Data Stream Algorithms Intro, Sampling, Entropy
Graham Cormode [email protected]
Data Stream Algorithms2
Outline
Introduction to Data Streams– Motivating examples and applications– Data Streaming models– Basic tail bounds
Sampling from data streams Sampling to estimate entropy
Data Stream Algorithms3
Data is growing faster than our ability to store or index it
There are 3 Billion Telephone Calls in US each day, 30 Billion emails daily, 1 Billion SMS, IMs.
Scientific data: NASA's observation satellites generate billions of readings each per day.
IP Network Traffic: up to 1 Billion packets per hour per router. Each ISP has many (hundreds) routers!
Whole genome sequences for many species now available: each megabytes to gigabytes in size
Data is Massive
Data Stream Algorithms4
Massive Data Analysis
Must analyze this massive data: Scientific research (monitor environment, species) System management (spot faults, drops, failures) Customer research (association rules, new offers) For revenue protection (phone fraud, service abuse)
Else, why even measure this data?
Data Stream Algorithms5
Example: Network Data
Networks are sources of massive data: the metadata per hour per router is gigabytes
Fundamental problem of data stream analysis: Too much information to store or transmit
So process data as it arrives: one pass, small space: the data stream approach.
Approximate answers to many questions are OK, if there are guarantees of result quality
Data Stream Algorithms6
IP Network Monitoring Application
24x7 IP packet/flow data-streams at network elements Truly massive streams arriving at rapid rates
– AT&T/Sprint collect ~1 Terabyte of NetFlow data each day
Often shipped off-site to data warehouse for off-line analysis
Source Destination Duration Bytes Protocol 10.1.0.2 16.2.3.7 12 20K http 18.6.7.1 12.4.0.3 16 24K http 13.9.4.3 11.6.8.2 15 20K http 15.2.2.9 17.1.2.1 19 40K http 12.4.3.8 14.8.7.4 26 58K http 10.5.1.3 13.0.0.1 27 100K ftp 11.1.0.6 10.3.4.5 32 300K ftp 19.7.1.2 16.5.5.8 18 80K ftp
Example NetFlow IP Session Data
DSL/CableNetworks
• Broadband Internet Access
Converged IP/MPLSCore
PSTNEnterpriseNetworks
• Voice over IP• FR, ATM, IP VPN
Network OperationsCenter (NOC)
SNMP/RMON,NetFlow records
Peer
Data Stream Algorithms8
Network Monitoring Queries
DBMS(Oracle, DB2)
Back-end Data Warehouse
Off-line analysis – slow, expensive
DSL/CableNetworks
EnterpriseNetworks
Peer
Network OperationsCenter (NOC)
What are the top (most frequent) 1000 (source, dest) pairs seen over the last month?
SELECT COUNT (R1.source, R2.dest)FROM R1, R2WHERE R1.dest = R2.source
SQL Join Query
How many distinct (source, dest) pairs have been seen by both R1 and R2 but not R3?
Set-Expression Query
PSTN
Extra complexity comes from limited space and time Will introduce solutions for these and other problems
R1
R2
R3
Data Stream Algorithms10
Other Streaming Applications
Sensor networks– Monitor habitat and environmental parameters– Track many objects, intrusions, trend analysis…
Utility Companies– Monitor power grid, customer usage patterns etc.– Alerts and rapid response in case of problems
Data Stream Algorithms11
Streams Defining Frequency Dbns.
We will consider streams that define frequency distributions– E.g. frequency of packets from source A to source B
This simple setting captures many of the core algorithmic problems in data streaming– How many distinct (non-zero) values seen?– What is the entropy of the frequency distribution?– What (and where) are the highest frequencies?
More generally, can consider streams that define multi-dimensional distributions, graphs, geometric data etc.
But even for frequency distributions, several models are relevant
Data Stream Algorithms12
Data Stream Models
We model data streams as sequences of simple tuples Complexity arises from massive length of streams Arrivals only streams:
– Example: (x, 3), (y, 2), (x, 2) encodesthe arrival of 3 copies of item x, 2 copies of y, then 2 copies of x.
