通過時間序列分析為人工智能賦能.pptx

1、Empowering AI Through Time Series AnalysisSalochina Oad,PhD Usxpress.IncSAgenda Why do we need time series analysis?What are the key expectations before starting forecasting?How to deal with variation of time series data?How can time series empower AI?The value of improving the performance of AI Mod

2、els Challenges in detecting and preventing fraudulent activitiesModelsBehavioral AnalyticsReal-time monitoring&alerts Model training&evaluationEfficiency&performanceTime seriesTrend/Seasonal Residuals SignalNoiseTime seriesMultiplicative TSAdditive TSTypes of time series data Time seriesDateCompanyC

3、losetCloset-112-27-2022Google87.9312-28-2022Google86.4687.9312-29-2022Google88.9586.4612-30-2022Google88.7388.9501-03-2023Google89.7088.7301-04-2023Google88.7189.7001-05-2023Google86.7788.7101-06-2023Google88.1686.77Pooled/Panel DateCompanyClosetCloset-112-27-2022Google87.9312-28-2022Google86.4687.9

4、312-29-2022Google88.9586.4612-30-2022Google88.7388.9512-27-2022AAPL89.7012-28-2022AAPL88.7189.7012-29-2022AAPL86.7788.7112-30-2022AAPL88.1686.77Time series use casesDemand forecast How many order a trucking company got this week-Cost planningSales forecastHow much revenue was generated serving speci

5、fic customer.-Financial outcomesRoadmap-Forecasting projectDetermine TargetHorizon of the forecastGather dataDevelop a modelDeploy to productionMonitor Time series pipelineETLEDAPreprocessingForecastingDiagnostics Read Impute Sampling Visualization StatisticsDetrendDestationariesFeature-EngineeringP

6、arametric-ModelsNonparametric-Models Autocorrelation Stationarity Normality ResidualsEDA-Components of a time seriesTrendSeasonal componentResiduals Other elements Holidays AnomaliesAirline passenger EDA-Components of a time seriesTrendSeasonal componentResiduals Other elements Holidays AnomaliesTre

7、nd and Seasonality Differencing LogPreprocessing-anomalies SLT plot Check standard deviation Back fill using meanPreprocessing-Special events Create dummy featureJuly 4July 4July 4Preprocessing-Feature engineeringImputation Backfill InterpolationTransformation Log Trend adj-differencingTemporal feat

8、ure Time-Based Features Lag features using shift Rolling Window Features Choosing the best technique for modelingNon-Parametric methodsParametric methodsForecasting-Parametric MethodsAssumption:stationarity Advantages Parametric methods are simpler,easier to understand and interpret results Requires

9、 smaller amount of data Computationally inexpensiveLimitations Inability to capture subtle patterns in time series dataARMAARIMA(p,d,q)SARIMA(p,d,q)(P,D,Q)mSteps to identify a modelStocks Gather dataStationary ACFAutocorrelationApply transformationRandom walkNoYesYesNoNot a random walkApplyNave fore

10、castingPrevious seasonal pointGather dataStationary ACFAutocorrelationNoYesNoNoApply transformationRandom walkYesAutocorrelation coef.abruptly non sig after lag qMAYesNoMoving average(q)Random walkNot MAAuto regressive(p)Gather dataStationary ACFAutocorrelationNoYesNoNot ARMA NoApply transformationR

11、andom walkYesAutocorrelation coef.abruptly non sig after lag qMAYesNoPACFNon sig.coef abruptlyNoARYesMoving average(q)Random walk ARMA(p,q)ARIMA(p,d,q)AR:auto regressive I:integration MA:moving averageARIMA(1,0,1)ARIMA(1,1,1)ARIMA(1,0,0)ACF and PACF plots show a decayingforecasting-Non-Parametric Me

12、thodsAdvantages High performance models More data neededLimitations Computationally expensiveNeural networks(LSTM)CART Regression tressfbprophet Ensemble-XgboostExtreme Gradient Boosting combines the predictions of multiple weak decision trees to create a strong predictive model Scalability&Speed Ea

13、sy to tune Handling Missing Values&Outliers Feature ImportanceFeature creationKmeans Segmented modelingDateStoreSalesStore Segment6/11/2014Savannah155476Super Store6/12/2014Savannah149291Super Store6/13/2014Savannah174930Super Store6/14/2014Savannah192542Super Store6/11/2014Portland43070Medium Store

14、6/12/2014Portland41014Medium Store6/13/2014Portland40263Medium Store6/14/2014Portland40553Medium Store6/11/2014Columbus23588Small Store6/12/2014Columbus25552Small Store6/13/2014Columbus22583Small Store6/14/2014Columbus27557Small StoreCluster1Cluster2Cluster3Data leakage Long short-term memory(LSTM)P

15、redominantly used to learn,process,and classify sequential data Hidden stateVanishing gradient problem in RNN241219Sequential data split-data windowingTraining setLSTM:Single-step modelLSTM:Multi-step modelFbprophet:automated modelAn additive model where non-linear trends are fit with yearly,weekly,

16、and daily seasonality,plus holiday effects.Change in trend Multiple seasonalityOptimization Grid search Diagnostic tools-residual analysis,Akaikes information criterion(AIC)Cross-validation-time based,Rolling window,Walk forward validation Xgboot:param_grid=learning_rate:0.01,0.1,0.2,max_depth:3,5,7

17、,subsample:0.8,0.9,1.0 Arima:param_space=dict(p=range(0,30),d=range(0,30),q=range(0,30)Prophet:param_grid=changepoint_prior_scale:0.001,0.01,0.1,0.5,seasonality_prior_scale:0.01,0.1,1.0,10.0,seasonality_mode:multiplicative,additive,growth:linear,logistic,yearly_seasonality:5,10,20,40,weekly_seasonal

18、ity:5,10,20,40,daily_seasonality:5,10,20,40,Automated platform for AIStore sales Store sales DateStoreSalesStore Segment6/11/2014Savannah155476Super Store6/12/2014Savannah149291Super Store6/13/2014Savannah174930Super Store6/14/2014Savannah192542Super Store6/11/2014Portland43070Medium Store6/12/2014P

19、ortland41014Medium Store6/13/2014Portland40263Medium Store6/14/2014Portland40553Medium Store6/11/2014Columbus23588Small Store6/12/2014Columbus25552Small Store6/13/2014Columbus22583Small Store6/14/2014Columbus27557Small StoreFeature Feature engineering and patterns and patternsConclusion Data quality enhancements can aid AI in learning meaningful patterns over time.Robustness to variability ensures AI models can effectively handle fluctuations in data.Continuous learning frameworks from new data streams help in refining models to reflect evolving trends.Thank YouThank YouSalochina Oad,PhD