Software maintenance is the process of modifying a software system to fix defects, improve performance, add new functionality, or adapt the system to a new environment. A maintenance task is often initiated by a bug report or a request for new functionality. Bug reports typically describe problems with incorrect behaviors or functionalities. These behaviors or functionalities are known as features. Even in very well-designed systems, the source code that implements features is often not completely modularized. The delocalized nature of features makes maintaining them challenging. Since maintenance tasks are expressed in terms of features, the goal of this dissertation is to support software maintenance at the feature-level. We focus on two tasks in particular: feature location and impact analysis via feature coupling.

Feature location is the process of identifying the source code that implements a feature, and it is an essential first step to any maintenance task. There are many existing techniques for feature location that incorporate various types of analyses such as static, dynamic, and textual. In this dissertation, we recognize the advantages of leveraging several types of analyses and introduce a new approach to feature location based on combining dynamic analysis, textual analysis, and web mining algorithms applied to software. The use of web mining for feature location is a novel contribution, and we show that our new techniques based on web mining are significantly more effective than the current state of the art.

After using feature location to identify a feature's source code, maintenance can be completed on that feature. Impact analysis should then be performed to revalidate the system and determine which other features may have been affected by the modifications. We define three feature coupling metrics that capture the relationship between features based on structural information, textual information, and their combination. Our novel feature coupling metrics can be used for impact analysis to quantify the strength of coupling between pairs of features. We performed three empirical studies on open-source software systems to assess the feature coupling metrics and established three major results. First, there is a moderate to strong statistically significant correlation between feature coupling and faults. Second, feature coupling can be used to correctly determine about half of the other features that would be affected by a change to a given feature. Finally, we found that the metrics align with developers' opinions about pairs of features that are actually coupled.