SAaaaS To Oracle Advanced Analytics Migration At Zagrebacka Banka . PDF Free Download

1y ago

73 Views

1 Downloads

6.50 MB

63 Pages

Report/dmca

Download PDF

Transcription

SAaaaS to Oracle AdvancedAnalytics Migration atZagrebacka Banka,UniCredit Group [CAS6499]Making Big Data Analytics SimpleCharlie Berger, MS Engineering, MBASr. Director Product Management, Advanced Analytics andMachine harlieDataMineSinisa Behin, Head of IT CRM function,Zagrebacka banka - UniCredit GroupCopyright 2016, Oracle and/or its affiliates. All rights reserved.

Safe Harbor StatementThe following is intended to outline our general product direction. It is intended forinformation purposes only, and may not be incorporated into any contract. It is not acommitment to deliver any material, code, or functionality, and should not be relied uponin making purchasing decisions. The development, release, and timing of any features orfunctionality described for Oracle’s products remains at the sole discretion of Oracle.Copyright 2016, Oracle and/or its affiliates. All rights reserved. 2

SAS to Oracle Advanced Analytics Migration at ZagrebackaBanka, UniCredit Group [CAS6499] Agenda1Oracle Advanced Analytics Quick Overview, Brief Demo(15 mins)2SAS to Oracle Advanced Analytics Migration atZagrebacka Banka, UniCredit Group(30 mins)Copyright 2016, Oracle and/or its affiliates. All rights reserved.

Machine Learning/Analytics Data Warehouse Hadoop Platform Sprawl– More Duplicated Data– More Data Movement Latency– More Security challenges– More Duplicated Storage– More Duplicated Backups– More Duplicated Systems– More Space and PowerCopyright 2016, Oracle and/or its affiliates. All rights reserved.

VisionAdvanced Analytics Big Data Machine Learning/Analytics Platform for the Era of BigData and Cloud–Make Big Data ML/Analytics Model Discovery Simple Any data size, on any computer infrastructure—on-premise and/or cloud Any variety of data (structured, unstructured, transactional, geospatial), in anycombination–Make Big Data ML/Analytics Model Deployment Simple As a service, as a platform, as an application On-premise and/or cloudCopyright 2016, Oracle and/or its affiliates. All rights reserved. Oracle Cloud Advanced AnalyticsOracle Database 12c

Traditional vs. Oracle Machine Learning/Predictive Analtyics Traditional— “Move the data”—“Don’t move the data!”Copyright 2016, Oracle and/or its affiliates. All rights reserved. Oracle Confidential – Internal/Restricted/Highly Restricted6

Traditional vs. Oracle Machine Learning/Predictive Analtyics Traditional— “Move the data”— “Move the algorithms”Simpler, Smarter Data Management Analytics / Machine Learning ArchitectureCopyright 2016, Oracle and/or its affiliates. All rights reserved. Oracle Confidential – Internal/Restricted/Highly Restricted7

Oracle’s Advanced Analytics (Machine Learning Platform)Multiple interfaces across platforms — SQL, R, GUI, Dashboards, AppsInformation ProducersR programmersData & Business AnalystsUsersR ClientSQL Developer/Oracle Data MinerInformation ConsumersBusiness Analysts/MgrsDomain End el,distributedalgorithmsOracle Database Enterprise EditionOracle Advanced Analytics - Database OptionSQL Data Mining, ML & Analytic Functions R Integrationfor Scalable, Distributed, Parallel in-DB ML ExecutionOracle CloudAdvanced AnalyticsCopyright 2016, Oracle and/or its affiliates. All rights reserved. Oracle Database 12c

What is Machine Learning, Data Mining & PredictiveAnalytics?Automatically sifting through large amounts of data tocreate models that find previously hidden patterns,discover valuable new insights and make predictions Identify most important factor (Attribute Importance) Predict customer behavior (Classification) Predict or estimate a value (Regression) Find profiles of targeted people or items (Decision Trees) Segment a population (Clustering) Find fraudulent or “rare events” (Anomaly Detection) Determine co-occurring items in a “baskets” (Associations)Copyright 2016, Oracle and/or its affiliates. All rights reserved. A1 A2 A3 A4 A5 A6 A7

