Mining Multiple-level Association Rules in Large Databases IEEE Transactions on Knowledge and Data Engineering, 1999 Authors : Jiawei Han Simon Fraser University, British Columbia. Yongjian Fu University of Missouri-Rolla, Missouri. Presented by Michael Johnson 1 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness ● Optimizations 3. Conclusions/Future Work 4. Exam Questions 2 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness ● Optimizations 3. Conclusions/Future Work 4. Exam Questions 3 What is MLDM? What's the difference between the following rules: Rule A →70% of customers who bought diapers also bought beer Rule B →45% of customers who bought cloth diapers also bought dark beer Rule C →35% of customers who bought Pampers also bought Samuel Adams beer. 4 What is MLDM? Rule A applies at a generic higher level of abstraction (product) Rule B applies at a more specific level of abstraction (category) Rule C applies at the lowest level of abstraction (brand). This process is called drilling down. 5 What Do We Gain? Concrete rules allow for More targeted marketing New marketing strategies More concrete relationships 6 Hierarchy Types Generalization/Specialization (is a relationships) Is a With Multiple Inheritance Whole-Part hierarchies (is-part-of; has-part) 7 Is A Relationship Generalization to Specialization Vehicle 4-Wheels Sedan SUV 2-Wheels Bicycle Motorcycle 8 Is A With Multiple Inheritance Vehicle Commuting Car Bicycle Recreational Snowmobile 9 Whole-Part Hierarchies Computer Motherboard RAM CPU Hard Drive RW Head Platter 10 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness 3. Conclusions/Future Work 4. Exam Questions 11 MLAR: Main Goal As usual we are trying to develop a method to extract non-trivial, interesting, and strong rules from our transactional database. A method which: Avoids trivial rules (Milk→Bread) Common sense Avoids coincidental rules (Toy→Milk) Low support 12 What Do We Need? 1. Data Representation At Different Levels Of Abstraction ● Explicitly stored in databases ● Provided by experts or users ● Generated via clustering (OLAP) 2. Efficient Methods for ML Rule Mining (focus of this paper) 13 Possible Methods Apply single-level Apriori Algorithm to each of the multiple levels under the same miniconf and minsup. Potential Problems? Higher Levels of abstraction will naturally have higher support, support decreases as we drill down What is the optimal minsup for all levels? Too high a minsup → too few itemsets for lower levels Too low a minsup → too many uninteresting rules 14 Possible Solutions Adapt minsup for each level Adapt minconf for each level Do both This paper studies a progressive deepening method developed by extension of the Apriori Algorithm, focused on minsup 15 Assumption If an item is non-frequent at one level, its descendants no longer figure in further analysis. Explore only descendants of frequent items as we drill down Are there problems that arise from making these assumptions? 16 Problems May eliminate possible interesting rules for itemsets at one level whose ancestors are not frequent at higher levels. If so, can be addressed by 2 workarounds 2 minsup values at higher levels – one for filtering infrequent items, the other for passing down frequent items to lower levels; latter called level passage threshold (lph) The lph may be adjusted by user to allow descendants of sub frequent items 17 Differences From Previous Research Other studies use same minsup across different levels of the hierarchy This study…. Uses different minsup values at different levels of the hierarchy Analyzes different optimization techniques Studies the use of interestingness measures 18 Requirements Transactional database must contain: 1. Item dataset containing item description: {<Ai>,<description>} 2. A transaction dataset T containing set of transactions.. {Tid*,{Ap…Aq}} *Tid is a transaction identifier (key) 19 Algorithm Flow At Level 1: Generate frequent itemsets Get table filtered for frequent itemsets T[2] At subsequent levels: Generate candidate subsets using Apriori Calculate support for generated candidates Union 'passing' subsets with existing rule set Repeat until no additional rules are generated, or desired level is reached 20 