Handling Data Quality in Entity Resolution

Transcription

1 Handling Data Quality in Entity Resolution Hector Garcia-Molina Stanford University Work with: Omar Benjelloun, Qi Su, Jennifer Widom, Tyson Condie, Nicolas Pombourcq

2 Reverse Talk Entity Resolution Problem Confidences Two Ideas Ask YOU for ideas! 2

3 Entity Resolution e1 e2 N: a A: b CC#: c Ph: e N: a Exp: d Ph: e Applications: mailing lists, customer files, counter-terrorism,... 3

4 Challenges (1) No keys! Value matching Kaddafi, Qaddafi, Kadafi, Kaddaffi... Record matching Nm: Tom Ad: 123 Main St Ph: (650) Ph: (650) Nm: Thomas Ad: 132 Main St Ph: (650)

5 Challenges (2) Merging records Nm: Tom Ad: 123 Main St Ph: (650) Ph: (650) Nm: Thomas Ad: 132 Main St Ph: (650) Zp: Nm: Tom Nm: Thomas Ad: 123 Main St Ph: (650) Ph: (650) Zp:

6 Challenges (3) Chaining Nm: Tom Ad: 123 Main BD: Jan 1, 85 Wk: IBM Nm: Thomas Ad: 123 Maim Oc: lawyer Nm: Tom Wk: IBM Oc: laywer Sal: 500K Nm: Tom Ad: 123 Main BD: Jan 1, 85 Wk: IBM Oc: lawyer Nm: Tom Ad: 123 Main BD: Jan 1, 85 Wk: IBM Oc: lawyer Sal: 500K 6

7 Un-merging Challenges (4) Nm: Tom Ad: 123 Main BD: Jan 1, 85 Wk: IBM Oc: lawyer Sal: 500K too young to make 500K at IBM!! 7

8 Taxonomy Pairwise snaps vs. clustering De-duplication vs. fidelity enhancement Schema differences Relationships Exact vs. approximate Generic vs application specific Confidences 8

9 Pair-Wise Snaps vs. Clustering r1 r2 r3 r4 s7 s8 s9 s10 r3 r1 r2 r7 r8 r9 r10 r5 r5 r4 r6 r6 9

10 Taxonomy Pairwise snaps vs. clustering De-duplication vs. fidelity enhancement Schema differences Relationships Exact vs. approximate Generic vs application specific Confidences 10

11 Taxonomy Pairwise snaps vs. clustering De-duplication vs. fidelity enhancement Schema differences No Relationships No Exact vs. approximate Generic vs application specific Confidences... later on 11

12 Model r1 Nm: Tom Ad: 123 Main BD: Jan 1, 85 Wk: IBM r2 Nm: Thomas Ad: 123 Maim Oc: lawyer r3 Nm: Tom Wk: IBM Oc: laywer Sal: 500K Nm: Tom Ad: 123 Main BD: Jan 1, 85 Wk: IBM Oc: lawyer r4:<r1, r2> M(r1, r2) M(r4, r3) Nm: Tom Ad: 123 Main BD: Jan 1, 85 Wk: IBM Oc: lawyer Sal: 500K <r4, r3> 12

13 Records: Example r1 = [a: {1, 2}, b:2, c:{5,6}], r2 = [a:3, b:{1, 2}, c:5, d:8],... Features: F1 = {a, b}, F2 = {c} Match function: M(r1, r2) = M F1 V M F2 Merge Function: <r1, r2>=[a:{1, 2, 3}, b:{1, 2}, c:{5,6}, d:8] 13

14 Question What is best sequence of match, merge calls that give us right answer? 14

15 Properties Commutativity: M(r1, r2) = M(r2, r1) <r1, r2> = <r2, r1> Idempotence: M(r1, r1) = true; <r1, r1> = r1 Associativity <r1, <r2, r3>> = <<r1, r2>, r3> 15

16 Representativity More Properties If <r1, r2> = r3, then for any r4 such that M(r1, r4) is true we also have M(r3, r4) = true. r1 r2 r4 r3 16

17 4 Properties Efficiency Commutativity Idempotence Associativity Representativity ER result is unique ER result independent of processing order 17

18 Example Feature F1: {a} [a: v1, b: w1] [a: v2, b: w2] [a: v3, b: w3]... [a: vn, b: wn] M( ri, rj ) = True answer: [ a:{v1,...,vn}, b:{w1,..., wn}] 18

19 Brute Force Algorithm not_done := true while not_done do [not_done := false; RP := RN := empty set; for each in R, s.t. do if M then not_done := true; RP = RP union { } RN = RN union { } R := R union RP minus RN ] RP = new records RN = no longer needed records 19

20 Back to Example... [a: v1, b: w1] [a: v2, b: w2] [a: v3, b: w3] [a: v4, b: w4] [a:{v1,v2}, b:{w1,w2}] [a:{v1,v3}, b:{w1,w3}] [a:{v1,v4}, b:{w1,w4}] [a:{v2,v3}, b:{w2,w3}] [a:{v2,v4}, b:{w2,w4}] [a:{v3,v4}, b:{w3,w4}] 20

21 Example Continued... [a:{v1,v2},...] [a:{v1,v3},...] [a:{v1,v4},...] [a:{v2,v3},...] [a:{v2,v4},...] [a:{v3,v4},...] [a:{v1,v2,v3},...] [a:{v1,v2,v4},...] [a:{v2,v3,v4},...] [a:{v1,v2,v4},...] [a:{v1,v2,v3,v4},...]... A lot of useless work! 21

