Clusterer is the entry point of ARC, WCA and Limbo. It parameterizes and standardizes all three of them, since they are effectively the same algorithm.
First, the Clusterer initializes a few data structures through the ClustererArguments inner class.
- An Architecture is initialized with all the input information: input data, such as element dependencies and concern data. The ClustererArguments identifies whether concern data was input at all, for similarity measures ("clustering techniques") that do not take semantic information into consideration.
- A StoppingCriterion is created with the decision algorithm of when to stop the clustering process.
- A SerializationCriterion is created to handle serialization decisions, i.e. if/when to serialize during the clustering process.
- The SimMeasure to be applied is identified. This is the only part of the clustering techniques that actually differs, and therefore is the critical argument for determining how the architecture is recovered.
The actual clustering process (computeArchitecture) initializes a SimilarityMatrix by feeding it the initialized Architecture object and the selected SimMeasure. The SimilarityMatrix will calculate the similarity value of every cluster pair and hold them in order.
Here the algorithm enters a loop, predicated on the selected and parameterized StoppingCriterion. The highest-similarity cluster pair is requested from the SimilarityMatrix and is merged, with the Cluster constructor handling all the necessary merge actions depending on the selected similarity measure. Finally, the SimilarityMatrix is updated to reflect the clusters being merged. At this point, the SerializationCriterion is queried, and the Architecture is serialized if appropriate. Once the StoppingCriterion accuses the end of the process, the loop terminates and the resulting Architecture is returned.
All single-step architecture recovery techniques based on a similarity value can be implemented with this algorithm:
- New similarity measures can be implemented as SimMeasures, modifying the manner in which the SimilarityMatrix is calculated.
- New StoppingCriteria are easy to implement.
- Input handling is relegated to the Architecture object constructor, such that different input formats can have handlers implemented for them by extending the Architecture class. Processing and post-processing methods are made available for overriding, with the processing initialization being dedicated for structural data and the post-processing initialization being dedicated for (optional) semantic data.
- The format of semantic data can be overriding by extending the Cluster class, such that new concern extraction methods can be implemented with minimal refactoring effort. If the semantic data is in the form of a feature vector, no changes are necessary: it suffices to create a parser to read them into the existing data structures.
Multistep architecture recovery techniques (that is, recovery techniques that involve the use of different similarity measures at different stages) are still not viable through this design. Hybrid architecture recovery techniques are viable by calculating the composing similarity measures and assigning them weights between 0..1 such that the sum of the weights is 1.