DATA MINING

LECTURE 3
Frequent Itemsets and Association Rules
This is how it all started…
• Rakesh Agrawal, Tomasz Imielinski, Arun N. Swami:
Mining Association Rules between Sets of Items in
Large Databases. SIGMOD Conference 1993: 207-
216
• Rakesh Agrawal, Ramakrishnan Srikant: Fast
Algorithms for Mining Association Rules in Large
Databases. VLDB 1994: 487-499

• These two papers are credited with the birth of Data

Mining
• For a long time people were fascinated with
Association Rules and Frequent Itemsets
• Some people (in industry and academia) still are.
 3

Market-Basket Data
• A large set of items, e.g., things sold in a
supermarket.
• A large set of baskets, each of which is a small
subset of the items, e.g., the things one customer
buys on one day.
Items: {Bread, Milk, Diaper, Beer, Eggs, Coke}
TID Items
1 Bread, Milk
Baskets: Transactions
2 Bread, Diaper, Beer, Eggs
3 Milk, Diaper, Beer, Coke
4 Bread, Milk, Diaper, Beer
5 Bread, Milk, Diaper, Coke
 4

Frequent itemsets
• Goal: find combinations of items (itemsets) that
occur frequently
• Called Frequent Itemsets

Support 𝑠 𝐼 : number of
TID Items
transactions that contain
1 Bread, Milk
itemset I
2 Bread, Diaper, Beer, Eggs Examples of frequent itemsets 𝑠 𝐼 ≥ 3
3 Milk, Diaper, Beer, Coke {Bread}: 4
4 Bread, Milk, Diaper, Beer {Milk} : 4
5 Bread, Milk, Diaper, Coke {Diaper} : 4
{Beer}: 3
{Diaper, Beer} : 3
{Milk, Bread} : 3
 5

Market-Baskets – (2)
• Really, a general many-to-many mapping
(association) between two kinds of things, where the
one (the baskets) is a set of the other (the items)
• But we ask about connections among “items,” not “baskets.”

• The technology focuses on common/frequent events,

not rare events (“long tail”).
 6

Applications – (1)
• Items = products; baskets = sets of products
someone bought in one trip to the store.

• Example application: given that many people buy

beer and diapers together:
• Run a sale on diapers; raise price of beer.
• Only useful if many buy diapers & beer.
 7

Applications – (2)
• Baskets = Web pages; items = words.

• Example application: Unusual words appearing

together in a large number of documents, e.g.,
“Brad” and “Angelina,” may indicate an interesting
relationship.
 8

Applications – (3)
• Baskets = sentences; items = documents
containing those sentences.

• Example application: Items that appear together

too often could represent plagiarism.
• Notice items do not have to be “in” baskets.
Definitions
• Itemset TID Items
• A collection of one or more items 1 Bread, Milk
• Example: {Milk, Bread, Diaper} 2 Bread, Diaper, Beer, Eggs
3 Milk, Diaper, Beer, Coke
• k-itemset
4 Bread, Milk, Diaper, Beer
• An itemset that contains k items
5 Bread, Milk, Diaper, Coke
• Support (s)
• Count: Frequency of occurrence of an itemset
• E.g. s({Milk, Bread,Diaper}) = 2
• Fraction: Fraction of transactions that contain an itemset
• E.g. s({Milk, Bread, Diaper}) = 40%

• Frequent Itemset
• An itemset 𝐼 whose support is greater than or equal to a
minsup threshold, 𝑠 𝐼 ≥minsup
Mining Frequent Itemsets task
• Input: Market basket data, threshold minsup
• Output: All frequent itemsets with support ≥ minsup

• Problem parameters:
• N (size): number of transactions
• Wallmart: billions of baskets per year
• Web: billions of pages
• d (dimension): number of (distinct) items
• Wallmart sells more than 100,000 items
• Web: billions of words
• w: max size of a basket
• M: Number of possible itemsets.
M = 2𝑑
 The itemset lattice
null Representation of all possible
itemsets and their relationships

A B C D E

AB AC AD AE BC BD BE CD CE DE

ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

ABCD ABCE ABDE ACDE BCDE

Given d items, there are 2d

ABCDE possible itemsets
 A Naïve Algorithm
• Brute-force approach: Every itemset is a candidate :
• Consider all itemsets in the lattice, and scan the data for each candidate to
compute the support
• Time Complexity ~ O(NMw) , Space Complexity ~ O(d)
• OR
• Scan the data, and for each transaction generate all possible itemsets. Keep
a count for each itemset in the data.
• Time Complexity ~ O(N2w) , Space Complexity ~ O(M)

• Expensive since M = 2d !!!

• No solution that considers all candidates is acceptable!

