Getting in Front on Data Quality

Transcription

1 Getting in Frnt n Data Quality Thmas C. Redman, Ph.D. the Data Dc, in Ciudad Real, Espana, July, 2016 Data Quality Slutins tmredman@dataqualityslutins.cm /Redman-ICIQ-June2016 DQS T.C. Redman, Page 1

2 But first, Muchas gracias, Alarcs Research Grup!!! /Redman-ICIQ-June2016 DQS T. C. Redman, Page 2

3 Agenda Data Quality in Practice Mst Imprtant Things t Knw: n Getting in Frnt Apprach n Rles and Respnsibilities The Friday Afternn Measurement: Please g d this! /Redman-ICIQ-June2016 DQS T. C. Redman, Page 3

4 Data Quality in AT&T s Access Management Department Tm Redman /Redman-ICIQ-June2016 DQS T. C. Redman, Page 4

5 Figure 1: Access Bill Verificatin at AT&T Telc (supplier) AT&C (custmer) Predicted Bill - Prducing Prcess Bill- Prducing Prcess Bills Cmpare Predicted Bills Find discrepancies Respnd File Claims Prcess rebates File Cunterclaims /Redman-ICIQ-June2016 DQS T. C. Redman, Page 5

6 A revelatin unflds, We need t clean up ur data every ther year. And its gd fr me I get a big bnus. But smehw it desn t feel right fr the cmpany! Twain: A man with a watch knws what time it is. A man with tw is never sure. This desn t wrk (this is stupid!) /Redman-ICIQ-June2016 DQS T. C. Redman, Page 6

7 Yu had me at Hell A Tracked Data Recrd The nly change here invlves frmat. Nt f cncern Step f Prcess Attribute A B C D E Name XYZ.1234 XYZ-1234 XYZ-1234 XYZ-1234 XYZ-1234 Billing Number Bill Cde 1 A A A A Office In this case, 1 and A bth mean yes. It reflects pr architecture, but is nt a data errr /Redman-ICIQ-June2016 DQS Changes in data values. Errrs f serius cncern. Nte we cannt be certain where the errrs actually ccurred. T. C. Redman, Page 7

8 Start with a Basic Time-Series Plt Clearly, this prcess is brken! /Redman-ICIQ-June2016 DQS T. C. Redman, Page 8

9 The Search fr Rt Causes /Redman-ICIQ-June2016 DQS T. C. Redman, Page 9

10 The Search fr Rt Causes, Anther Perspective /Redman-ICIQ-June2016 DQS T. C. Redman, Page 10

11 Cmbining the Tw Perspectives With this plt, we have identified ptential rt causes well enugh t charter specific imprvement prjects. Eliminating the tp three will reduce the verall errr rate by 80! /Redman-ICIQ-June2016 DQS T. C. Redman, Page 11

12 This has t be end-t-end I dn t understand the details, but I trust the lgic. Let s give it a try Yu knw Bb, we dn t cme in every mrning thinking f ways t ful yu up. What des timely and accurate mean? Hmmm. That s A gd questin? /Redman-ICIQ-June2016 DQS T. C. Redman, Page 12

13 (Prpsed) Access Financial Assurance Prcess This is what we are ging d! Tracking Results Telc (supplier) This is what We are ging d! AT&T (custmer) Custmer Requirements Step a Step B Step C Step D Step E Prcess management Bills Supplier Management Bth: Eliminate Bill Verificatin (and all it brings Audit Supplier Perfrmance /Redman-ICIQ-June2016 DQS T. C. Redman, Page 13

14 Business Results Data accuracy errrs reduced 90% Billing errrs reduced 98% Cycle time (bill perid clsure) reduced 67% Custmer csts reduced 73% Csts acrss supplier base reduced 20% Achieved Financial Assurance Created Financial Predictability Tm, I really d appreciate the $100M. What I appreciate even mre is that nw we can manage the business. That is wrth even mre. /Redman-ICIQ-June2016 DQS T. C. Redman, Page 14

