Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Machine Learning on Graph Data @ICML2019

Machine Learning on Graph Data @ICML2019

ICML2019からグラフ機械学習に関する論文を4本紹介します。

Ryosuke Kamesawa

July 21, 2019
Tweet

More Decks by Ryosuke Kamesawa

Other Decks in Technology

Transcript

  1. Machine Learning on Graph Data @ ICML 2019 亀澤諒亮 /

    DeNA ICLR’19 / ICML’19 読み会 @DeNA
  2. Graph ML @ ICML 2019 グラフデータを扱った論文はおよそ30本 (ICML’18ではおよそ15本)
 様々なタスクや手法をグラフに対しても適用できるように拡張したものが多い
 • Adversarial

    attacks • Disentanglement • Similarity measure • Self attention, etc. ニューラルネットワーク以外のアルゴリズムでグラフを扱うものも
 • Graph feature (kernel) / Gaussian process 応用
 • 回路設計の最適化
 • 楽譜からの演奏生成
 
 7
  3. Outline - Graph ML @ ICML 2019 - グラフについて -

    論文紹介 - A Persistent Weisfeiler–Lehman Procedure for Graph Classification - Bastian Rieck, Christian Bock, Karsten Borgwardt / ETH Zurich - Adversarial Attacks on Node Embeddings via Graph Poisoning - Aleksandar Bojchevski, Stephan Günnemann / Technical University of Munich - Simplifying Graph Convolutional Networks - Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr., Christopher Fifty, Tao Yu / Cornell - Position-aware Graph Neural Networks - Jiaxuan You, Rex Ying, Jure Leskovec / Stanford 8
  4. Graph - 頂点(node, vertex)と辺(link, edge)からなるデータ構造
 - 今回紹介するものは
 - 基本的に無向グラフ(辺の向きは考えない)
 -

    各頂点が特徴ベクトルをもつ場合が多い
 
 - 具体的には隣接行列として表現
 - 隣接行列
 
 - 例
 - Web
 - SNSネットワーク
 - 引用ネットワーク
 9 - タンパク質ネットワーク
 - 分子構造
 - etc.

  5. A Persistent Weisfeiler-Lehman Procedure for Graph Classification Bastian Rieck, Christian

    Bock, Karsten Borgwardt / ETH Zurich 10 http://proceedings.mlr.press/v97/rieck19a/rieck19a.pdf
  6. グラフの変化で構造(連結成分, 閉路, etc.) がどう変わるか?
 - 辺の重みによってグラフを整列
 
 
 
 


    バーコード
 
 Persistent Homology 19 ※Persistent homology における一般的な定義とは異なる https://www.math.upenn.edu/~ghrist/preprints/barcodes.pdf 連結成分の寿命
 CC persistence 閉路の寿命
 cycle persistence
  7. グラフの変化で構造(連結成分, 閉路, etc.) がどう変わるか?
 - 辺の重みによってグラフを整列
 
 
 
 


    バーコード
 
 Persistent Homology 20 ※Persistent homology における一般的な定義とは異なる https://www.math.upenn.edu/~ghrist/preprints/barcodes.pdf 連結成分の寿命
 CC persistence 閉路の寿命
 cycle persistence
  8. グラフの変化で構造(連結成分, 閉路, etc.) がどう変わるか?
 - 辺の重みによってグラフを整列
 
 
 
 


    バーコード
 
 Persistent Homology 21 ※Persistent homology における一般的な定義とは異なる https://www.math.upenn.edu/~ghrist/preprints/barcodes.pdf 連結成分の寿命
 CC persistence 閉路の寿命
 cycle persistence
  9. グラフの変化で構造(連結成分, 閉路, etc.) がどう変わるか?
 - 辺の重みによってグラフを整列
 
 
 
 


    バーコード
 
 Persistent Homology 22 ※Persistent homology における一般的な定義とは異なる https://www.math.upenn.edu/~ghrist/preprints/barcodes.pdf 連結成分の寿命
 CC persistence 閉路の寿命
 cycle persistence
  10. グラフの変化で構造(連結成分, 閉路, etc.) がどう変わるか?
 - 辺の重みによってグラフを整列
 