Transactions List of
Candidates
TID Items
1 Bread, Milk
2 Bread, Diaper, Beer, Eggs
N 3 Milk, Diaper, Beer, Coke M
4 Bread, Milk, Diaper, Beer
5 Bread, Milk, Diaper, Coke
w
 13

Computation Model
• Typically, data is kept in flat files rather than in a
database system.
• Stored on disk.
• Stored basket-by-basket.
• We can expand a basket into pairs, triples, etc. as we read
the data.
• Use k nested loops, or recursion to generate all itemsets of size k.

• Data is too large to be loaded in memory.

Example file: retail
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
30 31 32
33 34 35
36 37 38 39 40 41 42 43 44 45 46
38 39 47 48
38 39 48 49 50 51 52 53 54 55 56 57 58 Example: items are positive integers,
32 41 59 60 61 62
3 39 48
and each basket corresponds to a line in
63 64 65 66 67 68 the file of space-separated integers
32 69
48 70 71 72
39 73 74 75 76 77 78 79
36 38 39 41 48 79 80 81
82 83 84
41 85 86 87 88
39 48 89 90 91 92 93 94 95 96 97 98 99 100 101
36 38 39 48 89
39 41 102 103 104 105 106 107 108
38 39 41 109 110
39 111 112 113 114 115 116 117 118
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
48 134 135 136
39 48 137 138 139 140 141 142 143 144 145 146 147 148 149
39 150 151 152
38 39 56 153 154 155
 15

Computation Model – (2)

• The true cost of mining disk-resident data is
usually the number of disk I/O’s.
• In practice, association-rule algorithms read the
data in passes – all baskets read in turn.

• Thus, we measure the cost by the number of

passes an algorithm takes.
 16

Main-Memory Bottleneck
• For many frequent-itemset algorithms, main
memory is the critical resource.
• As we read baskets, we need to count something, e.g.,
occurrences of pairs.
• The number of different things we can count is limited
by main memory.
• Swapping counts in/out is too slow
The Apriori Principle
• Apriori principle (Main observation):
– If an itemset is frequent, then all of its subsets must also
be frequent
– If an itemset is not frequent, then all of its supersets
cannot be frequent
– The support of an itemset never exceeds the support of
its subsets
∀𝑋, 𝑌: 𝑋 ⊆ 𝑌 ⇒ 𝑠 𝑋 ≥ 𝑠(𝑌)

– This is known as the anti-monotone property of support

Illustration of the Apriori principle
Frequent
subsets

Found to be frequent
 Illustration of the Apriori principle
null

A B C D E

AB AC AD AE BC BD BE CD CE DE

Found to be
Infrequent
ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

ABCD ABCE ABDE ACDE BCDE

Infrequent supersets
ABCDE
Pruned
 The Apriori algorithm
Ck = candidate itemsets of size k
Level-wise approach Lk = frequent itemsets of size k

1. k = 1, C1 = all items
2. While Ck not empty
Frequent 3. Scan the database to find which itemsets in
itemset
generation Ck are frequent and put them into Lk
Candidate 4. Generate the candidate itemsets Ck+1 of
generation size k+1 using Lk
5. k = k+1
R. Agrawal, R. Srikant: "Fast Algorithms for Mining Association Rules",
Proc. of the 20th Int'l Conference on Very Large Databases, 1994.
Candidate Generation
• Apriori principle:
• An itemset of size k+1 is candidate to be frequent only if
all of its subsets of size k are known to be frequent

Candidate generation:
• Construct a candidate of size k+1 by combining
frequent itemsets of size k
• If k = 1, take the all pairs of frequent items
• If k > 1, join pairs of itemsets that differ by just one item
• For each generated candidate itemset ensure that all
subsets of size k are frequent.
 Generate Candidates Ck+1
• Assumption: The items in an itemset are ordered
• Integers ordered in increasing order, strings ordered lexicographicly
• The order ensures that if item y > x appears before x, then x is not in
the itemset
• The itemsets in Lk are also ordered

Create a candidate itemset of size k+1, by joining

two itemsets of size k, that share the first k-1 items
Item 1 Item 2 Item 3
1 2 3
1 2 5
1 4 5
 Generate Candidates Ck+1
• Assumption: The items in an itemset are ordered
• Integers ordered in increasing order, strings ordered in lexicographicly
• The order ensures that if item y > x appears before x, then x is not in
the itemset
• The itemsets in Lk are also ordered

Create a candidate itemset of size k+1, by joining

two itemsets of size k, that share the first k-1 items
Item 1 Item 2 Item 3
1 2 3
1 2 3 5
1 2 5
1 4 5
 Generate Candidates Ck+1
• Assumption: The items in an itemset are ordered
• Integers ordered in increasing order, strings ordered in lexicographicly
• The order ensures that if item y > x appears before x, then x is not in
the itemset
• The itemsets in Lk are also ordered