– Could represent eg. packets on a network; power usage Arrivals and departures:
– Example: (x, 3), (y,2), (x, -2) encodes final state of (x, 1), (y, 2).
– Can represent fluctuating quantities, or measure differences between two distributions
xy
xy
Data Stream Algorithms13
Approximation and Randomization
Many things are hard to compute exactly over a stream– Is the count of all items the same in two different streams?– Requires linear space to compute exactly
Approximation: find an answer correct within some factor– Find an answer that is within 10% of correct result– More generally, a (1 ) factor approximation
Randomization: allow a small probability of failure– Answer is correct, except with probability 1 in 10,000– More generally, success probability (1-)
Approximation and Randomization: (, )-approximations
Data Stream Algorithms14
Basic Tools: Tail Inequalities General bounds on tail probability of a random variable
(probability that a random variable deviates far from its expectation)
Basic Inequalities: Let X be a random variable with expectation and variance Var[X]. Then, for any >0
Probabilitydistribution
Tail probability
Chebyshev:
22εμ
Var[X]με)|μXPr(|
Markov:
ε1
1ε)μ)(1Pr(X
Data Stream Algorithms15
Tail Bounds
Markov Inequality:For a random variable Y which takes only non-negative values.
Pr[Y k] E(Y)/k(This will be < 1 only for k > E(Y))
Chebyshev’s Inequality:For a random variable Y:
Pr[|Y-E(Y)| k] Var(Y)/k2
Proof: Set X = (Y – E(Y))2
E(X) = E(Y2+E(Y)2–2YE(Y)) = E(Y2)+E(Y)2-2E(Y)2= Var(Y)
So: Pr[|Y-E(Y)| k] = Pr[(Y – E(Y))2 k2].
Using Markov: E(Y – E(Y))2/k2 = Var(Y)/k2
Data Stream Algorithms16
Outline
Introduction to Data Streams– Motivating examples and applications– Data Streaming models– Basic tail bounds
Sampling from data streams Sampling to estimate entropy
Data Stream Algorithms17
Sampling From a Data Stream
Fundamental prob: sample m items uniformly from stream– Useful: approximate costly computation on small sample
Challenge: don’t know how long stream is – So when/how often to sample?
Two solutions, apply to different situations:– Reservoir sampling (dates from 1980s?)– Min-wise sampling (dates from 1990s?)
Data Stream Algorithms18
Reservoir Sampling
Sample first m items Choose to sample the i’th item (i>m) with probability m/i If sampled, randomly replace a previously sampled item
Optimization: when i gets large, compute which item will be sampled next, skip over intervening items. [Vitter 85]
Data Stream Algorithms19
Reservoir Sampling - Analysis
Analyze simple case: sample size m = 1 Probability i’th item is the sample from stream length n:
– Prob. i is sampled on arrival prob. i survives to end
1 i i+1 n-2 n-1 i i+1 i+2 n-1 n
…
= 1/n
Case for m > 1 is similar, easy to show uniform probability Drawbacks of reservoir sampling: hard to parallelize
Data Stream Algorithms20
Min-wise Sampling
For each item, pick a random fraction between 0 and 1 Store item(s) with the smallest random tag [Nath et al.’04]
0.391 0.908 0.291 0.555 0.619 0.273
Each item has same chance of least tag, so uniform Can run on multiple streams separately, then merge
Data Stream Algorithms21
Sampling Exercises
What happens when each item in the stream also has a weight attached, and we want to sample based on these weights?1. Generalize the reservoir sampling algorithm to draw a
single sample in the weighted case.
2. Generalize reservoir sampling to sample multiple weighted items, and show an example where it fails to give a meaningful answer.
