Authors Name: 
Tomas Higgins
University: 
Dublin Institute of Technology
Category: 
Computer Sciences
Award winner

Identifying Mood by Analysing Keystroke Dynamics

The aim of this paper is to investigate the possibility that intelligent systems could be used to recognise different human moods. A large amount of research has been done in recent years that demonstrates how keystroke dynamics can be used to confirm an individual’s identity. Certain characteristics and timings of keystroke rhythm have been found to be as unique as handwriting. The question posed by this paper is whether or not these unique characteristics change in such a manner that it is then possible to identify mood. An application was developed to address this question. This application executes in the background and records a user’s keystroke data and corresponding mood. Ten participants were monitored over a number of weeks resulting in datasets being collected ranging from 10 to 60 thousand different keystrokes. The data obtained was then evaluated using a large variety of different classifiers. The eight best performing classifiers produced error rates ranging from 0.155 to 0.567 on three target features. The initial application was modified to use a combination of the three best performing classifiers to return a prediction of the individual’s mood on the fly. It was found that on average the J48 classifier could successfully classify mood with 63.63% accuracy, BaggingREPTree with 63.6% accuracy, RandomForest with 61.53% accuracy, ClassRegression with 63.11% accuracy and SVM with 58.5% accuracy. The classifiers implemented in the application’s final build were, RandomForest, BaggingREPTree and J48. Near the end of the project’s lifecycle the application was updated and released with the mood prediction functionality enabled. Users returned feedback which suggests that that application frequently correctly predicted their current mood. It can be concluded, based on the results obtained from the classifier evaluations done throughout this project that a user’s unique keystroke pattern does change in such a manner as to facilitate the identification of mood.