Create a candidate itemset of size k+1, by joining

two itemsets of size k, that share the first k-1 items
Item 1 Item 2 Item 3
Are we missing something?
1 2 3
What about this candidate?
1 2 5
1 2 4 5
1 4 5
Generating Candidates Ck+1 in SQL

• self-join Lk ‫‏‬
insert into Ck+1
select p.item1, p.item2, …, p.itemk, q.itemk
from Lk p, Lk q
where p.item1=q.item1, …, p.itemk-1=q.itemk-1, p.itemk < q.itemk
Example
• L3={abc, abd, acd, ace, bcd}
• Generating candidate set C4
• Self-join: L3*L3

item1 item2 item3 item1 item2 item3

a b c a b c
a b d a b d
a c d a c d
a c e a c e
b c d b c d

p.item1=q.item1,p.item2=q.item2, p.item3< q.item3

Example
• L3={abc, abd, acd, ace, bcd}
• Generating candidate set C4
• Self-join: L3*L3
item1 item2 item3 item1 item2 item3
a b c a b c
a b d a b d
a c d a c d
a c e a c e
b c d b c d

p.item1=q.item1,p.item2=q.item2, p.item3< q.item3

 Example
• L3={abc, abd, acd, ace, bcd}
• Generating candidate set C4
• Self-join: L3*L3 C4 ={abcd}
item1 item2 item3 item1 item2 item3
a b c a b c
a b d a b d
{a,b,c} {a,b,d}
a c d a c d
a c e a c e
{a,b,c,d}
b c d b c d

p.item1=q.item1,p.item2=q.item2, p.item3< q.item3

 Example
• L3={abc, abd, acd, ace, bcd}
• Generating candidate set C4
• Self-join: L3*L3 C4 ={abcd
acde}
item1 item2 item3 item1 item2 item3
a b c a b c
a b d a b d {a,c,d} {a,c,e}
a c d a c d
a c e a c e {a,c,d,e}
b c d b c d

p.item1=q.item1,p.item2=q.item2, p.item3< q.item3

Illustration of the Apriori principle
TID Items
1 Bread, Milk

minsup = 3 2
3
Bread, Diaper, Beer, Eggs
Milk, Diaper, Beer, Coke
Item Count Items (1-itemsets) 4 Bread, Milk, Diaper, Beer
Bread 4 5 Bread, Milk, Diaper, Coke
Coke 2
Milk 4 Itemset Count Pairs (2-itemsets)
Beer 3
{Bread,Milk} 3
Diaper 4
{Bread,Beer} 2 (No need to generate
Eggs 1
{Bread,Diaper} 3 candidates involving Coke
{Milk,Beer} 2 or Eggs)
{Milk,Diaper} 3
{Beer,Diaper} 3
Triplets (3-itemsets)
If every subset is considered,
6 6 6 Itemset Count
+ + = 6 + 15 + 20 = 41 {Bread,Milk,Diaper} 2
1 2 3
With support-based pruning,
6 4 Only this triplet has all subsets to be frequent
+ + 1 = 6 + 6 + 1 = 13
1 2 But it is below the minsup threshold
Generate Candidates Ck+1
• Are we done? Are all the candidates valid?

Item 1 Item 2 Item 3

1 2 3
1 2 3 5
1 2 5
1 4 5
Is this a valid candidate?

No. Subsets (1,3,5) and (2,3,5) should also be frequent

Apriori principle
• Pruning step:
• For each candidate (k+1)-itemset create all subset k-itemsets
• Remove a candidate if it contains a subset k-itemset that is
not frequent
 Example
{a,b,c} {a,b,d}
• L3={abc, abd, acd, ace, bcd}
• Self-joining: L3*L3 {a,b,c,d}

– abcd from abc and abd

abc abd acd bcd
– acde from acd and ace
   
• C4={abcd, acde}
{a,c,d} {a,c,e}
• Pruning:
– abcd is kept since all subset itemsets are {a,c,d,e}
in L3
– acde is removed because ade is not in L3 acd ace ade cde
  X
• C4={abcd}
Example II
Itemset Count
{Beer,Diaper} 3
{Bread,Diaper} 3
{Bread,Milk} 3
{Diaper, Milk} 3
Itemset
{Bread,Diaper,Milk}
Itemset Count
{Beer,Diaper} 3
{Bread,Diaper} 3
{Bread,Milk} 3 {Bread,Diaper} 
{Diaper, Milk} 3
{Bread,Milk} 
{Diaper, Milk} 
Generate Candidates Ck+1
• We have all frequent k-itemsets Lk
• Step 1: self-join Lk
• Create set Ck+1 by joining frequent k-itemsets that
share the first k-1 items
• Step 2: prune
• Remove from Ck+1 the itemsets that contain a subset
k-itemset that is not frequent
Computing Frequent Itemsets
• Given the set of candidate itemsets Ck, we need to compute
the support and find the frequent itemsets Lk.
• Scan the data, and use a hash structure to keep a counter
for each candidate itemset that appears in the data

