CPS216: Advanced Database
Systems
Data Mining
Slides created by Jeffrey Ullman,
Stanford
1
What is Data Mining?
Discovery of useful, possibly
unexpected, patterns in data.
Subsidiary issues:
 Data cleansing: detection of bogus data.
• E.g., age = 150.
 Visualization: something better than
megabyte files of output.
 Warehousing of data (for retrieval).
2
Typical Kinds of Patterns
1. Decision trees: succinct ways to classify by
testing properties.
2. Clusters: another succinct classification by
similarity of properties.
3. Bayes, hidden-Markov, and other statistical
models, frequent-itemsets: expose
important associations within data.
3
Example: Clusters
x
x
x
x x
x x
x xx x
x x x
x x
x
x
xx x
x x
x x x
x
xx x
x
x x
x x x x
x x x
x
4
Example: Frequent Itemsets
 A common marketing problem:
examine what people buy together to
discover patterns.
1. What pairs of items are unusually often
found together at Kroger checkout?
•
Answer: diapers and beer.
2. What books are likely to be bought by
the same Amazon customer?
5
Meaningfulness of Answers
A big risk when data mining is that you
will “discover” patterns that are
meaningless.
Statisticians call it Bonferroni’s
principle: (roughly) if you look in more
places for interesting patterns than your
amount of data will support, you are
bound to find “false patterns”.
6
Rhine Paradox --- (1)
David Rhine was a parapsychologist in the
1950’s who hypothesized that some
people had Extra-Sensory Perception.
He devised an experiment where subjects
were asked to guess 10 hidden cards --red or blue.
He discovered that almost 1 in 1000 had
ESP --- they were able to get all 10 right!
7
Rhine Paradox --- (2)
He told these people they had ESP and
called them in for another test of the
same type.
Alas, he discovered that almost all of
them had lost their ESP.
What did he conclude?
 Answer on next slide.
8
Rhine Paradox --- (3)
He concluded that you shouldn’t tell
people they have ESP; it causes them
to lose it.
9
“Association Rules”
Market Baskets
Frequent Itemsets
A-priori Algorithm
10
The Market-Basket Model
A large set of items, e.g., things sold in
a supermarket.
A large set of baskets, each of which is
a small set of the items, e.g., the things
one customer buys on one day.
11
Association Rule Mining
sales
records:
tran1
tran2
tran3
tran4
tran5
tran6
cust33
cust45
cust12
cust40
cust12
cust12
p2,
p5,
p1,
p5,
p2,
p9
p5, p8
p8, p11
p9
p8, p11
p9
market-basket
data
• Trend: Products p5, p8 often bought together
12
Support
Simplest question: find sets of items
that appear “frequently” in the baskets.
Support for itemset I = the number of
baskets containing all items in I.
Given a support threshold s, sets of
items that appear in > s baskets are
called frequent itemsets.
13
Example
Items={milk, coke, pepsi, beer, juice}.
Support = 3 baskets.
B1
B3
B5
B7
=
=
=
=
{m, c, b}
{m, b}
{m, p, b}
{c, b, j}
B2
B4
B6
B8
=
=
=
=
{m, p, j}
{c, j}
{m, c, b, j}
{b, c}
Frequent itemsets: {m}, {c}, {b}, {j},
{m, b}, {c, b}, {j, c}.
14
Applications --- (1)
Real market baskets: chain stores keep
terabytes of information about what
customers buy together.
 Tells how typical customers navigate
stores, lets them position tempting items.
 Suggests tie-in “tricks,” e.g., run sale on
diapers and raise the price of beer.
High support needed, or no $$’s .
15
Applications --- (2)
“Baskets” = documents; “items” =
words in those documents.
 Lets us find words that appear together
unusually frequently, i.e., linked concepts.
“Baskets” = sentences, “items” =
documents containing those sentences.
 Items that appear together too often could
represent plagiarism.
16
Applications --- (3)
“Baskets” = Web pages; “items” =
linked pages.
 Pairs of pages with many common
references may be about the same topic.
“Baskets” = Web pages p ; “items” =
pages that link to p .
 Pages with many of the same links may be
mirrors or about the same topic.
17
Important Point
“Market Baskets” is an abstraction that
models any many-many relationship
between two concepts: “items” and
“baskets.”
 Items need not be “contained” in baskets.
The only difference is that we count cooccurrences of items related to a
basket, not vice-versa.
18
Scale of Problem
WalMart sells 100,000 items and can
store billions of baskets.
The Web has over 100,000,000 words
and billions of pages.
19
Association Rules
If-then rules about the contents of
baskets.
{i1, i2,…,ik} → j means: “if a basket
contains all of i1,…,ik then it is likely to
contain j.
Confidence of this association rule is
the probability of j given i1,…,ik.
