style-analyzer: fixing code style inconsistencies with interpretable unsupervised algorithms

When to help?

Lookout

PR review comments on GitHub

dataset

Classes predicted

"White box" is the key

• Builds trust with the users
• Provides interpretability
• Allows editing
• Allows blacklisting

Labels predicted

We concatenate common classes together.

• ␣"
• ␣+␣+␣+␣+
• '↲
• ↲→-↲
• ∅

No BPE or similar reduction is required.

Annotated code snippet

∅function␣classesToArray∅(␣value␣)␣{↲
             ⇥   if␣(␣isArray∅(␣value␣)␣)␣{∅return␣value∅;∅}
                 if␣(␣typeof␣value␣===␣"string"␣)␣{↲
             ⇥       return␣value∅.∅match(␣rnothtml␣)␣||␣[]∅;↲
             ⇤   }↲
                 return␣[]∅;↲
             ⇤}∅
        

Augmented token stream

a = b * 2

Machine Learning

Machine Learning

• Feature selection (univariate, ANOVA F-criterion)
• Hyperparameter optimization (Bayesian)
• 80% + 20% split

Rules

• a≤5 Λ b≤1 Λ c ⇒ α
• a≤5 Λ 1<b<4 ⇒ β
• 5<a<10 Λ c ⇒ γ
• a>5 Λ c Λ b>2 ⇒ α

Redundant attributes elimination

• Trigger imbalance: <5% activation rate
• For each attribute comparison build the set of false positives
• Build graph by thresholding Jaccard similarity between each pair (0.98)
• Run multilevel community detection
• Leave only one representative (greedily)

Annotated code snippet

∅function␣classesToArray∅(␣value␣)␣{↲
             ⇥   if␣(␣isArray∅(␣value␣)␣)␣{∅return␣value∅;∅}
                 if␣(␣typeof␣value␣===␣"string"␣)␣{↲
             ⇥       return␣value∅.∅match(␣rnothtml␣)␣||␣[]∅;↲
             ⇤   }↲
                 return␣[]∅;↲
             ⇤}∅
        

Rule example

• -1.value = {
• -3.value not in {else}
• +1.internal_type not in {StringLiteral}
• +2.roles not in {LITERAL, COMMENT}
• ˆ1.roles in {IF, STATEMENT}

Inference

1. Apply rules
   • Fixes to old code?
   • AST breakage?
   • Identification?
2. Generate code
   • Indentation?
   • Multiple lines?

Debugging

Lookout comments on GitHub

Precision > Recall

Precision

Prediction Rate

Results

Enough quality for a product?

Other evaluation is required

Comparison to linters

But what if the project already uses a linter?

Mining Software Repositories

Clustering

Given codebases X and Y

style_distance(X, Y) - how similar are the formatting styles?

Common formatting rules analysis

Which formatting styles dominate on GitHub?

Summary

• style-analyzer infers code human-friendly formatting rules
• Reaches 94% precision at 95% prediction rate
• Is based on Lookout - assisted code review framework
• May be useful for general mining

Thank you

• arxiv.org/abs/1905.06782
• machine-learning@sourced.tech
• @sourcedtech @vadimlearning @warenlg
• src-d
• blog.sourced.tech
• Awesome #MLonCode

bit.ly/2LRkWdW