Transactions Hash Structure

Ck
TID Items
1 Bread, Milk
2 Bread, Diaper, Beer, Eggs
N 3 Milk, Diaper, Beer, Coke k
4 Bread, Milk, Diaper, Beer
5 Bread, Milk, Diaper, Coke
Buckets
A simple hash structure
• Create a dictionary (hash table) that stores the
candidate itemsets as keys, and the number of
appearances as the value.
• Initialize with zero
• Increment the counter for each itemset that you
see in the data
 Key Value
Example {3 6 7} 0
{3 4 5} 1
{1 3 6} 3
Suppose you have 15 candidate
itemsets of length 3: {1 4 5} 5
{2 3 4} 2
{1 5 9} 1
C3 = {
{3 6 8} 0
{1 4 5}, {1 2 4}, {4 5 7}, {1 2 5}, {4 5 8},
{4 5 7} 2
{1 5 9}, {1 3 6}, {2 3 4}, {5 6 7}, {3 4 5},
{6 8 9} 0
{3 5 6}, {3 5 7}, {6 8 9}, {3 6 7}, {3 6 8} {5 6 7} 3
} {1 2 4} 8
{3 5 7} 1
Hash table stores the counts of the {1 2 5} 0
candidate itemsets as they have been {3 5 6} 1
computed so far
{4 5 8} 0
 Key Value
Example {3 6 7} 0
{3 4 5} 1
{1 3 6} 3
A new tuple {1,2,3,5,6} generates the
following itemsets of length 3: {1 4 5} 5
{2 3 4} 2
{1 5 9} 1
{1 2 3}, {1 2 5}, {1 2 6}, {1 3 5}, {1 3 6},
{3 6 8} 0
{1 5 6}, {2 3 5}, {2 3 6}, {3 5 6},
{4 5 7} 2
{6 8 9} 0
Increment the counters for the itemsets {5 6 7} 3
in the dictionary
{1 2 4} 8
{3 5 7} 1
{1 2 5} 0
{3 5 6} 1
{4 5 8} 0
 Key Value
Example {3 6 7} 0
{3 4 5} 1
{1 3 6} 4
A new tuple {1,2,3,5,6} generates the
following itemsets of length 3: {1 4 5} 5
{2 3 4} 2
{1 5 9} 1
{1 2 3}, {1 2 5}, {1 2 6}, {1 3 5}, {1 3 6},
{3 6 8} 0
{1 5 6}, {2 3 5}, {2 3 6}, {3 5 6},
{4 5 7} 2
{6 8 9} 0
Increment the counters for the itemsets {5 6 7} 3
in the dictionary
{1 2 4} 8
{3 5 7} 1
{1 2 5} 1
{3 5 6} 2
{4 5 8} 0
The frequent itemset algorithm
Count All pairs Count
All the items of items the pairs
items from L1

C1 Filter L1 Construct C2 Filter L2 Construct C3

First Second
pass pass

Frequent Frequent
items pairs
 41

A-Priori for All Frequent Itemsets

• One pass for each k.
• Needs room in main memory to count each
candidate k -set.
• For typical market-basket data and reasonable
support (e.g., 1%), k = 2 requires the most
memory.
 42

Picture of A-Priori

Item counts Frequent items

Counts of
pairs of
frequent
items

Pass 1 Pass 2
 43

Details of Main-Memory Counting

• Two approaches:
1. Count all pairs, using a “triangular matrix” = one
dimensional array that stores the lower diagonal.
2. Keep a table of triples [i, j, c] = “the count of the pair
of items {i, j } is c.”
• (1) requires only 4 bytes/pair.
• Note: always assume integers are 4 bytes.
• (2) requires 12 bytes/pair, but only for those pairs
with count > 0.
 44

12 per
4 per pair
occurring pair

Method (1) Method (2)

 45

Triangular-Matrix Approach
• Number items 1, 2,…
• Requires table of size O(n) to convert item names to
consecutive integers.
• Count {i, j } only if i < j.
• Keep pairs in the order {1,2}, {1,3},…, {1,n }, {2,3},
{2,4},…,{2,n }, {3,4},…, {3,n },…{n -1,n }.
• Find pair {i, j } at the position
(i –1)(n –i /2) + j – i.

• Total number of pairs n (n –1)/2; total bytes about 2n2.

 46

A-Priori Using Triangular Matrix for Counts

Freq- Old
Item counts
quent item
items #’s

Counts of
pairs of
frequent
items

Pass 1 Pass 2
 47

Details of Approach #2

• Total bytes used is about 12p, where p is the

number of pairs that actually occur.
• Beats triangular matrix if no more than1/3 of possible
pairs actually occur.

