Causal attention-based multi-stream RNN for computer system metric prediction and influential events identification based on metric and event logs
A method for system metric prediction and influential events identification by concurrently employing metric logs and event logs is presented. The method includes concurrently modeling multivariate metric series and individual events in event series by a multi-stream recurrent neural network (RNN) to improve prediction of future metrics, where the multi-stream RNN includes a series of RNNs, one RNN for each metric and one RNN for each event sequence and modeling causality relations between the multivariate metric series and the individual events in the event series by employing an attention mechanism to identify target events most responsible for fluctuations of one or more target metrics.
1. A method for system metric prediction and influential events identification by concurrently employing metric logs and event logs, the method comprising:
concurrently modeling multivariate metric series and individual events in event series by a multi-stream recurrent neural network (RNN) to improve prediction of future metrics, where the multi-stream RNN includes a series of RNNs, one RNN for each metric and one RNN for each event sequence; and
modeling causality relations between the multivariate metric series and the individual events in the event series by employing an attention mechanism to identify target events responsible for fluctuations of one or more target metrics, the attention mechanism including a causal attention mechanism and an input attention mechanism, the casual attention mechanism used over hidden states to compute a context vector and a distinct attention layer is used on each RNN to identify that different metrics draw on different information from past input windows, where weights of the causal attention mechanism are independent of each other, and the input attention mechanism is applied over context vectors of every RNN as a dot-product operation to aggregate the information from the metrics resulting in an aggregated vector forwarded to metric-specific linear layers to generate respective predicted future values for each metric.
2. The method of claim 1 , wherein the multivariate metric series include log data.
3. The method of claim 1 , wherein the multi-stream RNN is trained by using historical metrics data and historical event log data.
4. The method of claim 1 , wherein the target events responsible for the fluctuations of the one or more target metrics are ranked and displayed in a ranked list.
5. The method of claim 1 , wherein a network loss is computed as a total mean squared error (MSE).
6. The method of claim 1 , wherein the causal attention mechanism assigns weights to different events to model influences of the different events on a prediction of the one or more target metrics.
7. The method of claim 1 , wherein the input attention mechanism assigns weights to different metrics to model influences of each of the different metrics on a prediction of the one or more target metrics.
8. The method of claim 1 , wherein, after a training phase, a user specifies a prediction horizon during which values of the one or more target metrics are to be predicted.
9. The method of claim 8 , wherein log data is retrieved in a context time window of a predetermined length preceding a current time.
10. A non-transitory computer-readable storage medium comprising a computer-readable program for system metric prediction and influential events identification by concurrently employing metric logs and event logs, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:
concurrently modeling multivariate metric series and individual events in event series by a multi-stream recurrent neural network (RNN) to improve prediction of future metrics, where the multi-stream RNN includes a series of RNNs, one RNN for each metric and one RNN for each event sequence; and
modeling causality relations between the multivariate metric series and the individual events in the event series by employing an attention mechanism to identify target events responsible for fluctuations of one or more target metrics, the attention mechanism including a causal attention mechanism and an input attention mechanism, the casual attention mechanism used over hidden states to compute a context vector and a distinct attention layer is used on each RNN to identify that different metrics draw on different information from past input windows, where weights of the causal attention mechanism are independent of each other, and the input attention mechanism is applied over context vectors of every RNN as a dot-product operation to aggregate the information from the metrics resulting in an aggregated vector forwarded to metric-specific linear layers to generate respective predicted future values for each metric.
11. The non-transitory computer-readable storage medium of claim 10 , wherein the multivariate metric series include log data.
12. The non-transitory computer-readable storage medium of claim 10 , wherein the multi-stream RNN is trained by using historical metrics data and historical event log data.
13. The non-transitory computer-readable storage medium of claim 10 , wherein the target events responsible for the fluctuations of the one or more target metrics are ranked and displayed in a ranked list.
14. The non-transitory computer-readable storage medium of claim 10 , wherein a network loss is computed as a total mean squared error (MSE).
15. The non-transitory computer-readable storage medium of claim 10 , wherein the causal attention mechanism assigns weights to different events to model influences of the different events on a prediction of the one or more target metrics.
16. The non-transitory computer-readable storage medium of claim 10 , wherein the input attention mechanism assigns weights to different metrics to model influences of each of the different metrics on a prediction of the one or more target metrics.
17. The non-transitory computer-readable storage medium of claim 10 , wherein, after a training phase, a user specifies a prediction horizon during which values of the one or more target metrics are to be predicted.
18. The non-transitory computer-readable storage medium of claim 17 , wherein log data is retrieved in a context time window of a predetermined length preceding a current time.
19. A system for system metric prediction and influential events identification by concurrently employing metric logs and event logs, the system comprising:
a memory; and
one or more processors in communication with the memory configured to:
concurrently model multivariate metric series and individual events in event series by a multi-stream recurrent neural network (RNN) to improve prediction of future metrics, where the multi-stream RNN includes a series of RNNs, one RNN for each metric and one RNN for each event sequence; and
model causality relations between the multivariate metric series and the individual events in the event series by employing an attention mechanism to identify target events responsible for fluctuations of one or more target metrics, the attention mechanism including a causal attention mechanism and an input attention mechanism, the casual attention mechanism used over hidden states to compute a context vector and a distinct attention layer is used on each RNN to identity that different metrics draw on different information from past input windows, where weights of the causal attention mechanism are independent of each other, and the input attention mechanism is applied over context vectors of every RNN as a dot-product operation to aggregate the information from the metrics resulting in an aggregated vector forwarded to metric-specific linear layers to generate respective predicted future values for each metric.
20. The system of claim 19 , wherein the causal attention mechanism assigns weights to different events to model influences of the different events on a prediction of the one or more target metrics and the input attention mechanism assigns weights to different metrics to model influences of each of the different metrics on a prediction of the one or more target metrics.