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness ● Optimizations 3. Conclusions/Future Work 4. Exam Questions 21 Definitions A pattern or an itemset A is one item Ai or a set of conjunctive items Ai Λ …. Λ Aj The support of a pattern is the number of transactions that contain A vs. the total number of transactions σ(A|S) The confidence of a rule A → B in S is given by: φ(A→B) = σ(AUB)/σ(A) (i.e. conditional probability) Specify 2 thresholds: minsup(σ’) and miniconf (φ’); different values at different levels 22 Definitions A pattern A is frequent in set S at if: the support of A is no less than the corresponding minimum support threshold σ’ A rule A → B is strong for a set S, if: a. each ancestor of every item in A and B is frequent at its corresponding level b. c. A Λ B is frequent at the current level and φ(A→B)≥ φ’ (miniconf criteria) This ensures that the patterns examined at the lower levels arise from itemsets that have a high support at higher levels 23 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness 3. Conclusions/Future Work 4. Exam Questions 24 Example: Taxonomy food milk Level 1: Level 2: Level 3: Dairyland bread chocolate 2% Foremost ... white ... Old Mills Wonder wheat ... ... Generalized ID System: 2% Foremost Milk Coded as GID:112 (1st item in Level 1, 1st item in level 2, 2nd item in level 3) 25 Example: Dataset Table 1: A sales-transaction Table Trans-id Bar_code_set 351428 {17325, 92108, ….} 653234 {23423, 56432,…} Table 2: A sales_item Description Relation Bar_code GID Category Brand Content Size Price 17325 112 milk Foremost 2% 1 Gal $3.31 ….. …… ….. ….. ….. ….. 26 Example: Preprocessing Join sales_transaction table to sales_item table to produce encoded transaction table T[1]: TID Items T1 T2 T3 T4 T5 T6 T7 {111 , 121 , 211 , 221} {111 , 211 , 222 , 323} {112 , 122 , 221 , 411} {111 , 121} {111 , 122 , 211 , 221 , 413} {211 , 323 , 524} {323 , 411 , 524 , 713} 27 Example: Step 1 Find Level-1 frequent itemsets Minsup = 4 T[1] Level-1 Frequent-1 Itemsets L[1,1] TID Items T1 {111 , 121 , 211 , 221} Itemset Support T2 {111 , 211 , 222 , 323} {1**} 5 T3 {112 , 122 , 221 , 411} {2**} 5 T4 {111 , 121} T5 {111 , 122 , 211 , 221 , 413} Level-1 Frequent 2-Itemsets T6 {211 , 323 , 524} L[1,2] T7 {323 , 411 , 524 , 713} Itemset Support {1**,2**} 4 28 Example: Step 2 Create T[2] by filtering T[1] w/ L[1,1] T[1] Filtered T[2] TID Items TID Items T1 {111 , 121 , 211 , 221} T1 {111 , 121 , 211 , 221} T2 {111 , 211 , 222 , 323} T2 {111 , 211 , 222} T3 {112 , 122 , 221 , 411} T3 {112 , 122 , 221} T4 {111 , 121} T4 {111 , 121} T5 {111 , 122 , 211 , 221 , 413} T5 {111 , 122 , 211 , 221} T6 {211 , 323 , 524} T6 {211} T7 {323 , 411 , 524 , 713} Itemset Support {1**} 5 {2**} 5 29 L[2,1] Example: Step 3 Find Level-2 Frequent Itemsets Minsup = 3 Filtered T[2] Itemset Support {11*} 5 {12*} 4 {21*} 4 {22*} 4 L[2,2] Itemset Support TID Items {11*, 12*} 4 T1 {111 , 121 , 211 , 221} {11*, 21*} 3 T2 {111 , 211 , 222} {11*,22*} 4 T3 {112 , 122 , 221} {12*, 22*} 3 T4 {111 , 121} {21*, 22*} 3 T5 {111 , 122 , 211 , 221} T6 {211} L[2,3] Itemset Support {11*, 12*, 22*} 3 {11*, 21*, 22*} 3 30 Example: Step 4 Find Level-3 Frequent Itemsets Minsup = 3 Filtered T[2] L[3,1] Itemset Support {111} 4 {211} 4 {221} 3 TID Items T1 {111 , 121 , 211 , 221} T2 {111 , 211 , 222} T3 {112 , 122 , 221} T4 {111 , 121} Itemset Support T5 {111 , 122 , 211 , 221} {111, 211*} 3 T6 {211} L[3,2] Stop: Lowest Level Reached 31 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness ● Optimizations 3. Conclusions/Future Work 4. Exam Questions 32 Are All Of The Strong Rules Interesting? MLDM creates unique challenges for rule pruning The paper defines two filters for interesting rules: 1. Removal of redundant rules 2. Removal of unnecessary rules 33 Redundant Rules Consider a strong rule at Level 1: Milk→Bread food milk chocolate 2% Dairyland bread Foremost ... white ... Old Mills Wonder wheat ... ... This rule is likely to have descendent rules which may or may not contain additional information, even if they met our minconf and minsup criteria at that level: 2% Milk→Wheat Bread, 2% Milk→White Bread, Chocolate Milk→Wheat Bread We need a way to distinguish between rules that add information and those that are redundant 34 Redundant Rules A rule is redundant