Oracle Advanced Analytics DB OptionIn-Database Machine Learning Algorithms*—SQL &Classification Decision Tree Logistic Regression (GLM) Naïve Bayes Support Vector Machine (SVM) Random ForestRegression Multiple Regression (GLM) Support Vector Machine (SVM) Stepwise Linear Regression Linear Model Generalized Linear Model Multi-Layer Neural NetworksAnomaly Detection 1-Class Support Vector MachineAdvanced Analytics& GUI AccessClusteringPredictive Queries Hierarchical k-Means Orthogonal Partitioning Clustering Expectation-MaximizationAttribute Importance Minimum Description Length Unsupervised pair-wise KL div.A1 A2 A3 A4 A5A6 A7Market Basket Analysis Apriori – Association RulesText Mining All OAA/ODM SQL ML support Explicit Semantic Analysis Clustering Regression Anomaly Detection Feature ExtractionFeature Extraction & Creation Nonnegative Matrix Factorization Principal Component Analysis Singular Value DecompositionTime Series Single & Double Exp. SmoothingOpen Source R Algorithms Ability to run any R package(9,000 )via Embedded R mode Ability to Mine Unstructured, Structured & Transactional data Partitioned ModelsCopyright 2016, Oracle and/or its affiliates. All rights reserved.

Oracle’s Advanced AnalyticsAdvanced AnalyticsFastest Way to Deliver Scalable Enterprise-wide ML/Predictive AnalyticsTraditional AnalyticsMajor BenefitsOracle Advanced AnalyticsData ImportData MiningModel “Scoring” Data remains in Database & Hadoop Model building and scoring occur in-database Use R packages with data-parallel invocations Leverage investment in Oracle ITData Prep. &TransformationavingsData MiningModel Building Eliminate data duplication Eliminate separate analytical serversData Prep &Transformation Deliver enterprise-wide applications GUI for ML/Predictive Analytics & code gen R interface leverages database as HPC engineData ExtractionModel “Scoring”Embedded Data PrepModel BuildingData PreparationHours, Days or WeeksCopyright 2016, Oracle and/or its affiliates. All rights reserved. Secs, Mins or Hours

Oracle Advanced Analytics 12.2Prelim/UnofficialModel Build Time PerformanceOAA 12.2 AlgorithmsRows (Ms)T7-4 (Sparc & Solaris)X5-4 (Intel and Linux)Model Build Time (Secs / Degree of Parallelism)Wow! That’s Fast!Attributes Importance64028s / 51244s / 72K Means ClusteringExpectation Maximization640159161s / 256455s / 512268s / 144588s / 144Naive Bayes ClassificationGLM ClassificationGLM Regression32064064017s / 256154s / 51255s / 51223s / 72363s / 14493s / 144Support Vector Machine (IPM solver)Support Vector Machine (SGD solver)640640404s / 51284s / 2561411s / 144188s / 72The way to read their results is that they compare 2 chips: X5 (Intel and Linux) and T7 (Sparc and Solaris). They are measuring scalability (time in seconds) with increasedegree of parallelism (dop). The data also has high 9K miningCopyright 2016, Oracleand/orwhichits affiliates.All rights attributes (when algorithms require explosion). Thereare no comparisons to 12.1 and it is fair to say that the 12.1 algorithms could not run on data of this size.

You Can Think of Oracle’s Advanced Analytics Like This Traditional SQLOracle Advanced Analytics - SQL &– “Human-driven” queries– Domain expertise– Any “rules” must be defined andmanagedSQL Queries– SELECT– DISTINCT– Automated knowledge discovery, modelbuilding and deployment– Domain expertise to assemble the “right”data to mine/analyze Analytical SQL “Verbs”– PREDICT– DETECT– AGGREGATE– CLUSTER– WHERE– CLASSIFY– AND OR– REGRESS– GROUP BY– PROFILE– ORDER BY– IDENTIFY FACTORS– RANK– ASSOCIATECopyright 2016, Oracle and/or its affiliates. All rights reserved. Advanced Analytics

Oracle TextNative Capability of every Oracle Database Oracle Text uses standard SQL to index, search, and analyze text anddocuments stored in the Oracle database, in files, and on the web. Oracle Text supports multiple languages and uses advanced relevanceranking technology to improve search quality. Oracle Advanced Analytics uses Oracle Text to pre-process (“tokenize”)unstructured data for the OAA SQL data mining functionsCopyright 2016, Oracle and/or its affiliates. All rights reserved.

