Capturing network dynamics using dynamic graph representation learning

1. A computer-implemented method for link prediction of a dynamic network, the method comprising:

generating a dynamic graph embedding model for capturing temporal patterns of dynamic graphs, each of the graphs being an evolved representation of the dynamic network over time steps for a temporal lookback period and having a set of vertices and edges, the edges representing network links, the temporal patterns including nonlinear interactions between graph vertices at each time step and across multiple time steps, wherein the dynamic graph embedding model is configured as a neural network modeled as an autoencoder-recurrent neural network combination using fully connected dense layers as a first part of an encoder of the neural network to generate a low dimensional vector representation of a sparse neighborhood vector of the evolved dynamic graphs, and using layers of long short term memory networks as a second part of an encoder of the neural network to generate the embedding from the low dimensional vector representation;

learning structural patterns in the dynamic network using the nonlinear layers;

learning, by the dynamic graph embedding model, a dynamic graph embedding for the temporal lookback period by optimizing a loss function that includes a weighting matrix for weighting reconstruction of observed edges higher than unobserved links, wherein the network embedding is a representation of each of the vertices in a series of low-dimensional vector space; and

predicting edges at a future time step based on parameters of the neural network, wherein the parameters are tuned by optimizing the loss function.

2. The computer-implemented method of claim 1 , wherein the weighting matrix includes a hyperparameter for controlling weight of penalizing observed edges.

3. The computer-implemented method of claim 1 , wherein the predicted edges include at least one edge representation of an unobserved network link.

4. The computer-implemented method of claim 1 , further comprising:

receiving, by the dynamic graph embedding model, inputs as a set of previous dynamic graphs for the temporal lookback period, the lookback period being used to control a length of temporal patterns learned; and

decoding the dynamic graph embedding to produce the output of the dynamic graph embedding model, wherein the output is a predicted graph for the next time step.

5. The computer-implemented method of claim 1 , wherein the long short term memory networks are sparsely connected, and wherein l long short term memory networks are connected in the first layer of the neural network, where l is the number of previous graphs for the temporal lookback period.

6. The computer-implemented method of claim 1 , further comprising:

using fully connected layers as a decoder for the neural network; and

decoding, by the decoder, the dynamic graph embedding to produce the output of the dynamic graph embedding model, wherein the output is a predicted graph for the next time step.

7. A system comprising:

a memory comprising computer readable instructions; and

a processing device for executing the computer readable instructions for executing steps of link prediction of a dynamic network, the steps comprising:

generating a dynamic graph embedding model for capturing temporal patterns of dynamic graphs, each of the graphs being an evolved representation of the dynamic network over time steps for a temporal lookback period and having a set of vertices and edges, the edges representing network links, the temporal patterns including nonlinear interactions between graph vertices at each time step and across multiple time steps, wherein the dynamic graph embedding model is configured as a neural network modeled as an autoencoder-recurrent neural network combination comprising:

an encoder comprising:

a first encoder part having fully connected dense layers to generate a low dimensional vector representation of a sparse neighborhood vector of the evolved dynamic graphs; and

a second encoder part having layers of long short term memory networks to generate the embedding from the low dimensional vector representation;

learning structural patterns in the dynamic network using the nonlinear layers;

learning, by the dynamic graph embedding model, a dynamic graph embedding for the temporal lookback period by optimizing a loss function that includes a weighting matrix for weighting reconstruction of observed edges higher than unobserved links, wherein the network embedding is a representation of each of the vertices in a series of low-dimensional vector space; and

predicting edges at a future time step based on parameters of the neural network, wherein the parameters are tuned by optimizing the loss function.

8. The system of claim 7 , wherein the predicted edges include at least one edge representation of an unobserved network link.

9. The system of claim 7 , wherein the steps further comprise:

receiving, by the dynamic graph embedding model, inputs as a set of previous dynamic graphs for the temporal lookback period, the lookback period being used to control a length of temporal patterns learned; and

decoding the dynamic graph embedding to produce the output of the dynamic graph embedding model, wherein the output is a predicted graph for the next time step.

10. The system of claim 7 , wherein the long short term memory networks are sparsely connected, and wherein l long short term memory networks are connected in a first layer of the neural network, where l is the number of previous graphs for the temporal lookback period.

11. The system of claim 7 , wherein the neural network further comprises

a decoder having fully connected layers for decoding the embedding to produce the output of the dynamic graph embedding model, wherein the output is a predicted graph for the next time step.