High dim.

Graph Infinite Machine

Apps
data data data learning

Locality Filtering
PageRank, Recommen
sensitive data SVM
SimRank der systems
hashing streams

Community Queries on Decision Association

Clustering
Detection streams Trees Rules

Dimensional Duplicate
Spam Web Perceptron,
ity document
Detection advertising kNN
reduction detection

2
¡ In many data mining situations, we do not
know the entire data set in advance

¡ Stream Management is important when the

input rate is controlled externally:
§ Google queries
§ Twitter or Facebook status updates
¡ We can think of the data as infinite and
non-stationary (the distribution changes
over time)
3
¡ Input elements enter at a rapid rate,
at one or more input ports (i.e., streams)
§ We call elements of the stream tuples

¡ The system cannot store the entire stream

accessibly

¡ Q: How do you make critical calculations

about the stream using a limited amount of
(secondary) memory?

4
¡ Stochastic Gradient Descent (SGD) is an
example of a stream algorithm
¡ In Machine Learning we call this: Online Learning
§ Allows for modeling problems where we have
a continuous stream of data
§ We want an algorithm to learn from it and
slowly adapt to the changes in data
¡ Idea: Do slow updates to the model
§ SGD (SVM, Perceptron) makes small updates
§ So: First train the classifier on training data.
§ Then: For every example from the stream, we slightly
update the model (using small learning rate)
5
 Ad-Hoc
Queries

. . . 1, 5, 2, 7, 0, 9, 3 Standing
Queries
. . . a, r, v, t, y, h, b Output
Processor
. . . 0, 0, 1, 0, 1, 1, 0
time

Streams Entering.
Each is stream is
composed of
elements/tuples
Limited
Working
Storage Archival
Storage

6
¡ Types of queries one wants on answer on
a data stream:
§ Sampling data from a stream
§ Construct a random sample
§ Queries over sliding windows
§ Number of items of type x in the last k elements
of the stream

7
¡ Types of queries one wants on answer on
a data stream:
§ Filtering a data stream
§ Select elements with property x from the stream
§ Counting distinct elements
§ Number of distinct elements in the last k elements
of the stream
§ Estimating moments
§ Estimate avg./std. dev. of last k elements
§ Finding frequent elements

8
¡ Mining query streams
§ Google wants to know what queries are
more frequent today than yesterday

¡ Mining click streams

§ Yahoo wants to know which of its pages are
getting an unusual number of hits in the past hour

¡ Mining social network news feeds

§ E.g., look for trending topics on Twitter, Facebook

9
¡ Sensor Networks
§ Many sensors feeding into a central controller
¡ Telephone call records
§ Data feeds into customer bills as well as
settlements between telephone companies
¡ IP packets monitored at a switch
§ Gather information for optimal routing
§ Detect denial-of-service attacks

10
As the stream grows the sample
also gets bigger
¡ Since we can not store the entire stream,
one obvious approach is to store a sample
¡ Two different problems:
§ (1) Sample a fixed proportion of elements
in the stream (say 1 in 10)
§ (2) Maintain a random sample of fixed size
over a potentially infinite stream
§ At any “time” k we would like a random sample
of s elements
§ What is the property of the sample we want to maintain?
For all time steps k, each of k elements seen so far has
equal prob. of being sampled
12
¡ Problem 1: Sampling fixed proportion
¡ Scenario: Search engine query stream
§ Stream of tuples: (user, query, time)
§ Answer questions such as: How often did a user
run the same query in a single days
§ Have space to store 1/10th of query stream
¡ Naïve solution:
§ Generate a random integer in [0..9] for each query
§ Store the query if the integer is 0, otherwise
discard
13
¡ Simple question: What fraction of queries by an
average search engine user are duplicates?
§ Suppose each user issues x queries once and d queries
twice (total of x+2d queries)
§ Correct answer: d/(x+d)
§ Proposed solution: We keep 10% of the queries
§ Sample will contain x/10 of the singleton queries and
2d/10 of the duplicate queries at least once
§ But only d/100 pairs of duplicates
§ d/100 = 1/10 · 1/10 · d
§ Of d “duplicates” 18d/100 appear exactly once
§ 18d/100 = ((1/10 · 9/10)+(9/10 · 1/10)) · d
!
"## 𝒅
§ So the sample-based answer is $ ! "&! =
% % 𝟏𝟎𝒙%𝟏𝟗𝒅
"# "## "##

