DATA MINING

LECTURE 5
Sketching, Locality Sensitive Hashing
 2

Jaccard Similarity
• The Jaccard similarity (Jaccard coefficient) of two sets S1,
S2 is the size of their intersection divided by the size of
their union.
• JSim (S1, S2) = |S1S2| / |S1S2|.

3 in intersection.
8 in union.
Jaccard similarity
= 3/8

• Extreme behavior:
• Jsim(X,Y) = 1, iff X = Y
• Jsim(X,Y) = 0 iff X,Y have no elements in common
• JSim is symmetric
Cosine Similarity

• Sim(X,Y) = cos(X,Y)
• The cosine of the angle between X and Y

• If the vectors are aligned (correlated) angle is zero degrees and

cos(X,Y)=1
• If the vectors are orthogonal (no common coordinates) angle is 90
degrees and cos(X,Y) = 0

• Cosine is commonly used for comparing documents, where we

assume that the vectors are normalized by the document length.
Application: Recommendations
• Recommendation systems
• When a user buys or rates an item we want to
recommend other items that the user may like
• Initially applied to books, but now recommendations are
everywhere: songs, movies, products, restaurants, hotels, etc.

• Commonly used algorithms:

• Find the k users most similar to the user at hand and
recommend items that they like.
• Find the items most similar to the items that the user
has previously liked, and recommend these items.
Application: Finding near duplicates
• Find duplicate and near-duplicate documents
from a web crawl.
• Why is it important:
• Identify mirrored web pages, and avoid indexing them,
or serving them multiple times
• Find replicated news stories and cluster them under a
single story.
• Identify plagiarism
• Near duplicate documents differ in a few
characters, words or sentences
Finding similar items
• The problems we have seen so far have a
common component
• We need a quick way to find highly similar items to a
query item
• OR, we need a method for finding all pairs of items that
are highly similar.
• Also known as the Nearest Neighbor problem, or
the All Nearest Neighbors problem
SKETCHING
AND
LOCALITY SENSITIVE
HASHING
Thanks to:
Rajaraman and Ullman, “Mining Massive Datasets”
Evimaria Terzi, slides for Data Mining Course.
Problem
• Given a (large) collection of documents find all
pairs of documents which are near duplicates
• Their similarity is very high

• What if we want to find identical documents?

Main issues
• What is the right representation of the document
when we check for similarity?
• E.g., representing a document as a set of characters
will not do (why?)
• When we have billions of documents, keeping the
full text in memory is not an option.
• We need to find a shorter representation
• How do we do pairwise comparisons of billions of
documents?
• If we wanted exact match it would be ok, can we
replicate this idea?
 10

Three Essential Techniques for Similar

Documents

1. Shingling : convert documents, emails, etc.,

to sets.

2. Minhashing : convert large sets to short

signatures, while preserving similarity.

3. Locality-Sensitive Hashing (LSH): focus on

pairs of signatures likely to be similar.
 11

The Big Picture

Candidate
pairs :
Locality-
Docu- those pairs
sensitive
ment of signatures
Hashing
that we need
to test for
The set Signatures : similarity.
of strings short integer
of length k vectors that
that appear represent the
in the doc- sets, and
ument reflect their
similarity
 12

Shingles

• A k -shingle (or k -gram) for a document is a

sequence of k characters that appears in the
document.
• Example: document = abcab. k=2
• Set of 2-shingles = {ab, bc, ca}.
• Option: regard shingles as a bag, and count ab twice.

• Represent a document by its set of k-shingles.

Shingling
• Shingle: a sequence of k contiguous characters
a rose is a rose is a rose
a rose is
rose is a
rose is a
ose is a r
se is a ro
e is a ros
is a rose
is a rose
s a rose i
a rose is
a rose is
Shingling
• Shingle: a sequence of k contiguous characters
a rose is a rose is a rose
a rose is a rose is
rose is a rose is a
rose is a rose is a
ose is a r ose is a r
se is a ro se is a ro
e is a ros e is a ros
is a rose is a rose
is a rose is a rose
s a rose i s a rose i
a rose is a rose is
a rose is
 15

Working Assumption
• Documents that have lots of shingles in common
have similar text, even if the text appears in
different order.
• Careful: you must pick k large enough, or most
documents will have most shingles.
• Extreme case k = 1: all documents are the same
• k = 5 is OK for short documents; k = 10 is better for long
documents.
• Alternative ways to define shingles:
• Use words instead of characters
• Anchor on stop words (to avoid templates)
 16

