3. Traditional Approach Two types of variation: Common cause of variation (noise) occur as normal part of manufacturing process Special cause of variation represents a potential problem 3
4. 3 s Shewhart control chart used by NIST for calibrating the standard KG. 4 Source: NIST
5. Shewhart / Deming Cycle Plan – identify opportunity or problem and make a plan. Do – implement the change on a small scale and collect the data. Check – perform a statistical analysis and check if there was an impact. Act – if there was an impact, broaden the scale and continuously improve your results. 5
6. Case Study 1. Data Center Thousands of servers Complex workloads Large variations are normal Problems make the front page 6
7. Case Study 2. Payments Network Billion+ cards 100+ million terminals Millions of merchants Thousands of transactions per second Thousands of member banks Data highly heterogeneous Variations among products Variations among cardholders Variations among merchants Variations among banks Variation among payment networks 7
8. The Challenge Today Many sources and data feeds Data is complex and highly heterogeneous High volume, streaming data from around the world Multiple parties involved, each of which can modify the data in subtle ways 8
14. Build separate model for each different type of merchant (over 800 types of merchants)
15. For each distinct cube, build a distinct modelGeospatialregion Type of Transaction Bank Modeling using Cubes of Models (MCM) 12
16. learning sets data updates 1. data collection 3. on-line scoring 2. off-line modeling Entity/Feature Database Data Mining System PMML models features Model Consumer Data Mining Mart events Rules candidate alerts Operational systems, data feeds, warehouses, … 4. reporting Dashboard engine reports 13
17. Augustus Augustus is an open source data mining platform: Used to estimate baselines for over 15,000 separate segmented models Used to score high volume operational data and issue alerts for follow up investigations Augustus is PMML compliant Augustus scales with Volume of data Real time transaction streams (15,000/sec+) Number of segmented models (10,000+) 14
23. To decrease alerts, remove breakpoint,order by number of decreased alerts, & select one or more breakpoints to removeOne model for each cell in data cube 15
26. Results Dozens of separate statistical baselines models developed and deployed. Effective for discovering nodes that are hindering effective use of OCC’s large data cloud. Dead nodes are easy to identify and remove. Removing just one or two “slow” nodes from a pool of 100 nodes can improve overall performance by 15% - 20+%. 18
28. Case Study 2 Account Issuing Bank Payments Network Merchant Acquiring Bank 20
29. Program Structure Strategic objective identified early: “Identify and ameliorate data interoperability issues to improve the approval rate of valid transactions and the disapproval rate of invalid transactions, ...” Report quarterly to CIOs’ council with third-party endorsed monetary benefits summarized on an executive dash board Introduced data governance program early in project Developed payment transaction monitor that produced candidate alerts Set up investigation process to screen alerts and investigate those of interest Developed reference models and appropriate standards 21
30. Results ROI 5.1x Year 1 (over 6 months) 7.3x Year 2 (12 months) 10.0x Year 3 (12 months) Over 15,500 separate statistical baselines models developed and deployed. Also developed appropriate rules-based models to make work of analysts more efficient. 22
32. Business Process Strategic Objective Dashboard Governance Modeling Process Reference Model Investigative Process Monitor - produces candidate alerts InvestigativeProcess candidate alerts events program alerts 24
33. Some Lesson Learned Business Processes Importance of “C”-level executive support, dashboard reports, and a data governance program Modeling Processes Critical to build as many statistical models as the data required; used open source Augustus software for this Architecture separated offline modeling and online scoring Post processing with business rules to control workflow to analysts Investigative Processes It is not about the models and alerts – it is about optimizing the analysts’ workload and derived business value Small changes in report designs had large impact in the effectiveness of the alerts 25
37. References Joseph Bugajski, Chris Curry, Robert L. Grossman, David Locke, Steve Vejcik, Detecting Changes in Large Data Sets of Payment Card Data: A Case Study, Proceedings of The Thirteenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD 2007), ACM, 2007 Joseph Bugajski and Robert L. Grossman, An Alert Management Approach to Data Quality: Lessons Learned from the Visa Data Authority Program, Proceedings of the 12th International Conference on Information Quality, (ICIQ 2007). Walter A. Shewhart, Statistical Method from the Viewpoint of Quality Control, Dover, 1986. H. Vincent Poor and Olympia Hadjiliadis, Quickest Detection, Cambridge University Press, 2009. Augustus is an open source system available from augustus.googlecode.com. 28