22 Record Swoosh Swoosh Algorithms Merges records as soon as they match Optimal in terms of record comparisons Feature Swoosh Remembers values seen for each feature Avoids redundant value comparisons 22

23 Swoosh Example [a: v1, b: w1] [a: v2, b: w2] [a: v3, b: w3] [a: v4, b: w4] M(r1,r2) [a:{v1,v2},...] M(r3, r12) [a:{v1,v2,v3},...] M(r4, r123) [a: v1, a: v2, a: v3, a: v4,...] 23

24 Swoosh Performance (I) 24

25 Swoosh Performance (II) 25

26 Swoosh Performance (III) 26

27 In data (0.8) Confidences Nm: Tom (0.9) Ad: 123 Main St (1.0) Ph: (650) (0.6) Ph: (650) (0.8) In value matching: sim(qadafi, Kadafi) = 0.95 In match rules: M( r1, r2 ) = T (0.9) In merge/fusion: Merge(Héctor, Ettore) = Hector (0.8) 27

28 Challenges What do confidences mean? (0.8) Nm: Tom (0.9) Ad: 123 Main St (1.0) Ph: (650) (0.6) Ph: (650) (0.8) 28

29 Challenges How do we operate on confidences? (0.9) Nm: Tom (0.9) Ad: 123 Main St (1.0) Ph: (650) (0.7) Ph: (650) (0.9) (0.9) Nm: Thomas (0.8) Ad: 132 Main St (0.8) Ph: (650) (1.0) Zp: (0.95) sim(tom, Thomas) = 0.92 M F{Nm, Ad} = ? 29

30 One Confidence Model [id1, a, b, c, d] [id2, a, c, e] [id3, a, b, f, g] [id1, a, b, c, d] [id1, a, b, d] [id1, a, x] [id1, b, y] [id2, a, b, c] [id2, a, c, e] [id2, a, c, e] [id2, a, c, e] [id3, a, b, c] [id3, a, b, d] [id3, a, b, f, g] [id3, a, b, f, g] shorthand 30

31 Records Are Evidence [id1, a, b, c, d] [id1, a, b, c, d] [id1, a, b, d] [id1, a, x] [id1, b, y] not 0.25 [id1, (3/4)a, (3/4)b, (1/4)c, (2/4)d, (1/4)x, (1/4)y] 31

32 New Evidence [id1, a, b, c, d] [id1, a, b, c, d] [id1, a, b, d] [id1, a, x] [id1, b, y] + [id1, a, b, c, d] [id1, (3/4)a, (3/4)b, (1/4)c, (2/4)d, (1/4)x, (1/4)y] + [id1, a, b, c, d] [id1, (4/5)a, (4/5)b, (2/5)c, (3/5)d, (1/5)x, (1/5)y] 32

33 No Ids [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, (1/2)c, (1/2)d] [a, x] [c, d, y] 33

34 No Ids [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, (1/2)c, (1/2)d] [a, x] [c, d, y] [a, b, (1/2)c, (1/2)d] [a, x] [c, d, y] [(2/3)a, (2/3)b, (2/3)c, (2/3)d, (1/3)y] [a, x] 34

35 Queries? [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, (1/2)c, (1/2)d] [a, x] [c, d, y] Threshold = 0.5; Support = 2 Maximal Record Example: [a, b, c, d] [a, b, (1/2)c, (1/2)d] [a, x] [c, d, y] [(2/3)a, (2/3)b, (2/3)c, (2/3)d, (1/3)y] [a, x] 35

36 Queries? [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, c] [a, b, d] [a, x] [c, d, y] [a, b, (1/2)c, (1/2)d] [a, x] [c, d, y] Threshold = 0.5; Support = 2 Maximal Record Example: [a, b, c, d] [a, b, (1/2)c, (1/2)d] [a, x] [c, d, y] [(2/3)a, (2/3)b, (2/3)c, (2/3)d, (1/3)y] [a, x] 36

37 Need Simpler Model? 37

38 0.7 [a, b] Alternate Worlds: Simple Confidence Model [a, b] [a, b] [a, b, c] [a, b] [a, b] [a, b] [a, b, d]????????? 38

39 Rules 0.7[a, b, c], 0.7[a, b, c] 0.7 [a, b, c] 0.7 [a, b], 0.5 [a, b] 0.7 [a, b] 0.7 [a, b, c], 0.5 [a, b] 0.7 [a, b, c] 0.7 [a, b, c], 0.9[a, b] 0.7 [a, b, c], 0.9[a, b] etc 39

40 Matches 0.9[a, b, c] 0.8[a, b, d] [a, x] [c, d, y] Match with confidence 0.5 worlds [a,b,c] [a,b,d] [a,b,c,d] 40

41 Matches 0.9[a, b, c] 0.8[a, b, d] [a, x] [c, d, y] 0.4[a,b,c,d] 0.9[a, b, c] 0.8[a, b, d] [a, x] [c, d, y] worlds [a,b,c] [a,b,d] [a,b,c,d] 41

42 Goal: C-Swoosh base records all possible merges eliminate dominated eliminate below threshold 42

43 Goal: C-Swoosh base records all possible merges eliminate dominated eliminate below threshold earlier 43

44 Questions Each model has drawbacks... Is there a better confidence model?? 44

45 Summary Entity resolution is critical Efficient resolution important Confidences are important, but how? ER is key aspect of info privacy check www-db.stanford.edu for Swoosh paper & forthcoming paper 45

46 Thanks. 46