Shingles: Compression Option

• To compress long shingles, we can hash them to

(say) 4 bytes.
ℎ: 𝑉 𝑘 → 0,1 64
• Represent a doc by the set of hash values of its k-
shingles.
• Shingle 𝑠 will be represented by the 64-bit integer ℎ(𝑠)
• From now on we will assume that shingles are
integers
• Collisions are possible, but very rare
Fingerprinting
• Hash shingles to 64-bit integers
Set of Shingles Set of 64-bit integers
Hash function
a rose is (Rabin’s fingerprints) 1111
rose is a 2222
rose is a 3333
ose is a r 4444
se is a ro 5555
e is a ros 6666
is a rose 7777
is a rose 8888
s a rose i 9999
a rose is 0000
 18

Basic Data Model: Sets

• Document: A document is represented as a set
shingles (more accurately, hashes of shingles)

• Document similarity: Jaccard similarity of the sets of

shingles.
• Common shingles over the union of shingles
• Sim (C1, C2) = |C1C2|/|C1C2|.

• Although we use the documents as our driving

example the techniques we will describe apply to any
kind of sets.
• E.g., similar customers or items.
Signatures
• Problem: shingle sets are still too large to be kept in memory.

• Key idea: “hash” each set S to a small signature Sig (S), such
that:

1. Sig (S) is small enough that we can fit a signature in main memory
for each set.

2. Sim (S1, S2) is (almost) the same as the “similarity” of Sig (S1) and
Sig (S2). (signature preserves similarity).

• Warning: This method can produce false negatives, and false

positives (if an additional check is not made).
• False negatives: Similar items deemed as non-similar
• False positives: Non-similar items deemed as similar
 20

From Sets to Boolean Matrices

• Represent the data as a boolean matrix M

• Rows = the universe of all possible set elements
• In our case, shingle fingerprints take values in [0…264-1]
• Columns = the sets
• In our case, documents, sets of shingle fingerprints
• M(r,S) = 1 in row r and column S if and only if r is a
member of S.

• Typical matrix is sparse.

• We do not really materialize the matrix
Example
• Universe: U = {A,B,C,D,E,F,G}

• X = {A,B,F,G} X Y
A 1 1
• Y = {A,E,F,G} B 1 0
C 0 0

3
D 0 0

• Sim(X,Y) = E 0 1
5 F 1 1
G 1 1
Example
• Universe: U = {A,B,C,D,E,F,G}

• X = {A,B,F,G} X Y
A 1 1
• Y = {A,E,F,G} B 1 0
C 0 0

3
D 0 0

• Sim(X,Y) = E 0 1
5 F 1 1
G 1 1

At least one of the columns has value 1

Example
• Universe: U = {A,B,C,D,E,F,G}

• X = {A,B,F,G} X Y
A 1 1
• Y = {A,E,F,G} B 1 0
C 0 0

3
D 0 0

• Sim(X,Y) = E 0 1
5 F 1 1
G 1 1

Both columns have value 1

 24

Minhashing
• Pick a random permutation of the rows (the
universe U).
• Define “hash” function for set S
• h(S) = the index of the first row (in the permuted order)
in which column S has 1.
same as:
• h(S) = the index of the first element of S in the permuted
order.
• Use k (e.g., k = 100) independent random
permutations to create a signature.
Example of minhash signatures
• Input matrix Random
Permutation
elem index elem
ent S1 S2 S3 S4 ent S1 S2 S3 S4
A
A 1 0 1 0 1 A 1 0 1 0
C
B 1 0 0 1 2 C 0 1 0 1
C 0 1 0 1 G 3 G 1 0 1 0
D 0 1 0 1 F 4 F 1 0 1 0
E 0 1 1 1 B 5 B 1 0 0 1
F 1 0 1 0 E 6 E 0 1 1 1
G 1 0 1 0 D 7 D 0 1 0 1
1 2 1 2
Example of minhash signatures
• Input matrix Random
Permutation
elem index elem
ent S1 S2 S3 S4 ent S1 S2 S3 S4
D
A 1 0 1 0 1 D 0 1 0 1
B
B 1 0 0 1 2 B 1 0 0 1
C 0 1 0 1 A 3 A 1 0 1 0
D 0 1 0 1 C 4 C 0 1 0 1
E 0 1 1 1 F 5 F 1 0 1 0
F 1 0 1 0 G 6 G 1 0 1 0
G 1 0 1 0 E 7 E 0 1 1 1
2 1 3 1
Example of minhash signatures
• Input matrix Random
Permutation
elem index elem
ent S1 S2 S3 S4 ent S1 S2 S3 S4
C
A 1 0 1 0 1 C 0 1 0 1
D
B 1 0 0 1 2 D 0 1 0 1
C 0 1 0 1 G 3 G 1 0 1 0
D 0 1 0 1 F 4 F 1 0 1 0
E 0 1 1 1 A 5 A 1 0 1 0
F 1 0 1 0 B 6 B 1 0 0 1
G 1 0 1 0 E 7 E 0 1 1 1