14
Solution:
¡ Pick 1/10th of users and take all their
searches in the sample

¡ Use a hash function that hashes the

user name or user id uniformly into 10
buckets

15
¡ Stream of tuples with keys:
§ Key is some subset of each tuple’s components
§ e.g., tuple is (user, search, time); key is user
§ Choice of key depends on application

¡ To get a sample of a/b fraction of the stream:

§ Hash each tuple’s key uniformly into b buckets
§ Pick the tuple if its hash value is at most a

Hash table with b buckets, pick the tuple if its hash value is at most a.
How to generate a 30% sample?
Hash into b=10 buckets, take the tuple if it hashes to one of the first 3 buckets
16
As the stream grows, the sample is of
fixed size
¡ Problem 2: Fixed-size sample
¡ Suppose we need to maintain a random
sample S of size exactly s tuples
§ E.g., main memory size constraint
¡ Why? Don’t know length of stream in advance
¡ Suppose at time n we have seen n items
§ Each item is in the sample S with equal prob. s/n
How to think about the problem: say s = 2
Stream: a x c y z k c d e g…
At n= 5, each of the first 5 tuples is included in the sample S with equal prob.
At n= 7, each of the first 7 tuples is included in the sample S with equal prob.
Impractical solution would be to store all the n tuples seen
so far and out of them pick s at random
18
¡ Algorithm (a.k.a. Reservoir Sampling)
§ Store all the first s elements of the stream to S
§ Suppose we have seen n-1 elements, and now
the nth element arrives (n > s)
§ With probability s/n, keep the nth element, else discard it
§ If we picked the nth element, then it replaces one of the
s elements in the sample S, picked uniformly at random

¡ Claim: This algorithm maintains a sample S

with the desired property:
§ After n elements, the sample contains each
element seen so far with probability s/n 19
¡ We prove this by induction:
§ Assume that after n elements, the sample contains
each element seen so far with probability s/n
§ We need to show that after seeing element n+1
the sample maintains the property
§ Sample contains each element seen so far with
probability s/(n+1)
¡ Base case:
§ After we see n=s elements the sample S has the
desired property
§ Each out of n=s elements is in the sample with
probability s/s = 1

20
¡ Inductive hypothesis: After n elements, the sample
S contains each element seen so far with prob. s/n
¡ Now element n+1 arrives
¡ Inductive step: For elements already in S,
probability that the algorithm keeps it in S is:
æ s ö æ s öæ s - 1 ö n
ç1 - ÷+ç ÷ç ÷=
è n + 1 ø Element
è n +n+11 øè Element
s ø n +1
in the
Element n+1 discarded
not discarded sample not picked
¡ So, at time n, tuples in S were there with prob. s/n
¡ Time n®n+1, tuple stayed in S with prob. n/(n+1)
𝒔 𝒏 𝒔
¡ So prob. tuple is in S at time n+1 = ⋅ =
𝒏 𝒏%𝟏 𝒏%𝟏
21
¡ A useful model of stream processing is that
queries are about a window of length N –
the N most recent elements received
¡ Interesting case: N is so large that the data
cannot be stored in memory, or even on disk
§ Or, there are so many streams that windows
for all cannot be stored
¡ Amazon example:
§ For every product X we keep 0/1 stream of whether
that product was sold in the n-th transaction
§ We want answer queries, how many times have we
sold X in the last k sales
23
¡ Sliding window on a single stream: N=6

qwertyuiopasdfghjklzxcvbnm

qwertyuiopasdfghjklzxcvbnm

qwertyuiopasdfghjklzxcvbnm

qwertyuiopasdfghjklzxcvbnm

Past Future

24
¡ Problem:
§ Given a stream of 0s and 1s
§ Be prepared to answer queries of the form
How many 1s are in the last k bits? where k ≤ N

¡ Obvious solution:
Store the most recent N bits
§ When new bit comes in, discard the N+1st bit
010011011101010110110110 Suppose N=6
Past Future