 
 
 


    バーコード
 
 Persistent Homology 23 ※Persistent homology における一般的な定義とは異なる https://www.math.upenn.edu/~ghrist/preprints/barcodes.pdf 連結成分の寿命
 CC persistence 閉路の寿命
 cycle persistence
  11. グラフの変化で構造(連結成分, 閉路, etc.) がどう変わるか?
 - 辺の重みによってグラフを整列
 
 
 
 


    バーコード
 
 Persistent Homology 24 ※Persistent homology における一般的な定義とは異なる https://www.math.upenn.edu/~ghrist/preprints/barcodes.pdf 連結成分の寿命
 CC persistence 閉路の寿命
 cycle persistence
  12. Persistent Homology グラフの変化で構造(連結成分, 閉路, etc.) がどう変わるか?
 - 辺の重みによってグラフを整列
 
 


    
 
 バーコード
 
 25 ※Persistent homology における一般的な定義とは異なる https://www.math.upenn.edu/~ghrist/preprints/barcodes.pdf 連結成分の寿命
 CC persistence 閉路の寿命
 cycle persistence
  13. Algorithm: Persistent subtree feature グラフ G=(V, E), 反復数 H Loop

    h = 1, …, H-1 隣接するラベルを集約
 辺の重み、PHを計算
 ラベル圧縮
 連結成分特徴
 
 閉路特徴
 
 
 26 http://www.jmlr.org/papers/volume12/shervashidze11a/shervashidze11a.pdf
  14. Algorithm: Persistent subtree feature 27 http://www.jmlr.org/papers/volume12/shervashidze11a/shervashidze11a.pdf グラフ G=(V, E), 反復数

    H Loop h = 1, …, H-1 隣接するラベルを集約
 辺の重み、PHを計算
 ラベル圧縮
 連結成分特徴
 
 閉路特徴
 
 

  15. Algorithm: Persistent subtree feature 28 http://www.jmlr.org/papers/volume12/shervashidze11a/shervashidze11a.pdf グラフ G=(V, E), 反復数

    H Loop h = 1, …, H-1 隣接するラベルを集約
 辺の重み、PHを計算
 ラベル圧縮
 連結成分特徴
 
 閉路特徴
 
 

  16. グラフ G=(V, E), 反復数 H Loop h = 1, …,

    H-1 隣接するラベルを集約
 辺の重み、PHを計算
 ラベル圧縮
 連結成分特徴
 
 閉路特徴
 
 
 Algorithm: Persistent subtree feature 29 http://www.jmlr.org/papers/volume12/shervashidze11a/shervashidze11a.pdf ラベル l i に対応するCC persistence の和
  17. Algorithm: Persistent subtree feature 30 グラフ G=(V, E), 反復数 H

    Loop h = 1, …, H-1 隣接するラベルを集約
 辺の重み、PHを計算
 ラベル圧縮
 連結成分特徴
 
 閉路特徴
 
 
 http://www.jmlr.org/papers/volume12/shervashidze11a/shervashidze11a.pdf 両辺どちらかの頂点のラベルが l i である cycle persistence の和
  18. Summary Weisfeiler-Lehman subtree は明示的にトポロジカルな特徴を扱わない
 
 Persistent Weisfeiler-Lehman Procedure を提案
 -

    明示的に閉路の情報を考慮
 - Persistent homology による表現を利用
 
 Graph feature / kernel の既存手法に比べて性能向上
 32
  19. Adversarial Attacks on Node Embeddings via Graph Poisoning Aleksandar Bojchevski,

    Stephan Günnemann / Technical University of Munich 33 http://proceedings.mlr.press/v97/bojchevski19a/bojchevski19a.pdf
  20. Abstract - DeepWalk(教師なしノード埋め込み)への攻撃を提案
 - 攻撃:グラフの辺の反転
 - 従来法[Zügner+, ICLR’19][Zügner+, KDD’18]は半教師あり学習に対する攻撃
 