3 1 3 1
Example of minhash signatures
• Input matrix
S1 S2 S3 S4 Signature matrix
A 1 0 1 0
S1 S2 S3 S4 We now have a
B 1 0 0 1 smaller dataset

≈
C 0 1 0 1 h1 1 2 1 2 with just 𝑘 rows
D 0 1 0 1 h2 2 1 3 1

E 0 1 1 1 h3 3 1 3 1

F 1 0 1 0
G 1 0 1 0
• Sig(S) = vector of hash values
• e.g., Sig(S2) = [2,1,1]
• Sig(S,i) = value of the i-th hash
function for set S
• E.g., Sig(S2,3) = 1
 29

A Subtle Point
• People sometimes ask whether the minhash
value should be the original number of the row, or
the number in the permuted order (as we did in
our example).
• Answer: it doesn’t matter.
• You only need to be consistent, and assure that
two columns get the same value if and only if
their first 1’s in the permuted order are in the
same row.
 30

Hash function Property

Pr(h(S1) = h(S2)) = Sim(S1,S2)

• where the probability is over all choices of

permutations.

• Why?
• The first row where one of the two sets has value 1
belongs to the union.
• Recall that union contains rows with at least one 1.
• We have equality if both sets have value 1, and this row
belongs to the intersection
Example
• Universe: U = {A,B,C,D,E,F,G}
• X = {A,B,F,G} Rows C,D could be anywhere
• Y = {A,E,F,G} they do not affect the probability

X Y X Y
• Union = A 1 1 D D 0 0
{A,B,E,F,G} B 1 0 *
C 0 0 *
• Intersection =
D 0 0 C C 0 0
{A,F,G} E 0 1 *
F 1 1 *
G 1 1 *
Example
• Universe: U = {A,B,C,D,E,F,G}
• X = {A,B,F,G} The * rows belong to the union
• Y = {A,E,F,G}

X Y X Y
• Union = A 1 1 D D 0 0
{A,B,E,F,G} B 1 0 *
C 0 0 *
• Intersection =
D 0 0 C C 0 0
{A,F,G} E 0 1 *
F 1 1 *
G 1 1 *
Example
• Universe: U = {A,B,C,D,E,F,G}
• X = {A,B,F,G} The question is what is the value
• Y = {A,E,F,G} of the first * element

X Y X Y
• Union = A 1 1 D D 0 0
{A,B,E,F,G} B 1 0 *
C 0 0 *
• Intersection =
D 0 0 C C 0 0
{A,F,G} E 0 1 *
F 1 1 *
G 1 1 *
Example
• Universe: U = {A,B,C,D,E,F,G}
• X = {A,B,F,G} If it belongs to the intersection
• Y = {A,E,F,G} then h(X) = h(Y)

X Y X Y
• Union = A 1 1 D D 0 0
{A,B,E,F,G} B 1 0 *
C 0 0 *
• Intersection =
D 0 0 C C 0 0
{A,F,G} E 0 1 *
F 1 1 *
G 1 1 *
Example
• Universe: U = {A,B,C,D,E,F,G}
• X = {A,B,F,G} Every element of the union is equally likely
to be the * element
• Y = {A,E,F,G} | A,F,G | 3
Pr(h(X) = h(Y)) = = = Sim(X,Y)
| A,B,E,F,G | 5

X Y X Y
• Union = A 1 1 D D 0 0
{A,B,E,F,G} B 1 0 *
C 0 0 *
• Intersection =
D 0 0 C C 0 0
{A,F,G} E 0 1 *
F 1 1 *
G 1 1 *
 36

Similarity for Signatures

• The similarity of signatures is the fraction of the
hash functions in which they agree.
S1 S2 S3 S4 Actual Sig
Signature matrix
A 1 0 1 0 (S1, S2) 0 0
S1 S2 S3 S4
B 1 0 0 1 (S1, S3) 3/5 2/3