20
Example
+ B1 = {m, c, b}
_
B3 = {m, b}
_
B5 = {m, p, b}
B7 = {c, b, j}
B2
B4
+ B6
B8
=
=
=
=
{m, p, j}
{c, j}
{m, c, b, j}
{b, c}
An association rule: {m, b} → c.
 Confidence = 2/4 = 50%.
21
Interest
The interest of an association rule is
the absolute value of the amount by
which the confidence differs from what
you would expect, were items selected
independently of one another.
22
Example
B1
B3
B5
B7
=
=
=
=
{m, c, b}
{m, b}
{m, p, b}
{c, b, j}
B2
B4
B6
B8
=
=
=
=
{m, p, j}
{c, j}
{m, c, b, j}
{b, c}
For association rule {m, b} → c, item c
appears in 5/8 of the baskets.
Interest = | 2/4 - 5/8 | = 1/8 --- not
very interesting.
23
Relationships Among Measures
Rules with high support and confidence
may be useful even if they are not
“interesting.”
 We don’t care if buying bread causes
people to buy milk, or whether simply a lot
of people buy both bread and milk.
But high interest suggests a cause that
might be worth investigating.
24
Finding Association Rules
A typical question: “find all association
rules with support ≥ s and confidence ≥ c.”
 Note: “support” of an association rule is the
support of the set of items it mentions.
Hard part: finding the high-support
(frequent ) itemsets.
 Checking the confidence of association rules
involving those sets is relatively easy.
25
Computation Model
Typically, data is kept in a “flat file”
rather than a database system.
 Stored on disk.
 Stored basket-by-basket.
 Expand baskets into pairs, triples, etc. as
you read baskets.
26
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.
27
Main-Memory Bottleneck
In many algorithms to find frequent
itemsets we need to worry about how
main memory is used.
 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 a disaster.
28
Finding Frequent Pairs
The hardest problem often turns out to
be finding the frequent pairs.
We’ll concentrate on how to do that,
then discuss extensions to finding
frequent triples, etc.
29
Naïve Algorithm
A simple way to find frequent pairs is:
 Read file once, counting in main memory
the occurrences of each pair.
• Expand each basket of n items into its
n (n -1)/2 pairs.
Fails if #items-squared exceeds main
memory.
30
Details of Main-Memory Counting
 There are two basic approaches:
1. Count all item pairs, using a triangular
matrix.
2. Keep a table of triples [i, j, c] = the count
of the pair of items {i,j } is c.
 (1) requires only (say) 4 bytes/pair;
(2) requires 12 bytes, but only for
those pairs with >0 counts.
31
4 per pair
Method (1)
12 per
occurring pair
Method (2)
32
Details of Approach (1)
Number items 1,2,…
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 2n 2.
33
Details of Approach (2)
You need a hash table, with i and j as the
key, to locate (i, j, c) triples efficiently.
 Typically, the cost of the hash structure can be
neglected.
Total bytes used is about 12p, where p is
the number of pairs that actually occur.
 Beats triangular matrix if at most 1/3 of
possible pairs actually occur.
34
A-Priori Algorithm --- (1)
A two-pass approach called a-priori
limits the need for main memory.
Key idea: monotonicity : if a set of
items appears at least s times, so does
every subset.
 Contrapositive for pairs: if item i does not
appear in s baskets, then no pair including
i can appear in s baskets.
35
A-Priori Algorithm --- (2)
Pass 1: Read baskets and count in main
memory the occurrences of each item.
 Requires only memory proportional to #items.
Pass 2: Read baskets again and count in
main memory only those pairs both of
which were found in Pass 1 to be frequent.
 Requires memory proportional to square of
frequent items only.
36
Picture of A-Priori
Item counts
Frequent items
Counts of
candidate
pairs
Pass 1
Pass 2
37
Detail for A-Priori
You can use the triangular matrix
method with n = number of frequent
items.
 Saves space compared with storing triples.
Trick: number frequent items 1,2,… and
keep a table relating new numbers to
original item numbers.
38
Frequent Triples, Etc.
For each k, we construct two sets of
k –tuples:
 Ck = candidate k – tuples = those that
might be frequent sets (support > s )
based on information from the pass for
k –1.
 Lk = the set of truly frequent k –tuples.
39
C1
Filter
First
pass
L1
Construct
C2
Filter
L2
Construct
C3
Second
pass
40
A-Priori for All Frequent
Itemsets
One pass for each k.
Needs room in main memory to count
each candidate k –tuple.
For typical market-basket data and
reasonable support (e.g., 1%), k = 2
requires the most memory.
41
Frequent Itemsets --- (2)
C1 = all items
L1 = those counted on first pass to be
frequent.
C2 = pairs, both chosen from L1.
In general, Ck = k –tuples each k –1 of
which is in Lk-1.
Lk = those candidates with support ≥ s.
42