15 Getting in Frnt: The Mst Imprtant Things T Knw /Redman-ICIQ-June2016 DQS T. C. Redman, Page 15

16 Rising Middle Manager VS /Redman-ICIQ-June2016 DQS T. C. Redman, Page 16

17 What Makes Data f High Quality Data are f high quality if they are fit fr their intended uses (by custmers) in peratins, decisin-making, analytics and planning (after Juran). free f defects: - accessible - accurate - up-t-date - etc. Mstly, are the data right? Data that s fit fr use Largely, the right data pssess desired features (attributes): - relevant - cmprehensive - prper level f detail - easy-t-interpret - etc. Custmers are the ultimate arbiters f quality!! /Redman-ICIQ-June2016 DQS T. C. Redman, Page 17

18 Day-in, day-ut definitin Meeting the mst imprtant needs f the mst imprtant custmers /Redman-ICIQ-June2016 DQS T. C. Redman, Page 18

19 Data Quality: The Nn-delegatable Chice T Clean Up The Lake, One Must First Eliminate The Surces Of Pllutant Unmanaged /Redman-ICIQ-June2016 DQS T. C. Redman, Page 19

20 Fr Data, Only Tw Mments Really Matter The Mment f Creatin Nte that they DO NOT ccur in IT The Mment f Use PATH FROM CREATOR TO CUSTOMER /Redman-ICIQ-June2016 DATA CREATOR The whle pint f data quality management is t cnnect the tw! DQS DATA CUSTOMER T. C. Redman, Page 20

21 Custmer Rles and Respnsibilities Recgnize that yu ve becme way t tlerant f junk. Recgnize yur hidden data factries and vw t put a stp t them. Srt ut and dcument yur mst imprtant needs. Cmmunicate them t suppliers. Wrk with suppliers t clse the mst imprtant gaps. /Redman-ICIQ-June2016 DQS T. C. Redman, Page 21

22 The mst imprtant tl in all f data quality management (and maybe all management) CUSTOMER-SUPPLIER MODEL requirements requirements Yur Business Prcess inputs utputs Suppliers Custmers feedback feedback T be clear, all must build these cmmunicatins channels with data custmers and suppliers. /Redman-ICIQ-June2016 DQS T. C. Redman, Page 22

23 Creatr Rles and Respnsibilities Recgnize that the data yu create impacts thers. Understand the mst imprtant needs f the mst imprtant custmers Measure quality against thse needs. Cnduct imprvement prjects t eliminate rt causes. Put in place cntrls t clse the gaps. /Redman-ICIQ-June2016 DQS T. C. Redman, Page 23

24 Prvcateurs disrupt the dynamic that leads t hidden data factries Dissatisfied with the status qu. Curage t try smething new. Great crprate citizens. Achieve real results within their spans f cntrl. At all levels! THERE IS A LITTLE PROVOCATEUR IN ALL OF US /Redman-ICIQ-June2016 DQS T. C. Redman, Page 24

25 Prvcateurs Can G Only S Far penetratin f DQ acrss rganizatin Prgress f a Typical Data Quality Prgram tractin real results plateau Order-f-magnitude imprvement n sme data next level = mre data time /Redman-ICIQ-June2016 DQS T. C. Redman, Page 25

26 Getting in Frnt Takes a Village: New Rles fr Everyne Data custmers and creatrs: Everyne tuches data, s these rles are fundamental! n Ideally, in a prcess framewrk. n Embedded data managers Prvcateurs DQ Teams: Day-in, day-ut faces f DQ n n Chief Data Architect Data Maestr Leadership Technlgists Either prvcateurs r DQ Team lead must help senir management understand what t d /Redman-ICIQ-June2016 DQS T. C. Redman, Page 26

27 The Prcess Management Cycle 2. Understand custmer needs 3. Describe prcess 4. Establish measurement system 1. Establish management respnsibilities The prcess management cycle prvides a pwerful, repeatable means t bring the tasks t bear 7. Make imprvements & sustain gains 6. Set and pursue target fr imprvement 5. Establish cntrl & check cnfrmance t requirements /Redman-ICIQ-June2016 DQS T. C. Redman, Page 27

28 Typical Result 1. Rqmts defined First-time, n-time results Fractin Perfect Recrds First Meas 4. Cntrl 3. Imprvements Mnth Prgram start Accuracy Rate ave lwer cntrl limit upper cntrl limit target Each errr nt made saves an average f $500. Quickly millins. But the imprved cmpe==ve psi=n may be wrth even mre! /Redman-ICIQ-June2016 DQS T. C. Redman, Page 28