• May require extra space for retrieval structure, e.g.,

a hash table.
ASSOCIATION RULES
 Association Rule Mining
• Given a set of transactions, find rules that will predict the
occurrence of an item based on the occurrences of other
items in the transaction

Market-Basket transactions
Example of Association Rules
TID Items
{Diaper}  {Beer},
1 Bread, Milk {Milk, Bread}  {Eggs,Coke},
2 Bread, Diaper, Beer, Eggs {Beer, Bread}  {Milk},
3 Milk, Diaper, Beer, Coke
4 Bread, Milk, Diaper, Beer Implication means co-occurrence,
5 Bread, Milk, Diaper, Coke not causality!
 Mining Association Rules
 Association Rule TID Items
– An implication expression of the form 1 Bread, Milk
X  Y, where X and Y are itemsets 2 Bread, Diaper, Beer, Eggs
– {Milk, Diaper}  {Beer} 3 Milk, Diaper, Beer, Coke
 Rule Evaluation Metrics 4 Bread, Milk, Diaper, Beer
– Support (s) 5 Bread, Milk, Diaper, Coke
 Fraction of transactions that contain both X
and Y = the probability P(X,Y) that X and Y
Example:
occur together {Milk, Diaper}  Beer
– Confidence (c)  (Milk , Diaper, Beer) 2
 How often Y appears in transactions that s   0.4
|T| 5
contain X = the conditional probability P(Y|X)
that Y occurs given that X has occurred.  (Milk, Diaper, Beer) 2
c   0.67
 (Milk , Diaper ) 3
 Problem Definition
– Input: Market-basket data, minsup, minconf values
– Output: All rules with items in I having s ≥ minsup and c≥ minconf
 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 partitioning of a frequent itemset into Left-
Hand-Side (LHS) and Right-Hand-Side (RHS)
Frequent itemset: {A,B,C,D}
E.g., Rule: ABCD
All Candidate rules:
BCD A, ACD B , ABD C, ABC D,
CD AB, BD AC, BC AD, AD  BC, AB CD, AC  BD,
D ABC, C ABD, B ACD, A BCD
Association Rule anti-monotonicity
• In general, confidence does not have an anti-
monotone property with respect to the size of the
itemset:
c(ABC D) can be larger or smaller than c(AB D)

• But confidence is anti-monotone w.r.t. number of

items on the RHS of the rule (or monotone with
respect to the LHS of the rule)

• e.g., L = {A,B,C,D}:

c(ABC  D)  c(AB  CD)  c(A  BCD)

 Rule Generation for Apriori Algorithm
ABCD=>{ } }
ABCD=>{
Low
Confidence
Rule
BCD=>A
BCD=>A ACD=>B
ACD=>B ABD=>C
ABD=>C ABC=>D
ABC=>D

CD=>AB
CD=>AB BD=>AC
BD=>AC BC=>AD
BC=>AD AD=>BC
AD=>BC AC=>BD
AC=>BD AB=>CD
AB=>CD

D=>ABC
D=>ABC C=>ABD
C=>ABD B=>ACD
B=>ACD A=>BCD
A=>BCD
Pruned
Rules
Lattice of rules created by the RHS
Rule Generation for APriori Algorithm
• Candidate rule is generated by merging two rules that
share the same prefix
in the RHS
CD->AB BD->AC

• join(CDAB,BDAC)
would produce the candidate
rule D  ABC

• Prune rule D  ABC if its

subset ADBC does not have D->ABC
high confidence

• Essentially we are doing APriori on the RHS

RESULT
POST-PROCESSING
Compact Representation of Frequent
Itemsets
• Some itemsets are redundant because they have identical
support as their supersets
TID A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 C1 C2 C3 C4 C5 C6 C7 C8 C9 C10
1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
4 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
5 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
6 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0
7 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0
8 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0
9 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0
10 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0
11 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
12 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1

10 
• Number of frequent itemsets  3    
10

k
k 1

• Need a compact representation

Maximal Frequent Itemsets
An itemset is maximal frequent if none of its immediate supersets is
frequent null

Maximal A B C D E
Itemsets
Maximal itemsets = positive border

AB AC AD AE BC BD BE CD CE DE

ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

ABCD ABCE ABDE ACDE BCDE

Infrequent
Itemsets Border
ABCDE

Maximal: no superset has this property

Negative Border
Itemsets that are not frequent, but all their immediate subsets are
frequent. null

A B C D E

AB AC AD AE BC BD BE CD CE DE

ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

ABCD ABCE ABDE ACDE BCDE

Infrequent
Itemsets Border
ABCDE

Minimal: no subset has this property

Border
• Border = Positive Border + Negative Border
• Itemsets such that all their immediate subsets are
frequent and all their immediate supersets are
infrequent.