    - 反転すべき辺を効率的に見つけるアルゴリズムを提案
 
 - 半教師あり学習でも攻撃によって性能劣化
 
 - 攻撃に用いたグラフはDeepWalk以外にも転用可能
 34
  21. DeepWalk [Perozzi+, KDD’14] ノード埋め込みを計算する手法
 - サンプルされたrandom walk 上での skip-gram -

    Random walkを文、ノードを単語とみなす
 - Skip-gramによってノード(単語)の埋め込みを計算
 実は特異値分解として解ける
 35 http://proceedings.mlr.press/v97/bojchevski19a/bojchevski19a.pdf
  22. Approximation - 勾配法
 - 特異値分解のためにナイーブには勾配を計算できない
 - Eigenvalue perturbationによって近似
 - 隣接行列の変化量は

    ±1のため誤差も大きくなる
 - 計算量 
 
 - 貪欲法
 - 辺を反転した場合の損失の変化量を効率的に計算できるように近似
 - 辺の追加
 - 候補となる頂点対をサンプリングし、最適な辺を選択
 - 辺の削除
 - 存在する辺の中から最適な辺を選択
 
 
 44
  23. Simplifying Graph Convolutional Networks Felix Wu, Tianyi Zhang, Amauri Holanda

    de Souza Jr., Christopher Fifty, Tao Yu / Cornell 48 http://proceedings.mlr.press/v97/wu19e/wu19e.pdf
  24. Abstract - Graph Convolutional Network (GCN) [Kipf+, ICLR’17] を単純化したモデルでも
 ノード分類において大きな性能低下が見られないことを確認


    • Simple Graph Convolution (SGC) 
 - SGCがグラフスペクトル上のローパスフィルタであることを指摘
 
 
 49 http://proceedings.mlr.press/v97/wu19e/wu19e.pdf
  25. Graph Convolutional Network [Kipf+, ICLR’17] 
 
 
 
 


    
 正規化隣接行列
 50 http://proceedings.mlr.press/v97/wu19e/wu19e.pdf (次数行列)
  26. Simple Graph Convolution (SGC) 
 
 
 
 
 


    
 GCNの活性化関数(ReLU)を除くことで単純化
 特徴量の伝播は単純な行列積に帰着
 51 http://proceedings.mlr.press/v97/wu19e/wu19e.pdf
  27. Graph Signal Processing 正規化グラフラプラシアン 
 ノード上の値(信号)
 グラフフーリエ変換
 逆グラフフーリエ変換
 
 


    
 
 
 
 53 http://web.media.mit.edu/~xdong/presentation/CDT_GuestLecture_GSP.pdf (固有値(周波数)小) (固有値(周波数)大) 固有値分解 グラフ上の固有ベクトル [Shuman+, ’13] http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.367.6064&rep=rep1&type=pdf
  28. Position-aware Graph Neural Networks Jiaxuan You, Rex Ying, Jure Leskovec

    / Stanford 56 http://proceedings.mlr.press/v97/you19b/you19b.pdf
  29. Motivation 既存のGNN(e.g. GCN)では同じ局所構造を持ったノードを区別できない
 
 
 
 
 
 
 


    
 → 基準となるノードの集合 Anchor-setを設定し、そこから距離を考慮した特徴伝播を行う
 58 http://proceedings.mlr.press/v97/you19b/you19b.pdf
  30. Algorithm: P-GNN 1. Anchor-set S 1 , ... ,S k

    のサンプリング
 2. 頂点v とAnchor-set S i に対してメッセージM v [i] を計算
 
 3. M v から出力ベクトルと次の層の入力ベクトルを計算
 
 
 
 
 
 
 59 http://proceedings.mlr.press/v97/you19b/you19b.pdf
  31. Algorithm: P-GNN 60 http://proceedings.mlr.press/v97/you19b/you19b.pdf 1. Anchor-set S 1 , ...

