[
    {
        "type": "text",
        "text": "图-解决模式识别（Pattern Recognize 问题）",
        "bbox": [
            144,
            93,
            495,
            109
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "Node occurance 边的关联性",
        "bbox": [
            146,
            112,
            376,
            126
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "节点的属性",
        "bbox": [
            146,
            130,
            238,
            143
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "边的信息",
        "bbox": [
            146,
            149,
            221,
            162
        ],
        "page_idx": 0
    },
    {
        "type": "equation",
        "text": "$$\nX _ {V} \\in \\mathbb {R} ^ {| V | \\times d}\n$$",
        "text_format": "latex",
        "bbox": [
            151,
            165,
            242,
            181
        ],
        "page_idx": 0
    },
    {
        "type": "equation",
        "text": "$$\nX _ {E} \\in \\mathbb {R} ^ {| E | \\times d _ {E}}\n$$",
        "text_format": "latex",
        "bbox": [
            149,
            187,
            245,
            204
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "动态图 在紧凑的表示中集成拓扑和时间信息能力，更多用于建模动态系统，如社交网络预测、推荐系统、交通预测、谣言传播等等，结合了顺序/时序数据处理和静态图学习的挑战。图被要求处理随时间变化的动态的拓扑结构（and/or）（边数）属性（节点属性边属性），为了对动态系统建模。",
        "bbox": [
            144,
            223,
            853,
            293
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "对顺序数据进行处理，捕获序列不同实体之间的依赖关系。",
        "bbox": [
            146,
            297,
            600,
            312
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "主要有 LSTM 的卷积架构所取代。",
        "bbox": [
            146,
            315,
            403,
            330
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "基于解码器/编码器框架的序列模型，如 Transformer、依赖于注意力机制",
        "bbox": [
            144,
            334,
            722,
            349
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "在静态图学习，克服图表示中缺乏节点排序所固有的置换不变性/等变性约束。",
        "bbox": [
            144,
            353,
            746,
            368
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "解决方法消息传递神经网络 MPNN",
        "bbox": [
            146,
            370,
            418,
            386
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "不同的 DG",
        "bbox": [
            146,
            390,
            233,
            404
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "离散的/连续的",
        "bbox": [
            146,
            407,
            263,
            423
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "边进化/节点进化/属性进化",
        "bbox": [
            146,
            426,
            357,
            441
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "同质与异质",
        "bbox": [
            146,
            445,
            238,
            460
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "学习环境：直推式还是归纳式",
        "bbox": [
            146,
            464,
            381,
            478
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "推断（前向预测）阶段",
        "bbox": [
            146,
            482,
            327,
            497
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "通过输入X预测Y",
        "bbox": [
            146,
            501,
            287,
            516
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "学习阶段（反向优化）",
        "bbox": [
            146,
            519,
            315,
            533
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "根据Y和Y‘，优化参数",
        "bbox": [
            146,
            538,
            326,
            552
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "如前所述，监督学习旨在学习输入空间 X和输出空间Y之间的映射函数。X可以表示的空间。本小节现在讨论空间Y。",
        "bbox": [
            144,
            556,
            847,
            590
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "由于动态图涉及来自图和时间/顺序数据的概念，因此必须考虑这两个方面来定义空间Y。",
        "bbox": [
            146,
            593,
            831,
            609
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "当学习序列数据的统计模型时，输入 X 可以由 T 个时间步长的 d 维特征表示，表示为X∈RTxd $\\boldsymbol { X } \\in \\mathbb { R } ^ { T \\times d }$ 。",
        "bbox": [
            146,
            612,
            847,
            648
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "输出Y的粒度可以是时间步长级别（每个时间步长一个输出，如在词性标记或更一般地序列标记中），或聚集的（一个输出用于许多输入，如在情感分类或更一般地序列分类中）。在静态图学习中，除了特征XV和XE之外，输入还具有拓扑信息。至于序列，输出 Y可以是局部的（每个节点或每条边一个标签）或全局的（每个子图或整个图一个标签）。",
        "bbox": [
            144,
            667,
            847,
            739
        ],
        "page_idx": 0
    },
    {
        "type": "image",
        "img_path": "images/296469f5d0a2010eb375ed7c66eed44792ebbceb0c9c733fa650c9c61c177c37.jpg",
        "image_caption": [
            ""
        ],
        "image_footnote": [],
