神经网络该如何看待社会网络

在过去的一年里,我们已经开始使用一种新颖的深度学习架构作为Lynx基于Spark的社交网络分析平台的一部分。神经网络应用的一个普遍担忧是它们是“黑匣子”:数据进入,好的预测出来,但是没有人能解释为什么。我们必须解决这个问题,不仅是为了让客户放心,而且还帮助我们调试和改进我们的面向图形的神经模型。我们回顾了特征可视化和属性的最新进展。
展开查看详情

1.How Neural Networks See Social Networks Daniel Darabos, Janos Maginecz #DLSAIS11

2.Context Lynx does social network analytics for telcos & others. We use graph convolutional neural networks. Good predictions of missing data and future behavior. We want to understand how the network does it. #DLSAIS11 2

3.Graph Convolutional Network #DLSAIS11 3

4.input network prediction #DLSAIS11 4

5. repeat n times state new state input network prediction #DLSAIS11 5

6. repeat n times state new state vertex features network for each known vertex labels prediction neighbor vertex state edges #DLSAIS11 6

7.“Neural Network That Learns From a Huge Graph” Spark Summit East 2017 #DLSAIS11 7

8.Running Examples #DLSAIS11 8

9.complete real data network prediction #DLSAIS11 9

10. ✔ Validated against ✔ known labels ARTIAL real data Pcomplete network prediction #DLSAIS11 10

11. ✔ Validated against ✔ known labels ARTIAL real data Pcomplete network prediction attribution & feature explanation visualization #DLSAIS11 11

12. ✔ Validated against ✔ known labels t he t ic complete real data network prediction syn ✔ Validated against ✔ known rules attribution & feature explanation visualization #DLSAIS11 12

13.Running example: Three Friends Real social network edges. positive Synthetic features: random “talent”, 0–1. Synthetic labels: “three friends are rockstars” or not. “rockstar” means “talent” > θ rockstar #DLSAIS11 13

14. node rockstar color is class subject node #DLSAIS11 14

15. direct neighbor Running example: Friends of Friends rockstar Real social network edges. Synthetic features: random “talent”, 0–1. Synthetic labels: “distance to a rockstar is 2” or not. “rockstar” means “talent” > θ positive #DLSAIS11 15

16.Running example: age prediction 10–20 Real social network edges. 29–69 No features. Real labels: age bucket (4 equal-sized buckets) 20–24 24–29 #DLSAIS11 16

17.Running example: gender prediction Real social network edges. No features. Real labels: gender female male #DLSAIS11 17

18.Feature Visualization #DLSAIS11 18

19.Feature visualization What fictional inputs exemplify a class the most? “Feature Visualization” Chris Olah, et al., Distill #DLSAIS11 19

20.Creating examples random noise optimized image backprop on input to maximize class probability “Visualizing Higher-Layer Features of a Deep Network” Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pascal Vincent #DLSAIS11 20

21.Creating graph examples Initial state: complete graph with 10 vertices features randomized, 0–1 Backprop: features adjacency matrix #DLSAIS11 21

22.Feature visualization: Friends of Friends #DLSAIS11 22

23.Feature visualization: Friends of Friends #DLSAIS11 23

24.Feature visualization: Three Friends #DLSAIS11 24

25.Feature visualization: age prediction 24–29 99.2% confidence #DLSAIS11 25

26.Feature visualization: age prediction 29–69 99.9998% confidence #DLSAIS11 26

27.Feature visualization: age prediction 99.99% 24–29 confidence 29–69 10–20 #DLSAIS11 27

28.Feature visualization: age prediction 99.99% 24–29 confidence 10–20 #DLSAIS11 28

29.Feature visualization: age prediction 99.99% 24–29 confidence 10–20 #DLSAIS11 29