3. Research problem: design new streaming algorithms for sampling in the weighted case, and analyze their properties.
Data Stream Algorithms22
Outline
Introduction to Data Streams– Motivating examples and applications– Data Streaming models– Basic tail bounds
Sampling from data streams Sampling to estimate entropy
Data Stream Algorithms23
Application of Sampling: Entropy
Given a long sequence of characters
S = <a1, a2, a3… am> each aj {1… n}
Let fi = frequency of i in the sequence Compute the empirical entropy:
H(S) = - i fi/m log fi/m = - i pi log pi
Example: S = < a, b, a, b, c, a, d, a>– pa = 1/2, pb = 1/4, pc = 1/8, pd = 1/8
– H(S) = ½ + ¼ 2 + 1/8 3 + 1/8 3 = 7/4 Entropy promoted for anomaly detection in networks
Data Stream Algorithms24
Challenge
Goal: approximate H(S) in space sublinear (poly-log) in m (stream length), n (alphabet size)– (,) approx: answer is (1§)H(S) w/prob 1-
Easy if we have O(n) space: compute each fi exactly More challenging if n is huge, m is huge, and we have
only one pass over the input in order– (The data stream model)
Data Stream Algorithms25
Sampling Based Algorithm
Simple estimator: – Randomly sample a position j in the stream
– Count how many times aj appears subsequently = r
– Output X = -(r log (r/m) – (r-1) log((r-1)/m))
Claim: Estimator is unbiased – E[X] = H(S)– Proof: prob of picking j = 1/m, sum telescopes correctly
Variance of estimate is not too large – Var[X] = O(log2 m)– Observe that |X| ≤ log m– Var[X] = E[(X – E[X])2] < (max(X) – min(X))2 = O(log2 m)
Data Stream Algorithms26
Analysis of Basic Estimator
A general technique in data streams:– Repeat in parallel an unbiased estimator with bounded
variance, take average of estimates to improve variance
– Var[ 1/k (Y1 + Y2 + ... Yk) ] = 1/k Var[Y]
– By Chebyshev, need k repetitions to be Var[X]/2E2[X] – For entropy, this means space k = O(log2m/2H2(S))
Problem for entropy: when H(S) is very small?– Space needed for an accurate approx goes as 1/H2!
22εμ
Var[X]με)|μXPr(|
Data Stream Algorithms27
Low Entropy
But... what does a low entropy stream look like? – aaaaaaaaaaaaaaaaaaaaaaaaaaaaaabaaaaa
Very boring most of the time, we are only rarely surprised Can there be two frequent items?
– aabababababababaababababbababababababa– No! That’s high entropy (¼ 1 bit / character)
Only way to get H(S) =o(1) is to have only one character with pi close to 1
Data Stream Algorithms28
Removing the frequent character
Write entropy as – -pa log pa + (1-pa) H(S’)
– Where S’ = stream S with all ‘a’s removed Can show:
– Doesn’t matter if H(S’) is small: as pa is large, additive error on H(S’) ensures relative error on (1-pa)H(S’)
– Relative error (1-pa) on pa gives relative error on pa log pa
– Summing both (positive) terms gives relative error overall
Data Stream Algorithms29
Finding the frequency character
Ejecting a is easy if we know in advance what it is– Can then compute pa exactly
Can find online deterministically– Assume pa > 2/3 (if not, H(S) > 0.9, and original alg works)
– Run a ‘heavy hitters’ algorithm on the stream (see later)
– Modify analysis, find a and pa § (1-pa)
But... how to also compute H(S’) simultaneously if we don’t know a from the start... do we need two passes?
Data Stream Algorithms30
Always have a back up plan...
Idea: keep two samples to build our estimator– If at the end one of our samples is ‘a’, use the other– How to do this and ensure uniform sampling?
Pick first sample with ‘min-wise sampling’: At end of the stream, if the sampled character = ‘a’, we
want to sample from the stream ignoring all ‘a’s This is just “the character achieving the smallest label
distinct from the one that achieves the smallest label” Can track information to do this in a single pass,
constant space
Data Stream Algorithms31
Sampling Two Tokens
Repeats:
0.62
7
0.54
9
0.22
8
0.36
6
0.77
0
0.19
1
0.40
8
Tags:
BCDBBB
AAA
0.20
2
0.17
3
0.08
2
0.21
7
0.81
5
AAAAACStream:
min tag
Assign tags, choose first token as before Delete all occurrences of first token Choose token with min remaining tag; count repeats Implementation: keep track of two triples
(min tag, corresponding token, number of repeats)
second smallest tag, but we don’t want this; same
token as min tag!