29 Dn t Frget the Metadata Metadata are data abut data Yu have t take gd care f at least sme metadata. The principles f data quality management apply A Chief Data Architect, shuld lead wrk n cmmn definitins. Embeds play huge rles here! METADATA /Redman-ICIQ-June2016 DQS T. C. Redman, Page 29

30 The Friday Afternn Measurement /Redman-ICIQ-June2016 DQS T. C. Redman, Page 30

31 The Friday afternn measurement aims t help answer the questin, D I need t wrry abut data quality? /Redman-ICIQ-June2016 DQS T. C. Redman, Page 31

32 Friday Afternn Measurement Prtcl Assemble last 100 recrds Assemble 2-3 experts Mark bviuslyerred data in red Summarize and interpret results /Redman-ICIQ-June2016 DQS T. C. Redman, Page 32

33 Assemble the last 100 recrds Select mst imprtant attributes Attribute 1: Attribute 2: Size Attribute 3: Amunt Recrd A Jane De Null $ Recrd B Jhn Smith Medium $ C Stuart Madnick XXXL Null D Thams Jnes etc Recrd 100 James Olsen One Lcked Place $76.24 After Fig 18.2, Redman, Data Quality: /Redman-ICIQ-June2016 DQS The T. C. Field Redman, Guide Page 33

34 Mark the bvius errrs Attribute 1: Attribute 2: Size Attribute 3: Amunt Recrd A Jane De Null $ Recrd B Jhn Smith Medium $ C Stuart Madnick XXXL Null D Thams Jnes etc Recrd 100 James Olsen One Lcked Place $76.24 After Fig 18.2, Redman, Data Quality: /Redman-ICIQ-June2016 DQS The T. C. Field Redman, Guide Page 34

35 Rate the recrd as perfect r nt Attribute 1: Attribute 2: Size Attribute 3: Amunt etc recrd perfect? (y/n) Recrd A Jane De Null $ n Recrd B Jhn Smith Medium $ y C Stuart Madnick XXXL Null n D Thams Jnes n Recrd 100 James Olsen One Lcked Place $76.24 n After Fig 18.2, Redman, Data Quality: /Redman-ICIQ-June2016 DQS The T. C. Field Redman, Guide Page 35

36 Cunt the perfects Attribute 1: Attribute 2: Size Attribute 3: Amunt etc recrd perfect? (y/ n) Recrd A Jane De Null $ n Recrd B Jhn Smith Medium $ y C Stuart Madnick XXXL Null n D Thams Jnes n Recrd 100 James Olsen One Lcked Place $76.24 n Cunt perfect After Fig 18.2, Redman, Data Quality: /Redman-ICIQ-June2016 DQS The T. C. Field Redman, Guide Page 36 67

37 Data Quality = 67% Here the interpretatin is a full third f recent custmer rders had a serius DQ issue. A wrry indeed! /Redman-ICIQ-June2016 DQS T. C. Redman, Page 37

38 Summary Order-f-magnitude imprvements are pssible. Such imprvements bring rich rewards Sner r later, yu have t getting in frnt. The mst imprtant rles are data custmers and data creatrs. There is a little prvcateur in all f us. Time t let it ut. Making an initial measurement is nt s hard using the Friday Afternn Measurement. /Redman-ICIQ-June2016 DQS T. C. Redman, Page 38

39 Questins? Thmas C. Redman, Ph.D. the Data Dc /Redman-ICIQ-June2016 DQS T. C. Redman, Page 39

40 Clinician: Thmas C. Redman, the Data Dc Ph.D., Statistics, Flrida State, Cnceived and led the Data Quality Lab at AT&T Bell Labs. Frmed Data Quality Slutins in Latest and greatest: Data s Credibility Prblem, HBR, Dec, Data Driven: Prfiting frm Yur Mst Imprtant Business Asset, Harvard Business Schl Press, Knwn bias: Data are quite bviusly the key asset f the Infrmatin Age. Yet tday s rganizatins are unfit fr data. Further, despite enrmus ptential, few are yet cnsidering hw they will cmpete with data. Finally, high-quality data is pre-requisite. These crystallize THE management challenges f the 21st century. /Redman-ICIQ-June2016 DQS Picture T.C. Redman, Page 40

41 The Statistician in View 1980 /Redman-ICIQ-June2016 Tday DQS T. C. Redman, Page 41