• To get the results that answer to research questions in the paper, first run runexp.py to get measurement scores for all planners

• Then run the specific rqx.py to get results for the corresponding RQ.

• A sample result for all 3 RQs is placed under results directory.

A defect predictor can classify clean and buggy codes within a software project by learning from class-level CK code metrics. TimeLIME aims to interpret such predictor to gain information on what is the most critical factor contributing to a defect classification. Then TimeLIME will propose maintainable and achievable plans to each individual class within the project. Such plans can be mapped onto some simple code refactoring methods, as summarized in refactoring.guru. The following is an example illustrating how TimeLIME's plans can be applied by practitioners.

#Example

LIME will report instance-based feature importance weights. Positive weights indicate a positive linear correlation between the feature value and predicted defect-proneness, and vice versa.

Frequent itemset mining will find each unique combination of changes that occured in the project from previous release to current release. Such result will be used to filter out plans if they are not precedented according to the historical records. TimeLIME only mines changes that happended to actionable features.

TimeLIME generate plans by finding the maximum conjunction of actionable changes with precedence support. For the instance used in this example, the plan could be: {decrease max_cc; decrease avg_cc; increase dam}.

One code refactoring method that can match the plan proposed by TimeLIME is extract method. This method can be used to split one method with relatively high cyclomatic complexity into 2 simpler methods.

The effect of extract method is shown as follows. By extracting the rather complex piece of code into an individual method, maintainability and readability will get improved for future maintenance.

• Before :
  
• After :
  

Apparently, after the extract method both avg_cc and max_cc are reduced and dam increased since the new method is private.