Chapter 4 better.pptx
Chapter 4 better.pptx
 2. AGENDA
• Abstract
• Background and Problem Definition
• Supervised Graph Neural Networks
• General Frame work of Graph Neural Networks (GNN)
• Graph Convolutional Networks (GCN)
• Graph Attention Networks
• Neural Message Passing Networks(MPNN)
• Continuous Graph Neural Networks
• MultiScale Spectral Graph Convolutional Networks
• Unsupervised Graph Neural Networks
• Variational Graph AutoEncoders
• Deep Graph Infomax
• Oversmoothing Problem  4. BACKGROUND AND PROBLEM DEFINITION
• A fundamental task on graphs is node classification, which tries to
classify the nodes into a few predefined categories.
• To make effective prediction, a critical problem is to have very effective
node representations, which largely determine the performance of node
classification.  5. BACKGROUND AND PROBLEM DEFINITION
• N is the total number of nodes.
• C is the number of features for each node.
• F is the dimension of node representations.
The goal of graph neural networks is
to learn effective node representations
(denoted as H ∈ R N×F) by combining
the graph structure information and the node attributes, which are further used for
node classification.  6. Supervised Graph Neural Networks
• General Frame work of Graph Neural Networks (GNN)
• Graph Convolutional Networks (GCN)
• Graph Attention Networks
• Neural Message Passing Networks(MPNN)
• Continuous Graph Neural Networks
• MultiScale Spectral Graph Convolutional Networks  7. 1 General Frame work of Graph Neural Networks (GNN)
• The essential idea of graph neural networks is to iteratively update the
node representations by combining the representations of their neighbors
and their own representations.  8. 1 General Frame work of Graph Neural Networks (GNN)
Starting from the initial node representation H0 = X,
in each layer we have two important functions:
• AGGREGATE : which tries to aggregate the information from the neighbours
of each node;
• COMBINE : which tries to update the node representations by combining the
aggregated information from neighbors with the current node
representations.  10. 2 Graph Convolutional Networks (GCN)
• Which is now the most popular graph neural network architecture due to its
simplicity and effectiveness in a variety of tasks and applications.
• Specifically, the node representations in each layer is updated according to a
propagation rule.  11. 2 Graph Convolutional Networks (GCN)
• AGGREGATE function is defined as the weighted average of the
neighbor node representations. The weight of the neighbor j is
determined by the weight of the edge between i and j (i.e. Ai j
normalized by the degrees of the two nodes)
• COMBINE function is defined as the summation of the aggregated
messages and the node representation itself, in which the node
representation is normalized by its own degree.  12. 3 Graph Attention Networks
• In GCNs, for a target node i, the importance of a neighbor j is determined
by the weight of their edge Ai j (normalized by their node degrees).
• However, in practice, the input graph may be noisy.
• The edge weights may not be able to reflect the true strength between two
nodes.
• As a result, a more principled approach might be to automatically learn the
importance of each neighbor. Graph Attention Networks (a.k.a. GAT) is built
on this idea and try to learn the importance of each neighbor based on the
Attention mechanism.  13. 3 Graph Attention Networks
• Graph Attention Layer: defnes how to transfer the hidden node representations at layer k −1 (denoted
as H k−1) to the new node representations H.
• In order to guarantee suffcient expressive power to transform the lowerlevel node representations to
higherlevel node representations, a shared linear transformation is applied to every node.
• Afterwards, selfattention is defned on the nodes, which measures the attention coeffcients for any
pair of nodes through a shared attentional mechanism (indicates the relationship strength between
node i and j).
• And to make the coeffcients comparable across different nodes, the attention coeffcients are usually
normalized with the softmax function:
• The new node representation is a linear combination of the neighboring node representations with the
weights determined by the attention coeffcients (with a potential nonlinear transformation)  14. 4 Neural Message Passing Networks(MPNN)
• MPNN is actually very general, provides a general framework of
graph neural networks, and can be used for the task of node
classifcation as well.
• The essential idea of MPNN is formulating existing graph neural
networks as a general framework of neural message passing
among nodes.
• In MPNNs, there are two important functions including
Message and Updating function:  15. 4 Neural Message Passing Networks(MPNN)
• The MPNN framework is very similar to the general framework
• The AGGREGATE function defned here is simply a summation of all the
messages from the neighbors.
• The COMBINE function is the same as the node Updating function: which
combines the aggregated messages from the neighbors and the node
representation itself.  16. 5 Continuous Graph Neural Networks
• The above graph neural networks iteratively update the node representations with different
kinds of graph convolutional layers. Essentially, these approaches model the discrete
dynamics of node representations with GNNs.
• Xhonneux et al (2020) proposed the continuous graph neural networks (CGNNs), which
generalizes existing graph neural networks with discrete dynamics to continuous settings, i.e.,
trying to model the continuous dynamics of node representations.
• The key idea is how to characterize the continuous dynamics of node representations, i.e. the
derivatives of node representation w.r.t. time.
• The derivatives of the node representations are defned as a combination of the node
representation itself, the representations of its neighbors, and the initial status of the nodes.  17. 5 Continuous Graph Neural Networks
• Specifcally, two different variants of node dynamics are
introduced.
• The frst model assumes that different dimensions of node
presentations (a.k.a. feature channels) are independent;
• The second model is more fexible, which allows different feature
channels to interact with each other.  20. VARIATIONAL GRAPH AUTOENCODERS
The VGAE model consists of an encoder, a decoder, and a prior .
• Encoder
The goal of the encoder is to learn a distribution of latent variables associated with each node
conditioning on the node features X and the adjacency matrix A. (qφ (ZX,A)).
• Decoder
Given sampled latent variables, the decoder aims at predicting the connectivities among nodes.
Aopts a simple dotproduct based predictor.
• Prior
The prior distributions over the latent variables are simply set to independent zeromean Gaussians with unit variances,
This prior is fixed throughout the learning.  21. DEEP GRAPH INFOMAX
• Deep Graph Infomax is an unsupervised learning framework that
learns graph representations via the principle of mutual information
maximization.  22. OVERSMOOTHING PROBLEM
• Training deep graph neural networks by stacking multiple layers of graph neural networks
usually yields inferior results, which is a common problem observed in several different
graph neural network architectures.
• This is mainly due to the problem of oversmoothing.
• The graph convolutional network is a special case of Laplacian smoothing:
• The Laplacian smoothing will push nodes belonging to the same clusters to take similar
representations, which are benefcial for downstream tasks such as node classifcation.
• However, when the GCNs go deep, the node representations suffer from the problem of
oversmoothing, i.e., all the nodes will have similar representations. As a result, the
performance on downstream tasks suffer as well.  23. OVERSMOOTHING PROBLEM
• PairNorm: a method for alleviating the problem of oversmoothing
when GNNs go deep.
• The essential idea of PairNorm is to keep the total pairwise squared
distance (TPSD) of node representations unchanged, which is the
same as that of the original node feature X.
Disasters Expo USA, is proud to be supported by Inergency for their next upcoming edition on March 6th & 7th 2024!
The leading event mitigating the world’s most costly disasters is returning to the Miami Beach
So don’t miss out and register today: https://shorturl.at/aikrW
And in case you missed it, here is our ultimate road trip playlist is the perfect mix of podcasts, and hidden gems that will keep you energized for the entire journey
