1. 1 Classification Using Statistically Significant Rules Sanjay Chawla School of IT University of Sydney (joint work with Florian Verhein and Bavani Arunasalam)

2. 2 Overview Data Mining Tasks Associative Classifiers Support and Confidence for Imbalanced Datasets The use of exact tests to mine rules Experiments Conclusion

3. 3 Data Mining Data Mining research has settled into an equilibrium involving four tasks Pattern Mining (Association Rules) Classification Clustering Anomaly or Outlier Detection Associative Classifier DB ML

4. 4 Outlier Detection Outlier Detection (Anomaly Detection) can be studied from two aspects –Unsupervised Nearest Neighbor or K-Nearest Neighbor Problem –Supervised Classification for imbalanced data set –Fraud Detection –Medical Diagnosis

5. 5 Association Rule Mining In terms of impact nothing rivals association rule mining within the data mining community –SIGMOD 93 (~4100 citations) Agrawal, Imielinski, Swami –VLDB 94 (~4900 Citations) Agrawal, Srikant –C4.5 93 (~7000 citations) Ross Quinlan –Gibbs Sampling 84 (IEEE PAMI, ~5000 citations) Geman & Geman –Content Addressable Network (~3000) Ratnasamy, Francis, Hadley, Karp

6. 6 Association Rules (Agrawal, Imielinksi and Swami, 93 SIGMOD) Example: –An implication expression of the form X  Y, where X and Y are itemsets –Example: {Milk, Diaper}  {Beer} Rule Evaluation Metrics –Support (s) Fraction of transactions that contain both X and Y –Confidence (c) Measures how often items in Y appear in transactions that contain X From “Introduction to Data Mining”, Tan,Steinbach and Kumar

7. 7 Association Rule Mining (ARM) Task Given a set of transactions T, the goal of association rule mining is to find all rules having –support ≥ minsup threshold –confidence ≥ minconf threshold Brute-force approach: –List all possible association rules –Compute the support and confidence for each rule –Prune rules that fail the minsup and minconf thresholds  Computationally prohibitive!

8. 8 Mining Association Rules Two-step approach: 1.Frequent Itemset Generation –Generate all itemsets whose support  minsup 2.Rule Generation –Generate high confidence rules from each frequent itemset, where each rule is a binary partitioning of a frequent itemset Frequent itemset generation is still computationally expensive

9. 9 Frequent Itemset Generation Given d items, there are 2 d possible candidate itemsets From “Introduction to Data Mining”, Tan,Steinbach and Kumar

10. 10 Reducing Number of Candidates Apriori principle: –If an itemset is frequent, then all of its subsets must also be frequent Apriori principle holds due to the following property of the support measure: –Support of an itemset never exceeds the support of its subsets –This is known as the anti-monotone property of support

11. 11 Found to be Infrequent Illustrating Apriori Principle Pruned supersets From “Introduction to Data Mining”, Tan,Steinbach and Kumar

12. 12 Classification using ARM TIDItemsGender 1Bread, MilkF 2Bread, Diaper, Beer, EggsM 3Milk Diaper, Beer, CokeM 4Bread, Milk, Diaper, BeerM 5Bread, Milk, Diaper, CokeF In a Classification task we want to predict the class label (Gender) using the attributes A good (albeit stereotypical) rule is {Beer,Diaper}  Male whose support is 60% and confidence is 100%

13. 13 Classification Based on Association (CBA): Liu, Hsu and Ma (KDD 98) Mine association rules of the form A  c on the training data Prune or Select rules using a heuristic Rank rules –Higher confidence; higher support; smallest antecedent New data is passed through the ordered rule set –Apply first matching rule or variation thereof

14. 14 Several Variations AcronymAuthorsForumComments CMARLi, Han, PeiICDM 01Use Chi^2 --Antonie, Zaiane DMKD 04Pos and Neg rules FARMERCong et. alSIGMOD 04row enumeration Top-KCong et. alSIGMOD 05Limit no. of rules CCCSArunasalam et. Al. SIGKDD 06Support-free

15. 15 Downsides of Support (1) High support does not necessarily mean the rule is statistically significant. –Meggido and Srikant (KDD 98) claim that high support & confidence filter out non-significant rules. –However the null hypothesis is that true support s = minsup –Alternative hypothesis is s > minsup –Assumes rules with high support are significant (provides no evidence that they are)

16. 16 Downside of Support (2) High support does not mean good classification performance: –Many good rules have low support! –Evidenced by requirement of low support in CBA, CMAR, etc.

