In complex skill-training systems, trainees are required to master multiple skills in a limited time, which may produce a large mental workload. Increased workload often affects performance, and trainees may get distracted or overloaded during training. Attention control is a critical activity in time-sharing environments with automatic tasks, and psychologists found that better attention control strategies can develop through training. Even though attention management is a key skill has to be acquired, it has not been considered to assess as a user model content sufficiently. In this paper, we propose an approach for attention-control modeling by detecting regular behavioral patterns that potentially explain the interdependency between primary and subtask performance. We can detect trainees' attention shift between tasks by interpreting the serial episodes of behaviors that have been uncovered. As a high attention needing training domain, we used Space Fortress game in which continuous input stream of ship maneuvering and intermittent event data are the source of the user model. We found the dependencies between these heterogeneous, multi-time streams and the point of attention shift. Domain experts or training coaches can infer the trainees' attention-control skill based on the detected rules of pattern that help them to instruct desirable strategies to handle multi subtasks.