1. Is Sampling Useful in Data Mining? A Case in the Maintenance of Discovered Association Rules S.D. Lee, David W. Cheung, Ben Kao The University of Hong Kong Data Mining and Knowledge Discovery, 1998 Revised and Presented by Matthew Starbuck : April 2, 2014 1

2.  Definitions  Old Algorithms Apriori FUP 2  New Algorithm: DELI Design Pseudo code Sampling Techniques  Experiments(Comparisons) showing DELI is better  Consecutive runs/Conclusions/Exam Questions Outline 2

3. Definitions(1) ‏ D = Transaction Set I = Full Item Set T1T1 T2T2 T3T3 3

4. Definitions(2) ‏ For X= σ X = Support count = 4 Support = 4/5 = 80% Support threshold: s% 4

5. Definitions(3)  K-itemset: itemset containing k items.  Large itemset (L k ): itemset with support larger than support threshold. 5

6. Old Algorithm (1) ‏ Apriori 6

7. Pseudo code of Apriori get C 1 ; k = 1; until (C k is empty || L k is empty)‏ do { Get L k from C k using minimum support count; Use apriori_gen() to generate C k+1 from L k ; k++; } ; return union(all L k ); 7

8. C k = Candidate Set L k = Large Set apriori_gen() ‏ s% = 40% 8

9. Use Apriori in Maintenance  Simply apply the algorithm to the updated database again; Not efficient; Fails to reuse the results of previous mining; Very cost-expensive. 9

10.  FUP 2 works similarly to Apriori by generating large itemsets iteratively;  It scans only the updated part of the database for old large itemsets;  For the rest, it scans the whole database. Old Algorithm 2: FUP 2 10

11. Δ - : set of deleted transactions Δ +: set of added transactions D: old database D': updated database D*: set of unchanged transactions σ X : support count of itemset X σ ’ X : new support count of itemset X δ X - : support count of itemset X in Δ - δ X + : support count of itemset X in Δ + 11

12. Pseudo code of FUP 2 get C 1 ; k = 1; until (C k is empty || L k ’ is empty)‏ do { divide C k into two partitions: P k = C k ۸ L k and Q k = C k – P k ; For X in P k, calculate σ’ X = σ X - δ X - + δ X + and get part 1 of L k ’ ; For X in Q k, eliminate candidates with δ X + - δ X - < (Δ + -Δ - )s% ; For the remaining candidates X in Q k, scan D* to get part 2 of L k ’ ; Use apriori_gen() to generate C k+1 from L k ’; k++; }; return union(all L k ’); CkCk LkLk PkPk QkQk Δ - (δ - X )‏ Δ + (δ + X )‏ D* D(σX)‏D(σX)‏ D’(σ’X)‏D’(σ’X)‏ 12

13. An Example on FUP 2 13

14. DELI Algorithm  Difference Estimation for Large Itemsets  Key idea: It examines samples of the database when the update is not too much; 14

15. Basic pseudo code of DELI get C 1 ; k = 1; until (C k is empty || L k ’ is empty)‏ do { divide C k into two partitions: P k = C k ۸ L k and Q k = C k – P k For X in P k, calculate σ’ X = σ X - δ X - + δ X + and get part 1 of L k ’ For X in Q k, eliminate candidates with δ X + - δ X - < (Δ + -Δ - )s%, For the remaining candidates X in Q k, scan D* to get part 2 of L k ’ Use apriori_gen() to generate C k+1 from L k ’; k++; }; return union(all L k ’); A sample subset of D* 15

16. Binomial Distribution  Assume 5% of the population is green-eyed.  You pick 500 people randomly with replacement.  The total number of green-eyed people you pick is a random variable X which follows a binomial distribution with n = 500 and p = 0.05. 16

17. Binomial Distribution http://en.wikipedia.org/wiki/Image:Binomial_distribution_pmf.png 17

18. Sampling Techniques (1) ‏  Consider an arbitrary itemset X;  Randomly select m transactions from D with replacement;  T X = the total number of X out of m;  T X is binomially distributed with p = σ X / |D| n = m Mean = np = (m / |D|) σ X Variance = np(1-p) 18

19. Sampling Techniques (2) ‏  T X approximates normally distributed with Mean = (m / |D|) σ X Variance = mp(1 - p) ‏  Define: σ X ^ = |D| / m * T X  σ X ^ is normally distributed with Mean = σ X Variance = σ X (|D| - σ X )/m 19

20. Confidence Interval axax bxbx Mean = σ X α/2 20

21. Sampling Techniques (3) ‏  We can obtain a 100(1-α)% confidence interval [a x, b x ] for σ X where  Typical Values: For α= 0.1, z α/2 =1.645 For α= 0.05,z α/2 =1.960 For α= 0.01,z α/2 =2.576 21

22. Sampling Techniques (4) ‏  The width of this interval is  The widths of all confidence intervals are no more than  Suppose we want the widths not to exceed 22

23. Sampling Techniques (5) ‏  If s = 2 and α= 0.05, then z α/2 =1.96  Solving the above inequality gives m ≥ 18823.84.  This value is independent of the size of the database D! Note*: D may contain billions of transactions. A sample of around 19 thousand is large enough for the desired accuracy in this example 23