    ,S k のサンプリング
 2. 頂点v とAnchor-set S i に対してメッセージM v [i] を計算
 
 3. M v から出力ベクトルと次の層の入力ベクトルを計算
 
 
 
 
 
 

  32. Algorithm: P-GNN 61 http://proceedings.mlr.press/v97/you19b/you19b.pdf 1. Anchor-set S 1 , ...

    ,S k のサンプリング
 2. 頂点v とAnchor-set S i に対してメッセージM v [i] を計算
 
 3. M v から出力ベクトルと次の層の入力ベクトルを計算
 
 
 
 
 
 

  33. Anchor-Set 距離空間
 
 
 Bourgain Theorem [Bourgain, 1985] 距離空間
 


    
 具体的構成
 Anchor-set S i,j は各頂点を確率1/2iで独立にサンプリング
 
 62 のdistortionが である
 ⇔ 
 に対してdistortionが
 となるような
 が存在する。

  34. Anchor-Set 距離空間
 
 
 Bourgain Theorem [Bourgain, 1985] 距離空間
 


    
 具体的構成
 Anchor-set S i,j は各頂点を確率1/2iで独立にサンプリング
 
 63 のdistortionが である
 が小さいほど距離を保つ ⇔ 
 に対してdistortionが
 となるような
 が存在する。

  35. Anchor-Set 距離空間
 
 
 Bourgain Theorem [Bourgain, 1985] 距離空間
 


    
 具体的構成
 Anchor-set S i,j は各頂点を確率1/2iで独立にサンプリング
 
 64 のdistortionが である
 ⇔ 
 に対してdistortionが
 となるような
 が存在する。

  36. Anchor-Set 距離空間
 
 
 Bourgain Theorem [Bourgain, 1985] 距離空間
 


    
 具体的構成
 Anchor-set S i,j は各頂点を確率1/2iで独立にサンプリング
 → S i,j を使うことでグラフ上の距離を保ったまま特徴空間にマッピングできる
 
 65 のdistortionが である
 ⇔ 
 に対してdistortionが
 となるような
 が存在する。

  37. Whole Summary - グラフを扱う論文は増えてきている (昨年からおよそ倍増)
 - タスクも手法も応用も多様化
 
 - 論文紹介


    - A Persistent Weisfeiler–Lehman Procedure for Graph Classification - Adversarial Attacks on Node Embeddings via Graph Poisoning - Simplifying Graph Convolutional Networks - Position-aware Graph Neural Networks 
 69
  38. References (ICML’19) B. Rieck, C. Bock, and K. Borgwardt, “A

