{
  "schema": "auturnal.deep-research-review.v1",
  "slug": "graph-neural-networks",
  "title": "图神经网络的表达与传播：聚合可辨性、平滑和评估泄漏",
  "category": "图学习",
  "dek": "用确定性路径图观察传播，再区分过平滑、信息挤压和数据划分问题。",
  "english": "Message-passing graph networks are analyzed through normalized aggregation, permutation behavior and multiset distinguishability. A six-node path calculation illustrates smoothing for one declared propagation operator. It is not a universal oversmoothing theorem or model benchmark. Graph-level and temporal splits, feature provenance and inductive access rules are required for credible evaluation.",
  "kind": "ai_assisted_research_guide",
  "author": "Codex（AI）",
  "original_language": "zh-Hans",
  "published_date": "2026-10-22",
  "edition": 5,
  "version": 6,
  "previous_version": "/guides/graph-neural-networks/v5/",
  "refs": [
    "arxiv-1609.02907",
    "arxiv-1810.00826",
    "arxiv-1706.02216"
  ],
  "sections": [
    [
      "摘要",
      "图神经网络通过邻域聚合传播信息，但聚合器的可辨性、重复传播的混合与图划分方式共同决定结果。本文比较GCN、GIN与邻域采样思想，区分过平滑和过挤压，并用六节点路径图计算重复平均后的方差变化。配图没有训练参数，不代表真实图任务准确率。",
      [
        "arxiv-1609.02907",
        "arxiv-1810.00826",
        "arxiv-1706.02216"
      ]
    ],
    [
      "1. 邻域聚合应尊重节点重编号",
      "节点编号通常不应改变图的含义，因此节点表示应随重编号相应变换，图级输出则应保持不变。归一化邻接传播与共享参数可以满足这种结构要求，但特征中的ID或划分信息仍可能引入泄漏。GCN中的自环和度归一化会改变传播尺度，应保存是否有向、是否加权和自环规则，不能只用“用了GCN”描述实现。",
      [
        "arxiv-1609.02907"
      ]
    ],
    [
      "2. 聚合器能否区分多重集合很关键",
      "平均、求和和最大值对不同邻域可能产生相同结果；例如相同均值的不同多重集合会被均值聚合混同。GIN研究了消息传递网络与图可辨性之间的联系，其结论依赖函数类和假设，不能理解为任意训练后的GIN都能区分所有图。结构不可辨与优化未学会也应分开：增加训练轮数不能修复某个固定聚合器固有的信息丢失。",
      [
        "arxiv-1810.00826"
      ]
    ],
    [
      "3. 重复混合可能导致表示趋同",
      "在固定线性传播下，邻域平均会逐步混合节点信息；特定连通与谱条件下可分析其长期行为。实际网络包含非线性、残差、归一化和训练权重，不能把简单平均的方差曲线当成所有GNN的定理。本篇六节点路径加自环后采用行归一化，直接迭代一个固定输入向量，展示这一算子下的趋同现象；没有测分类质量。",
      [
        "arxiv-1609.02907"
      ]
    ],
    [
      "4. 过平滑与过挤压不是同一个瓶颈",
      "表示趋同强调不同节点难以区分；过挤压强调大量远处信息需要经过有限宽度或少量边压缩。某个图任务可能只需局部信息，也可能需要跨图长距离关系。增加层数会扩大感受野，同时改变优化和信息混合。应构造能分离局部与远程需求的测试，并将结构、宽度、残差与采样分别消融，不能看到深层效果差就归为单一原因。",
      [
        "arxiv-1810.00826",
        "arxiv-1706.02216"
      ]
    ],
    [
      "5. 归纳与传导评估的可见信息不同",
      "传导设置可能允许训练时看到未标注测试节点的图结构，归纳设置则需要对新节点或新图泛化。GraphSAGE研究采样和聚合的归纳表示方式，但实现必须明确训练时能访问哪些边和特征。若时间后的边或目标标签派生特征进入训练，分数可能被泄漏抬高。随机节点划分不能自动代表新图、未来关系或新机构上的泛化。",
      [
        "arxiv-1706.02216"
      ]
    ],
    [
      "6. 图实验需要结构与系统双重复核",
      "保存图版本、去重、自环、边方向、特征来源及划分，按图家族或时间隔离相关样本。报告多个种子、类别分层及对简单特征基线的增益。大图采样还需测覆盖、通信和内存，并区分采样近似与全邻域结果。本文只计算传播算子，不声称训练了新的图模型或取得基准成绩。",
      [
        "arxiv-1609.02907",
        "arxiv-1810.00826",
        "arxiv-1706.02216"
      ]
    ]
  ],
  "equations": [
    {
      "tex": "H^{\\prime}=\\sigma(\\tilde D^{-1/2}\\tilde A\\tilde D^{-1/2}HW),\\quad\\tilde A=A+I",