• Either the positive, or the negative border is

sufficient to summarize all frequent itemsets.
Closed Itemsets
• An itemset is closed if none of its immediate supersets
has the same support as the itemset

TID Items Itemset Support Itemset Support

1 {A,B} {A} 4 {A,B,C} 2
2 {B,C,D} {B} 5 {A,B,D} 3
3 {A,B,C,D} {C} 3 {A,C,D} 2
4 {A,B,D} {D} 4 {B,C,D} 3
{A,B} 4 {A,B,C,D} 2
5 {A,B,C,D}
{A,C} 2
{A,D} 3
{B,C} 3
{B,D} 4
{C,D} 3
Maximal vs Closed Itemsets
null Transaction
Ids
124 123 1234 245 345
TID Items A B C D E

1 ABC
2 ABCD 12 124 24 4 123 2 3 24 34 45
AB AC AD AE BC BD BE CD CE DE
3 BCE
4 ACDE
5 DE 12 2 24 4 4 2 3 4
ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

2 4
ABCD ABCE ABDE ACDE BCDE

Not supported
by any ABCDE

transactions
 Maximal vs Closed Frequent Itemsets
Closed but not
null
Minimum support = 2 maximal
124 123 1234 245 345
A B C D E
Closed
and
12 124 24 123
maximal
4 2 3 24 34 45
AB AC AD AE BC BD BE CD CE DE

12 2 24 4 4 2 3 4
ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

2 4
ABCD ABCE ABDE ACDE BCDE
# Closed = 9
# Maximal = 4
ABCDE
Maximal vs Closed Itemsets
Frequent
Itemsets

Closed
Frequent
Itemsets

Maximal
Frequent
Itemsets
Pattern Evaluation
• Association rule algorithms tend to produce too many rules but
many of them are uninteresting or redundant
• Redundant if {A,B,C}  {D} and {A,B}  {D} have same support &
confidence
• Summarization techniques
• Uninteresting, if the pattern that is revealed does not offer useful
information.
• Interestingness measures: a hard problem to define

• Interestingness measures can be used to prune/rank the

derived patterns
• Subjective measures: require human analyst
• Objective measures: rely on the data.

• In the original formulation of association rules, support &

confidence are the only measures used
 Computing Interestingness Measure
• Given a rule X  Y, information needed to compute rule
interestingness can be obtained from a contingency table

Contingency table for X  Y

𝑌 𝑌 f11: support of X and Y
𝑋 f11 f10 f1+ f10: support of X and Y
𝑋 f01 f00 f0+ f01: support of X and Y
f+1 f+0 N f00: support of X and Y

𝑋: itemset X appears in tuple Used to define various measures

𝑌: itemset Y appears in tuple
𝑋: itemset X does not appear in tuple  support, confidence, lift, Gini,
𝑌: itemset Y does not appear in tuple J-measure, etc.
Drawback of Confidence Number of people that
drink tea
Number of people that
Coffee Coffee drink coffee and tea
Tea 15 5 20 Number of people that
Tea 75 5 80 drink coffee but not tea
90 10 100 Number of people that
drink coffee
Association Rule: Tea  Coffee

15
Confidence= 𝑃(Coffee|Tea) = = 0.75
20

Although confidence is high, rule is misleading

90
• 𝑃(Coffee) = = 0.9
100

• 𝑃(Coffee|Tea) = 0.9375
Statistical Independence
• Population of 1000 students
• 600 students know how to swim (S)
• 700 students know how to bike (B)
• 420 students know how to swim and bike (S,B)

• P(S,B) = 420/1000 = 0.42

• P(S)  P(B) = 0.6  0.7 = 0.42

• P(S,B) = P(S)  P(B) => Statistical independence

Statistical Independence
• Population of 1000 students
• 600 students know how to swim (S)
• 700 students know how to bike (B)
• 500 students know how to swim and bike (S,B)

• P(S,B) = 500/1000 = 0.5

• P(S)  P(B) = 0.6  0.7 = 0.42

• P(S,B) > P(S)  P(B) => Positively correlated

Statistical Independence
• Population of 1000 students
• 600 students know how to swim (S)
• 700 students know how to bike (B)
• 300 students know how to swim and bike (S,B)

• P(S,B) = 300/1000 = 0.3

• P(S)  P(B) = 0.6  0.7 = 0.42

• P(S,B) < P(S)  P(B) => Negatively correlated

Statistical-based Measures
• Measures that take into account statistical dependence

• Lift/Interest/PMI

𝑃(𝑌|𝑋) 𝑃(𝑋, 𝑌)
Lift = = = Interest
𝑃(𝑌) 𝑃 𝑋 𝑃(𝑌)

