Using Eye Gaze Data to Recognize Task-Relevant Source Code Better and More Fine-Grained

Models to assess a source code element's relevancy for a given change task are the basis of many software engineering tools, such as recommender systems, for code comprehension. To improve such relevancy models and to aid developers in finding relevant parts in the source code faster, we studied developer's fine-grained navigation patterns with eye tracking technology. By combining the captured eye gaze data with interaction data of 12 developers working on a change task, we were able to identify relevant methods with high accuracy and improve precision and recall compared to the widely used click frequency technique by 77% and 24% respectively. Furthermore, we were able to show that the captured gaze data enables to retrace which source code lines developers found relevant. Our results thus provide evidence that eye gaze data can be used to improve existing models in terms of accuracy and granularity.