≈
1 2 1 2
C 0 1 0 1 (S1, S4) 1/7 0
2 1 3 1
D 0 1 0 1 (S2, S3) 0 0
3 1 3 1
E 0 1 1 1 (S2, S4) 3/4 1
F 1 0 1 0 (S3, S4) 0 0
Zero similarity is preserved
G 1 0 1 0
High similarity is well approximated
• With multiple signatures we get a good approximation
• Why? What is the expected value of the fraction of agreements?
Is it now feasible?
• Assume a billion rows
• Hard to pick a random permutation of 1…billion
• Even representing a random permutation
requires 1 billion entries!!!
• How about accessing rows in permuted order?
•

• Instead of permutations we will consider hash

functions that map the N rows to N buckets
• Some collisions may happen, but with well chosen
functions they are rare.
 Being more practical
Approximating row permutations: pick k=100 hash
functions (h1,…,hk) In practice this means selecting the
function parameters
for each row r In practice only the rows (shingles)
that appear in the data
for each hash function hi
compute hi (r ) hi (r) = index of shingle r in permutation

for each column S that has 1 in row r S contains shingle r

if hi (r ) is a smaller value than Sig(S,i) then
Sig(S,i) = hi (r); Find the shingle r with minimum index

Sig(S,i) will become the smallest value of hi(r) among all rows
(shingles) for which column S has value 1 (shingle belongs in S);
i.e., hi (r) gives the min index for the i-th permutation
 39

Example Sig1 Sig2

h(0) = 1 1 -
g(0) = 3 3 -
x Row S1 S2 h(x) g(x)
0 A 1 0 1 3 h(1) = 2 1 2
1 B 0 1 2 0 g(1) = 0 3 0
2 C 1 1 3 2
3 D 1 0 4 4 h(2) = 3 1 2
g(2) = 2 2 0
4 E 0 1 0 1
h(3) = 4 1 2
g(3) = 4 2 0
h(x) = x+1 mod 5 g(x) = 2x+1 mod 5

Row S1 S2 Row S1 S2 h(4) = 0 1 0

E 0 1 B 0 1 g(4) = 1 2 0
A 1 0 E 0 1
B 0 1 C 1 0
C 1 1 A 1 1
D 1 0 D 1 0
 40

Implementation – (4)
• Often, data is given by column, not row.
• E.g., columns = documents, rows = shingles.
• If so, sort matrix once so it is by row.
• And always compute hi (r ) only once for each
row.
 41

Finding similar pairs

• Problem: Find all pairs of documents with
similarity at least t = 0.8
• While the signatures of all columns may fit in
main memory, comparing the signatures of all
pairs of columns is quadratic in the number of
columns.
• Example: 106 columns implies 5*1011 column-
comparisons.
• At 1 microsecond/comparison: 6 days.
 42

Locality-Sensitive Hashing
• What we want: a function f(X,Y) that tells whether or not X
and Y is a candidate pair: a pair of elements whose
similarity must be evaluated.

• A simple idea: X and Y are a candidate pair if they have

the same min-hash signature. ! Multiple levels of Hashing!
• Easy to test by hashing the signatures.
• Similar sets are more likely to have the same signature.
• Likely to produce many false negatives.
• Requiring full match of signature is strict, some similar sets will be lost.

• Improvement: Compute multiple signatures; candidate

pairs should have at least one common signature.
• Reduce the probability for false negatives.
 43

Signature matrix reminder

Prob(Sig(S,i) == Sig(S’,i)) = sim(S,S’)

Sig(S,i)
Sig(S’,i)

hash function i

n hash functions

Sig(S):
signature for set S

signature for set S’ Matrix M

 44

Partition into Bands – (1)

• Divide the signature matrix Sig into b bands of r

rows.
• Each band is a mini-signature with r hash functions.
 45

Partitioning into bands

n = b*r hash functions Matrix Sig

r rows
per band

b bands
b mini-signatures

One
signature
 46

Partition into Bands – (2)

• Divide the signature matrix Sig into b bands of r

rows.
• Each band is a mini-signature with r hash functions.
• For each band, hash the mini-signature to a hash
table with k buckets.
• Make k as large as possible so that mini-signatures that
hash to the same bucket are almost certainly identical.
 47

Hash Table Columns 2 and 6

are (almost certainly) identical.

