Refine anomaly detection features

[JohT]

🚀 Feature

• Remove constant features from anomaly detection: Some of the newly introduced features may turn out to be constant and therefore not informative. Such features are now removed during the data preparation step.
• Consider abstract classes to have 70% abstractness: When calculating abstractness (according to Robert C. Martin’s object-oriented design metrics), interfaces and abstract classes were previously treated as equal. Since abstract classes often contain concrete implementations for inheriting classes, they are now weighted with 70% abstractness.
• 🌟 Add abstractness to anomaly detection features: Adds “abstractness” (Robert C. Martin’s Object-Oriented Design Metrics) as an additional feature for anomaly detection. This introduces semantic information to better distinguish, for example, between interfaces and concrete classes.
• 🌟 Add weakly/strongly connected components to anomaly detection features: The graph algorithm “strongly connected components” produces a directed acyclic graph (DAG) whose members contain cycles, which is often the case for code elements such as Java types. Additionally, “weakly connected components” group one or more strongly connected components into an “island.” Together, they add hierarchical and structural information and enable topological sorting (which requires a DAG).

⚙️ Optimization

• Optimize Neo4j start: Assumes a faster Neo4j startup by initially waiting 4 seconds instead of 16, followed by several 1-second waits. This results in a better balance between polling frequency and startup delay.
• Skip PCA when exploring anomaly detection solely on node embeddings: When exploring anomaly detection based only on node embeddings, the exploratory notebook includes an experiment that relies exclusively on embeddings. In this case, principal component analysis (PCA) is skipped, as it is only required to balance the number of embedding-based features with other feature types (which are not present here).
• Improve readability of anomaly detection feature plots: Embedding plots do not require axes. The dimensionally reduced x and y values have no inherent meaning beyond relative position and distance between code units.
• Add topological sort max distance from source for strongly connected components to anomaly detection features: This feature provides an architectural view of code layers based on dependencies. Layer zero, for example, represents the most general or abstract layer, containing code units that do not depend on others but are frequently used by them.

🛠 Fix

• Fix typo in feature existence query for article to page rank feature: This typo caused the PageRank-to-ArticleRank difference to be recalculated every time instead of reusing precomputed values.