One-Sentence Summary
Where2Change extends Palomba et al.'s work on mining user reviews for information that helps maintain and evolve mobile apps.
Background
App developers often rely on user reviews and ratings to maintain their apps and locate bugs. Previous research has applied IR-based fault localization techniques, but important gaps remain. First, most prior work focused on desktop applications rather than mobile apps. Second, existing techniques cannot fully automate user-review analysis, which means developers may miss newly reported problems.
Related Work
Palomba et al. proposed CHANGEADVISOR, which clusters reviews by similar requirements and identifies the source-code classes that need to be changed for each review cluster. However, this technique depends on reviews that contain detailed information; otherwise, it cannot map a review to any class.
Method
- First, extract informative sentences, or user feedback, from user reviews.
- Second, use tool-SURF to automatically classify user feedback into five categories, then cluster the feedback with HDP, which achieves the best clustering performance among six popular clustering algorithms.
- Third, treat each user-feedback cluster as a query to search for source-code classes that should be changed.
- Fourth, propose a more accurate similarity metric, called weight-selection-based cosine similarity, to measure the similarity between an enriched comment cluster and the source code.
- Finally, for each user-feedback cluster, Where2Change returns a ranked list of classes that may need to be changed.
Experiment
Discussion
Overall, the researchers proposed a weight-selection-based cosine similarity metric to compute the similarity between enriched user-feedback clusters and source code. For future work, the authors plan to explore better ways to select user feedback related to real faults and feature requests reported in issue trackers. They are also interested in developing a new method to locate source-code classes that should be changed for each individual user-feedback entry, rather than for an entire feedback cluster.
Reading Summary
What is the motivation?
To propose a framework that automates user-review analysis and addresses a key limitation of CHANGEADVISOR: its lack of robustness when reviews are sparse or insufficiently detailed.
What is the novelty?
The proposed framework implements a weight-selection-based cosine similarity metric.
What is the contribution?
It overcomes an important limitation of CHANGEADVISOR by improving the mapping between user feedback and source-code classes.