    Persistent Weisfeiler-Lehman Procedure for Graph Classification,” in International Conference on Machine Learning, 2019, pp. 5448–5458. A. Bojchevski and S. Günnemann, “Adversarial Attacks on Node Embeddings via Graph Poisoning,” in International Conference on Machine Learning, 2019, pp. 695–704. G. Zhang, H. He, and D. Katabi, “Circuit-GNN: Graph Neural Networks for Distributed Circuit Design,” in International Conference on Machine Learning, 2019, pp. 7364–7373. Y. Yu, J. Chen, T. Gao, and M. Yu, “DAG-GNN: DAG Structure Learning with Graph Neural Networks,” in International Conference on Machine Learning, 2019, pp. 7154–7163. J. Ma, P. Cui, K. Kuang, X. Wang, and W. Zhu, “Disentangled Graph Convolutional Networks,” in International Conference on Machine Learning, 2019,pp.4212–4221. F. Gao, G. Wolf, and M. Hirn, “Geometric Scattering for Graph Data Analysis,” in International Conference on Machine Learning, 2019, pp. 2122–2131. E. Smith, S. Fujimoto, A. Romero, and D. Meger, “GEOMetrics: Exploiting Geometric Structure for Graph-Encoded Objects,” in International Conference on Machine Learning, 2019, pp. 5866–5876. M. Qu, Y. Bengio, and J. Tang, “GMNN: Graph Markov Neural Networks,” in International Conference on Machine Learning, 2019, pp. 5241–5250. I. Walker and B. Glocker, “Graph Convolutional Gaussian Processes,” in International Conference on Machine Learning, 2019, pp. 6495–6504. F. Alet, A. K. Jeewajee, M. B. Villalonga, A. Rodriguez, T. Lozano-Perez, and L. Kaelbling, “Graph Element Networks: adaptive, structured computation and memory,” in International Conference on Machine Learning, 2019, pp. 212–222. Y. Li, C. Gu, T. Dullien, O. Vinyals, and P. Kohli, “Graph Matching Networks for Learning the Similarity of Graph Structured Objects,” in International Conference on Machine Learning, 2019, pp. 3835–3845. D. Jeong, T. Kwon, Y. Kim, and J. Nam, “Graph Neural Network for Music Score Data and Modeling Expressive Piano Performance,” in International Conference on Machine Learning, 2019, pp. 3060–3070. J. Hendrickx, A. Olshevsky, and V. Saligrama, “Graph Resistance and Learning from Pairwise Comparisons,” in International Conference on Machine Learning, 2019, pp. 2702–2711. H. Gao and S. Ji, “Graph U-Nets,” in International Conference on Machine Learning, 2019, pp. 2083–2092. A. Grover, A. Zweig, and S. Ermon, “Graphite: Iterative Generative Modeling of Graphs,” in International Conference on Machine Learning, 2019, pp. 2434–2444. R. Suzuki, R. Takahama, and S. Onoda, “Hyperbolic Disk Embeddings for Directed Acyclic Graphs,” in International Conference on Machine Learning, 2019, pp. 6066–6075. C. Zhu, S. Storandt, K.-Y. Lam, S. Han, and J. Bi, “Improved Dynamic Graph Learning through Fault-Tolerant Sparsification,” in International Conference on Machine Learning, 2019, pp. 7624–7633. S. Zhang, X. He, and S. Yan, “LatentGNN: Learning Efficient Non-local Relations for Visual Recognition,” in International Conference on Machine Learning, 2019, pp. 7374–7383. L. Franceschi, M. Niepert, M. Pontil, and X. He, “Learning Discrete Structures for Graph Neural Networks,” in International Conference on Machine Learning, 2019, pp. 1972–1982. L. Guo, Z. Sun, and W. Hu, “Learning to Exploit Long-term Relational Dependencies in Knowledge Graphs,” in International Conference on Machine Learning, 2019, pp. 2505–2514. D. Baranchuk, D. Persiyanov, A. Sinitsin, and A. Babenko, “Learning to Route in Similarity Graphs,” in International Conference on Machine Learning, 2019, pp. 475–484. S. Abu-El-Haija et al., “MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing,” in International Conference on Machine Learning, 2019, pp. 21–29. H. Kajino, “Molecular Hypergraph Grammar with Its Application to Molecular Optimization,” in International Conference on Machine Learning, 2019, pp. 3183–3191. V. Titouan, N. Courty, R. Tavenard, C. Laetitia, and R. Flamary, “Optimal Transport for structured data with application on graphs,” in International Conference on Machine Learning, 2019, pp. 6275–6284. J. You, R. Ying, and J. Leskovec, “Position-aware Graph Neural Networks,” in International Conference on Machine Learning, 2019, pp. 7134–7143. R. Murphy, B. Srinivasan, V. Rao, and B. Ribeiro, “Relational Pooling for Graph Representations,” in International Conference on Machine Learning, 2019, pp. 4663–4673. J. Lee, I. Lee, and J. Kang, “Self-Attention Graph Pooling,” in International Conference on Machine Learning, 2019, pp. 3734–3743. F. Wu, A. Souza, T. Zhang, C. Fifty, T. Yu, and K. Weinberger, “Simplifying Graph Convolutional Networks,” in International Conference on Machine Learning, 2019, pp. 6861–6871. P. Mercado, F. Tudisco, and M. Hein, “Spectral Clustering of Signed Graphs via Matrix Power Means,” in International Conference on Machine Learning, 2019, pp. 4526–4536. N. Mehta, L. C. Duke, and P. Rai, “Stochastic Blockmodels meet Graph Neural Networks,” in International Conference on Machine Learning, 2019, pp. 4466–4474. 70
  39. References (ICML’18) H. Dai et al., “Adversarial Attack on Graph

