TensorFlow 这下更牛了，现已支持图神经网络

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"18日，TensorFlow 项目组发布了"},{"type":"link","attrs":{"href":"https:\/\/github.com\/tensorflow\/gnn","title":null,"type":null},"content":[{"type":"text","text":"图神经网络（GNN）库"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"，使开发者可以通过 TensorFlow 轻松处理图结构化数据。从2019年开始，图神经网络就被视作 AI 领域的核心研究方向，但从 2019 - 2021，该领域的主要动态集中在论文层面，在 TensorFlow GNN 发布前，工业环境还未出现大规模支持。此外，一直有许多开发者研究基于 TensorFlow 的 GNN 实现，这次"},{"type":"link","attrs":{"href":"https:\/\/blog.tensorflow.org\/2021\/11\/introducing-tensorflow-gnn.html","title":null,"type":null},"content":[{"type":"text","text":"发布"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"也相当于正式给出官方答案。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"TensorFlow GNN 基本信息"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"TensorFlow GNN主要提供了几类主要服务："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"numberedlist","attrs":{"start":null,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"建模 API；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"基于张量的图形数据结构；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"数据处理管道；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":4,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"实例模型。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"其工作流程及组件示意如下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/ef\/ef836db92518cd27bea2a23414aa78d6.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/github.com\/tensorflow\/gnn","title":null,"type":null},"content":[{"type":"text","text":"TF-GNN"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"项目组在官方博客上详细列举了初始版本的程序和功能，包括："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"用于创建可轻松与其他类型模型组合的 GNN 模型的高级 Keras 样式 API。GNN 通常与排序、深度检索（双编码器）结合使用或与其他类型的模型（图像、文本等）混合使用；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":1,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"用于异构图的 GNN API。我们在 Google 和现实世界中处理的许多图问题都包含不同类型的节点和边。因此，我们选择提供一种简单的方法来对此进行建模；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"一个定义良好的模式来声明一个图的拓扑结构，以及验证它的工具。该模式描述了其训练数据的形状，并用于指导其他工具；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"一种GraphTensor复合张量类型，它保存图形数据，可以批处理，并具有可用的图形操作例程；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"GraphTensor结构上的操作库："}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":1,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"节点和边上的各种高效广播和池化操作，以及相关工具。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":1,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"标准烘焙卷积库，机器学习工程师\/研究人员可以轻松扩展。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":1,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"产品工程师可以快速构建 GNN 模型而不必担心其细节的高级 API。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"磁盘上图形训练数据的编码，以及用于将此数据解析为数据结构的库，您的模型可以从中提取各种特征。"}]}]}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"图数据逐渐成为主流"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"图包含四大类基本信息：节点、边、全局信息（节点数等）、连通性（两个节点间是否是连通的）。要把图的信息以矩阵的形式输入机器学习模型是非常复杂的，具体可以通过下图体验下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/8a\/8ab5308ad63459803e920385c71fed7d.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"而图神经网络，就是对图的所有属性（节点、边、全局上下文）进行的可优化转换，同时保留对称性。一般认为 GNN 可以分为图卷积网络（Graph Convolution Networks，GCN）、 图注意力网络（Graph Attention Networks）、图自编码器（ Graph Autoencoders）、图生成网络（ Graph Generative Networks） 和图时空网络（Graph Spatial-temporal Networks）等几大类。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"不仅是 GNN，“Graph（图）”这个词，本身在 IT 圈的出现频率也越来越高。比如图数据库，使用图形化的模型进行查询的数据库，通过节点、边和属性等方式来表示和存储数据，支持增删改查（CRUD）等操作。图数据库的相关产品有"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/qfgGZtr6VxclyxKtLqae","title":null,"type":null},"content":[{"type":"text","text":"Nebula Graph"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]},{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"、Neo4j、Janus、HugeGraph 等，不但社区内关注度很高，资本的投入也越来越大。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"以开源的 Nebula Graph 为例，在腾讯、微信、美团都有大规模的应用。在腾讯，图数据库被用于安全、知识图谱等业务场景；在微信，则被用于解决大数据集存储、秒级回滚等问题。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"图数据在交通预测、舆情勘测、疾病传播建模、物理模拟等领域都有比较广泛的应用。举个例子，今年 8 月 8 日，康奈尔大学的专家发表了一篇论文，用循环图神经网络实现对在线论坛的谣言预测，GitHub 代码："},{"type":"link","attrs":{"href":"https:\/\/github.com\/google-research\/social_cascades","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/github.com\/google-research\/social_cascades"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"当我们要表达真实世界对象与对象间的数据关系时，图数据更加自然。在接下来的几年，与图数据有关的处理和存储软件，可能也会越来越多。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":" "}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}],"text":"参考链接："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/blog.tensorflow.org\/2021\/11\/introducing-tensorflow-gnn.html","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/blog.tensorflow.org\/2021\/11\/introducing-tensorflow-gnn.html"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/distill.pub\/2021\/gnn-intro\/","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/distill.pub\/2021\/gnn-intro\/"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/arxiv.org\/abs\/2108.03548","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/arxiv.org\/abs\/2108.03548"}],"marks":[{"type":"color","attrs":{"color":"#494949","name":"user"}}]}]}]}