Emotion Prediction using Keystroke Dynamics: A Comparative Study of Machine Learning Models

The keystroke dynamics has increasingly attracted attention as an efficient method for inferring emotional states from users' typing activity. The feasibility and effectiveness of predicting user emotions with keystroke patterns and self-reported labels are explored in this research using a web interface, It is shown that the consistency and better trade-off between the precision and recall, maintaining high accuracy overall of Random Forest classifier in comparison to other two classifiers. In this paper, the methodology is depicted which involves dataset preparation, feature extractions and model evaluations then finally competitive analysis. The results indicate that keystroke dynamics along with self-reported emotional tags can be a trustworthy noninvasive modality compared to the ground truth for basic emotion prediction. To assess the ability of keystroke-based features to predict suicidal ideation, three machine learning models (Random Forest, SVM, and Logistic Regression) were implemented for evaluation. The class imbalance problem was dealt by initial pre-processing and balancing the dataset using Synthetic Minority Oversampling Technique (SMOTE). The three different models have been assessed in terms of performance metrics-accuracy, precision, recall and F1-score based on the analysis the Random Forest machine learning model outperformed well with accuracy (95.4%), precision, recall and F1-score.