    Structured Data,” in International Conference on Machine Learning, 2018, pp. 1115–1124. D. Bacciu, F. Errica, and A. Micheli, “Contextual Graph Markov Model: A Deep and Generative Approach to Graph Processing,” in International Conference on Machine Learning, 2018, pp. 294–303. A. Sanchez-Gonzalez et al., “Graph Networks as Learnable Physics Engines for Inference and Control,” in International Conference on Machine Learning, 2018, pp. 4470–4479. J. You, R. Ying, X. Ren, W. Hamilton, and J. Leskovec, “GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models,” in International Conference on Machine Learning, 2018, pp. 5708–5717. D. Calandriello, A. Lazaric, I. Koutis, and M. Valko, “Improved large-scale graph learning through ridge spectral sparsification,” in International Conference on Machine Learning, 2018, pp. 688–697. W. Jin, R. Barzilay, and T. Jaakkola, “Junction Tree Variational Autoencoder for Molecular Graph Generation,” in International Conference on Machine Learning, 2018, pp. 2323–2332. H. Dai, Z. Kozareva, B. Dai, A. Smola, and L. Song, “Learning Steady-States of Iterative Algorithms over Graphs,” in International Conference on Machine Learning, 2018, pp. 1106–1114. A. Douik and B. Hassibi, “Low-Rank Riemannian Optimization on Positive Semidefinite Stochastic Matrices with Applications to Graph Clustering,” in International Conference on Machine Learning, 2018, pp. 1299–1308. A. Bojchevski, O. Shchur, D. Zügner, and S. Günnemann, “NetGAN: Generating Graphs via Random Walks,” in International Conference on Machine Learning, 2018, pp. 610–619. K. Levin, F. Roosta, M. Mahoney, and C. Priebe, “Out-of-sample extension of graph adjacency spectral embedding,” in International Conference on Machine Learning, 2018, pp. 2975–2984. J. Xu, “Rates of Convergence of Spectral Methods for Graphon Estimation,” in International Conference on Machine Learning, 2018, pp. 5433–5442. K. Xu, C. Li, Y. Tian, T. Sonobe, K. Kawarabayashi, and S. Jegelka, “Representation Learning on Graphs with Jumping Knowledge Networks,” in International Conference on Machine Learning, 2018, pp. 5453–5462. A. Loukas and P. Vandergheynst, “Spectrally Approximating Large Graphs with Smaller Graphs,” in International Conference on Machine Learning, 2018, pp. 3237–3246. J. Chen, J. Zhu, and L. Song, “Stochastic Training of Graph Convolutional Networks with Variance Reduction,” in International Conference on Machine Learning, 2018, pp. 942–950. 71
  40. References N. Shervashidze, P. Schweitzer, E. J. van Leeuwen, K.

    Mehlhorn, and K. M. Borgwardt, “Weisfeiler-Lehman Graph Kernels,” Journal of Machine Learning Research, vol. 12, no. Sep, pp. 2539–2561, 2011. B. Perozzi, R. Al-Rfou, and S. Skiena, “DeepWalk: Online Learning of Social Representations,” Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining - KDD ’14, pp. 701–710, 2014. H. Edelsbrunner and J. Harer. Persistent homology — a survey. Surveys on Discrete and Computational Geometry. Twenty Years Later, eds. J. E. Goodman, J. Pach and R. Pollack, Contemporary Mathematics 453, 257–282, Amer. Math. Soc., Providence, Rhode Island, 2008. J. Qiu, Y. Dong, H. Ma, J. Li, K. Wang, and J. Tang, “Network Embedding as Matrix Factorization: Unifying DeepWalk, LINE, PTE, and node2vec,” Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining - WSDM ’18, pp. 459–467, 2018. T. N. Kipf and M. Welling, “Semi-Supervised Classification with Graph Convolutional Networks,” arXiv:1609.02907 [cs, stat], Sep. 2016. J. Bourgain, “On lipschitz embedding of finite metric spaces in Hilbert space,” Israel Journal of Mathematics, vol. 52, no. 1, pp. 46–52, Mar. 1985. D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains,” IEEE Signal Processing Magazine, vol. 30, no. 3, pp. 83–98, May 2013. 72