25
¡ You can not get an exact answer without
storing the entire window

¡ Real Problem:
What if we cannot afford to store N bits?
§ E.g., we’re processing 1 billion streams and
N = 1 billion 0 1 0 0 1 1 0 1 1 1 0 1 0 1 0 1 1 0 1 1 0 1 1 0
Past Future

¡ But we are happy with an approximate

answer
26
 ¡ Q: How many 1s are in the last N bits?
¡ A simple solution that does not really solve
our problem: Uniformity assumption
N
010011100010100100010110110111001010110011010
Past Future

¡ Maintain 2 counters:
§ S: number of 1s from the beginning of the stream
§ Z: number of 0s from the beginning of the stream
𝑺
¡ How many 1s are in the last N bits? 𝑵 1
𝑺%𝒁
¡ But, what if stream is non-uniform?
§ What if distribution changes over time?
27
 [Datar, Gionis, Indyk, Motwani]

¡ DGIM solution that does not assume

uniformity

¡ We store 𝑶(log𝟐𝑵) bits per stream

¡ Solution gives approximate answer,

never off by more than 50%
§ Error factor can be reduced to any fraction > 0,
with more complicated algorithm and
proportionally more stored bits

28
 ¡ Solution that doesn’t (quite) work:
§ Summarize exponentially increasing regions
of the stream, looking backward
§ Drop small regions if they begin at the same point
Window of as a larger region
width 16
has 6 1s 6 10
? 4
3 2
1 2
1 0
010011100010100100010110110111001010110011010
N
We can reconstruct the count of the last N bits, except we
are not sure how many of the last 6 1s are included in the N
29
¡ Stores only O(log2N ) bits
§ 𝑶(log 𝑵) counts of log 𝟐 𝑵 bits each

¡ Easy update as more bits enter

¡ Error in count no greater than the number

of 1s in the “unknown” area

30
 ¡ As long as the 1s are fairly evenly distributed,
the error due to the unknown region is small
– no more than 50%
¡ But it could be that all the 1s are in the
unknown area at the end
¡ In that case, the error is unbounded!
6 10
? 4
3 2
1 2
1 0
010011100010100100010110110111001010110011010
N
31
 [Datar, Gionis, Indyk, Motwani]

¡ Idea: Instead of summarizing fixed-length

blocks, summarize blocks with specific
number of 1s:
§ Let the block sizes (number of 1s) increase
exponentially

¡ When there are few 1s in the window, block

sizes stay small, so errors are small
1001010110001011010101010101011010101010101110101010111010100010110010
N

32
¡ Each bit in the stream has a timestamp,
starting 1, 2, …

¡ Record timestamps modulo N (the window

size), so we can represent any relevant
timestamp in 𝑶(𝒍𝒐𝒈𝟐 𝑵) bits

33
 ¡ A bucket in the DGIM method is a record
consisting of:
§ (A) The timestamp of its end [O(log N) bits]
§ (B) The number of 1s between its beginning and
end [O(log log N) bits]

¡ Constraint on buckets:
Number of 1s must be a power of 2
§ That explains the O(log log N) in (B) above
1001010110001011010101010101011010101010101110101010111010100010110010
N
34
¡ Either one or two buckets with the same
power-of-2 number of 1s

¡ Buckets do not overlap in timestamps

¡ Buckets are sorted by size

§ Earlier buckets are not smaller than later buckets

¡ Buckets disappear when their

end-time is > N time units in the past

35
 At least 1 of 2 of 2 of 1 of 2 of
size 16. Partially size 8 size 4 size 2 size 1
beyond window.

1001010110001011010101010101011010101010101110101010111010100010110010

Three properties of buckets that are maintained:

- Either one or two buckets with the same power-of-2 number of 1s
- Buckets do not overlap in timestamps
- Buckets are sorted by size
36
¡ When a new bit comes in, drop the last
(oldest) bucket if its end-time is prior to N
time units before the current time

¡ 2 cases: Current bit is 0 or 1

¡ If the current bit is 0:

no other changes are needed

37
¡ If the current bit is 1:
§ (1) Create a new bucket of size 1, for just this bit
§ End timestamp = current time
§ (2) If there are now three buckets of size 1,
combine the oldest two into a bucket of size 2
§ (3) If there are now three buckets of size 2,
combine the oldest two into a bucket of size 4
§ (4) And so on …

