A comparison of two ways to read a user’s emotional state, from heart-rate signals and from everyday computer activity, so that calendar apps can adapt to stress and mood.

Authors

Feiting Yang (University of Toronto), Antoine Moevus (DIRO, Université de Montréal), Steve Lévesque (ÉTS)

Venue

IEEE IISA 2025, Mytilene, Greece

Year

2025

Pages

1–8

DOI

10.1109/IISA66859.2025.11311192

Preprint

arXiv:2506.19280 [cs.AI]

Keywords

Machine Learning, Deep Learning, UI/UX, HCI, Calendar Optimization, Computer Activity, Emotion Detection

Status

Published

Read the paper on IEEE Xplore

Overview

Human-computer interaction increasingly includes emotion recognition. This paper, written with Feiting Yang and Antoine Moevus, looks at bringing emotion detection into calendar applications so that the interface can react to a user’s emotional state and stress level, with the aim of improving both productivity and engagement.

Approach

We evaluate two complementary approaches on real-world datasets:

• a biometric approach, where heart rate extracted from ECG signals feeds LSTM and GRU networks that predict valence, arousal and dominance
• a behavioral approach, where several machine learning models classify emotions from fine-grained computer activity such as mouse movements, clicks and keystroke patterns

Findings

Both approaches work, but the computer-activity approach proved more consistent and more accurate, especially for mouse interactions, which reached about 90% accuracy. In the biometric approach, GRU networks outperformed LSTM models, with valence prediction reaching 84.38% accuracy.

Publication details

Presented at the International Conference on Information, Intelligence, Systems & Applications (IISA 2025), held July 10–12, 2025, in Mytilene, on the island of Lesvos, Greece, and published by IEEE. A preprint has been available on arXiv since June 2025.

Read the paper on IEEE XploreRead the preprint on arXivView on Google Scholar