Rapidly Build, Evaluate & Deploy Analytical MethodologiesLeveraging a Variety of Data Sources and TypesTransactionalPOS dataSQL Joins and arbitrary SQLtransforms & queries – power of SQLModelingApproachesInline predictivemodel to augmentinput dataAdvanced AnalyticsConsider: DemographicsGenerates SQL scripts Past purchasesand workflow API for Recent purchasesdeployment Comments&tweetsUnstructured dataalso mined byalgorithmsCopyright 2016, Oracle and/or its affiliates. All rights reserved.

More Data Variety—Better Predictive ModelsEngineered Features – Derived attributes/variable thatreflect domain knowledge—key to best models100% Increasing sources ofrelevant data can boostmodel accuracyNaïve Guess orRandomRespondersModel with “Big Data” andhundreds -- thousands of inputvariables including: Demographic data Purchase POS transactionaldata “Unstructured data”, text &comments Spatial location data Long term vs. recent historicalbehavior Web visits Sensor data etc.100%Model with 20 variablesModel with 75 variablesModel with 250 variables0%Population SizeCopyright 2016, Oracle and/or its affiliates. All rights reserved.

Oracle Data Miner GUIEasy to Use for “Citizen Data Scientist” Easy to use todefine analyticalmethodologiesthat can beshared SQL DeveloperExtension Workflow APIand generatesSQL code forimmediatedeploymentCopyright 2016, Oracle and/or its affiliates. All rights reserved. 19

OAA Oracle Data Mining SQL Sample ProgramsSample program for the DBMS DATA MINING package – Decision Tree-- Given demographic data about a set of customers, predict the-- customer response to an affinity card program using a classifier-- based on Decision Trees algorithm.-- CREATE A NEW MODEL--- Build a DT modelBEGINDBMS DATA MINING.CREATE MODEL(model name 'DT SH Clas sample',mining function dbms data mining.classification,data table name 'mining data build v',case id column name 'cust id',target column name 'affinity card',settings table name 'dt sh sample settings');END;/SELECT T.cust id, S.prediction, S.probability, S.costFROM (SELECT cust id,PREDICTION SET(dt sh clas sample COST MODEL USING *) psetFROM mining data apply vWHERE cust id 100011) T,TABLE(T.pset) SORDER BY cust id, S.prediction;Copyright 2016, Oracle and/or its affiliates. All rights reserved.

R—Widely PopularR is a statistics language similar to Base SAS or SPSS statisticsR environment Strengths– Powerful & Extensible– Graphical & Extensive statistics– Free—open source Challenges– Memory constrained– Single threaded– Outer loop—slows down process– Not industrial strengthCopyright 2016, Oracle and/or its affiliates. All rights reserved.

R: Transparency through function overloadingInvoke in-database Data Mining model (Support Vector Machine)Oracle PL/SQLOracle Advanced AnalyticsBEGINORE Client Packages svm mod - ore.odmSVM(BUY INCOME YRS CUST MARITAL STATUS,data CUST,"classification", kernel "linear")Transparency Layer summary(svm mod)Call:ore.odmSVM(formula BUY INCOME YRS CUST MARITAL STATUS, data CUST,type "classification", kernel.function "linear")Settings:valueprep.autoonactive.learning al.enablecomplexity.factor oefficients:classvariable valueestimate10INCOME5.204561e-0520 MARITAL STATUSM -4.531359e-0530 MARITAL STATUSS 4.531359e-0540YRS .032340e-0571 MARITAL STATUSM 2.636552e-0681 MARITAL STATUSS -2.636555e-0691YRS CUST-1.588211e-04101(Intercept)-9.999324e-01DBMS DATA MINING.CREATE MODEL(model name ’SVM MOD’,mining function dbms data mining.classification.Oracle DatabaseIn-dbMiningModelCUSTDatabase ServerCopyright 2016, Oracle and/or its affiliates. All rights reserved.

