04.深度學(xué)習(xí)pytorch6網(wǎng)絡(luò)節(jié)點(diǎn)分類

1、標(biāo)題：網(wǎng)絡(luò)節(jié)點(diǎn)分類(Pea主講:王)大綱網(wǎng)絡(luò)節(jié)點(diǎn)分類簡(jiǎn)介經(jīng)典網(wǎng)絡(luò)節(jié)點(diǎn)分類算法介紹圖卷積神經(jīng)網(wǎng)絡(luò)介紹網(wǎng)絡(luò)（）數(shù)據(jù)Graphs are everywhere!網(wǎng)絡(luò)節(jié)點(diǎn)分類針對(duì)包含的數(shù)據(jù)進(jìn)行分類。其是如何利用這種“”關(guān)系In ciion networks: predict the topic of prs from properties of citedpInrsnetworks: predictanizational roles from communicationpatternsIn sofriendsl networks: predict theal preferenfrom characte

2、ristics ofIn biological networks: predict the protein function frompatternseraction經(jīng)典網(wǎng)絡(luò)節(jié)點(diǎn)分類算法Weighted-vote relational neighbor (wvRN)The consistency (of label propagation) methodWeighted-vote relational neighbor (wvRN)wvRN 的: 節(jié)點(diǎn)的應(yīng)該由其鄰居投票決定：14是節(jié)點(diǎn)的鄰居集合？23wvRN算法實(shí)現(xiàn)如下：Fix the label of labeled nodesIterat

3、ively re-calculates probabilities based on current be of neighbors until convergencefsMacskassy S A, Provost F. A simple relational classifierR. NEW YORK UNIV NY STERN SCHOOL OF BUSINESS, 2003.wvRN的例子對(duì)所有的無第一次迭代:節(jié)點(diǎn)初始化：對(duì)于，對(duì)于，對(duì)于,wvRN的例子進(jìn)行5次迭代的結(jié)果：Nodes 5, 8, 9 are + ()Node 3 is ()Node 4 is in betn ()傳統(tǒng)網(wǎng)

4、絡(luò)節(jié)點(diǎn)分類的缺點(diǎn)未考慮節(jié)點(diǎn)上的屬性信息（比如文本網(wǎng)絡(luò)上的文本）1.2.都是淺層模型，模型容量比較小，在復(fù)雜網(wǎng)絡(luò)上表現(xiàn)一般3.基本都是線性模型，不能表達(dá)圖中的非線性信息圖神經(jīng)網(wǎng)絡(luò)GCN回顧：卷積神經(jīng)網(wǎng)絡(luò)的卷積操作回顧：卷積神經(jīng)網(wǎng)絡(luò)的卷積操作不能直接用到Graph上：這是因?yàn)镚raph是不規(guī)則的數(shù)據(jù)，即每個(gè)節(jié)點(diǎn)的鄰居結(jié)構(gòu)不一樣，導(dǎo)致不能用一個(gè)卷積核（濾波器）圖卷積神經(jīng)網(wǎng)絡(luò)Thomas N. Kipf, Max Welling, Semi-Supervised Classification with Graph Convolutional Networks (ICLR 2017)Graph Convolutional Networks (GCN) 參考了圖像卷積()的，將其應(yīng)用到圖結(jié)構(gòu)這類帶的數(shù)據(jù)上。圖卷積神經(jīng)網(wǎng)絡(luò)GCN的主要貢獻(xiàn)：1.2.3.增加卷積時(shí)的“自回路”，即節(jié)點(diǎn)卷積的時(shí)候要考慮自身節(jié)點(diǎn)的信息避免“梯度”設(shè)計(jì)高效的圖卷積模式GCN算法1.2.3.給每個(gè)節(jié)點(diǎn)加上自連接，設(shè)鄰接矩陣，得到對(duì)進(jìn)行歸一化：那么第層圖卷積可以寫成：輸入的原始節(jié)點(diǎn)特征，比如節(jié)點(diǎn)文本：第卷積層參數(shù)：用于A歸一化的輔助矩陣注意：也有人用另外一種歸一化方法GCN算法神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)：GCN只使用了兩層圖卷積神經(jīng)網(wǎng)