        "bbox": [
            152,
            740,
            423,
            789
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "Inference ",
        "bbox": [
            216,
            799,
            315,
            815
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "Parameter Updates ",
        "bbox": [
            524,
            741,
            707,
            758
        ],
        "page_idx": 0
    },
    {
        "type": "equation",
        "text": "$$\n\\mathbf {\\Theta} ^ {\\prime} = \\operatorname {u p d a t e} (\\mathbf {\\Theta}, \\hat {\\mathbf {Y}}, \\mathbf {Y})\n$$",
        "text_format": "latex",
        "bbox": [
            463,
            759,
            766,
            791
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "许多最新的模型遵循编码器/解码器原理，其中可变长度的输入信号被编码成潜在表示，然后由解码器使用该潜在表示来计算下游任务的输出信号",
        "bbox": [
            144,
            834,
            847,
            868
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "这里的 Z 就是 embedding",
        "bbox": [
            146,
            872,
            354,
            888
        ],
        "page_idx": 0
    },
    {
        "type": "text",
        "text": "再把 embedding 通过解码器得到预测",
        "bbox": [
            146,
            890,
            448,
            906
        ],
        "page_idx": 0
    },
    {
        "type": "image",
        "img_path": "images/e2949c5b136177889bb7cc594ee5765d36bef4e6c8b7e90fe074c101bfb8a6ea.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            147,
            91,
            845,
            174
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "输入和输出序列的单元可能具有不同的顺序。",
        "bbox": [
            144,
            185,
            494,
            200
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "学习潜在表示（也称为嵌入）被称为表示学习。",
        "bbox": [
            144,
            204,
            510,
            219
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "当有多个节点/边的类型（项目和用户），用类型映射函数",
        "bbox": [
            144,
            223,
            600,
            239
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "GNN/GAT/K-GNN ",
        "bbox": [
            144,
            259,
            294,
            273
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "DG 动态图——两种形式",
        "bbox": [
            146,
            278,
            339,
            293
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "一种离散形式的",
        "bbox": [
            146,
            297,
            275,
            312
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "DTDG ",
        "bbox": [
            146,
            316,
            200,
            329
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "只要记住每一个时间段的快照（snapshot->ti）",
        "bbox": [
            146,
            334,
            507,
            350
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "DTDG $=$ （G1,G2…,GT）",
        "bbox": [
            146,
            353,
            322,
            368
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "一种连续形式的",
        "bbox": [
            146,
            370,
            275,
            386
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "CTDG 连续时间动态图",
        "bbox": [
            146,
            390,
            324,
            404
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "基于事件流的形式",
        "bbox": [
            146,
            407,
            294,
            423
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "每一次事件的变化记录下来",
        "bbox": [
            146,
            426,
            363,
            442
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "瞬时交互两个节点的（u，v）和时间 t",
        "bbox": [
            146,
            445,
            445,
            460
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "Contact -Sequence ={(ui,vi,ti)} ",
        "bbox": [
            146,
            464,
            394,
            480
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "Event-Based={(ui,vi,ti,deltai)} 起始(出现 )时间 ti 和持续时间 delta i，两个节点关联的边的存在的时刻",
        "bbox": [
            144,
            482,
            847,
            516
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "Interval-Graph={(ui,vi,Te);Te=((t1,t1’),(t2,t2’))}，用 Te 来表示边的所有活动",
        "bbox": [
            144,
            520,
            734,
            535
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "Graph stream 大规模图，用于边的添加和减少 delta i=1 add or -1deletion 标志位不考虑持续时间",
        "bbox": [
            144,
            538,
            848,
            571
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "Graph Stream={（ui，vi，ti，delta i）} ",
        "bbox": [
            144,
            575,
            468,
            590
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "什么事件？",