Oracle R Advanced Analytics for HadoopAdvanced AnalyticsUsing Hadoop and HIVE Integration, plus R Engine and Open-Source R PackagesHadoop Clusterwith Oracle R Advanced Analytics for Hadoop (ORAAH)R interface to HQL Basic Statistics,Data Prep, Joins and View creationParallel, distributed algorithms:MLP Neural Nets*, GLM*, LM, PCA,k-Means, NMF, LMFR ClientHQLR AnalyticsOracle R AdvancedAnalytics for HadoopOracle Databasewith Advanced Analytics option* Spark-Caching enabledSQL ClientUse of Open-source R packages viacustom R Mappers / ReducersSQL DeveloperOther SQL AppsCopyright 2016, Oracle and/or its affiliates. All rights reserved.

Oracle R Advanced Analytics for HadoopAA Algorithms in a Hadoop Cluster: Map-Reduce and Spark (2.6)ClassificationGLM ORAAHLogistic Regression ORAAHRegressionMLP Neural Networks ORAAHLASSOFeature ExtractionNon-negative MatrixFactorizationRandom ForestsRidge RegressionDecision TreesSupport Vector MachinesSupport Vector MachinesRandom ForestAttribute ImportanceClusteringLinear RegressionPrincipal ComponentsAnalysisHierarchical k-MeansHierarchical k-MeansCollaborative Filtering(LMF)Basic StatisticsCorrelation/CovarianceCopyright 2016, Oracle and/or its affiliates. All rights reserved. Principal ComponentsAnalysis32

Getting started: OAA Links and ResourcesOracle Advanced Analytics Overview: OAA presentation— Big Data Analytics with Oracle Advanced Analytics or just watch Watch YouTube videopresentation and demo(s) Big Data Analytics with Oracle Advanced Analytics: Making Big Data and Analytics Simple white paper on OTN Oracle Internal OAA Product Management Wiki and Workspace Oracle Advanced Analytics Customer SuccessesYouTube recorded OAA Presentations and Demos: Oracle Advanced Analytics and Data Mining at the YouTube Movies (6 OAA “live” Demos on ODM’r 4.0 NewFeatures, Retail, Fraud, Loyalty, Overview, etc.)Getting Started: Link to OAA/Oracle Data Miner Workflow GUI Online (free) Tutorial Series on OTNLink to OAA/Oracle R Enterprise (free) Tutorial Series on OTNLink to Free Oracle Advanced Analytics "Test Drives" on Oracle Cloud via Vlamis PartnerLink to Getting Started w/ ODM blog entryLink to New OAA/Oracle Data Mining 2-Day Instructor Led Oracle University course.Oracle Data Mining Sample Code ExamplesSend email now to charlie.berger@oracle.com andyou’ll get my “away message” with these links.Additional Resources: Oracle Advanced Analytics Option on OTN pageOAA/Oracle Data Mining on OTN page, ODM Documentation & ODM BlogOAA/Oracle R Enterprise page on OTN page, ORE Documentation & ORE BlogOracle SQL based Basic Statistical functions on OTNOracle R Advanced Analytics for Hadoop (ORAAH) on OTN Business Intelligence, Warehousing & Analytics—BIWA Summit’17, Jan 31, Feb 1 & 2,2017 at Oracle HQ Conference Center (w/ links to customer presentations)Copyright 2016, Oracle and/or its affiliates. All rights reserved.

