Name: egads
Owner: Yahoo Inc.
Description: Extendible Generic Anomaly Detection System
Created: 2015-05-06 17:47:52.0
Updated: 2018-05-24 13:02:57.0
Pushed: 2018-01-02 23:12:13.0
Size: 1342
Language: Java
GitHub Committers
User | Most Recent Commit | # Commits |
---|
Other Committers
User | Most Recent Commit | # Commits |
---|
EGADS (Extensible Generic Anomaly Detection System) is an open-source Java package to automatically detect anomalies in large scale time-series data. EGADS is meant to be a library that contains a number of anomaly detection techniques applicable to many use-cases in a single package with the only dependency being Java. EGADS works by first building a time-series model which is used to compute the expected value at time t. Then a number of errors E are computed by comparing the expected value with the actual value at time t. EGADS automatically determines thresholds on E and outputs the most probable anomalies. EGADS library can be used in a wide variety of contexts to detect outliers and change points in time-series that can have a various seasonal, trend and noise components.
EGADS was designed as a self contained library that has a collection of time-series and anomaly detection models that are applicable to a wide-range of use cases. To compile the library into a single jar, clone the repo and type the following:
clean compile assembly:single
You may have to set your JAVA_HOME
variable to the appropriate JVM. To do this run:
rt JAVA_HOME=/usr/lib/jvm/{JVM directory for desired version}
To run a simple example type:
-Dlog4j.configurationFile=src/test/resources/log4j2.xml -cp target/egads-*-jar-with-dependencies.jar com.yahoo.egads.Egads src/test/resources/sample_config.ini src/test/resources/sample_input.csv
which produces the following picture (Note that you can enable this UI by setting OUTPUT
config key to GUI
in sample_config.ini
).
One can also specify config parameters on a command line. For example to do anomaly detection using Olympic Scoring as a time-series model and a density based method as an anomaly detection model use the following.
-Dlog4j.configurationFile=src/test/resources/log4j2.xml -cp target/egads-*-jar-with-dependencies.jar com.yahoo.egads.Egads "MAX_ANOMALY_TIME_AGO:999999999;AGGREGATION:1;OP_TYPE:DETECT_ANOMALY;TS_MODEL:OlympicModel;AD_MODEL:ExtremeLowDensityModel;INPUT:CSV;OUTPUT:STD_OUT;BASE_WINDOWS:168;PERIOD:-1;NUM_WEEKS:3;NUM_TO_DROP:0;DYNAMIC_PARAMETERS:0;TIME_SHIFTS:0" src/test/resources/sample_input.csv
To run anomaly detection using no time-series model with an auto static threshold for anomaly detection, use the following:
-Dlog4j.configurationFile=src/test/resources/log4j2.xml -cp target/egads-*-jar-with-dependencies.jar com.yahoo.egads.Egads "MAX_ANOMALY_TIME_AGO:999999999;AGGREGATION:1;OP_TYPE:DETECT_ANOMALY;TS_MODEL:NullModel;AD_MODEL:SimpleThresholdModel;SIMPLE_THRESHOLD_TYPE:AdaptiveMaxMinSigmaSensitivity;INPUT:CSV;OUTPUT:STD_OUT;AUTO_SENSITIVITY_ANOMALY_PCNT:0.2;AUTO_SENSITIVITY_SD:2.0" src/test/resources/sample_input.csv
While rapid advances in computing hardware and software have led to powerful applications, still hundreds of software bugs and hardware failures continue to happen in a large cluster compromising user experience and subsequently revenue. Non-stop systems have a strict uptime requirement and continuous monitoring of these systems is critical. From the data analysis point of view, this means non-stop monitoring of large volume of time-series data in order to detect potential faults or anomalies. Due to the large scale of the problem, human monitoring of this data is practically infeasible which leads us to automated anomaly detection. An anomaly, or an outlier, is a data point which is significantly different from the rest of the data. Generally, the data in most applications is created by one or more generating processes that reflect the functionality of a system.
When the underlying generating process behaves in an unusual way, it creates outliers. Fast and efficient identification of these outliers is useful for many applications including: intrusion detection, credit card fraud, sensor events, medical diagnoses, law enforcement and others. Current approaches in automated anomaly detection suffer from a large number of false positives which prohibit the usefulness of these systems in practice. Use-case, or category specific, anomaly detection models may enjoy a low false positive rate for a specific application, but when the characteristics of the time-series change, these techniques perform poorly without proper retraining.
EGADS (Extensible Generic Anomaly Detection System) enables the accurate and scalable detection of time-series anomalies. EGADS separates forecasting and anomaly detection two separate components which allows the person to add her own models into any of the components.
The EGADS framework consists of two main components: the time-series modeling module (TMM), the anomaly detection module (ADM). Given a time-series the TMM component models the time-series producing an expected value later consumed by the ADM that computes anomaly scores. EGADS was built as a framework to be easily integrated into an existing monitoring infrastructure. At Yahoo, our internal Yahoo Monitoring Service (YMS) processes millions of data-points every second. Therefore, having a scalable, accurate and automated anomaly detection for YMS is critical. For this reason, EGADS can be compiled into a single light-weight jar and deployed easily at scale.
