The interaction logs which researchers collected during their information-architecture evaluation process contain detailed proof which shows how users select between successful and unsuccessful navigation routes. The predictive signal displays its initial appearance during task execution yet users exhibit different navigation patterns depending on their current task and interface they are using. The researchers of this study developed an early navigation failure prediction system which uses public interaction data to create task-specific prefix classification models. The study analyzes data from an open dataset which includes 180 participants completing 1800 tasks across six testing conditions that evaluate tree testing and highfidelity prototype navigation. A prefix-structural encoder works together with a regularized task-conditioned logistic model which predicts success based on the first k navigation actions. The researchers assessed model performance through participant-specific validation using three different machine learning techniques which included random forest, extra trees, and gradient boosting. The optimal configuration achieved 0.7833 accuracy, 0.7513 balanced accuracy, 0.8350 F1-score, and 0.7949 ROC–AUC performance at k = 3. The horizon analysis demonstration shows that predictive signals become accessible after users complete their first three actions. The ablation study proves that task conditioning functions as an essential component. The study results demonstrate that early trace analytics enable quick identification of navigation failures in information-architecture research while providing a useful method for customized assessment during usability testing.
Read MoreDoi: https://doi.org/10.54216/JCHCI.110101
Vol. 11 Issue. 1 PP. 01–09, (2026)