Emotion classification from keystroke dynamics: a dual-input deep learning approach

Accurate emotion recognition from human-computer interactions is essential for adaptive and user-centered applications. Traditional affect recognition methods often overlook individual behavioral patterns, thereby limiting their effectiveness across diverse user populations. We propose a personalized deep learning framework for emotion recognition from keystroke dynamics that addresses this limitation. Our dual-input neural network architecture combines temporal features extracted from keystroke synchronization patterns and user-specific metadata via Feature-wise Linear Modulation (FiLM) and Low-Rank Adaptation (LoRA). This design enables the model to capture both universal typing patterns and individual-specific behavioral characteristics. Trained and evaluated on the EmoSurv dataset, the proposed model achieves around 98% accuracy across five emotion classes: neutral, happy, sad, calm, and angry. Experimental results demonstrate that the personalized dual-input architecture with FiLM and LoRA conditioning improves performance compared to user-agnostic baselines. Per-class analysis further indicates more consistent recognition across emotions. By employing parameter-efficient FiLM and LoRA modules for user-specific conditioning, the framework introduces personalization while maintaining a compact model design, supporting its applicability to real-time keystroke-based emotion recognition scenarios.