min tag amongst remaining tokens
BBBB
Data Stream Algorithms32
Putting it all together
Can combine all these pieces Build an estimator based on tracking this information,
deciding whether there is a frequent character or not A more involved Chernoff bounds argument improves
number of repetitions of estimator from O(-2Var[X]/E2[X]) to O(-2Range[X]/E[X]) = O(-2 log m)
In O(-2 log m log 1/) space (words) we can compute an (,) approximation to H(S) in a single pass
Data Stream Algorithms33
Entropy Exercises
As a subroutine, we need to find an element that occurs more than 2/3 of the time and estimate its weight1. How can we find a frequently occurring item?
2. How can we estimate its weight p with (1-p) error?
3. Our algorithm uses O(-2 log m log 1/) space, could this be improved or is it optimal (lower bounds)?
4. Our algorithm updates each sampled pair for every update, how quickly can we implement it?
5. (Research problem) What if there are multiple distributed streams and we want to compute the entropy of their union?
Data Stream Algorithms34
Outline
Introduction to Data Streams– Motivating examples and applications– Data Streaming models– Basic tail bounds
Sampling from data streams Sampling to estimate entropy
Data Stream Algorithms Frequency Moments
Graham Cormode [email protected]
Data Stream Algorithms36
Frequency Moments
Introduction to Frequency Moments and Sketches Count-Min sketch for Fand frequent items
AMS Sketch for F2
Estimating F0
Extensions: – Higher frequency moments– Combined frequency moments
Data Stream Algorithms37
Last Time
Introduced data streams and data stream models– Focus on a stream defining a frequency distribution
Sampling to draw a uniform sample from the stream Entropy estimation: based on sampling
Data Stream Algorithms38
This Time: Frequency Moments
Given a stream of updates, let fi be the number of times that item i is seen in the stream
Define Fk of the stream as i (fi)k – the k’th Frequency Moment
“Space Complexity of the Frequency Moments” by Alon, Matias, Szegedy in STOC 1996 studied this problem– Awarded Godel prize in 2005– Set the pattern for much streaming algorithms to follow– Frequency moments are at the core of many streaming
problems
Data Stream Algorithms39
Frequency Moments
F0 : count 1 if fi 0 – number of distinct items
F1 : length of stream, easy
F2 : sum the squares of the frequencies – self join size
Fk : related to statistical moments of the distribution
F : (really lim k Fk1/k) dominated by the largest fk, finds
the largest frequency
Different techniques needed for each one.– Mostly sketch techniques, which compute a certain kind of
random linear projection of the stream
Data Stream Algorithms40
Sketches
Not every problem can be solved with sampling– Example: counting how many distinct items in the stream– If a large fraction of items aren’t sampled, don’t know if
they are all same or all different Other techniques take advantage that the algorithm can
“see” all the data even if it can’t “remember” it all (To me) a sketch is a linear transform of the input
– Model stream as defining a vector, sketch is result of multiplying stream vector by an (implicit) matrix
linear projection
Data Stream Algorithms41
Trivial Example of a Sketch
Test if two (asynchronous) binary streams are equal d= (x,y) = 0 iff x=y, 1 otherwise
To test in small space: pick a random hash function h Test h(x)=h(y) : small chance of false positive, no chance
of false negative. Compute h(x), h(y) incrementally as new bits arrive
(e.g. h(x) = xiti mod p for random prime p, and t < p) – Exercise: extend to real valued vectors in update model
1 0 1 1 1 0 1 0 1 …
1 0 1 1 0 0 1 0 1 …
Data Stream Algorithms42
Frequency Moments
Introduction to Frequency Moments and Sketches Count-Min sketch for Fand frequent items
AMS Sketch for F2
Estimating F0
Extensions: – Higher frequency moments– Combined frequency moments
Data Stream Algorithms43
Count-Min Sketch
Simple sketch idea, can be used for as the basis of many different stream mining tasks.