SAS to Oracle Advanced AnalyticsMigration at Zagrebačka banka - UniCreditGroup memberSiniša Behin, Head of ICT CRM functionOracle Open World 2016San Francisco 22.09.2016

Agenda UniCredit and Zagrebačka banka(ZaBa) key figures Reasons for migration Project scope and final goals Challenges during project Commercial modelling Results and lessons learned Next steps37/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

UniCredit Group – key structural figures*Over 143.000employees7.839branchesInternationalnetwork thatspans 50 markets Banking operations in 17 countries – Europe CEE marketleader Global player in asset management: 218.7 bn in managedassets Customers: over 25 million Market leader in Central and Eastern Europe leveraging onthe region's structural strengths One of the global systemically important financialinstitutions identified by the Financial Stability Board38/26Data as per FY 2015.Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Zagrebačka banka (ZaBa) – key structural figures1 Member of UniCredit Group since March 2002 26,4 % market share - 1st Croatian bank Leading bank in terms of products, services qualityand technology innovations Total assets of EUR 106 billion 4.200 employees, 125 branches, 867 ATMs 1 million customers Undisputed market leader Total loans – 27,8% Total deposits – 26,4% Investment funds – 23,9% Pension funds – obligatory 39,9%; voluntary53,4%40/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved1 figures2015as of Dec

Awards - selectionZagrebačka bankaprofile2015EuromoneyGlobal Banking & Finance review6/26Best Private banking in CroatiaBest Asset Management (for Private banking clients)Best Investment banking quality (for Private banking clients)Best Bank for International clientsBest bank for Planning inheritance consultingBest Trade Finance Bank in CroatiaBest Investment Bank in CroatiaGlobal FinanceBest Trade Finance Group in CEEEMEA FinanceBest Bank in Croatia in 2014MergermarketCEE Financial Adviser of the year

Reasons for migration Part of DWH consolidation phase Increase models development efficiency Improve slow preselection process Create standardized way of model development More flexible commercial activities – time tomarket Efficient risk mitigation Expand Exadata X4-2 purpose Integrate models within data itself Simpler architecture Easier DBA maintenance Simpler scoring process Model development flexibility – ORE/R Overall TCO optimization42/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Project scopeWhat? Data consolidation DWH Functional datamarts Different RISK portfolio basedmodels CRM commercial modelsWhen?How? 3 phases in 6 months Prior to the project - onemonth pilot project Strict compliance rules Keep the very same process ofmodels development like in SAS Get the same models (numbers)as with models created using SAS43/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Model development methodology – CRISP DM1234Initial StepsPreparationModelDevelopment10%50%30%Goals definitionBusiness analysisRegression buildingData cleaningVariable preselectionModel ing the dots” MonitoringMandatory requirements Information value44/26Copyright 2016 UniCredit Group and its affiliates. All rights reservedCalibration

Data preparation phase (1/2)Risk stream Used for preselection of 7.000 variables from riskand retail ABTs Descriptive statisticsPearson and Spearman correlationQuantiles (1, 5, 95, 99)AUROC (C-statistic)Kolmogorov-Smirnov testInformation value (IV) Avoid capturing noise in the data Reduce 7.000 variables to 1.000 variables Originally developed in SAS Base using SAS EGby ZaBa modelers45/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Data preparation phase (2/2)Risk stream Developed and migrated to Oracle using Oracle R Enterprise(ORE) Oracle maximum of 1000 variables per table Using transparency layer of ORE to speedup Saving results of computation in result table in Oracle DB Failure resistant computation due to saved results after eachcomputation Modular - it is easy to add any new R statistic to compute Computation time reduced from 20h to 5h46/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Clustering variables Risk stream Used to arrange variables by clusters based on theirsimilarity Similar to correlation matrix VARCLUS procedure in SAS base - Varclass forORE Developed by Oracle development team specially forZaBa project Results saved in ORE data-store First version in only 5 days from request Now explained on ORE bloghttps://blogs.oracle.com/R/entry/variable selection with ore varclushttps://blogs.oracle.com/R/entry/variable selection with ore varclus147/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Variable ImportanceRisk streamWoE Information value (IV) Used in SAS as measure of variable prediction Standard in Risk industry Similar to attribute importance node in OAA To calculate IV Variable (numeric) need to be binned and WoE foreach bin is calculated IV is calculated from WoE Exist in klaR package on CRAN Created version for preselection which use OREtransparency layer to compute IV and WoE much faster Stored in Oracle DB and used from ORE and Oracle Dataminer (ODM)48/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

