An Explanation Method for Models of Code
summary
Presents WheaCha, a method and tool (HuoYan) that explains predictions of code models by splitting an input program into ‘wheat’ (defining features) and ‘chaff’. Applies HuoYan to code2vec, seq-GNN, GGNN, and CodeBERT, reports per-program explanation times under ~20 seconds, finds models rely mainly on syntactic/lexical cues like identifier names, and shows WheaCha outperforms SIVAND, Counterfactual Explanations, Integrated Gradients, and SHAP at identifying faulty or biased trained models.
AI-written summary · about
related talks
18mEvaluating the effectiveness of Deep Learning Models for Foundational Program Analysis(…)
31mBeyond the Black Box: Interpreting ML Models with SHAP
35mTaking a peek under the hood - Interpreting black box model
16mBuilding Trust through Explainable AI
28mExplaining Machine Learning Models with LLMs
37mExplaining model explainability
15mAdversarial Examples for Models of Code
46mExplaining a Machine Learning blackbox