Model input stream as a vector x of dimension U Creates a small summary as an array of w d in size Use d hash function to map vector entries to [1..w] Works on arrivals only and arrivals & departures streams
W
dArray: CM[i,j]
Data Stream Algorithms44
Count-Min Sketch Structure
Each entry in vector x is mapped to one bucket per row. Merge two sketches by entry-wise summation Estimate x[j] by taking mink CM[k,hk(j)]
– Guarantees error less than F1 in size O(1/ log 1/)– Probability of more error is less than 1-
+c
+c
+c
+c
h1(j)
hd(j)
j,+c
d=log 1/
w = 2/
[C, Muthukrishnan ’04]
Data Stream Algorithms45
Approximation of Point Queries
Approximate point query x’[j] = mink CM[k,hk(j)]
Analysis: In k'th row, CM[k,hk(j)] = x[j] + Xk,j
– Xk,j = x[i] | hk(i) = hk(j)
– E(Xk,j) = i j x[i]*Pr[hk(i)=hk(j)] Pr[hk(i)=hk(k)] * i x[i]= F1/2 by pairwise independence of h
– Pr[Xk,j F1] = Pr[Xk,j 2E(Xk,j)] 1/2 by Markov inequality
So, Pr[x’[j] x[j] + F1] = Pr[ k. Xk,j > F1] 1/2log 1/ =
Final result: with certainty x[j] x’[j] and with probability at least 1-, x’[j] < x[j] + F1
Data Stream Algorithms46
Applications of Count-Min to F
Count-Min sketch lets us estimate fi for any i (up to F1)
F asks to find maxi fi
Slow way: test every i after creating sketch Faster way: test every i after it is seen in the stream, and
remember largest estimated value Alternate way:
– keep a binary tree over the domain of input items, where each node corresponds to a subset
– keep sketches of all nodes at same level– descend tree to find large frequencies, discarding
branches with low frequency
Data Stream Algorithms47
Count-Min Exercises
1. The median of a distribution is the item so that the sum of the frequencies of lexicographically smaller items is ½ F1. Use CM sketch to find the (approximate) median.
2. Assume the input frequencies follow the Zipf distribution so that the i’th largest frequency is (i-z) for z>1. Show that CM sketch only needs to be size -1/z to give same guarantee
3. Suppose we have arrival and departure streams where the frequencies of items are allowed to be negative. Extend CM sketch analysis to estimate these frequencies (note, Markov argument no longer works)
4. How to find the large absolute frequencies when some are negative? Or in the difference of two streams?
Data Stream Algorithms48
Frequency Moments
Introduction to Frequency Moments and Sketches Count-Min sketch for Fand frequent items
AMS Sketch for F2
Estimating F0
Extensions: – Higher frequency moments– Combined frequency moments
Data Stream Algorithms49
F2 estimation
AMS sketch (for Alon-Matias-Szegedy) proposed in 1996– Allows estimation of F2 (second frequency moment)
– Used at the heart of many streaming and non-streaming mining applications: achieves dimensionality reduction
Here, describe AMS sketch by generalizing CM sketch. Uses extra hash functions g1...glog 1/ {1...U} {+1,-1}
Now, given update (j,+c), set CM[k,hk(i)] += c*gk(j)
linear projection
AMS sketch
Data Stream Algorithms50
F2 analysis
Estimate F2 = mediank i CM[k,i]2
Each row’s result is i g(i)2x[i]2 + h(i)=h(j) 2 g(i) g(j) x[i] x[j]
But g(i)2 = -12 = +12 = 1, and i x[i]2 = F2
g(i)g(j) has 1/2 chance of +1 or –1 : expectation is 0 …
+c*g1(j)
+c*g2(j)
+c*g3(j)
+c*g4(j)
h1(j)
hd(j)
j,+c
d=8log 1/
w = 4/
Data Stream Algorithms51
F2 Variance
Expectation of row estimate Rk = i CM[k,i]2 is exactly F2
Variance of row k, Var[Rk], is an expectation:
– Var[Rk] = E[ (buckets b (CM[k,b])2 – F2)2 ]
– Good exercise in algebra: expand this sum and simplify– Many terms are zero in expectation because of terms like
g(a)g(b)g(c)g(d) (degree at most 4)– Requires that hash function g is four-wise independent: it
behaves uniformly over subsets of size four or smaller Such hash functions are easy to construct
Data Stream Algorithms52
F2 Variance
Terms with odd powers of g(a) are zero in expectation– g(a)g(b)g2(c), g(a)g(b)g(c)g(d), g(a)g3(b)
Leaves Var[Rk] i g4(i) x[i]4
+ 2 j i g2(i) g2(j) x[i]2 x[j]2 + 4 h(i)=h(j) g2(i) g2(j) x[i]2 x[j]2 - (x[i]4 + j i 2x[i]2 x[j]2) F2
2/w Row variance can finally be bounded by F2
2/w– Chebyshev for w=4/2 gives probability ¼ of failure– How to amplify this to small probability of failure?