The TMM and ADM can be found under main/java/com/yahoo/egads/models
.
The example of the models supported by TMM and ADM can be found in in the two table below. We expect this collection of models to grow as more contribution is put forward by the community.
Below are the various configuration parameters supported by EGADS.
ly show anomalies no older than this.
this is set to 0, then only output an anomaly
it occurs on the last time-stamp.
ANOMALY_TIME_AGO 99999
notes how much should the time-series be aggregated by.
set to 1 or less, this setting is ignored.
EGATION 1
_TYPE specifies the operation type.
tions: DETECT_ANOMALY,
UPDATE_MODEL,
TRANSFORM_INPUT
YPE DETECT_ANOMALY
_MODEL specifies the time-series
del type.
tions: AutoForecastModel
DoubleExponentialSmoothingModel
MovingAverageModel
MultipleLinearRegressionModel
NaiveForecastingModel
OlympicModel
PolynomialRegressionModel
RegressionModel
SimpleExponentialSmoothingModel
TripleExponentialSmoothingModel
WeightedMovingAverageModel
SpectralSmoother
NullModel
ODEL OlympicModel
_MODEL specifies the anomaly-detection
del type.
tions: ExtremeLowDensityModel
AdaptiveKernelDensityChangePointDetector
KSigmaModel
NaiveModel
DBScanModel
SimpleThresholdModel
ODEL ExtremeLowDensityModel
pe of the simple threshold model.
tions: AdaptiveMaxMinSigmaSensitivity
AdaptiveKSigmaSensitivity
MPLE_THRESHOLD_TYPE
ecifies the input src.
tions: STDIN
CSV
T CSV
ecifies the output src.
tions: STD_OUT,
ANOMALY_DB
GUI
PLOT
UT STD_OUT
RESHOLD specifies the threshold for the
omaly detection model.
mment to auto-detect all thresholds.
tions: mapee,mae,smape,mape,mase.
RESHOLD mape#10,mase#15
#################################
Olympic Forecast Model Config ###
#################################
e possible time-shifts for Olympic Scoring.
_SHIFTS 0,1
e possible base windows for Olympic Scoring.
_WINDOWS 24,168
riod specifies the periodicity of the
me-series (e.g., the difference between successive time-stamps).
tions: (numeric)
0 - auto detect.
-1 - disable.
OD -1
M_WEEKS specifies the number of weeks
use in OlympicScoring.
WEEKS 8
M_TO_DROP specifies the number of
ghest and lowest points to drop.
TO_DROP 0
dynamic parameters is set to 1, then
ADS will dynamically vary parameters (NUM_WEEKS)
produce the best fit.
MIC_PARAMETERS 0
###############################################
ExtremeLowDensityModel & DBScanModel Config ###
###############################################
notes the expected % of anomalies
your data.
_SENSITIVITY_ANOMALY_PCNT 0.01
fers to the cluster standard deviation.
_SENSITIVITY_SD 3.0
########################
NaiveModel Config ###
########################
ndow size where the spike is to be found.
OW_SIZE 0.1
###################################################
AdaptiveKernelDensityChangePointDetector Config ###
###################################################
ange point detection parameters
WINDOW_SIZE 48
_WINDOW_SIZE 48
IDENCE 0.8
###########################
SpectralSmoother Config ###
###########################
NDOW_SIZE should be greater than the size of longest important seasonality.
default it is set to 192 = 8 * 24 which is worth of 8 days (> 1 week) for hourly time-series.
OW_SIZE 192
LTERING_METHOD specifies the filtering method for Spectral Smoothing
tions: GAP_RATIO (Recommended: FILTERING_PARAM = 0.01)
EIGEN_RATIO (Recommended: FILTERING_PARAM = 0.1)
EXPLICIT (Recommended: FILTERING_PARAM = 10)
K_GAP (Recommended: FILTERING_PARAM = 8)
VARIANCE (Recommended: FILTERING_PARAM = 0.99)
SMOOTHNESS (Recommended: FILTERING_PARAM = 0.97)
ERING_METHOD GAP_RATIO
ERING_PARAM 0.01
mvn package
and adding additional tests.We actively welcome contributions. If you don't know where to start, try checking out the issue list and fixing up the place. Or, you can add a model - a goal of this project is to have a robust, lightweight and dependency-free set of models to choose from that are ready to be deployed in production.
Generic and Scalable Framework for Automated Time-series Anomaly Detection by Nikolay Laptev, Saeed Amizadeh, Ian Flint , KDD 2015 (August 10, 2015)
If you use EGADS in your projects, please cite: Generic and Scalable Framework for Automated Time-series Anomaly Detection by Nikolay Laptev, Saeed Amizadeh, Ian Flint , KDD 2015
BibTeX:
roceedings{laptev2015generic,
title={Generic and Scalable Framework for Automated Time-series Anomaly Detection},
author={Laptev, Nikolay and Amizadeh, Saeed and Flint, Ian},
booktitle={Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
pages={1939--1947},
year={2015},
organization={ACM}
Code licensed under the GPL License. See LICENSE file for terms.