if the confidence for a rule falls in a certain range and the items in the rule are descendents of a different rule. 35 Redundant Rules Applying Redundant Rule reduction eliminates 40-70% of discovered Strong Rules 36 Unnecessary Rules Consider the following rules: R: Milk→Bread (minsup = 80%) R': Milk, Butter → Bread (minsup = 80%) How much additional information do we gain from the R'? MLDM can produced very complex rules that meet our minsup and minconf criteria, but do contain much unique/useful information. We need a way to distinguish between rules that add information and those that are Unnecessary 37 Unnecessary Rule A rule R is unnecessary if there is a simpler rule R' and φ(R) is within a given range of φ(R') 38 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness ● Optimizations 3. Conclusions/Future Work 4. Exam Questions 39 Hardware Setup Hardware: Sun Microsystems SPARCstation 20 1. 32MB RAM 2. 100Mhz Clock 3. CLI 40 Algorithm Optimizations Authors proposed 3 iterations of the original algorithm 1)ML_T1LA Use only one encoded table T[1] ● 2)ML_TML1 Generate T[1], T[2], … T[n+1] ● 3)ML_T2LA Uses T[2], but calculates down level support with a single scan ● 41 ML_T1LA Instead of generating T[2] from T[1], ML_T1LA algorithm generates support for all levels of hierarchy in a single scan from T[1] Pros: Avoids Limits generation of new transaction table number of scans to the size of the largest transaction Cons: Scanning T[1] requires scanning all items, even infrequent ones Performance may suffer for DB w/ many infrequent itemsets Large memory required (32MB RAM = page swapping!!!) 42 ML_TML1 Instead of using only T[2] for rule mining, ML_TML1 algorithm generates a table for each level, using L[i,1] to filter T[i] and create T[i+1] Pros: Saves significant processing time if only a small portion of the data is frequent at each level Allows for creation of T[i] and L[i,1] in parallel ● Cons: May not be efficient if only a small number of items is filtered at each level 43 ML_T2LA Like the base algorithm, ML_T2LA creates T[2] table from the frequent itemsets in T[1]. However, it allows for parallel creation of L[i,k]. Pros: Saves time by limiting the number of scans Cons: May not be efficient if only a small number of items is filtered at each level 44 Experimental Results While the figures show that T2LA is best for most of the time, the authors preferred ML_T1LA 45 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness ● Optimizations 3. Conclusions/Future Work 4. Exam Questions 46 Conclusions This paper demonstrated: Extending association rules from single-level to multiplelevel. A top-down progressive deepening technique for mining multiple-level association rules. Filtering of uninteresting association rules Performance optimization techniques (not covered) 47 Future Work Develop efficient algorithms for mining multiple-level sequential patterns Cross-level associations Improve interestingness of rules 48 Outline What is MLAR? 1. ● Concepts ● Motivation A Method For Mining M-L Association Rules 2. ● Problems/Solutions ● Definitions ● Algorithm Example ● Interestingness ● Optimizations 3. Conclusions/Future Work 4. Exam Questions 49 Exam Question 1 What is a major drawback to multiple-level data mining using the same minsup at all levels of a concept hierarchy? A. Large support exists at higher levels of the hierarchy; smaller support at lower levels. In order to insure that sufficiently strong association rules are generated at the lower levels, we must reduce the support at higher levels which, in turn, could result in generation of many uninteresting rules at higher levels. Thus we are faced with the problem of determining which is the optimal minsup at all levels Q. 50 Exam Question 2 Q. Give an example of a multiple level association rule A. High level: 80% of people who buy cereal also buy milk Low Level: 25% of people who buy Cheerios cereal buy Hood 2% Milk 51 Exam Question 3 Q. There were 3 examples of hierarchy types in multiple level rule mining. Pick one and draw an example Is-A Is-A Vehicle 4-Wheels Whole-Part Computer Multiple Inheritance Motherboard 2-Wheels Hard Drive Vehicle Sedan SUV Bike Motorcycle Commuting Car Bicycle Recreational RAM CPU RW Head Platter Snowmobile 52