Data Stream Algorithms53
Tail Inequalities for Sums We derive stronger bounds on tail probabilities for the sum
of independent Bernoulli trials via the Chernoff Bound:
– Let X1, ..., Xm be independent Bernoulli trials s.t. Pr[Xi=1] = p
(Pr[Xi=0] = 1-p).
– Let X = i=1m Xi ,and = mp be the expectation of X.
– Then, for any >0,
2
με2
2expμε)|μXPr(|
Data Stream Algorithms54
Applying Chernoff Bound
Each row gives an estimate that is within relative error with probability p > ¾
Take d repetitions and find the median. Why the median?
– Because bad estimates are either too small or too large– Good estimates form a contiguous group “in the middle”– At least d/2 estimates must be bad for median to be bad
Apply Chernoff bound to d independent estimates, p=3/4– Pr[ More than d/2 bad estimates ] < 2exp(d/8)– So we set d = (ln ) to give probability of failure
Same outline used many times in data streams
Data Stream Algorithms55
Aside on Independence Full independence is expensive in a streaming setting
– If hash functions are fully independent over n items, then we need (n) space to store their description
– Pairwise and four-wise independent hash functions can be described in a constant number of words
The F2 algorithm uses a careful mix of limited and full independence – Each hash function is four-wise independent over all n items– Each repetition is fully independent of all others – but there
are only O(log 1/) repetitions.
Data Stream Algorithms56
AMS Sketch Exercises
1. Let x and y be binary streams of length n. The Hamming distance H(x,y) = |{i | x[i] y[i]}|Show how to use AMS sketches to approximate H(x,y)
2. Extend for strings drawn from an arbitrary alphabet
3. The inner product of two strings x, y is x y = i=1n
x[i]*y[i]Use AMS sketches to estimate x y– Hint: try computing the inner product of the sketches.
Show the estimator is unbiased (correct in expectation)– What form does the error in the approximation take?– Use Count-Min Sketches for the same problem and
compare the errors.– Is it possible to build a (1) approximation of x y?
Data Stream Algorithms57
Frequency Moments
Introduction to Frequency Moments and Sketches Count-Min sketch for Fand frequent items
AMS Sketch for F2
Estimating F0
Extensions: – Higher frequency moments– Combined frequency moments
Data Stream Algorithms58
F0 Estimation
F0 is the number of distinct items in the stream – a fundamental quantity with many applications
Early algorithms by Flajolet and Martin [1983] gave nice hashing-based solution– analysis assumed fully independent hash functions
Will describe a generalized version of the FM algorithm due to Bar-Yossef et. al with only pairwise indendence
Data Stream Algorithms59
F0 Algorithm Let m be the domain of stream elements
– Each item in stream is from [1…m] Pick a random hash function h: [m] [m3]
– With probability at least 1-1/m, no collisions under h
For each stream item i, compute h(i), and track the t distinct items achieving the smallest values of h(i)– Note: if same i is seen many times, h(i) is same
– Let vt = t’th smallest value of h(i) seen.
