How to make blazing fast time series prediction ? A popular framework in ARIMA. But when the time series is periodic this algorithm is very slow. I explain here how to challenge this problem.
3. 3
We use ai to constantly
monitor and correlate
business performance,
providing real time alerts and
forecasts, reducing time to
detection and resolution,
dramatically impacting
bottom lines.
4. 44
THE COMPREHENSIVE SOLUTION
Automatically analyzes all data continuously, interacts with us only when relevant
CORRELATE ACT
● All the data
● Cross silo
● Continuous learning of 100%
● Adjusts to changes
● Root cause guidance
● Multiple correlation algorithms
● Realtime
● Significance learning
ANALYZECOLLECT
USER
PRODUCT
SYS. HEALTH
INFRA
14. 1414
NOT ALL METRICS ARE CREATED EQUAL
SMOOTH IRREGULAR SAMPLING
MULTIMODAL
SPARSEDISCRETE
“STEP”
15. 1515
REAL TIME ANOMALY DETECTION STRATEGY
Anomaly =
Too far from the
expected behavior
Prediction range
16. 1616
Business problem Mathematical / algorithms problem
REAL – TIME ANOMALY
DETECTION PROBLEM
Learning: Time series
modeling
Inference: Real time
prediction calculation
REAL TIME ANOMALY DETECTION STRATEGY
CONDITIONAL PREDICTION PROBLEM
17. 1717
TIME SERIES MODELING
● The modeling process in
finding a function, f, that
explain the present based on
the past: self history and
prediction error
● The model f can be linear,
neural networks, LSTM ...
19. 1919
WHY
ARIMA?
(S)ARIMA is the best linear time series model,
consequence of the Wold theorem. It has a very
strong theoretical basis
One model for season, trend, level and noise,
no need to model each part separately.
One to rule them all
Winner model in lots of prediction competitions
22. 22
LEARNING ARIMA : THE BIRD EYE VIEW STRATEGY
Maximum likelihood estimation (MLE)
Likelihood :
The loss function
Maximum
finding
Newton algorithmKalman filtering Need to compute
the gradient of the
Kalman filter !
25. 2525
• The system find the daily pattern (and not the 3h sub-pattern)
• The conditional prediction begins then to be very accurate.
• We are now able to catch even the smallest anomaly.
SEASON IS A KEY COMPONENT
26. 2626
ANODOT’s scale
5.8 BILLION
DATA POINTS PER DAY
120 MILLION
UNIQUE METRICS
240 MILLION
MODELS
500 MILLION
CORRELATIONS
14 MILLION
SEASONAL METRICS
30 TYPES
OF LEARNING ALGORITHMS
27. 2727
Probably the fastest
implementation
available in the market
YOU SAID SCALE ?
ANODOT’S PERFORMANCE BREAKTHROUGHS
• Likelihood = O(Ns6)
• BFGS
• Fast Lyapunov equation solver ► Likelihood = O(Ns3)
• Sparse Kalman Filter ► Likelihood = O(Ns2)
• T ~ 10 min
• Fast Hessian approximation, use Newton algorithm
• T ~ 10 sec
• Kalman convergence in likelihood and Gradient ► Likelihood =
O(Ns1.1)
• T ~ 750 ms
• Down sampled SARIMA estimation, Uncoupled estimation
• T ~ 80 ms, Independent of the length of the season !!! O(Ns0)
> 480000x
(4.8 E5)
28. 2828
SARIMA ESTIMATION : WHERE IS THE TRICK ?
▪ We developed a divide and conquer strategy
Estimation of the slow (seasonal) dynamic
Downsampling
the series
Estimate the
downsampled model
Resample the model1.a. 1.b. 1.c.
1.
Estimation of the fast (non seasonal) dynamic2.
Model
merging
(Convolution)
3. Estimate the
residual
model
4.
Pending US patent
32. 3232
SUMMARY
● Anomaly detection require time series forecasting
● ARIMA is best in class technique for time series analysis
● Seasonal ARIMA is extremely challenging at scale
● Anodot break the state of the art for learning seasonal model
● As a result, the anomaly detection system sensitivity has been
drastically improved