      "explanation": "GCN式对称归一化示意；方向、权重、自环和度定义应与实现一致。"
    },
    {
      "tex": "h_v^{\\prime}=\\operatorname{MLP}\\left((1+\\epsilon)h_v+\\sum_{u\\in N(v)}h_u\\right)",
      "explanation": "GIN式求和聚合；可辨性分析有函数类条件，不是实际训练的万能保证。"
    },
    {
      "tex": "x^{(k)}=P^kx^{(0)},\\quad P=D^{-1}(A+I)",
      "explanation": "本篇图使用六节点路径和这个行随机算子，区别于上面的对称归一化，避免混淆。"
    }
  ],
  "table": {
    "headers": [
      "方法或判断",
      "控制变量与测量",
      "失效条件及解释"
    ],
    "rows": [
      [
        "归一化传播",
        "保存方向、自环与度规则",
        "同名架构可能采用不同算子"
      ],
      [
        "均值与求和",
        "构造多重集合可辨性对照",
        "结构不可辨不靠更多训练解决"
      ],
      [
        "深层传播",
        "分测趋同和远程信息需求",
        "过平滑与过挤压不可混同"
      ],
      [
        "图划分",
        "明确传导/归纳和时间可见性",
        "随机节点分割可能泄漏"
      ]
    ]
  },
  "protocol": [
    "冻结图、特征来源、边方向、自环与去重版本。",
    "明确传导或归纳访问边界，按时间或图家族划分。",
    "用重编号和小图手算核对聚合输出。",
    "比较简单特征基线、不同聚合器、深度和残差，分离结构因素。",
    "使用局部/远程任务与多种子，报告分层质量及失败图。",
    "大图下联合测采样覆盖、峰值内存和时间，保留全邻域参考。"
  ],
  "diagram": [
    "图与特征来源",
    "可见边及划分边界",
    "邻域采样与归一化",
    "共享聚合 / 消息传播",
    "可辨性与信息瓶颈",
    "独立图或时间泛化"
  ],
  "analytic": "graph-smoothing",
  "limitations": [
    "路径图传播没有训练权重。",
    "方差下降不是所有GNN的定理。",
    "表达上界不保证优化成功。",
    "传导成绩不能自动外推归纳任务。"
  ],
  "keywords": [
    "graph-neural-networks",
    "message-passing",
    "oversmoothing"
  ],
  "peer_reviewed": false,
  "journal_publication": false,
  "disclosure": {
    "authorship": "AI辅助研究综述；Codex（AI）依据所列一手文献撰写并进行解析推导。",
    "figures": "概念封面与科学图解分开；科学图由公开公式和明示假设计算，原始论文结果均注明归属。",
    "review": "本站没有本篇独立同行评审或原创模型实验记录。"
  },
  "section_kinds": [
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review",
    "source_based_review"
  ],
  "revision_note": "阅读器性能修订：新近 HEAD 返回 200 且 ETag 相同时也复用已读页面。正文、公式、来源和科学图的数据保持上一版本内容。",
  "sources": [
    {
      "id": "arxiv-1609.02907",
      "title": "Semi-Supervised Classification with Graph Convolutional Networks",
      "authors": ",",
      "url": "https://arxiv.org/abs/1609.02907",
      "date": "2016-09-09",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; 原文方法、理论和实验章节见作者原始论文链接",
      "note": "一手论文；原文研究归属原作者。本文的推导示例与建议实验方案另行标注，不冒充原论文或本站实测。"
    },
    {
      "id": "arxiv-1810.00826",
      "title": "How Powerful are Graph Neural Networks?",
      "authors": ",  ,  ,",
      "url": "https://arxiv.org/abs/1810.00826",
      "date": "2018-10-01",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; 原文方法、理论和实验章节见作者原始论文链接",
      "note": "一手论文；原文研究归属原作者。本文的推导示例与建议实验方案另行标注，不冒充原论文或本站实测。"
    },
    {
      "id": "arxiv-1706.02216",
      "title": "Inductive Representation Learning on Large Graphs",
      "authors": ",  ,",
      "url": "https://arxiv.org/abs/1706.02216",
      "date": "2017-06-07",
      "accessed_at": "2026-10-21",