38
Current state of the stream:
1001010110001011010101010101011010101010101110101010111010100010110010

Bit of value 1 arrives

0010101100010110101010101010110101010101011101010101110101000101100101
Two orange buckets get merged into a yellow bucket
0010101100010110101010101010110101010101011101010101110101000101100101
Next bit 1 arrives, new orange bucket is created, then 0 comes, then 1:
0101100010110101010101010110101010101011101010101110101000101100101101

Buckets get merged…

0101100010110101010101010110101010101011101010101110101000101100101101

State of the buckets after merging

0101100010110101010101010110101010101011101010101110101000101100101101

39
¡ To estimate the number of 1s in the most
recent N bits:
1. Sum the sizes of all buckets but the last
(note “size” means the number of 1s in the bucket)
2. Add half the size of the last bucket

¡ Remember: We do not know how many 1s

of the last bucket are still within the wanted
window

40
 At least 1 of 2 of 2 of 1 of 2 of
size 16. Partially size 8 size 4 size 2 size 1
beyond window.

1001010110001011010101010101011010101010101110101010111010100010110010

41
 ¡ Why is error 50%? Let’s prove it!
¡ Suppose the last bucket has size 2r
¡ Then by assuming 2r-1 (i.e., half) of its 1s are
still within the window, we make an error of
at most 2r-1
¡ Since there is at least one bucket of each of
the sizes less than 2r, the true sum is at least
1 + 2 + 4 + .. + 2r-1 = 2r -1
¡ Thus, error at most 50% At least 16 1s

111111110000000011101010101011010101010101110101010111010100010110010
N
42
¡ Instead of maintaining 1 or 2 of each size
bucket, we allow either r-1 or r buckets (r > 2)
§ Except for the largest size buckets; we can have
any number between 1 and r of those
¡ Error is at most O(1/r)
¡ By picking r appropriately, we can tradeoff
between number of bits we store and the
error

43
 ¡ Can we use the same trick to answer queries
How many 1’s in the last k? where k < N?
§ A: Find earliest bucket B that at overlaps with k.
Number of 1s is the sum of sizes of more recent
buckets + ½ size of B
1001010110001011010101010101011010101010101110101010111010100010110010
k

¡ Can we handle the case where the stream is

not bits, but integers, and we want the sum
of the last k elements?
44
 ¡ Stream of positive integers
¡ We want the sum of the last k elements
§ Amazon: Avg. price of last k sales
¡ Solution:
§ (1) If you know all have at most m bits
§ Treat m bits of each integer as a separate stream
§ Use DGIM to count 1s in each integer ci …estimated count for i-th bit
§ The sum is = ∑CDE
AFG 𝑐A 2
A

§ (2) Use buckets to keep partial sums

§ Sum of elements in size b bucket is at most 2b
2 5 7 1 3 8 4 6 7 9 1 3 7 6 5 3 5 7 1 3 3 1 2 2 6 Idea: Sum in each
bucket is at most
2 5 7 1 3 8 4 6 7 9 1 3 7 6 5 3 5 7 1 3 3 1 2 2 6 3 2b (unless bucket
has only 1 integer)
2 5 7 1 3 8 4 6 7 9 1 3 7 6 5 3 5 7 1 3 3 1 2 2 6 3 2 Bucket sizes:
2 5 7 1 3 8 4 6 7 9 1 3 7 6 5 3 5 7 1 3 3 1 2 2 6 3 2 5 16 8 4 2 1
45
¡ Sampling a fixed proportion of a stream
§ Sample size grows as the stream grows
¡ Sampling a fixed-size sample
§ Reservoir sampling
¡ Counting the number of 1s in the last N
elements
§ Exponentially increasing windows
§ Extensions:
§ Number of 1s in any last k (k < N) elements
§ Sums of integers in the last N elements

46
¡ More algorithms for streams:
§ (1) Filtering a data stream: Bloom filters
§ Select elements with property x from stream
§ (2) Counting distinct elements: Flajolet-Martin
§ Number of distinct elements in the last k elements
of the stream
§ (3) Estimating moments: AMS method
§ Estimate std. dev. of last k elements
§ (4) Counting frequent items

48
¡ Each element of data stream is a tuple
¡ Given a list of keys S
¡ Determine which tuples of stream are in S

