disentangled graph convolutional networks
play

Disentangled Graph Convolutional Networks Jianxin Ma, Peng Cui, Kun - PowerPoint PPT Presentation

Disentangled GCNs (ICML19) Disentangled Graph Convolutional Networks Jianxin Ma, Peng Cui, Kun Kuang, Xin Wang, Wenwu Zhu Tsinghua University Disentangled GCNs (ICML19) Motivation The neighborhood of a node is formed due to many


  1. Disentangled GCNs (ICML’19) Disentangled Graph Convolutional Networks Jianxin Ma, Peng Cui, Kun Kuang, Xin Wang, Wenwu Zhu Tsinghua University

  2. Disentangled GCNs (ICML’19) Motivation • The neighborhood of a node is formed due to many latent factors . 𝑤 % 𝑤 $ 𝑤 # 𝑣 𝑤 % 𝑤 ) 𝑤 $ 𝑤 & 𝑤 ) 𝑤 # 𝑣 𝑤 * 𝑤 ' 𝑤 ( 𝑣 𝑤 * 𝑣 Latent factor: Family Latent factor: Hobby 𝑤 Latent factor: Work & 𝑤 ' 𝑤 ( • Existing GCNs convolute the neighborhood as a whole. • They do not distinguish between the latent factors. • Their node representations are thus not robust , and hardly interpretable.

  3. Disentangled GCNs (ICML’19) Disentangled GCNs • Disentangled representation learning aims to identify and separate the underlying explanatory factors behind the observed data (Bengio et al., 2013). Feed back to improve neighborhood routing. 𝑿 𝟐 convolution 𝑤 % 𝑤 % 𝑤 $ 𝑤 $ 𝑤 % 𝑤 # 𝑤 $ 𝑤 % 𝑤 # 𝑤 $ 𝑤 # 𝑤 # convolution 𝑿 𝟑 𝑤 ) 𝑤 ) 𝑤 ) 𝑤 * 𝑤 ) 𝑤 * 𝑣 𝑣 𝑤 * 𝑤 * convolution 𝑿 𝟒 𝑤 𝑤 & & 𝑤 & 𝑤 ( 𝑤 ' 𝑤 & 𝑤 ( 𝑤 ' concatenate 𝑤 ' 𝑤 ' 𝑤 ( 𝑤 ( Layer 𝑣 Layer Input Extract features specific to each factor. Neighborhood Routing Output • We identify the latent factors, and segment the neighborhood accordingly. • Each segment is related with an isolated factor, and is convoluted separately.

  4. Disentangled GCNs (ICML’19) Neighborhood Routing • We propose neighborhood routing , to segment a neighborhood. • Dynamic & differentiable. Similar to capsule networks’ dynamic routing. • Phase I: • Phase II: • To extract factor-specific features. • To infer the factor that causes the link between node 𝑣 and a neighbor 𝑤 . § For node 𝑗 ∈ 𝑣 ∪ 𝑤: 𝑤, 𝑣 ∈ 𝐻 , and factor 𝑙 ∈ 1,2, … , 𝐿 , § Initialize 𝒅 < ← 𝒜 J,< for each factor 𝑙 . A 𝒚 C D𝒄 @ ) § Iterate for 𝑈 ≈ 5 times, >(𝑿 @ 𝒜 ;,< = § A 𝒚 C D𝒄 @ ) A >(𝑿 @ PQR 𝒜 S,@ 𝒅 @ /U 𝑞 O,< ← § G § which describes node 𝑗 ’s aspect 𝑙 . A ∑ @W PQR 𝒜 S,@W 𝒅 @W /U 𝒜 X,@ D∑ S: S,X ∈Y Z S,@ 𝒜 S,@ 𝒅 < ← § 𝒜 X,@ D∑ S: S,X ∈Y Z S,@ 𝒜 S,@ G § 𝒅 < describes the neighborhood’s aspect 𝑙 .

  5. Disentangled GCNs (ICML’19) Intuitions & Theories • The two intuitions behind neighborhood routing: • 𝑞 Factor 𝑙 is the one that causes the links between node 𝑣 and a segment ∝ The segment contains a large number of nodes that are similar w.r.t. aspect 𝑙 . • 𝑞 Factor 𝑙 is the one that causes the link between node 𝑣 and a neighbo r ∝ Node 𝑣 and the neighbor are similar w.r.t. aspect 𝑙 . • Neighborhood routing is equivalent to an EM algorithm that performs inference under a von Mises-Fisher subspace clustering model. • It finds one large cluster in each of the 𝐿 subspaces.

  6. Disentangled GCNs (ICML’19) Results: Multi-label Node Classification

  7. Disentangled GCNs (ICML’19) Results: Disentangled Node Representations • Correlations between the 64 dimensions, on a graph with eight factors. 1 1 1 1 2 2 3 3 4 4 5 5 6 6 7 0.9 7 0.9 8 8 9 9 10 10 11 11 12 12 13 13 0.8 0.8 14 14 15 15 16 16 17 17 18 18 19 19 0.7 0.7 20 20 21 21 22 22 23 23 24 24 25 25 26 26 0.6 0.6 27 27 28 28 29 29 30 30 31 31 32 32 0.5 0.5 33 33 34 34 35 35 36 36 37 37 38 38 39 0.4 39 0.4 40 40 41 41 42 42 43 43 44 44 45 45 0.3 0.3 46 46 47 47 48 48 49 49 50 50 51 51 0.2 0.2 52 52 53 53 54 54 55 55 56 56 57 57 58 58 0.1 0.1 59 59 60 60 61 61 62 62 63 63 64 64 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9

Recommend


More recommend