        "bbox": [
            146,
            593,
            231,
            608
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "Ti时刻下，ui和vi的边增加了/减少了",
        "bbox": [
            144,
            612,
            443,
            627
        ],
        "page_idx": 1
    },
    {
        "type": "image",
        "img_path": "images/c746830d302e3e44284fd48b73c0af69a11d0acf4ee6059f4189b80414d69c7f.jpg",
        "image_caption": [
            "Static Graphs ",
            "Weighted directed graph "
        ],
        "image_footnote": [],
        "bbox": [
            221,
            664,
            315,
            734
        ],
        "page_idx": 1
    },
    {
        "type": "image",
        "img_path": "images/6f1bb89499913ea46bf89c5766e785d288ad81eb6d1d4371993b350c787bd28d.jpg",
        "image_caption": [
            "Heterogeneous graph "
        ],
        "image_footnote": [],
        "bbox": [
            223,
            755,
            315,
            825
        ],
        "page_idx": 1
    },
    {
        "type": "image",
        "img_path": "images/3e4e316277ba51154368e0a57961bebe492938e98fcdc466f4510639ac5629e6.jpg",
        "image_caption": [
            "Dynamic Graphs ",
            "DTDG (or Snapshot vision) "
        ],
        "image_footnote": [],
        "bbox": [
            369,
            665,
            584,
            731
        ],
        "page_idx": 1
    },
    {
        "type": "image",
        "img_path": "images/8ef90f4621dfc41510d57405f3c6bb283fb266f06ec85835cad4eb31d5ae5683.jpg",
        "image_caption": [
            "CTDG (or Event vision) "
        ],
        "image_footnote": [],
        "bbox": [
            369,
            757,
            584,
            825
        ],
        "page_idx": 1
    },
    {
        "type": "image",
        "img_path": "images/0681e7b4f29f3c5eba8a879c82bd10665da1edd827bdf1c903ae625ea3f30b1c.jpg",
        "image_caption": [
            "(Edge-oriented) ESG "
        ],
        "image_footnote": [],
        "bbox": [
            601,
            678,
            885,
            722
        ],
        "page_idx": 1
    },
    {
        "type": "image",
        "img_path": "images/c758284426fe875cbb2097ef354f0d4a80e757498dc69b3b2a64c06ba310b2a0.jpg",
        "image_caption": [
            "(Node-oriented) ESG "
        ],
        "image_footnote": [],
        "bbox": [
            601,
            768,
            875,
            810
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "如何将动态图转换成等价的静态图ESG？",
        "bbox": [
            144,
            852,
            460,
            868
        ],
        "page_idx": 1
    },
    {
        "type": "list",
        "sub_type": "text",
        "list_items": [
            "1、基于边的形式 节点的备份 $^ +$ 边的联系（节点间 $^ +$ 时间戳，加入了节点之间的关联性）",
            "2、基于节点的形式"
        ],
        "bbox": [
            144,
            871,
            811,
            904
        ],
        "page_idx": 1
    },
    {
        "type": "text",
        "text": "GNNDelete如何删除边的关联性",
        "bbox": [
            176,
            93,
            436,
            108
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "定义动态度：",
        "text_level": 1,
        "bbox": [
            144,
            112,
            247,
            126
        ],
        "page_idx": 2
    },
    {
        "type": "list",
        "sub_type": "text",
        "list_items": [
            "1、边和顶点都固定，fix V,E 时空图（STGCN 时空模型，既要学时间上的， 又要学空间上）",
            "2、只固定了顶点,fix V ",
            "3、都变化，vary ",
            "4、边集不变，节点集变化，没有意义"
        ],
        "bbox": [
            146,
            130,
            847,
            219
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "输出是有粒度：",
        "text_level": 1,
        "bbox": [
            146,
            241,
            265,
            256
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "预测单一时刻的结果",
        "text_level": 1,
        "bbox": [
            146,
            260,
            310,
            275
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "还是多个时刻的",
        "text_level": 1,
        "bbox": [
            146,
            279,
            275,
            293
        ],
        "page_idx": 2
    },
    {
        "type": "list",
        "sub_type": "text",
        "list_items": [
            "：时间步长级别的（ 每个时间步长一个输出）",
            "：聚合的，一个输出多个输入（未来五个时间的拓扑结构）"
        ],
        "bbox": [
            147,
            297,
            598,
            331
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "拓扑结构也有局部（子图）和全局的",
        "bbox": [
            146,
            334,
            433,
            349
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "直推式 原来的拓扑结构不变 做出预测的时候没有在训练过程中没出现过的节点归纳式 做出推断的时候训练过程中没有看到",
        "bbox": [
            146,