¡ Obvious solution: Hash table

§ But suppose we do not have enough memory to
store all of S in a hash table
§ E.g., we might be processing millions of filters
on the same stream

50
¡ Example: Email spam filtering
§ We know 1 billion “good” email addresses
§ If an email comes from one of these, it is NOT
spam

¡ Publish-subscribe systems
§ You are collecting lots of messages (news articles)
§ People express interest in certain sets of keywords
§ Determine whether each message matches user’s
interest

51
¡ Given a set of keys S that we want to filter
¡ Create a bit array B of n bits, initially all 0s
¡ Choose a hash function h with range [0,n)
¡ Hash each member of sÎ S to one of
n buckets, and set that bit to 1, i.e., B[h(s)]=1
¡ Hash each element a of the stream and
output only those that hash to bit that was
set to 1
§ Output a if B[h(a)] == 1

52
 Output the item since it may be in S.
Item hashes to a bucket that at least
one of the items in S hashed to.
Item

Hash
func h

0010001011000 Bit array B

Drop the item.

It hashes to a bucket set
to 0 so it is surely not in S.

¡ Creates false positives but no false negatives

§ If the item is in S we surely output it, if not we may
still output it
53
¡ |S| = 1 billion email addresses
|B|= 1GB = 8 billion bits
¡ If the email address is in S, then it surely
hashes to a bucket that has the big set to 1,
so it always gets through (no false negatives)
¡ Approximately 1/8 of the bits are set to 1, so
about 1/8th of the addresses not in S get
through to the output (false positives)
§ Actually, less than 1/8th, because more than one
address might hash to the same bit
54
¡ More accurate analysis for the number of
false positives

¡ Consider: If we throw m darts into n equally

likely targets, what is the probability that
a target gets at least one dart?

¡ In our case:
§ Targets = bits/buckets
§ Darts = hash values of items

55
¡ We have m darts, n targets
¡ What is the probability that a target gets at
least one dart?
Equals 1/e
Equivalent
as n ®∞

n( m / n)
1 - (1 – 1/n)
1 – e–m/n
Probability some
target X not hit
Probability at
by a dart least one dart
hits target X
56
¡ Fraction of 1s in the array B =
= probability of false positive = 1 – e-m/n

¡ Example: 109 darts, 8·109 targets

§ Fraction of 1s in B = 1 – e-1/8 = 0.1175
§ Compare with our earlier estimate: 1/8 = 0.125

57
¡ Consider: |S| = m, |B| = n
¡ Use k independent hash functions h1 ,…, hk
¡ Initialization:
§ Set B to all 0s
§ Hash each element sÎ S using each hash function hi,
set B[hi(s)] = 1 (for each i = 1,.., k) (note: we have a
single array B!)
¡ Run-time:
§ When a stream element with key x arrives
§ If B[hi(x)] = 1 for all i = 1,..., k then declare that x is in S
§ That is, x hashes to a bucket set to 1 for every hash function hi(x)
§ Otherwise discard the element x
58
¡ What fraction of the bit vector B are 1s?
§ Throwing k·m darts at n targets
§ So fraction of 1s is (1 – e-km/n)

¡ But we have k independent hash functions

and we only let the element x through if all k
hash element x to a bucket of value 1

¡ So, false positive probability = (1 – e-km/n)k

59
 m = 1 billion, n = 8 billion
0.2

¡ 0.18

§ k = 1: (1 – e-1/8) = 0.1175 0.16

False positive prob.

0.14

§ k = 2: (1 – e-1/4)2 = 0.0493 0.12

0.1

0.08

¡ What happens as we 0.06

0.04

keep increasing k? 0.02

0 2 4 6 8 10 12 14 16 18 20

Number of hash functions, k

¡ “Optimal” value of k: n/m ln(2)

§ In our case: Optimal k = 8 ln(2) = 5.54 ≈ 6
§ Error at k = 6: (1 – e-1/6)2 = 0.0235
60
¡ Bloom filters guarantee no false negatives,
and use limited memory
§ Great for pre-processing before more
expensive checks
¡ Suitable for hardware implementation
§ Hash function computations can be parallelized

¡ Is it better to have 1 big B or k small Bs?

