Node classification in dynamic networks using graph factorization
Methods and systems for detecting and responding to anomalous nodes in a network include inferring temporal factors, using a computer-implemented neural network, that represent changes in a network graph across time steps, with a temporal factor for each time step depending on a temporal factor for a previous time step. An invariant factor is inferred that represents information about the network graph that does not change across the time steps. The temporal factors and the invariant factor are combined into a combined temporal-invariant representation. It is determined that an unlabeled node is anomalous, based on the combined temporal-invariant representation. A security action is performed responsive to the determination that unlabeled node is anomalous.
1. A method for detecting and responding to anomalous nodes in a network, comprising:
inferring temporal factors, using a computer-implemented neural network, that represent changes in a network graph across a plurality of time steps, with a temporal factor for each time step depending on a temporal factor for a previous time step;
inferring an invariant factor that represents information about the network graph that does not change across the plurality of time steps;
combining the temporal factors and the invariant factor into a combined temporal-invariant representation;
determining that an unlabeled node is anomalous, based on the combined temporal-invariant representation; and
performing a security action responsive to the determination that unlabeled node is anomalous.
2. The method of claim 1 , further comprising aggregating neighbors of the unlabeled node in the network graph at each of the plurality of time steps.
3. The method of claim 2 , wherein aggregating neighbors of the unlabeled node comprises determining neighbors within k hops of the unlabeled node.
4. The method of claim 2 , wherein inferring the temporal factors and inferring the invariant factor each operate on the aggregated neighbors of each unlabeled node.
5. The method of claim 1 , wherein inferring the temporal factors uses a Markovian model that bases each temporal factor on a temporal factor at a previous time step.
6. The method of claim 5 , wherein a temporal factor at a first time step is based on a previous time step temporal factor of zero.
7. The method of claim 1 , further comprising combining the temporal factors together into a combined temporal representation using an attentive temporal aggregator that assigns different levels of attention to different sub-sequences.
8. The method of claim 1 , wherein the invariant factor is based on features of all of the plurality of time steps in the network graph.
9. The method of claim 1 , wherein the security action is selected from the group consisting of shutting down devices, stopping or restricting a type of network communication, enabling or disabling a connection between two devices, raising an alert to a system administrator, and changing a security policy level.
10. A system for detecting and responding to anomalous nodes in a network, comprising:
a hardware processor;
a memory, configured to store a temporal graph factorization network that is executed by the processor, wherein the temporal graph factorization network is configured to infer temporal factors that represent changes in a network graph across a plurality of time steps, with a temporal factor for each time step depending on a temporal factor for a previous time step, to infer an invariant factor that represents information about the network graph that does not change across the plurality of time steps, to combine the temporal factors and the invariant factor into a combined temporal-invariant representation, and to determine that an unlabeled node is anomalous, based on the combined temporal-invariant representation; and
a security console, configured to perform a security action responsive to the determination that unlabeled node is anomalous.
11. The system of claim 10 , wherein the temporal graph factorization network is further configured to aggregate neighbors of the unlabeled node in the network graph at each of the plurality of time steps.
12. The system of claim 11 , wherein the temporal graph factorization network is further configured to determine neighbors within k hops of the unlabeled node.
13. The system of claim 11 , wherein the temporal graph factorization network is further configured to infer the temporal factors and to infer the invariant factor based on the aggregated neighbors of each unlabeled node.
14. The system of claim 10 , wherein the temporal graph factorization network is further configured to infer the temporal factors using a Markovian model that bases each temporal factor on a temporal factor at a previous time step.
15. The system of claim 14 , wherein a temporal factor at a first time step is based on a previous time step temporal factor of zero.
16. The system of claim 10 , wherein the temporal graph factorization network is further configured to combine the temporal factors together into a combined temporal representation using an attentive temporal aggregator that assigns different levels of attention to different sub-sequences.
17. The system of claim 10 , wherein the invariant factor is based on features of all of the plurality of time steps in the network graph.
18. The method of claim 10 , wherein the security action is selected from the group consisting of shutting down devices, stopping or restricting a type of network communication, enabling or disabling a connection between two devices, raising an alert to a system administrator, and changing a security policy level.