Intelligent Test Evaluation: Risk Dimension Mining, Admission Assessment, Multi‑Dimensional Activity Data Mining, and Model‑Based Risk Evaluation

Previous articles introduced intelligent research and practice for test input, execution, and analysis; this chapter focuses on the intelligent practice of the test evaluation stage.

Test evaluation collects system performance data generated by quality‑assurance activities, analyzes it to assess overall quality risk after all activities, and differs from conventional test reports; it decides whether additional quality activities are needed, a step often overlooked and prone to missed risks when relying solely on automated reports.

Intelligent test evaluation combines data, algorithms, and engineering to predict project risk and determine release readiness; research spans risk introduction, activity data mining, and risk assessment, with Baidu QA actively exploring these areas.

1. Mining Quality Risk Dimensions Based on Development Behavior – Risk is introduced during development; the goal is to identify who, under what context, writes risky code, structure these factors, and build models to quantify risk probability and impact, providing a basis for risk control and decision‑making.

Risk dimensions include project risk (duration, module count, change count), personnel risk (bugs per KLOC, project familiarity, test rejection count), and code risk (change lines, complexity, impact on interfaces/UI, user density).

Using these data, models and rules estimate risk occurrence probability and severity, helping decide the level of QA involvement such as self‑test, unattended testing, or full QA review.

Project and personnel risk rely on data retained throughout the delivery process; code risk is derived from historical commits, AST analysis, and deep‑learning techniques.

Dependency risk can be captured via RPC + mesh architectures to assess service‑level impact of code changes.

2. Admission Risk Assessment and Application – Consists of risk‑dimension data, risk decision, and task type to determine the appropriate quality activities (full automation, RD self‑test, QA involvement). A unified data‑mining pipeline builds risk portraits for projects, personnel, and code, quantifies risk, and generates rule‑ and model‑based decisions, forming a closed‑loop feedback system.

3. Multi‑Dimensional Activity Process Data Mining – After admission assessment, quality activities leave traces that are mined for white‑box coverage, log coverage, business‑request coverage, and simulation fidelity. White‑box coverage records execution of statements, branches, functions via instrumentation; log coverage tracks exception‑log occurrence without instrumentation; business‑request coverage maps test cases to a knowledge graph of request paths; simulation fidelity measures environment and traffic realism.

4. Model‑Based Risk Assessment – Traditional risk assessment relies on tester experience, which has blind spots; a model‑based system uses risk‑dimension and activity data to predict final project risk, deciding release or additional testing. Logistic regression is chosen for interpretability and speed; features combine risk‑introduction (code, personnel) and risk‑removal (test completeness, monitoring). Decisions use a risk‑matrix considering probability and impact, with visual reports closing the feedback loop.

Results show the risk assessment identified over 1,000 non‑self‑test projects, intercepted 300+ bugs, and managed 4,000+ self‑test projects, saving more than 2,000 person‑days; Baidu QA will continue research across these four dimensions to further enhance risk‑driven quality assurance.