§ It is the same: (1 – e-km/n)k vs. (1 – e-m/(n/k))k
§ But keeping 1 big B is simpler

61
¡ Problem:
§ Data stream consists of a universe of elements
chosen from a set of size N
§ Maintain a count of the number of distinct
elements seen so far

¡ Obvious approach:
Maintain the set of elements seen so far
§ That is, keep a hash table of all the distinct
elements seen so far

63
¡ How many different words are found among
the Web pages being crawled at a site?
§ Unusually low or high numbers could indicate
artificial pages (spam?)

¡ How many different Web pages does each

customer request in a week?

¡ How many distinct products have we sold in

the last week?

64
¡ Real problem: What if we do not have space
to maintain the set of elements seen so far?

¡ Estimate the count in an unbiased way

¡ Accept that the count may have a little error,

but limit the probability that the error is large

65
¡ Pick a hash function h that maps each of the
N elements to at least log2 N bits

¡ For each stream element a, let r(a) be the

number of trailing 0s in h(a)
§ r(a) = position of first 1 counting from the right
§ E.g., say h(a) = 12, then 12 is 1100 in binary, so r(a) = 2
¡ Record R = the maximum r(a) seen
§ R = maxa r(a), over all the items a seen so far

¡ Estimated number of distinct elements = 2R

66
¡ Very very rough and heuristic intuition why
Flajolet-Martin works:
§ h(a) hashes a with equal prob. to any of N values
§ Then h(a) is a sequence of log2 N bits,
where 2-r fraction of all as have a tail of r zeros
§ About 50% of as hash to ***0
§ About 25% of as hash to **00
§ So, if we saw the longest tail of r=2 (i.e., item hash
ending *100) then we have probably seen
about 4 distinct items so far
§ So, it takes to hash about 2r items before we
see one with zero-suffix of length r
67
¡ Now we show why Flajolet-Martin works

¡ Formally, we will show that probability of

finding a tail of r zeros:
§ Goes to 1 if 𝒎 ≫ 𝟐𝒓
§ Goes to 0 if 𝒎 ≪ 𝟐𝒓
where 𝒎 is the number of distinct elements
seen so far in the stream
¡ Thus, 2R will almost always be around m!

68
¡ What is the probability that a given h(a) ends
in at least r zeros is 2-r
§ h(a) hashes elements uniformly at random
§ Probability that a random number ends in
at least r zeros is 2-r
¡ Then, the probability of NOT seeing a tail
of length r among m elements:
𝟏 − 𝟐D𝒓 𝒎

Prob. all end in Prob. that given h(a) ends

fewer than r zeros. in fewer than r zeros

69
 -r m - r 2r ( m 2- r ) - m 2- r
¡ Note: (1 - 2 ) = (1 - 2 ) »e
¡ Prob. of NOT finding a tail of length r is:
§ If m << 2r, then prob. tends to 1
-r m
§ (1 - 2 ) » e - m 2- r
= 1 as m/2r® 0
§ So, the probability of finding a tail of length r tends to 0
§ If m >> 2r, then prob. tends to 0
-r m
§ (1 - 2 ) » e - m 2- r
= 0 as m/2r ® ¥
§ So, the probability of finding a tail of length r tends to 1

¡ Thus, 2R will almost always be around m!

70
¡ E[2R] is actually infinite
§ Probability halves when R ® R+1, but value doubles
¡ Workaround involves using many hash
functions hi and getting many samples of Ri
¡ How are samples Ri combined?
§ Average? What if one very large value 𝟐𝑹𝒊 ?
§ Median? All estimates are a power of 2
§ Solution:
§ Partition your samples into small groups
§ Take the median of groups
§ Then take the average of the medians
71
¡ Suppose a stream has elements chosen
from a set A of N values

¡ Let mi be the number of times value i occurs

in the stream

¡ The kth moment is

å iÎA
( mi ) k

73
 å iÎA
( mi ) k

¡ 0thmoment = number of distinct elements

§ The problem just considered
¡ 1st moment = count of the numbers of
elements = length of the stream
§ Easy to compute
¡ 2nd moment = surprise number S =
a measure of how uneven the distribution is

