Paper Number
ICIS2025-2553
Paper Type
Short
Abstract
Tracking user behaviors and engagement through devices such as the computer mouse or keyboard has been used to predict important user outcomes, including cognitive conflict, product interest, and fraud, to name a few. For this, raw behavior data (e.g., xand y-position of the mouse cursor over time) is transformed into metrics (e.g., movement speed or deviation) to predict these outcomes. However, these classical features take each trial independently, overlooking higher-order links among users, stimuli, and behaviors. We introduce a Tensor-Enrichment Pipeline, applying Tucker, Higher-Order Singular Value Decomposition, and CANDECOMP/PARAFAC Decomposition, to extract five-dimensional user and cursor embeddings and append them to the original metrics. Using a random forest classifier, we compare baseline and enriched feature sets in a study (N=98) where participants’ buying decisions of 30 products were tracked. Tensor enrichments yield substantial gains: boosting accuracy from 35% to 69%, with higher recall and precision.
Recommended Citation
Coors, Christopher; Jenkins, Jeffrey; Valacich, Joseph S.; and Weinmann, Markus, "Decomposing the Mouse Cursor: Empirically-Derived Multi-Dimensional Metrics for Predicting User Behavior and Engagement" (2025). ICIS 2025 Proceedings. 31.
https://aisel.aisnet.org/icis2025/user_behav/user_behav/31
Decomposing the Mouse Cursor: Empirically-Derived Multi-Dimensional Metrics for Predicting User Behavior and Engagement
Tracking user behaviors and engagement through devices such as the computer mouse or keyboard has been used to predict important user outcomes, including cognitive conflict, product interest, and fraud, to name a few. For this, raw behavior data (e.g., xand y-position of the mouse cursor over time) is transformed into metrics (e.g., movement speed or deviation) to predict these outcomes. However, these classical features take each trial independently, overlooking higher-order links among users, stimuli, and behaviors. We introduce a Tensor-Enrichment Pipeline, applying Tucker, Higher-Order Singular Value Decomposition, and CANDECOMP/PARAFAC Decomposition, to extract five-dimensional user and cursor embeddings and append them to the original metrics. Using a random forest classifier, we compare baseline and enriched feature sets in a study (N=98) where participants’ buying decisions of 30 products were tracked. Tensor enrichments yield substantial gains: boosting accuracy from 35% to 69%, with higher recall and precision.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
16-UserBehavior