The study addresses the challenge of interpreting human activity traces surrounding workplace agents. It emphasizes that behavioral interpretation depends on temporal resolution, with different questions requiring different levels of detail.
A vocabulary of semantically normalized operators, motifs, episodes, and rhythms is constructed, applied to data from a large productivity suite. This yields a taxonomy of behavioral patterns that is stable across samples and predictive of future activity.
The approach demonstrates that no single resolution is optimal for all questions. Instead, behavioral trace interpretation should be multi-resolution and query-conditioned, allowing agents to access the appropriate temporal grain for each inquiry.
This method improves the understanding and forecasting of user behavior, which is relevant for designing more effective workplace agents.
Source: https://arxiv.org/abs/2609.04556