Explainable and Efficient Deep Learning for Personality Analysis: A Comparative Behavioral Intelligence Framework

Abstract

Automated personality prediction systems based on deep learning have achieved state-of-the-art performance but remain largely opaque offering no interpretable account of how predictions are reached. This opacity creates significant barriers to deployment in high-stakes domains including clinical assessment, employment screening, and adaptive education, where legal requirements (GDPR Article 22), ethical obligations (explainability, fairness), and scientific requirements (construct validity) mandate transparent, auditable model behaviour. This paper presents the Explainable Multimodal Deep Learning (E-MMDL) framework an enhanced personality analysis system that augments state-of-the-art trimodal prediction with a comprehensive, multi-method explainability layer integrating SHAP DeepSHAP, LIME, Gradient-weighted Class Activation Mapping (Grad-CAM), Integrated Gradients, and cross-modal attention visualisation. Three efficiency optimisation strategies structured pruning (30% sparsity, −29.6% parameters), knowledge distillation (45M student model, 3.8 ms GPU inference), and INT8 quantisation (7.2 ms GPU inference) are systematically evaluated for accuracy-efficiency trade-offs. A novel Behavioural Intelligence Analysis maps SHAP and LIME attributions to documented psychological personality-behaviour associations, confirming that computational features align with personality science theory and providing the first comprehensive cross-modal behavioural signal interpretation for Big Five personality AI. A structured ethical risk matrix addresses privacy, bias, fairness, and responsible deployment across seven application domains. Results confirm that the full E-MMDL maintains state-of-the-art accuracy (macro F1=0.921, AUC=0.969) while delivering comprehensive multi-level interpretability establishing explainability as a compatible, not competing, property of high-performance personality AI.