            351,
            757,
            386
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "Transductive ",
        "text_level": 1,
        "bbox": [
            152,
            390,
            359,
            413
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "(1) Transductive-fixv,E Learning Inference ",
        "bbox": [
            152,
            419,
            386,
            458
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/bbcc93ac856be6d7c461092f4c7067219dd077d1ab5b3a264c057c10808b45aa.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            156,
            458,
            263,
            529
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/db4d98da46d1e6c6e158c05a5859fb732876cbad79b0288d64c35862f9805c4c.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            268,
            458,
            381,
            529
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "(2) Transductive-fixy Learning Inference ",
        "bbox": [
            416,
            419,
            638,
            458
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/5d77a9975b12967c827a904f0e653a6dd20b616eb17170ea86c227b4bff8c8af.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            418,
            458,
            524,
            529
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/fae214f407c72c98ec8787121c73e96333ea90b02082385935af350087cea3a4.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            534,
            458,
            645,
            529
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "(3) Transductive-vary Learning Inference ",
        "bbox": [
            667,
            419,
            890,
            458
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/e69e65651cde1ce432bc78ee711dfe7f2ab620c85a18ec5abadd2b7dc03c09ce.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            670,
            458,
            776,
            529
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/d3f2b68951d63e00609989797d595e11f69387b3ae4a9d49ddf6e70f76b236b8.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            784,
            458,
            895,
            529
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "Inductive ",
        "text_level": 1,
        "bbox": [
            152,
            545,
            304,
            568
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "（4） Inductivev",
        "bbox": [
            152,
            571,
            329,
            593
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/37efc40a1a19ed74b9949a8171bf9ac0dee51adfcf523d66b3d6c78e6feb4573.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            156,
            594,
            400,
            681
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "(5) InductiveDG ",
        "bbox": [
            416,
            571,
            606,
            594
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/cf63288df9720a4b9ad1e8a1ad001c61714dd67c023f799376cc80871a30dffe.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            418,
            594,
            556,
            680
        ],
        "page_idx": 2
    },
    {
        "type": "image",
        "img_path": "images/8df199df3a9c6c47dd44f0eda1352bfe46da210844d723b3f9bda3fbbf91d868.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            559,
            594,
            695,
            681
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "离散的时间直推式学习中（DTDG,Transductive），transfix VE ：交通流量，传染病病例，犯罪数量，作物产量",
        "bbox": [
            144,
            686,
            855,
            720
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "图/全局任务：每个快照的分类 睡眠阶段分类；整个DG的预测，如骨骼的动作STG；Tarnsfix V：电信网络中连通性，会议中个人联系",
        "bbox": [
            144,
            722,
            801,
            756
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "离散的归纳式学习（DTDG Inductive）中 社交网络未来快照的节点分类或链接预测图集输出 基于社交网络传播树的快照对真是和假新闻分类",
        "bbox": [
            144,
            760,
            794,
            793
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "连续时间（CTDG）",
        "bbox": [
            146,
            797,
            292,
            812
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "推荐系统、社交网络",
        "bbox": [
            146,
            816,
            310,
            831
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "CTCG再其最小时间单元下不能访问全局/整个图信息，全局的时间步标签是无意义的但是全局聚合任务可以通过聚合不同时间步长的节点来实现，如连续时间内谣言检测CTDG也可以定期拍摄快照转换为DTDG",
        "bbox": [
            144,
            834,
            803,
            885
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "动态时间下的任务",
        "bbox": [
            146,
            890,
            290,
            904
        ],
        "page_idx": 2