74
¡ Stream of length 100
¡ 11 distinct values

¡ Item counts: 10, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9

Surprise S = 910

¡ Item counts: 90, 1, 1, 1, 1, 1, 1, 1 ,1, 1, 1

Surprise S = 8,110

75
 [Alon, Matias, and Szegedy]

¡ AMS method works for all moments

¡ Gives an unbiased estimate
¡ We will just concentrate on the 2nd moment S
¡ We pick and keep track of many variables X:
§ For each variable X we store X.el and X.val
§ X.el corresponds to the item i
§ X.val corresponds to the count of item i
§ Note this requires a count in main memory,
so number of Xs is limited
¡ Our goal is to compute 𝑺 = ∑𝒊 𝒎𝟐𝒊
76
¡ How to set X.val and X.el?
§ Assume stream has length n (we relax this later)
§ Pick some random time t (t<n) to start,
so that any time is equally likely
§ Let at time t the stream have item i. We set X.el = i
§ Then we maintain count c (X.val = c) of the number
of is in the stream starting from the chosen time t
¡ Then the estimate of the 2nd moment (∑𝒊 𝒎𝟐𝒊 ) is:
𝑺 = 𝒇(𝑿) = 𝒏 (𝟐 · 𝒄 – 𝟏)
§ Note, we will keep track of multiple Xs, (X1, X2,… Xk)
and our final estimate will be 𝑺 = 𝟏/𝒌 ∑𝒌𝒋 𝒇(𝑿𝒋 )
77
 Count: 1 2 3 ma

Stream: a a b b b a b a
𝟐
¡ 2nd moment is 𝑺 = ∑𝒊 𝒎𝒊
¡ ct … number of times item at time t appears
from time t onwards (c1=ma , c2=ma-1, c3=mb)
𝟏 𝒏
¡ 𝑬 𝒇(𝑿) = ∑𝒕F𝟏 𝒏(𝟐𝒄𝒕 − 𝟏) m … total count of
𝒏 i
item i in the stream
𝟏
∑𝒊 𝒏
(we are assuming
= (𝟏 + 𝟑 + 𝟓 + ⋯ + 𝟐𝒎𝒊 − 𝟏) stream has length n)
𝒏
Time t when Time t when
Time t when the penultimate the first i is
Group times
the last i is i is seen (ct=2) seen (ct=mi)
by the value
seen (ct=1)
seen
78
 Count: 1 2 3 ma

Stream: a a b b b a b a
E
¡ 𝐸 𝑓(𝑋) = ∑A 𝑛 (1 + 3 + 5 + ⋯ + 2𝑚A − 1)
a
§ Little side calculation: 1 + 3 + 5 + ⋯ + 2𝑚A − 1 =
Ch Ch Ch %E
∑AFE(2𝑖 − 1) = 2 − 𝑚A = (𝑚A )i
i
𝟏 𝟐
¡ Then 𝑬 𝒇(𝑿) = ∑𝒊 𝒏 𝒎𝒊
𝒏

¡ So, 𝐄 𝐟(𝐗) = ∑𝒊 𝒎𝒊 𝟐 = 𝑺
¡ We have the second moment (in expectation)!

79
¡ For estimating kth moment we essentially use the
same algorithm but change the estimate:
§ For k=2 we used n (2·c – 1)
§ For k=3 we use: n (3·c2 – 3c + 1) (where c=X.val)
¡ Why?
§ For k=2: Remember we had 1 + 3 + 5 + ⋯ + 2𝑚A − 1
and we showed terms 2c-1 (for c=1,…,m) sum to m2
§ ∑C C i C
mFE 2𝑐 − 1 = ∑mFE 𝑐 − ∑mFE 𝑐 − 1
i
= 𝑚i
§ So: 𝟐𝒄 − 𝟏 = 𝒄𝟐 − 𝒄 − 𝟏 𝟐
§ For k=3: c3 - (c-1)3 = 3c2 - 3c + 1
¡ Generally: Estimate = 𝑛 (𝑐 n − 𝑐 − 1 n )
80
¡ In practice:
§ Compute 𝒇(𝑿) = 𝒏(𝟐 𝒄 – 𝟏) for
as many variables X as you can fit in memory
§ Average them in groups
§ Take median of averages

¡ Problem: Streams never end

§ We assumed there was a number n,
the number of positions in the stream
§ But real streams go on forever, so n is
a variable – the number of inputs seen so far

