Technical Summary: Computational Personality Assessment via Engineered Behavioral Features

Problem Statement

The classification of personality types, specifically the introversion–extroversion dimension, is a critical research area bridging psychology and computational science. While traditional assessment relies on self-report questionnaires (e.g., MBTI), these methods suffer from subjectivity, social desirability bias, and an inability to capture dynamic behavioral patterns. Existing machine learning approaches often rely on raw questionnaire data or social media text without sufficient theoretical grounding in feature construction. Furthermore, prior studies frequently lack rigorous statistical validation of performance differences between models or fail to demonstrate the specific contribution of feature engineering over raw data. This study addresses the need for an objective, data-driven pipeline that utilizes engineered behavioral metrics to improve classification accuracy and interpretability.

Methodology

The research employs a comprehensive machine learning pipeline applied to a dataset of 2,900 behavioral instances (1,491 extroverts, 1,409 introverts) sourced from a public repository. The methodology is structured around three core components:

1. Data Preprocessing and Feature Engineering

The dataset initially contains eight behavioral attributes (e.g., Time Spent Alone, Social Event Attendance, Stage Fear). The authors introduce a novel feature engineering step to create three composite features designed to operationalize latent psychological constructs:

• Social Ratio: Calculated as Social Event Attendance / (Time Spent Alone + 1). This captures the balance between social engagement and solitude preference.
• Social Engagement: Calculated as Social Event Attendance × Going Outside. This represents the multiplicative interaction of internal motivation and external manifestation.
• Recovery Need: Calculated as Drained After Socializing × Time Spent Alone. This quantifies the intensity of social exhaustion and the subsequent need for solitude.

Missing values (approx. 2.66%) were handled via median imputation for continuous variables and mode imputation for binary variables. All numerical features were standardized using StandardScaler.

2. Algorithmic Evaluation

The study evaluates 13 distinct machine learning algorithms spanning linear models, ensemble methods, neural networks, and probabilistic classifiers. The models include:

• Linear/Probabilistic: Logistic Regression, Linear SVM, LDA, QDA, Gaussian Naive Bayes.
• Tree/Ensemble: Decision Tree, Random Forest, AdaBoost, Gradient Boosting, XGBoost.
• Neural/Distance-based: MLP Neural Network, K-Nearest Neighbors (K-NN), RBF SVM.

The workflow utilizes 5-fold stratified cross-validation. Performance is assessed using Accuracy, Precision, Recall, F1-Score, and AUC-ROC.

3. Statistical and Ablation Analysis

To validate the efficacy of the proposed approach, the authors conduct:

• Ablation Studies: Comparing model performance on (a) original features only, (b) engineered features only, and (c) the full feature set (original + engineered).
• Statistical Significance Testing: Paired t-tests with Bonferroni correction are applied to determine if performance differences among top models are statistically significant.
• Feature Importance: Permutation importance analysis is used to rank predictors.
• Cluster Analysis: t-SNE visualization is employed to examine the separability of classes in the engineered feature space.

Key Results

Performance Metrics

• Top Performer: K-Nearest Neighbors (K-NN) achieved the highest accuracy of 91.90%, followed closely by RBF SVM and Gaussian Naive Bayes (both 91.72%).
• Baseline Comparison: When using only the original 8 features, the best model (RBF SVM) achieved 79.50% accuracy. The inclusion of engineered features improved the best model's accuracy to 91.90%, representing a 12.4% improvement.
• Engineered Features Alone: Using only the three engineered features yielded an accuracy of 85.30%, demonstrating their strong standalone predictive power.
• Feature Importance: The engineered features ranked among the top predictors, with Social Ratio being the most influential feature overall (Importance Score: 0.124).

Statistical and Structural Findings

• Statistical Equivalence: Paired t-tests revealed no statistically significant difference (p > 0.05) between the top three models (K-NN, RBF SVM, Gaussian Naive Bayes). This suggests practitioners can select models based on computational efficiency or interpretability rather than accuracy alone.
• K-NN Superiority: The authors attribute K-NN's success to the "well-clustered nature" of the engineered feature space. The composite features create distinct, compact clusters for introverts and extroverts, a structure that K-NN exploits effectively through local pattern recognition without assuming a specific data distribution.
• Misclassification Analysis: 63.2% of errors occurred near the decision boundary, suggesting that misclassified individuals likely exhibit mixed personality traits, a phenomenon consistent with psychological theory.

Significance and Contributions

The paper positions its contributions as a bridge between psychological theory and computational practice:

1. Theory-Guided Feature Engineering: The study demonstrates that creating composite features based on psychological constructs (e.g., social equilibrium, recharging needs) yields significant performance gains over raw data, validating the hypothesis that latent constructs can be effectively operationalized.
2. Comprehensive Benchmarking: By evaluating 13 diverse algorithms under identical conditions, the study provides a robust benchmark for personality classification, moving beyond isolated model comparisons.
3. Rigorous Validation: The inclusion of ablation studies and statistical significance testing addresses gaps in prior literature, providing empirical evidence for the value of specific features and the equivalence of top-performing models.
4. Practical Deployment Guidelines: The paper offers practical insights for model selection, noting that while K-NN is top-performing, Gaussian Naive Bayes offers superior computational efficiency (0.01s training time) with statistically equivalent accuracy, making it suitable for resource-constrained environments.

The authors conclude that their approach advances computational personality assessment by maintaining methodological rigor and theoretical grounding, offering validated models with immediate applicability in human resources, mental health screening, and adaptive systems.