    },
    {
        "type": "text",
        "text": "为了评估给定任务的统计模型的性能，传统机器学学习 ",
        "bbox": [
            144,
            93,
            574,
            109
        ],
        "page_idx": 3
    },
    {
        "type": "text",
        "text": "分类任务",
        "bbox": [
            146,
            112,
            221,
            127
        ],
        "page_idx": 3
    },
    {
        "type": "text",
        "text": "回归任务",
        "bbox": [
            146,
            130,
            221,
            145
        ],
        "page_idx": 3
    },
    {
        "type": "text",
        "text": "节点排名任务",
        "bbox": [
            146,
            148,
            255,
            164
        ],
        "page_idx": 3
    },
    {
        "type": "text",
        "text": "动态任务 沿着时间轴计算静态度量",
        "bbox": [
            144,
            167,
            418,
            183
        ],
        "page_idx": 3
    },
    {
        "type": "text",
        "text": "动态图的嵌入",
        "bbox": [
            146,
            186,
            255,
            200
        ],
        "page_idx": 3
    },
    {
        "type": "text",
        "text": "从编码器-解码器的角度来看，深度学习统计模型首先将原始输入映射到表示为 Z的嵌入中，然后利用 Z 来预测输出[12，4]。图形可以在节点/边缘级别或（子）图形级别嵌入[13，5].节点级嵌入有利于广泛的节点相关任务，并允许保留更完整的输入信息以供稍后计算[5]以同样的方式，时间步长级嵌入比时间聚合嵌入保留更多的信息。类似地，当在顺序数据上学 习 时 [7] 。",
        "bbox": [
            144,
            203,
            855,
            296
        ],
        "page_idx": 3
    },
    {
        "type": "image",
        "img_path": "images/c3a808e669a8b7be887e2835a878309fb07a5f2690536f0e5cc4d52868f7a209.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            152,
            303,
            823,
            618
        ],
        "page_idx": 3
    },
    {
        "type": "equation",
        "text": "$$\nZ \\in \\mathbb {R} ^ {| V | \\times | T | \\times d}, z _ {v} ^ {t} \\in \\mathbb {R} ^ {d}\n$$",
        "text_format": "latex",
        "bbox": [
            147,
            627,
            319,
            646
        ],
        "page_idx": 3
    },
    {
        "type": "text",
        "text": "对不同的动态图， 如何设计 encoder 的 settings",
        "bbox": [
            144,
            667,
            532,
            686
        ],
        "page_idx": 3
    },
    {
        "type": "image",
        "img_path": "images/46cb1d54b5358d567f25b2593621409ab8028885478293acb70c4cf8f849ab88.jpg",
        "image_caption": [
            "(A) Discrete, transductive "
        ],
        "image_footnote": [],
        "bbox": [
            152,
            111,
            475,
            291
        ],
        "page_idx": 4
    },
    {
        "type": "image",
        "img_path": "images/d2df25f6574b3cf24a96ef20688baf948c6322632b4acd9e8e97597ec9a56188.jpg",
        "image_caption": [
            "(B) Discrete, inductive "
        ],
        "image_footnote": [],
        "bbox": [
            526,
            111,
            828,
            288
        ],
        "page_idx": 4
    },
    {
        "type": "image",
        "img_path": "images/92b7d0fd700043ee860b01b2e882046208c24d7d786f335fd8b74b18954e36f3.jpg",
        "image_caption": [
            "(C) Continuous, transductive "
        ],
        "image_footnote": [],
        "bbox": [
            152,
            335,
            473,
            489
        ],
        "page_idx": 4
    },
    {
        "type": "image",
        "img_path": "images/731d08cee749b45a4111054e790163f2db076ce12de24586a28e861c309eb247.jpg",
        "image_caption": [
            "(D) Continuous, inductive "
        ],
        "image_footnote": [],
        "bbox": [
            527,
            335,
            828,
            483
        ],
        "page_idx": 4
    },
    {
        "type": "text",
        "text": "为了将动态图信息化地编码为张量或向量列表，DGNN必须捕获结构信息及其随时间的演化 。 因 此 ， 为 了 分 别 处 理 拓 扑 和 时 间 ，DG 通常被 分 解 或转换 为 等效静 态 （ 子 ） 图[74，89，75]，随机游走[113，124，123，121，122]，或者是矩阵序列[95，86，93]。",
        "bbox": [
            144,
            518,
            850,
            574
        ],
        "page_idx": 4
    },
    {
        "type": "text",
        "text": "在文献中，通过将不同的静态图编码器 fG（·）与时态数据的fT（·）相结合，出现了许多方法。大量的图和时态数据编码器是本节中回顾的 DG编码器的基础。这些编码器在附录B和C中进行了描述。",
        "bbox": [
            144,
            611,
            848,
            665
        ],
        "page_idx": 4
    },
    {
        "type": "text",
        "text": "在下面的小节中，我们提出了一个DGNN模型的分类，它依赖于五个类别。我们的分类，如图7所示，是基于处理时间和结构信息的策略。",
        "bbox": [
            144,
            703,
            838,
            739
        ],
        "page_idx": 4
    },
    {
        "type": "list",
        "sub_type": "text",
        "list_items": [
            "1.通过拓扑对时间边缘进行建模并对 ESG 进行编码，表示为 TE（第 3.1 节）。",
            "2.对隐藏状态进行顺序编码，记为 enc（H）（第 3.2 节）。",
            "3.对 DGNN 参数进行顺序编码，表示为 enc（Θ）（第 3.3 节）。",