17. 17 Downsides of Support (3) Support is biased towards the majority class –Eg: classes = {yes, no}, sup({yes})=90% –minSup > 10% wipes out any rule predicting “no” –Suppose X  no has confidence 1 and support 3%. Rule discarded if minSup > 3% even though it perfectly predicts 30% of the instances in the minority class! In summary, support has many downsides –especially for classification.

18. 18 Downside of Confidence(1) 20525 70575 9010100 Conf(A  C) = 20/25 = 0.8 Support(A  C) = 20/100 = 0.2 Correlation between A and C: Thus, when the data set is imbalanced a high support and high confidence rule may not necessarily imply that the antecedent and the consequent are positively correlated.

19. 19 Downside of Confidence (2) Reasonable to expect that for “good rules” the antecedent and consequent are not independent! Suppose –P(Class=Yes) = 0.9 –P(Class=Yes|X) = 0.9

20. 20 Complement Class Support(CCS) The following are equivalent for a rule A  C 1.A and C are positively correlated 2.The support of the antecedent(A) is less than CCS(A  C) 3.Conf(A  C) is greater than the support of Consequent(C)

21. 21 Downsides of Confidence (3) Another useful observation Higher confidence (support) for a rule in the minority class implies higher correlation, and lower correlation in the minority class implies lower confidence, but neither of these apply for the majority class. Confidence (support) tends to bias the majority class.

22. 22 Statistical Significant Rules Support is a computationally efficient measure (anti-monotonic) –Tendency to “force” a statistical interpretation on support Lets start with a statistically correct approach –And “force” it to be computationally efficient

23. 23 Exact Tests Let the class variable be {0,1}. Suppose we have two rules X  1 and Y  1 We want to determine if the two rules are different, i.e., they have different effect on “causing” or ‘associating” with 1 or 0 –e.g., medicine and placebo

24. 24 Exact Tests We assume X and Y are binomial random variables with the same parameter p We want to determine, the probability that a specific table instance occurs “purely by chance” Table[a,b;c,d]

25. 25 Exact Tests We can calculate the exact probability of a specific table instance without resorting to asymptotic approximations. This can be used to calculate the p-value of [a,b;c,d]

26. 26 Fisher Exact Test Given a table, [a,b;c,d], Fisher Exact Test will find the probability (p-value) of obtaining the given table or a more positively associated table under the assumption that X and Y come from the same distribution.

27. 27 Forcing “anti-monotonic” We test a rule X  1 against all its immediate generalizations {X-z  1; z in X} The The rule is significant if –P value < significance level (typically 0.05) Use a bottom up approach and only test rules whose immediate generalizations are significant Webb[06] has used Fisher Exact tests for generic association rule mining

28. 28 Example Suppose we have already determined that the rules (A = a1)  1 and (A = a2)  1 are significant. Now we want to test if –X=(A =a1) ^ (A=a2)  1 is significant Then we carry out a FET on X and X –{A=a1} and X and X-{A=a2}. If the minimum of their p-value is less than the significance level we keep the X  1 rule, otherwise we discard it.

29. 29 Contigency Table

30. 30 Ranking Rules We have already observed that –high confidence rule for the majority class may be “more” negatively correlated than the same rule predicting the other class –A high positively correlated rule that predicts the minority class may have a lower confidence than the same rule predicting the other class

31. 31 Experiments: Random Dataset Attributes independent and uniformly distributed. Makes no sense to find any rules – other than by chance However minSup=1% and minConf=0.5 mines 4149/13092 -- over 31% of all possible rules. Using our FET technique with standard significance level we find only 11 (0.08%)

32. 32 Experiments: Balanced Dataset Similar performance (within 1%)

33. 33 Experiments: Balanced Dataset But mines only 0.06% the number of rules By searching only 0.5% of the search space And using only 0.4% of the time.

34. 34 Experiments: Imbalanced Dataset Higher performance than support-confidence techniques Using 0.07% of the search space and time, 0.7% the number of rules.

35. 35 Contributions Strong evidence and arguments against the use of support and confidence for imbalanced classification Simple technique for using Fisher’s Exact test for finding positively associated and statistically significant rules. –Uses on average 0.4% of the time, searches only 0.5% of the search space, finds only 0.06% of the rules as support-confidence techniques. –Similar performance on balanced datasets, higher on imbalanced datasets. Parameter free (except for significance level)

36. 36 References Verhein and Chawla, Classification Using Statistically Significant Rules http://www.it.usyd.edu.au/~chawla Arunasalam and Chawla, CCCS: A Top- Down Associative Classifier for Imbalanced Class Distribution [ACM SIGKDD 2006; pp 517-522]