In text mining it is called: Pointwise Mutual Information

• Piatesky-Shapiro

PS = 𝑃 𝑋, 𝑌 − 𝑃 𝑋 𝑃(𝑌)

• All these measures measure deviation from independence

• The higher, the better (why?)
Example: Lift/Interest

Coffee Coffee
Tea 15 5 20
Tea 75 5 80
90 10 100

Association Rule: Tea  Coffee

Confidence= P(Coffee|Tea) = 0.75

but P(Coffee) = 0.9
 Lift = 0.75/0.9= 0.8333 (< 1, therefore is negatively associated)
= 0.15/(0.9*0.2)
 Another Example
of the of, the
Fraction of P(of, the) ≈ P of P(the)
0.9 0.9 0.8
documents

If I was creating a document by picking words randomly, (of, the) have

more or less the same probability of appearing together by chance No correlation

hong kong hong, kong

Fraction of P hong, kong ≫ P hong P(kong)
0.2 0.2 0.19
documents
(hong, kong) have much lower probability to appear together by chance.
The two words appear almost always only together Positive correlation

obama karagounis obama, karagounis P obama, karagounis ≪

Fraction of
0.2 0.2 0.001 P obama P(karagounis)
documents
(obama, karagounis) have much higher probability to appear together by chance.
The two words appear almost never together Negative correlation
Drawbacks of Lift/Interest/Mutual Information
honk konk honk, konk
Fraction of
0.0001 0.0001 0.0001
documents

0.0001
𝑀𝐼 ℎ𝑜𝑛𝑘, 𝑘𝑜𝑛𝑘 = = 10000
0.0001 ∗ 0.0001

hong kong hong, kong

Fraction of
0.2 0.2 0.19
documents

0.19
𝑀𝐼 ℎ𝑜𝑛𝑔, 𝑘𝑜𝑛𝑔 = = 4.75
0.2 ∗ 0.2

Rare co-occurrences are deemed more interesting.

But this is not always what we want
ALTERNATIVE FREQUENT
ITEMSET COMPUTATION
Slides taken from Mining Massive Datasets course by
Anand Rajaraman and Jeff Ullman.
 Finding the frequent pairs is usually
the most expensive operation

Count All pairs Count

All the items of items the pairs
items from L1

C1 Filter L1 Construct C2 Filter L2 Construct C3

First Second
pass pass

Frequent Frequent
items pairs
 76

Picture of A-Priori

Item counts Frequent items

Counts of
pairs of
frequent
items

Pass 1 Pass 2
 77

PCY Algorithm
Item counts
• During Pass 1 (computing frequent
items) of Apriori, most memory is idle.

• Use that memory to keep a hash table

where pairs of items are hashed.
• The hash table keeps just counts of the
number of pairs hashed in each bucket,
Pass 1
not the pairs themselves.
 78

Needed Extensions
1. Pairs of items need to be generated from the
input file; they are not present in the file.
2. Memory organization:
• Space to count each item.
• One (typically) 4-byte integer per item.
• Use the rest of the space for as many integers,
representing buckets, as we can.
 79

Picture of PCY

Item counts

Hash
table

Pass 1
 80

Picture of PCY

Item counts

Bucket Counts

Pass 1
 81

PCY Algorithm – Pass 1

FOR (each basket) {

FOR (each item in the basket)
add 1 to item’s count;
FOR (each pair of items in the basket)
{
hash the pair to a bucket;
add 1 to the count for that bucket
}
}
 82

Observations About Buckets

• A bucket is frequent if its count is at least the support
threshold.

• A bucket that a frequent pair hashes to is surely frequent.

• We cannot use the hash table to eliminate any member of this
bucket.
• Even without any frequent pair, a bucket can be frequent.
• Again, nothing in the bucket can be eliminated.
• But in the best case, the count for a bucket is less than
the support s.
• Now, all pairs that hash to this bucket can be eliminated as
candidates, even if the pair consists of two frequent items.

• On Pass 2 (frequent pairs), we only count pairs that hash

to frequent buckets.
 83

PCY Algorithm – Between Passes

• Replace the buckets by a bit-vector:
• 1 means the bucket is frequent; 0 means it is not.

• 4-byte integers are replaced by bits, so the bit-

vector requires 1/32 of memory.

• Also, find which items are frequent and list them

for the second pass.
• Same as with Apriori
 84

Picture of PCY

Item counts Frequent items

Bitmap

Hash
table Counts of
candidate
pairs

Pass 1 Pass 2
 85

PCY Algorithm – Pass 2

• Count all pairs {i, j } that meet the conditions
for being a candidate pair:
1. Both i and j are frequent items.
2. The pair {i, j }, hashes to a bucket number whose bit
in the bit vector is 1.