WoE and IV integration in OAARisk stream Substitution for IGN node in SAS Calculating WoE values and WOE graph withSQL Node and in-database R script Binning with Transformation IV/WOE are implemented by custom made Rscripts: woe.iv() – IV calculation woe.woe() – WoE calculation for given variable woe.plot() – grapical WoE presentation woe.apply() – WoE values substitution select * from SYS.RQ SCRIPT Calculating IV with SQL Node and in-database Rscriptfunction(x, grouping column, .) {library(klaR)x[is.na(x)] - "N/A"x - do.call(data.frame, lapply(x, as.factor))gc - x[[grouping column]]x[[grouping column]] - NULLwoe - woe(x, grouping gc, appont FALSE, .) wreturn(do.call(rbind, lapply(seq(woe),function(x) data.frame(variable names(woe[x]),var level names(woe[[x]]), WOE woe[[x]]))))}50/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Model build processRisk stream Final process in model build is done usingOracle Data Miner workflows Include steps: Binning of variables with TransformationnodeCalculation of WOE, IV and analyzing WoEGraphRepeating until satisfied WOE graph and IVfor each variableReplacing variable values with WoE values inWOEApplyTGTrainBuilding model on replaced data-set51/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Model build processRisk stream Final process in model build is done usingOracle Data Miner workflows Include steps: Binning of variables with TransformationnodeCalculation of WOE, IV and analyzing WoEGraphRepeating until satisfied WOE graph and IVfor each variableReplacing variable values with WoE values inWOEApplyTGTrainBuilding model on replaced data-set52/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Deployment of modelsRisk stream Models are deployed to productionenvironment Separated from development Deployed using ODM code generationoption to PL/SQL code In-database mining Oracle Data Miner ORE an R Test and Production modeling DBsModel movement automatizationCentralized work repositoryResults used in daily Campaigns53/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Data consolidationCommercial stream New client centric data infrastructure for CRManalytics and modelling Simple maintenance and further by focusingon three layer of best practices: DM vs ABT – business area layersExadata capabilitiesODI facilitiesBusiness mapping revisionNew product and trxcategorizationIndustry, market, strategycustomizationTrend variablesLogic and statistic relevance55/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Commercial predictive analysisCommercial streamSetbusinessgoalCommunicationTargetclient The best from two worlds SAS and Oracletechnologies combined SAS procs’ changed with ODM/PLSQL, OREand R Utilization of R language for its advancedanalytical and graphical options Significantly faster data processing of largeamounts of data correlation calculation with 20 millions ofrecords and 200 variables 20min vs 2 days!Next-bestCalculateoffer Optimized data preparation and preselection More flexible process than highly regulated Risk Paradigm change - client definesproducts/services56/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Results Centralized developing workflow repository New VARCLUS ORE package – standard in ORE 1.5version Information Value (IV) calculation in ORE and ODM Integrated production models on same machine as data –simpler scoring process High performance in-database modeling architecture New customer centric data infrastructure Efficient commercial realization 1Q2015 cash loans sold within CRM campaign basedon predictive model made up to 38% of total cashloan sales in this period) 1Q2015 – 57% credit card issues covered by crosssell model 1Q2015 – 10% live insurance sales as result of crossselling58/26Copyright 2016 UniCredit Group and its affiliates. All rights reserved

Next steps Analytics architecture expansion – OracleBDA Smarter, faster and more detailed analysis In-house data with new approach,outside data with new tools Go Big! Expand area of usage60/26 Fraud prevention Operational efficiency Utilizing social network communicationchannel Merchant analysis Geospatial campaigns – push notificationProduct recommendationsCopyright 2016 UniCredit Group and itsNonaffiliates.bankingAll rights reserved

SAS to Oracle Advanced Analytics Migration at Zagrebacka Banka, UniCredit Group [CAS6499] Agenda . Using Hadoop and HIVE Integration, plus R Engine and Open-Source R Packages R Analytics Oracle R Advanced Analytics for Hadoop R Client SQL Developer Other SQL Apps