System event analyzer and outlier visualization

Patent metadata

Field	Value
Publication number	US-10284453-B2
Application number	US-201514847656-A
Country	US
Kind code	B2
Filing date	Sep 8, 2015
Priority date	Sep 8, 2015
Publication date	May 7, 2019
Grant date	May 7, 2019

Abstract

An event analysis system receives events in a time-series from a set of monitored systems and identifies a set of alert threshold values for each of the types of events to identify outliers in the time-series at an evaluated time. Portions of historic event data is selected to identify windows of event data near the evaluated time at a set of seasonally-adjusted times to predict the value of the event type. The alert threshold value may also account for a prediction based on recent, higher-frequency events. Using the alert threshold values for a plurality of event types, the event data is compared with the alert threshold values to determine an alert level for the data. The event data types are also clustered and displayed with the alert levels to provide a visualization of the event data and identify outliers when the new event data is received.

First claim

What is claimed is:

1. A method comprising:
identifying a time-series sequence of event data for a type of event;
identifying an evaluation time at which to determine an outlier in the time-series sequence;
identifying a seasonal adjustment for the event data by: identifying a period over which the time-series sequence displays a similar pattern of event data;
selecting a set of windows of the time-series sequence, each window in the set of windows identifying a portion of time including the evaluation time as adjusted by the seasonal adjustment, wherein the set of windows does not include data immediately previous to the evaluation time;
determining a first predicted value of the time-series sequence at the evaluation time based on the set of windows of the time-series sequence;
determining an alert threshold value based on the first predicted value of the time-series sequence;
predicting a second value of the time-series sequence at the evaluation time based on a set of high-frequency data immediately previous to the evaluation time; and
modifying the alert threshold value based on a combination of the first predicted value of the time-series sequence and the second predicted value, wherein modifying the alert threshold value comprises multiplying the second predicted value by a scalar; and adding the scalar-multiplied second predicted value to the alert threshold value;
receiving subject event data of the time-series sequence corresponding to the evaluation time;
comparing, by the event analysis system, the subject event data to the alert threshold value; and
identifying an alert level when the subject event data exceeds the alert threshold value.

2. The method of claim 1 , wherein the evaluation time is the time at which the event data is received from a monitored system that generates the subject event data.

3. The method of claim 1 , wherein the time-series sequence of event data includes event data received from a plurality of monitored systems that generate events.

4. The method of claim 1 , wherein the seasonal adjustment is one week.

5. The method of claim 1 , further comprising:
determining a second alert threshold value by subtracting the scalar-multiplied second predicted value from the first predicted value of the time-series sequence;
comparing the subject event data to the second alert threshold value; and
identifying a second alert level when the subject event data is below the second alert threshold value.

6. The method of claim 1 , wherein the time-series sequence of event data comprises a set of data tiles aggregating the event data received at an update frequency, each data tile in the set of data tiles including a plurality of periods of the update frequency.

7. The method of claim 6 , wherein each window of the set of windows includes a portion of the set of data tiles, and further wherein predicting the value of the time-series sequence at the evaluation time is based on summary statistics of the portion of the set of data tiles for each window.

8. The method of claim 1 , further comprising:
predicting a third value of the time-series sequence at the evaluation time based on a set of recent data immediately previous to the evaluation time based on a gradient of the recent data; and
modifying the alert threshold value based on the third predicted value.

9. A non-transitory computer-readable medium having instructions stored thereon, the instructions executable by a processor and when executed causing the processor to:
identify a time-series sequence of event data for a type of event;
identify an evaluation time at which to determine an outlier in the time-series sequence;
identify a seasonal adjustment for the event data identifying a period over which the time-series sequence displays a similar pattern of event data;
select a set of windows of the time-series sequence, each window in the set of windows identifying a portion of time including the evaluation time as adjusted by the seasonal adjustment, wherein the set of windows does not include data immediately previous to the evaluation time;
determine a first predicted value of the time-series sequence at the evaluation time based on the set of windows of the time-series sequence;
determine an alert threshold value based on the first predicted value of the time-series sequence;
predict a second value of the time-series sequence at the evaluation time based on a set of high-frequency data immediately previous to the evaluation time; and
modify the alert threshold value based on a combination of the first predicted value of the time-series sequence and the second predicted value, wherein modifying the alert threshold value comprises multiplying the second predicted value by a scalar; and adding the scalar-multiplied second predicted value to the alert threshold value;
receive subject event data of the time-series sequence corresponding to the evaluation time;
compare, by the event analysis system, the subject event data to the alert threshold value; and
identify an alert level when the subject event data exceeds the alert threshold value.

10. The computer-readable medium of claim 9 , wherein the evaluation time is the time at which the event data is received from a monitored system that generates the subject event data.

11. The computer-readable medium of claim 9 , wherein the time-series sequence of event data includes event data received from a plurality of monitored systems that generate events.

12. The computer-readable medium of claim 9 , wherein the seasonal adjustment is one week.

13. The computer-readable medium of claim 9 , the instructions further causing the processor to:
determine a second alert threshold value by subtracting the scalar-multiplied second predicted value from the first predicted value of the time-series sequence;
compare the subject event data to the second alert threshold value; and
identify a second alert level when the subject event data is below the second alert threshold value.

14. The computer-readable medium of claim 9 , wherein the time-series sequence of event data comprises a set of data tiles aggregating the event data received at an update frequency, each data tile in the set of data tiles including a plurality of periods of the update frequency.

15. The computer-readable medium of claim 14 , wherein each window of the set of windows includes a portion of the set of data tiles, and further wherein predicting the value of the time-series sequence at the evaluation time is based on summary statistics of the portion of the set of data tiles for each window.

16. The computer-readable medium of claim 9 , the instructions further causing the processor to:
predict a third value of the time-series sequence at the evaluation time based on a set of recent data immediately previous to the evaluation time based on a gradient of the recent data; and
modify the alert threshold value based on the third predicted value.

Assignees

• Uber Technologies Inc

Inventors

• Bell Franziska
• Purdy David
• Korsos Laszlo
• He Shan

Classifications

• H04L41/069

  using logs of notifications; Post-processing of notifications · CPC title

• H04L41/064

  involving time analysis · CPC title

• H04L43/16 Primary

  Threshold monitoring · CPC title

• H04L41/0622

  based on time · CPC title