      "locator": "Abstract; 原文方法、理论和实验章节见作者原始论文链接",
      "note": "一手论文；原文研究归属原作者。本文的推导示例与建议实验方案另行标注，不冒充原论文或本站实测。"
    }
  ],
  "figures": {
    "mechanism": {
      "url": "/guides/assets/v5-mechanism-graph-neural-networks-89ba3c09c81b6939.png",
      "kind": "conceptual_method_flow",
      "data": [
        "图与特征来源",
        "可见边及划分边界",
        "邻域采样与归一化",
        "共享聚合 / 消息传播",
        "可辨性与信息瓶颈",
        "独立图或时间泛化"
      ]
    },
    "analysis": {
      "url": "/guides/assets/v5-analysis-graph-neural-networks-0f707ebfbc27987e.png",
      "kind": "analytical_or_constructed_example",
      "data_url": "/guides/assets/v5-data-graph-neural-networks-bc33b81ace7b4d1a.json",
      "assumptions": "Six-node undirected path, self-loops, row normalization, initial signal[1,0,0,0,0,-1]. Not a trained GCN or universal theorem."
    }
  },
  "cover_provenance_url": "/guides/assets/v5-provenance-graph-neural-networks-069f21ee54424ab4.json",
  "source_revision": "6251a250b27bb8b449f59b392f81abd355383770",
  "pdf": {
    "document_id": "dc6e702d-1799-52f9-8374-c78491e67f15",
    "version_id": "d2a938dc-9f83-5375-9e7c-d2381b55c50b",
    "media_type": "application/pdf",
    "authentication_required": true,
    "download_audited": true
  },
  "presentation_change": "Complete version-bound translated reading views and selectable authenticated original/translated PDFs; original scientific text and evidence retained.",
  "source_locale": "zh",
  "source_version_id": "728ae0c9-3067-57e7-9e2a-9a3ef0007bad",
  "translations": [
    {
      "locale": "en",
      "translation_id": "bd4b831c-a777-5ef8-a62f-ebb2250b8665",
      "translated_content_sha256": "3585035ef3e4b24ad455a25f1336ba0a501b4d5742efc5f187d5d74f980dca30",
      "sha256": "c949b0aac57e16817d8cb14f9fb694ace688b12f025fe9038812b001f0deab04",
      "byte_length": 670574
    },
    {
      "locale": "es",
      "translation_id": "e3d33493-b75a-5db9-9bc8-24f7026aac45",
      "translated_content_sha256": "a840be16546609ea5a3a00f78a6154579ae723662e2aaad171a299a8253c7f6d",
      "sha256": "25a070183da95069fff39201b9b959f7e78063785b7b57b94da8820dc0fcdccc",
      "byte_length": 677126
    },
    {
      "locale": "ja",
      "translation_id": "f7888c3a-86a5-544c-ac1c-e6587b4e1639",
      "translated_content_sha256": "e8680642cbca198ef6e92700a8beeb6b2bf12a05ad8a735f2b1b8a727747879c",
      "sha256": "aa3d296747fb6912a52f0d7b8e2d5493fca53484a28968748882f9274d741e90",
      "byte_length": 777061
    },
    {
      "locale": "ar",
      "translation_id": "40a4489b-96b2-55fa-9077-b71e7c32545b",
      "translated_content_sha256": "90231126024754be8b75792cb3b16788e04da156a8b0af84be08a670a30500af",
      "sha256": "cac3bee594d80abacfd58cfd8d8a5e76f5de21e1f49667f4884c88d1a4a8af56",
      "byte_length": 654656
    }
  ]
}