            "4.通过属性嵌入发生时间t作为ESG的边缘特征，表示为emb（t）（第3.4节）。",
            "5.抽样因果游走，表示为 CauseRW（第 3.5 节）。"
        ],
        "bbox": [
            144,
            741,
            781,
            832
        ],
        "page_idx": 4
    },
    {
        "type": "text",
        "text": "1. TE ",
        "bbox": [
            226,
            92,
            295,
            112
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/497a6bfcb699c59d3b4a9b42e24e4270536c5f1681807f903e0e85dbc4a67ff8.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            169,
            124,
            357,
            199
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/36a04b2c156c0dbfeeb1f8a5e3a2caa42895d4c195620b069cf075a54dca2717.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            171,
            210,
            344,
            230
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/84576c405221d1001ad1ab5a0f3de8a24f61ef9923085f789c09f31cc6a792f9.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            171,
            239,
            324,
            259
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "2. Enc(H) ",
        "bbox": [
            416,
            92,
            541,
            116
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/1bb2a53245bd5c1de66e86cdf425f6fca9fbf5d9068fee5c6037b2afd2fff0d9.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            388,
            122,
            579,
            263
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "3. Enc(0) ",
        "bbox": [
            636,
            92,
            759,
            116
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/11776d45b7f1d6ffcfced0128e76a2d7ea793afc930bd40b6c3b614f0c6b1320.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            603,
            122,
            794,
            263
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "4. Emb(t) ",
        "bbox": [
            858,
            92,
            978,
            116
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/28df9d27ec3d3faac8c88b3acad757af84feff54fee07c071e078c0dff6e42e6.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            813,
            120,
            1000,
            267
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "请注意，这五种方法并不是唯一的，也就是说，它们可以组合在一起并用于同一 DG。",
        "bbox": [
            144,
            278,
            805,
            294
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "1、对时序的边进行建模——转换为静态图",
        "bbox": [
            146,
            297,
            482,
            312
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "FixV,E 考虑了时间信息（embedding ET）",
        "bbox": [
            176,
            315,
            500,
            332
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/22d8dcba7544ae7f4b9a780b95130b617616df0c54a80f0a285449e5d6235839.jpg",
        "image_caption": [
            "Snapshots "
        ],
        "image_footnote": [],
        "bbox": [
            186,
            357,
            378,
            472
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "ASTTN ",
        "bbox": [
            176,
            482,
            236,
            495
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/8169e54de8b83dfaaa4b484f35da48c4abb52c1e0df89fb0f62c7e9903e8e394.jpg",
        "image_caption": [
            "STG with Er "
        ],
        "image_footnote": [],
        "bbox": [
            415,
            357,
            603,
            445
        ],
        "page_idx": 5
    },
    {
        "type": "image",
        "img_path": "images/6984145b6f133ae0040e209cde6ece2a90e61b23f8c1a1eedced479836a1b650.jpg",
        "image_caption": [
            "ST-GCN Module "
        ],
        "image_footnote": [],
        "bbox": [
            638,
            355,
            900,
            476
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "2、对隐藏状态进行顺序编码",
        "bbox": [
            144,
            519,
            374,
            535
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "对拓扑结构和时间戳信息进行编码",
        "bbox": [
            176,
            538,
            445,
            552
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "RST ",
        "bbox": [
            176,
            557,
            211,
            570
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "DyReG ",
        "bbox": [
            178,
            575,
            236,
            590
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "Dyn-GCN ",
        "bbox": [
            178,
            594,
            257,
            609
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "把通过 f(G)编码的信息经过 LSTM 或者 GRU（Sequential endocding）进行时间序列的学习",
        "bbox": [
            174,
            611,
