.png){kind=icon}
.png){kind=icon}
Return
Multi-Objective Explainable AI for Risk-Aware Test Selection in Multi-Layer Continuous Integration Systems
.png){kind=icon}
OA
AI
DOI：10.3390/electronics15173854.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Continuous Integration (CI) pipelines for multi-layer software systems face a tension between validation completeness and execution efficiency. Regression test selection and learned prioritization address this tension but typically operate as black boxes, offering no interpretable justification for their decisions. This paper introduces MOX, a preliminary decision-support framework for risk-aware test selection that jointly optimizes four objectives: execution cost, fault-detection impact, affected-test coverage, and a proposed explainability score, formalized as a first-class, numerically optimizable objective (a structural traceability proxy, not a validated measure of human interpretability) and traded off within a Pareto-based selection mechanism governed by a configurable impact budget. We evaluate MOX retrospectively on two public industrial CI datasets from ABB, iofrol (1941 tests, 320 cycles) and paintcontrol (89 tests, 352 cycles), containing execution logs of integration-level regression tests for industrial robot software. No new tests are developed or executed: selections are replayed per historical cycle against the recorded pass/fail verdicts. Because the logs contain no code changes, each cycle’s change event is inferred from its failing tests, so detection figures must be read as retrospective retention of failed tests under a selection budget, not prospective fault detection. Under this protocol MOX retains 96.7% (iofrol) and 98.2% (paintcontrol) of recorded failures at 53.9–57.5% of full-suite cost over the evaluated failure-containing cycles and raises the Average Percentage of Faults Detected (APFD) from 0.507 to 0.626 and from 0.596 to 0.815; paired Wilcoxon tests are highly significant ( p < 10 − 13 ), with MOX ordering the recorded failures earlier than the informative prioritizing baselines, including the input-matched single-objective analogue, in roughly 70–80% of cycles. Within the protocol, an objective-removal ablation associates the ordering gains with the explainability objective, and the cheapest configurations also score highest on the proposed traceability measure.
Keywords:
continuous integration
explainable AI
multi-objective optimization
Pareto frontier
test selection
risk-aware optimization
DevOps
test case prioritization
regression testing
software quality assurance
Journal
IF:
2.6
Papers:
9.5K
Citations:
4.7W