Paper Reading #31

Identifying emotional states using keystroke dynamics

 

Authors:	Clayton Epp	University of Saskatchewan, Saskatoon, Saskatchewan, Canada
	Michael Lippold	University of Saskatchewan, Saskatoon, Saskatchewan, Canada
	Regan L. Mandryk	University of Saskatchewan, Saskatoon, Saskatchewan, Canada

 

Proceeding 

CHI '11 Proceedings of the 2011 annual conference on Human factors in computing systems 

 

Summary

• Hypothesis - Traditionally, emotion detection requires sensors to directly measure physical data of a participant. The researchers believe that their system can reliably detect participant emotions based on the dynamics of how they type.

• Content - This paper avoids other types of machine learning processes and data collection to determine user emotional state.  Here, the only data collected is solely they keystroke dynamics and feedback for the machine learning via questionnaires. Previously, non-fixed typing data collection was unable to accurately determine user emotional state, so fixed-length typing sets were giving for users for data analysis. Free text was too variable according to the researchers to accurately accomplish the researchers goals.
• Methods - Instead of gathering keystroke data in a laboratory setting, the data was collected at home in a daily basis. 26 participants were asked to fill out a questionnaire about how they felt emotionally and then have them write a fixed length piece of text. The data was collected regularly for 4 weeks. The information collected was in 4 categories:

1. Keystroke features - keystroke duration and latency.
2. Content features - which characters are being typed including capital letters and numbers.
3. Emotional state classes - data collected from the questionnaires that included 15 questions about their emotional state.
4. Additional data points - the active process name for each collected keystroke.

Also since people often took pauses during the tests, further computation was performed on the keystroke data. If a pause was longer than the standard deviation of the data, the data point was thrown away, which took out 0.07% of all the data.

• Results - Models were created from the data that detected confidence, hesitance, nervousness, relaxation, sadness, and tired states. Some states were temporary, such as excitement and anger, and were not classified properly within the fixed length text. However, only a few features were necessary to classify emotions accurately.

Discussion
The researchers accomplished their goal adequately but I think their system should have been designed to allow more flexibility, especially over time. To require that a user types a fixed length text is prohibitive to deploying their models to any free-form typing. However, since the content of the words typed are not analyzed itself, this may have simply been impossible. Otherwise, this type of machine learning could be applied to operating systems to act differently around users that give different emotional states through how they type. Also another data point they left out for data analysis is mouse movement. This data could further model generation and possibly make it more accurate.