24.  L k » : large in D and D’ ;  L k > : not large in D, large in D’ with a certain confidence;  L k ≈ : not large in D, maybe large in D’ ;  L k ’ : approximation of new L k. L k ’ =L k »  L k >  L k ≈ LkLk Lk’Lk’ CkCk QkQk PkPk Lk»Lk» Lk>Lk> Lk≈Lk≈ Obtain the estimated set of L k 24

25.  Degree of uncertainty u k = L k ≈ /L k ’, uncertainty factor u k - is a user-specified threshold If u k ≥ u k -, then DELI halts and FUP 2 is needed  Amount of changes (symmetric difference) ‏ η k = |L k – L k ’ | ξ k = |L k (>) | + |L k (≈) | d k - is a user-specified threshold If d k ≥ d k -, then DELI halts and FUP 2 is needed Criteria met to perform a full update 25

26. Pseudo code of DELI get C 1 ; k = 1; until (C k is empty || L k is empty)‏ do { divide C k into two partitions: P k = C k ۸ L k and Q k = C k – P k For X in P k, calculate σ’ X = σ X - δ X - + δ X + and get part 1 of L k ’ For X in Q k, eliminate candidates with δ X + - δ X - < (Δ + -Δ - )s%, For the remaining candidates X in Q k, scan a sample subset of D* to get part 2 of L k ’ Use apriori_gen() to generate C k+1 from L k ’; If any criteria is met, then terminate and go to FUP 2 ; k++; }; return union(all L k ’); 26

27. An Improvement  Store the support counts of all 1- itemsets  Extra storage: O(|I|) ‏ 27

28. Experiment Preparation  Synthetic databases – generate D, Δ +, Δ -  1%-18% of the large itemsets are changed by the updates.  u k - = ∞  d k - = ∞ 28

29. Experimental Results (1) ‏ α= 0.05 z α/2 =1.960 |Δ + |=|Δ - | = 5000 |D| = 100000 s% = 2% 29

30. Experimental Results (2) ‏ α= 0.05 z α/2 =1.960 |Δ + |=|Δ - | = 5000 |D| = 100000 s% = 2% 30

31. Experimental Results (3) ‏ m=20000 |Δ + |=|Δ - | = 5000 |D| = 100000 s% = 2% 31

32. Experimental Results (4) ‏ m=20000 |Δ + |=|Δ - | = 5000 |D| = 100000 s% = 2% 32

33. Experimental Results (5) ‏ α= 0.05 z α/2 =1.960 m=20000 |D| = 100000 s% = 2% 33

34. Experimental Results (6) ‏ α= 0.05 z α/2 =1.960 m=20000 |D| = 100000 s% = 2% 34

35. Experimental Results (7) ‏ α= 0.05 z α/2 =1.960 |Δ - |= 5000 m = 20000 |D| = 100000 s% = 2% 35

36. Experimental Results (8) ‏ α= 0.05 z α/2 =1.960 |Δ - |= 5000 m = 20000 |D| = 100000 s% = 2% 36

37. Experimental Results (9) ‏ α= 0.05 z α/2 =1.960 |Δ + |= 5000 m = 20000 |D| = 100000 s% = 2% 37

38. Experimental Results (10) ‏ α= 0.05 z α/2 =1.960 |Δ + |= 5000 m = 20000 |D| = 100000 s% = 2% 38

39. Experimental Results (11) ‏ α= 0.05 z α/2 =1.960 |Δ + |= |Δ-| = 5000 m = 20000 |D| = 100000 39

40. Experimental Results (12) ‏ α= 0.05 z α/2 =1.960 |Δ + |= |Δ-| = 5000 m = 20000 |D| = 100000 40

41. Experimental Results (13) ‏ α= 0.05 z α/2 =1.960 |Δ + |= |Δ-| = 5000 m = 20000 s% = 2% 41

42. Experimental Summary  u c - < 0.036, very low;  when | Δ - | < 10000, d c - < 0.1;  when | Δ - | = 20000, d c - < 0.21;  (Suggested) u - = 0.05, d - = 0.1 42

43. Consecutive Runs:  Say we use Apriori to find association rules in a database  Later, 1 st batch of updates arrives, use DELI to make rules (r) if necessary  If r = F then use old association rules  When 2 nd batch comes, check both batches for significant changes  Sense the 2 nd batch is repeating work from 1 st batch we must try to afford some storage space  To get storage space we must keep every δ X + and δ X -.  Repeat for each updated batch, so that every update has resources stored from the previous batch 43

44. Conclusions  Real-world databases get updated constantly, therefore the knowledge extracted from them changes too.  The authors proposed DELI algorithm to determine if the change is significant so that it knows when to update the extracted association rules.  The algorithm applies sampling techniques and statistic methods to efficiently estimate an approximate large itemsets. 44

45. Final Exam Questions  Q1: Compare and contrast FUP 2 and DELI Both algorithms are used in Association Analysis; Goal: DELI decides when to update the association rules while FUP 2 provides an efficient way of updating them; Technique: DELI scans a small portion of the database (sample) and approximates the large itemsets whereas FUP 2 scans the whole database and returns the large itemsets exactly; DELI saves machine resources and time. 45

46. Final Exam Questions  Q2: Difference Estimation for Large Itemsets  Q3 Difference between Apriori and FUP 2 : Apriori scans the whole database to find association rules, and does not use old data mining results; For most itemsets, FUP 2 scans only the updated part of the database and takes advantage of the old association analysis results. 46

47. Thank you! Now it is discussion time! 47