            847,
            645
        ],
        "page_idx": 5
    },
    {
        "type": "equation",
        "text": "$$\nZ ^ {t} = f _ {T} \\left(\\left(f _ {G} \\left(X ^ {t}\\right)\\right)\\right), o r f _ {T} \\left(X ^ {T}\\right) \\oplus f _ {G} \\left(X ^ {t}\\right)\n$$",
        "text_format": "latex",
        "bbox": [
            176,
            648,
            492,
            665
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "解码的时候也要用sequential去阶码",
        "bbox": [
            176,
            686,
            468,
            702
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "3、对于新增的节点很难预测，及没有将 fG 和 fT 结合起来",
        "bbox": [
            144,
            705,
            600,
            720
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "对应的是 transformer 直接把时间戳的信息编码进去了",
        "bbox": [
            144,
            722,
            574,
            739
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "通过（fg参数）时间步来学习时间信息",
        "bbox": [
            144,
            741,
            453,
            758
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "将时间信息的编码加入到 fg 里",
        "bbox": [
            144,
            760,
            386,
            776
        ],
        "page_idx": 5
    },
    {
        "type": "equation",
        "text": "$$\n\\Theta_ {f _ {o}} ^ {t} = L S T M \\left(\\Theta^ {t - 1} _ {f _ {o}}\\right)\n$$",
        "text_format": "latex",
        "bbox": [
            154,
            777,
            322,
            795
        ],
        "page_idx": 5
    },
    {
        "type": "equation",
        "text": "$$\n\\Theta^ {t} _ {f _ {G}} = G R U \\left(H ^ {t}, \\Theta^ {t - 1} _ {f _ {G}}\\right)\n$$",
        "text_format": "latex",
        "bbox": [
            147,
            799,
            331,
            818
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "再经过编码，就没有时间上的序列性了",
        "bbox": [
            144,
            834,
            448,
            850
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "4、",
        "bbox": [
            146,
            854,
            168,
            866
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "还把时间作为嵌入信息，如何编码时间信息",
        "bbox": [
            144,
            871,
            485,
            887
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "Temproal point process ",
        "bbox": [
            146,
            890,
            346,
            906
        ],
        "page_idx": 5
    },
    {
        "type": "text",
        "text": "在每个节点上加一个 t 的 embedding",
        "bbox": [
            146,
            93,
            438,
            110
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "3、随机游走的方法",
        "text_level": 1,
        "bbox": [
            146,
            112,
            305,
            127
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "对于异质图的方法",
        "bbox": [
            146,
            130,
            292,
            145
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "怎么设计一个 GNN",
        "bbox": [
            146,
            149,
            300,
            162
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "设计 Workflow",
        "bbox": [
            146,
            168,
            268,
            181
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "如何优化框架",
        "bbox": [
            146,
            186,
            257,
            200
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "时序图神经网络",
        "text_level": 1,
        "bbox": [
            146,
            205,
            273,
            219
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "1、创建数据，根据快照的方式构建图",
        "text_level": 1,
        "bbox": [
            147,
            223,
            445,
            237
        ],
        "page_idx": 6
    },
    {
        "type": "image",
        "img_path": "images/215cfded78261b39f639e96eaae2b4b9cfafc37a05f9efefaa0359da4fcafbee.jpg",
        "image_caption": [],
        "image_footnote": [],
        "bbox": [
            178,
            237,
            947,
            545
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "表示节点的特征，如果没有特征用 one -hot 表示",
        "bbox": [
            176,
            556,
            559,
            571
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "计算 attention 系数",
        "bbox": [
            178,
            575,
            334,
            590
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "时间注意力",
        "bbox": [
            178,
            594,
            268,
            609
        ],
        "page_idx": 6
    },
    {
        "type": "text",
        "text": "如何计算 loss？",
        "bbox": [
            178,
            612,
            295,
            627
        ],
        "page_idx": 6
    }
]