Self-Healing Cloud Infrastructure Systematic Review — Data Package

This deposit is the data layer for a PRISMA 2020 systematic review of machine learning approaches to self-healing cloud infrastructure. The corpus is 344 peer-reviewed and arXiv-indexed studies (2022–2026), identified from 1,537 unique post-deduplication records across seven academic databases (Web of Science was targeted in the protocol but inaccessible without institutional credentials). The deposit contains the canonical extraction dataset, the full screening trail, the methodology (locked rubrics, protocol, dated deviations log, PRISMA 2020 checklist), the synthesis tables and gap verdicts, the figures and their source code, the per-paper extraction Markdown files, and the reproducibility scripts. The accompanying journal manuscript is not part of this deposit; the publisher hosts the version of record. The canonical CSV (canonical_studies.csv) records per-study CASP item scores (Q1–Q6, 0–2 each), PROBAST domain ratings (Data Selection, Feature Engineering, Model Evaluation, Analysis Reporting; Low/High/Unclear with worst-domain rule for overall), inclusion rationale, primary phase classification (DETECT, DIAGNOSE, DECIDE, RECOVER, Enabling Tech, Cross-Phase Integration), ML technique with controlled-vocabulary and free-text detail, deployment context, evaluation methodology, dataset descriptor, and headline performance metrics. The dual-path sensitivity framework (full 344 corpus + 128-paper Low-ROB subset) can be reproduced by applying the named predicates in rq_to_answer_mapping.md to canonical_studies.csv. Two post-extraction audits are recorded in protocol_deviations.md: a duplicate-record audit collapsed 19 within-corpus duplicates missed at Phase 1 deduplication (audit_exclusion: dup_phase1), and a scope audit dropped 6 pre-2022 records retained in error after Refinement A (audit_exclusion: scope_refinement_a). The arithmetic 369 − 19 − 6 = 344 is enforced as an assertion in figures/prisma_flow.py and as a stale-value drift detector in audit/cross_doc_audit.py (61 claims, all reconcile to canonical_studies.csv). Methodology highlights:- Four-pass single-author screening (title, abstract, re-screen, full-text adjudication) with intra-rater reliability quantified via Cohen's κ and prevalence-adjusted bias-adjusted kappa (PABAK 0.905 three-category, exceeding the pre-specified 0.80 lock threshold).- CASP item-level scoring deposited per paper (casp_rubric.md).- PROBAST-adapted four-domain risk-of-bias scoring with worst-domain rule deposited per paper (probast_rubric.md).- Hybrid pipelines decomposed into nine architectural sub-classes (H1 LLM-orchestrated, H2 GNN+temporal, H3 Federated, H4 sequence+RL, H5 autoencoder+sequence, H6 metaheuristic+ML, H7 multi-modal fusion, H8 statistical+ML, H9 other multi-component) by a deterministic strict-priority classifier (hybrid_decomposition.md).- Manifest-driven cross-document numerical audit (audit/cross_doc_audit.py) plus stale-value drift detector against prior-version numbers. See README.md for a full file-by-file inventory.本数据集为面向自愈云基础设施机器学习方法的系统评价与Meta分析首选报告条目（PRISMA 2020）系统评价的数据层。本次语料库包含2022至2026年间的344篇经同行评议且被arXiv收录的研究，这些研究从7个学术数据库的1537条去重后唯一记录中筛选得到（检索方案中预设覆盖Web of Science数据库，但因无机构权限无法访问该库资源）。 本数据集包含规范提取数据集、完整筛选追踪记录、研究方法学文档（含锁定的评价准则、检索方案、带时间戳的偏差日志及PRISMA 2020核查清单）、综合分析表格与研究缺口判定结果、配图及其源代码、单篇文献提取的Markdown文件，以及可复现的实验脚本。随附的期刊手稿不属于本数据集范畴，正式出版版本由对应期刊平台托管。 规范CSV文件（"canonical_studies.csv"）记录了每篇研究的临床鉴定技能项目（CASP）条目得分（Q1至Q6，每项取值0至2分）、预测模型偏倚风险评估工具（PROBAST）领域评级（涵盖数据选择、特征工程、模型评估、分析报告4个领域，整体评级采用最差域规则，分为低、高、不明确三类）、文献纳入理由、主要研究阶段分类（检测（DETECT）、诊断（DIAGNOSE）、决策（DECIDE）、恢复（RECOVER）、使能技术、跨阶段集成）、带受控词汇表与自由文本细节的机器学习技术、部署场景、评估方法学、数据集描述，以及核心性能指标。双路径敏感性分析框架（完整344篇语料库+128篇低偏倚风险（ROB）子集）可通过将"rq_to_answer_mapping.md"中指定的谓词应用于"canonical_studies.csv"实现复现。 两项提取后审核记录于"protocol_deviations.md"：其一为重复记录审核，剔除了阶段1去重时遗漏的19篇语料库内部重复文献（audit_exclusion: dup_phase1）；其二为范围审核，剔除了优化阶段A后误保留的6篇2022年前的记录（audit_exclusion: scope_refinement_a）。公式369−19−6=344已作为断言写入"figures/prisma_flow.py"，并作为失效值漂移检测器嵌入"audit/cross_doc_audit.py"（共包含61条数值校验声明，所有声明均与"canonical_studies.csv"文件内容一致）。 方法学亮点如下： 1. 四轮单作者筛选流程（标题筛选、摘要筛选、复筛、全文判定），组内信度通过科恩κ系数（Cohen's κ）与患病率校正偏倚校正Kappa（PABAK 0.905，三分制）量化，结果超过预设的0.80锁定阈值。 2. 单篇文献级别的CASP条目评分已存入"casp_rubric.md"。 3. 单篇文献级别的经PROBAST适配的四领域偏倚风险评分已存入"probast_rubric.md"，评分采用最差域规则。 4. 混合流水线通过确定性严格优先级分类器（"hybrid_decomposition.md"）划分为9个架构子类：H1 大语言模型（LLM）编排、H2 图神经网络（GNN）+时序模型、H3 联邦学习、H4 序列模型+强化学习（RL）、H5 自编码器+序列模型、H6 元启发式算法+机器学习、H7 多模态融合、H8 统计方法+机器学习、H9 其他多组件架构。 5. 清单驱动的跨文档数值审核（"audit/cross_doc_audit.py"）及针对旧版本号的失效值漂移检测器。 完整的逐文件清单请参阅"README.md"。