effects of neural inversion and improving discriminative edge selection for both AR and VV dimensions.

What the Improved AI System Can Do:

• Achieve higher cross-subject generalization (e.g., >70% for AR and >80% for VV) by harmonizing individual neural fingerprints.

• Detect and correct systematic inversions in real-time, providing reliable learning-style classification for unseen individuals.

• Offer personalized confidence scores and explanations (e.g., which brain networks drive the decision) to support adaptive learning environments.

• Scale to multi-class FSLSM dimensions (e.g., Sensing–Intuitive, Sequential–Global) with reduced overfitting and improved robustness.