• Notice both these conditions are necessary for

the pair to have a chance of being frequent.
 86

All (Or Most) Frequent Itemsets

in less than 2 Passes

• A-Priori, PCY, etc., take k passes to find

frequent itemsets of size k.
• Other techniques use 2 or fewer passes for all
sizes:
• Simple sampling algorithm.
• SON (Savasere, Omiecinski, and Navathe).
• Toivonen.
 87

Simple Sampling Algorithm – (1)

• Take a random sample of the market baskets.

• Run Apriori or one of its improvements (for sets

of all sizes, not just pairs) in main memory, so
you don’t pay for disk I/O each time you increase
the size of itemsets.
• Make sure the sample is such that there is enough
space for counts.
 88

Main-Memory Picture

Copy of
sample
baskets

Space
for
counts
 89

Simple Algorithm – (2)

• Use as your support threshold a suitable,

scaled-back number.
• E.g., if your sample is 1/100 of the baskets, use
s /100 as your support threshold instead of s.

• You could stop here (single pass)

• What could be the problem?
 90

Simple Algorithm – Option

• Optionally, verify that your guesses are truly
frequent in the entire data set by a second
pass (eliminate false positives)

• But you don’t catch sets frequent in the whole

but not in the sample. (false negatives)
• Smaller threshold, e.g., s /125, helps catch more
truly frequent itemsets.
• But requires more space.
 91

SON Algorithm – (1)

• First pass: Break the data into chunks that can be
processed in main memory.
• Read one chunk at the time
• Find all frequent itemsets for each chunk.
• Threshold = s/number of chunks

• An itemset becomes a candidate if it is found to

be frequent in any one or more chunks of the
baskets.
 92

SON Algorithm – (2)

• Second pass: count all the candidate itemsets
and determine which are frequent in the entire
set.

• Key “monotonicity” idea: an itemset cannot be

frequent in the entire set of baskets unless it is
frequent in at least one subset.
• Why?
 93

SON Algorithm – Distributed Version

• This idea lends itself to distributed data
mining.
• If baskets are distributed among many nodes,
compute frequent itemsets at each node, then
distribute the candidates from each node.
• Finally, accumulate the counts of all
candidates.
 94

Toivonen’s Algorithm – (1)

• Start as in the simple sampling algorithm, but
lower the threshold slightly for the sample.
• Example: if the sample is 1% of the baskets, use s /125
as the support threshold rather than s /100.
• Goal is to avoid missing any itemset that is frequent in
the full set of baskets.
 95

Toivonen’s Algorithm – (2)

• Add to the itemsets that are frequent in the

sample the negative border of these itemsets.
• An itemset is in the negative border if it is not
deemed frequent in the sample, but all its
immediate subsets are.
 96

Reminder: Negative Border

• Itemset ABCD is in the negative border if

and only if:
1. It is not frequent in the sample, but
2. All of ABC, BCD, ACD, and ABD are.
• Item A is in the negative border if and only if
it is not frequent in the sample.
 Because the empty set is always frequent.
 Unless there are fewer baskets than the support
threshold (silly case).
 97

Picture of Negative Border

Negative Border
…

triples

pairs

singletons
Frequent Itemsets
from Sample
 98

Toivonen’s Algorithm – (3)

• In a second pass, compute the support for all
candidate frequent itemsets from the first pass,
and also for their negative border.
• If no itemset from the negative border turns out to
be frequent, then the candidates found to be
frequent in the whole data are exactly the
frequent itemsets.
 99

Toivonen’s Algorithm – (4)

• What if we find that something in the negative

border is actually frequent?
• We must start over again!

• Try to choose the support threshold so the

probability of failure is low, while the number of
itemsets checked on the second pass fits in main-
memory.
 100

If Something in the Negative Border is

Frequent . . .
We broke through the
negative border. How
far does the problem
go?
…
Negative Border
tripletons

doubletons

singletons
Frequent Itemsets
from Sample
 101

Theorem:
• If there is an itemset that is frequent in the whole,
but not frequent in the sample, then there is a
member of the negative border for the sample
that is frequent in the whole.
 102

Proof: Suppose not; i.e.;

1. There is an itemset S frequent in the whole but not
frequent in the sample, and
2. Nothing in the negative border is frequent in the
whole.
• Let T be a smallest subset of S that is not
frequent in the sample.
• T is frequent in the whole (S is frequent +
monotonicity).
• T is in the negative border (else not
“smallest”).
Example
null

A B C D E

AB AC AD AE BC BD BE CD CE DE

ABC ABD ABE ACD ACE ADE BCD BCE BDE CDE

ABCD ABCE ABDE ACDE BCDE

ABCDE
Border
FREQUENT ITEMSET
RESEARCH