Columns 6 and 7 are

surely different.
Matrix M

1 2 3 4 5 6 7

r rows b bands
 48

Partition into Bands – (2)

• Divide the signature matrix Sig into b bands of r

rows.
• Each band is a mini-signature with r hash functions.
• For each band, hash the mini-signature to a hash table
with k buckets.
• Make k as large as possible so that mini-signatures that hash
to the same bucket are almost certainly identical.
• Candidate column pairs are those that hash to the
same bucket for at least 1 band.
• I.e., they have at least one mini-signature in common.
• Tune b and r to catch most similar pairs, but few non-
similar pairs.
 49

Analysis of LSH – What We Want

Probability
= 1 if s > t

Probability No chance
of sharing if s < t
a bucket

Similarity s of two sets

 50

What One Band of One Row Gives You

Single hash signature

Probability Remember:
of sharing probability of
a bucket equal hash-values
= similarity

t Prob(Sig(S,i) == Sig(S’,i)) = sim(S,S’)

Similarity s of two sets

 51

What b Bands of r Rows Gives You

At least
one band No bands
identical identical

t ~ (1/b)1/r 1 - (1 - s r )b
Probability
of sharing
a bucket

Some row All rows

of a band of a band
t unequal are equal

Similarity s of two sets

 52

Example: b = 20; r = 5
t = 0.5

s 1-(1-sr)b
.2 .006
.3 .047
.4 .186
.5 .470
.6 .802
.7 .975
.8 .9996
 53

Suppose S1, S2 are 80% Similar

• We want all 80%-similar pairs. Choose 20 bands of 5
integers/band.

• Probability S1, S2 identical in one particular band:

(0.8)5 = 0.328.

• Probability S1, S2 are not similar in any of the 20 bands:

(1-0.328)20 = 0.00035
• i.e., about 1/3000-th of the 80%-similar column pairs are false negatives.

• Probability S1, S2 are similar in at least one of the 20

bands:
1-0.00035 = 0.999
 54

Suppose S1, S2 Only 40% Similar

• Probability S1, S2 identical in any one particular

band:
(0.4)5 = 0.01 .

• Probability S1, S2 are not identical in any of the

20 bands:
20
1 − 0.01 = 0.81

• False positive probability = 0.19. But false

positives much lower for similarities << 40%.
 55

LSH Summary

• Tune to get almost all pairs with similar

signatures, but eliminate most pairs that do not
have similar signatures.
• Check in main memory that candidate pairs
really do have similar signatures.
• Optional: In another pass through data, check
that the remaining candidate pairs really
represent similar sets .
Locality-sensitive hashing (LSH)
• Big Picture: Construct hash functions h: Rd U such
that for any pair of objects p,q, for distance function D
we have:
• If D(p,q)≤r, then Pr[h(p)=h(q)] is high
• Close (similar) objects have high probability to be hashed together
• If D(p,q)≥cr, then Pr[h(p)=h(q)] is small
• Distant (dissimilar) objects have small probability of being hashed
together
• Then, we can find close pairs by hashing

• LSH is a general framework: for a given distance

function D we need to find the right h
 57

LSH for Cosine Distance

• For cosine distance, there is a technique

analogous to minhashing for generating a
Locality Sensitive Hashing functions
• Using random hyperplanes.
 58

Random Hyperplanes

• Pick a random vector 𝑣, which determines a

hash function ℎ𝑣 with two buckets.
• ℎ𝑣 𝑥 = +1 if 𝑣 ⋅ 𝑥 > 0;
• ℎ𝑣 𝑥 = −1 if 𝑣 ⋅ 𝑥 < 0.

• LS-family H = set of all functions derived from

any vector.

• Claim:
• Prob[h(x)=h(y)] = 1 – (angle between x and y)/180
 59

Proof of Claim Look in the plane of x and y.

For a random vector v the values of the

hv(x) = +1 hash functions hv(x) and hv(y) depend
on where the vector v falls
hv(x) = -1
x
θ
hv(x) ≠ hv(y) when v falls into the
shaded area.
θ What is the probability of this for
a randomly chosen vector v?
θ

hv(y) = -1
y P[hv(x) ≠ hv(y)] = 2θ/360 = θ/180

hv(y) = +1 P[hv(x) = hv(y)] = 1- θ/180

 60

Signatures for Cosine Distance

• Pick some number of vectors, and hash your

data for each vector.
• The result is a signature (sketch ) of +1’s and –
1’s that can be used for LSH like the minhash
signatures for Jaccard distance.
 61

Simplification
• We need not pick from among all possible vectors
v to form a component of a sketch.
• It suffices to consider only vectors v consisting of
+1 and –1 components.