cs224w(图机器学习)学习笔记1 Introduction and the Bowtie Structure of the Web

Churcee    2022-11-24 10:13

  1. 课程先导
    1. 证明可用方法:proof by contrapositive类比证明、by contradiction举反例、by cases举例子、by induction数学归纳法
    2. 数学算数基础:
      1. 微积分calculus:\(e^x=lim_{n\to\infty}(1+x/n)^n\),特例:\(e=lim_{n\to\infty}(1+1/n)^n,1/e=lim_{n\to\infty}(1-1/n)^n\)
      2. 概率论probablitity:条件概率conditional probability、随机变量random variables、期望和方差expectation and variance
      3. 线性代数linear algebra和矩阵运算matrix operations
  2. Why Graph
    1. 选择图的原因:图是用于描述并分析有关联/互动的实体的一种普适语言。它不将实体视为一系列孤立的点,而认为其互相之间有关系。它是一种很好的描述领域知识的方式。
    2. 网络与图的分类
      1. networks / natural graphs:自然表示为图
        1. Social networks: Society is a collection of 7+ billion individuals
        2. Communication and transactions: Electronic devices, phone calls, financial transactions
        3. Biomedicine: Interactions between genes/proteins regulate life(大概是基因或蛋白质之间互动从而调节生理活动的过程)
        4. Brain connections: Our thoughts are hidden in the connections between billions of neurons
      2. graphs:作为一种表示方法
        1. Information/knowledge are organized and linked
        2. Software can be represented as a graph
        3. Similarity networks: Connect similar data points
        4. Relational structures: Molecules, Scene graphs, 3D shapes, Particle-based physics simulations
      3. 有时network和graph之间的差别是模糊的
      4. 复杂领域会有丰富的关系结构,可以被表示为关系图relational graph,通过显式地建模关系,可以获得更好的表现
      5. 但是现代深度学习工具常用于建模简单的序列sequence(如文本、语音等具有线性结构的数据)和grid(图片具有平移不变性,可以被表示为fixed size grids或fixed size standards),这些传统工具很难用于图的建模,其难点在于网络的复杂:
        1. 任意大小和复杂拓扑Arbitrary size and complex topological structure (i.e.没有空间局部性no spatial locality like grids)
        2. 没有基准点,没有节点固定的顺序,没有那种上下左右的方向
        3. 网络是动态的并且具有多功能dynamic and have multi-model features
  3. main qusetion:how do we take advantage of relational structure for better prediction我们如何用这种结构优势作出更好更准确的预测——将神经网络模型使用范围扩展到图上(嵌入)
    machine learning with graphs
  4. 有监督学习流程
    1.  在传统机器学习流程中,我们需要对原始数据进行特征工程feature engineering(比如手动提取特征等),但是现在我们使用表征学习representation learning的方式来自动学习到数据的特征,直接应用于下游机器学习带来更好的预测。
      superviesd machine learning lifecycle
    2. 图的表示学习:大致来说就是将原始的节点(或链接、或图)表示为向量(嵌入embedding),图中相似的节点会被embed得靠近(指同一实体,在节点空间上相似,在向量空间上就也应当相似)
      嵌入式机器学习
  5. 课程聚焦——图结构数据的机器学习和表征学习machine learning and the representation learning for graph structure data
    1. Traditional methods: Graphlets, Graph Kernels
    2. Methods for node embeddings节点嵌入: DeepWalk, Node2Vec
    3. Graph Neural Networks: 卷积神经网络GCN, GraphSAGE, GAT(graph attention network), Theory of GNNs
    4. Knowledge graphs and reasoning: TransE, BetaE
    5. Deep generative models for graphs
    6. Applications to Biomedicine, Science, Industry as well.  
the outline of the course
 
Last Modified: 2022-11-24 19:06
Views: 517

[[total]] comments

Post your comment
  1. [[item.time]]
    [[item.user.username]] [[item.floor]]Floor
  2. Click to load more...
  3. Post your comment