81
¡ (1) The variables X have n as a factor –
keep n separately; just hold the count in X
¡ (2) Suppose we can only store k counts.
We must throw some Xs out as time goes on:
§ Objective: Each starting time t is selected with
probability k/n
§ Solution: (fixed-size sampling!)
§ Choose the first k times for k variables
§ When the nth element arrives (n > k), choose it with
probability k/n
§ If you choose it, throw one of the previously stored
variables X out, with equal probability
82
 ¡ New Problem: Given a stream, which items
appear more than s times in the window?
¡ Possible solution: Think of the stream of
baskets as one binary stream per item
§ 1 = item present; 0 = not present
§ Use DGIM to estimate counts of 1s for all items
6 10
4
3 2
1 2
1 0
010011100010100100010110110111001010110011010
N
84
¡ In principle, you could count frequent pairs
or even larger sets the same way
§ One stream per itemset

¡ Drawbacks:
§ Only approximate
§ Number of itemsets is way too big

85
¡ Exponentially decaying windows: A heuristic
for selecting likely frequent item(sets)
§ What are “currently” most popular movies?
§ Instead of computing the raw count in last N elements
§ Compute a smooth aggregation over the whole stream
¡ If stream is a1, a2,… and we are taking the sum
of the stream, take the answer at time t to be:
𝒕
= 𝒊F𝟏 𝒂𝒊 𝟏 − 𝒄 𝒕D𝒊
∑
§ c is a constant, presumably tiny, like 10-6 or 10-9
¡ When new at+1 arrives:
Multiply current sum by (1-c) and add at+1
86
¡ If each ai is an “item” we can compute the
characteristic function of each possible
item x as an Exponentially Decaying Window
§ That is: ∑𝒕𝒊F𝟏 𝜹𝒊 ⋅ 𝟏 − 𝒄 𝒕D𝒊
where δi=1 if ai=x, and 0 otherwise
§ Imagine that for each item x we have a binary
stream (1 if x appears, 0 if x does not appear)
§ New item x arrives:
§ Multiply all counts by (1-c)
§ Add +1 to count for element x
¡ Call this sum the “weight” of item x
87
...

1/c
¡ Important property: Sum over all weights
∑𝒕 𝟏 − 𝒄 𝒕 is 1/[1 – (1 – c)] = 1/c

88
¡ What are “currently” most popular movies?
¡ Suppose we want to find movies of weight > ½
§ Important property: Sum over all weights
∑q 1 − 𝑐 q is 1/[1 – (1 – c)] = 1/c
¡ Thus:
§ There cannot be more than 2/c movies with
weight of ½ or more
¡ So, 2/c is a limit on the number of
movies being counted at any time

89
¡ Count (some) itemsets in an E.D.W.
§ What are currently “hot” itemsets?
§ Problem: Too many itemsets to keep counts of
all of them in memory
¡ When a basket B comes in:
§ Multiply all counts by (1-c)
§ For uncounted items in B, create new count
§ Add 1 to count of any item in B and to any itemset
contained in B that is already being counted
§ Drop counts < ½
§ Initiate new counts (next slide)
90
¡ Start a count for an itemset S ⊆ B if every
proper subset of S had a count prior to arrival
of basket B
§ Intuitively: If all subsets of S are being counted
this means they are “frequent/hot” and thus S has
a potential to be “hot”
¡ Example:
§ Start counting S={i, j} iff both i and j were counted
prior to seeing B
§ Start counting S={i, j, k} iff {i, j}, {i, k}, and {j, k}
were all counted prior to seeing B

91
¡ Counts for single items < (2/c)·(avg. number
of items in a basket)

¡ Counts for larger itemsets = ??

¡ But we are conservative about starting

counts of large sets
§ If we counted every set we saw, one basket
of 20 items would initiate 1M counts

92
¡ More algorithms for streams:
§ (1) Filtering a data stream: Bloom filters
§ Select elements with property x from stream
§ (2) Counting distinct elements: Flajolet-Martin
§ Number of distinct elements in the last k elements
of the stream
§ (3) Estimating moments: AMS method
§ Estimate std. dev. of last k elements
§ (4) Counting frequent items

93