If F0 < t, give exact answer, else estimate F’0 = tm3/vt
– vt/m3 fraction of hash domain occupied by t smallest
m30m3 vt
Data Stream Algorithms60
Analysis of F0 algorithm
Suppose F’0 = tm3/vt > (1+) F0 [estimate is too high]
So for stream = set S 2[m], we have – |{ s S | h(s) < tm3/(1+)F0 }| > t
– Because < 1, we have tm3/(1+)F0 (1-/2)tm3/F0
– Pr[ h(s) < (1-/2)tm3/F0] 1/m3 * (1-/2)tm3/F0 = (1-/2)t/F0
– (this analysis outline hides some rounding issues)
m3tm3/(1+)F00m3 vt
Data Stream Algorithms61
Chebyshev Analysis
Let Y be number of items hashing to under tm3/(1+)F0
– E[Y] = F0 * Pr[ h(s) < tm3/(1+)F0] = (1-/2)t
– For each item i, variance of the event = p(1-p) < p
– Var[Y] = s S Var[ h(s) < tm3/(1+)F0] < (1-/2)t We sum variances because of pairwise independence
Now apply Chebyshev: – Pr[ Y > t ] Pr[|Y – E[Y]| > t/2]
4Var[Y]/2t2 4t/(2t2)
– Set t=20/2 to make this Prob 1/5
Data Stream Algorithms62
Completing the analysis
We have shownPr[ F’0 > (1+) F0 ] < 1/5
Can show Pr[ F’0 < (1-) F0 ] < 1/5 similarly– too few items hash below a certain value
So Pr[ (1-) F0 F’0 (1+)F0] > 3/5 [Good estimate]
Amplify this probability: repeat O(log 1/) times in parallel with different choices of hash function h– Take the median of the estimates, analysis as before
Data Stream Algorithms63
F0 Issues
Space cost: – Store t hash values, so O(1/2 log m) bits– Can improve to O(1/2 + log m) with additional tricks
Time cost: – Find if hash value h(i) < vt
– Update vt and list of t smallest if h(i) not already present
– Total time O(log 1/ + log m) worst case
Data Stream Algorithms64
Range Efficiency
Sometimes input is specified as a stream of ranges [a,b]– [a,b] means insert all items (a, a+1, a+2 … b)– Trivial solution: just insert each item in the range
Range efficient F0 [Pavan, Tirthapura 05]– Start with an alg for F0 based on pairwise hash functions
– Key problem: track which items hash into a certain range– Dives into hash fns to divide and conquer for ranges
Range efficient F2 [Calderbank et al. 05, Rusu,Dobra 06]– Start with sketches for F2 which sum hash values
– Design new hash functions so that range sums are fast
Data Stream Algorithms65
F0 Exercises
Suppose the stream consists of a sequence of insertions and deletions. Design an algorithm to approximate F0 of the current set.– What happens when some frequencies are negative?
Give an algorithm to find F0 of the most recent W arrivals
Use F0 algorithms to approximate Max-dominance: given a stream of pairs (i,x(i)), approximate i max(i, x(i)) x(i)
Data Stream Algorithms66
Frequency Moments
Introduction to Frequency Moments and Sketches Count-Min sketch for Fand frequent items
AMS Sketch for F2
Estimating F0
Extensions: – Higher frequency moments– Combined frequency moments
Data Stream Algorithms67
Higher Frequency Moments
Fk for k>2. Use sampling trick as with Entropy [Alon et al 96]:– Uniformly pick an item from the stream length 1…n– Set r = how many times that item appears subsequently
– Set estimate F’k = n(rk – (r-1)k)
E[F’k]=1/n*n*[ f1k - (f1-1)k + (f1-1)k - (f1-2)k + … + 1k-0k]+…
= f1k + f2
k + … = Fk
Var[F’k]1/n*n2*[(f1k-(f1-1)k)2 + …]
– Use various bounds to bound the variance by k m1-1/k Fk2
– Repeat k m1-1/k times in parallel to reduce variance Total space needed is O(k m1-1/k) machine words
Data Stream Algorithms68
Improvements
[Coppersmith and Kumar ‘04]: Generalize the F2 approach
– E.g. For F3, set p=1/m, and hash items onto {1-1/p, -1/p} with probability {1/p, 1-1/p} respectively.
– Compute cube of sum of the hash values of the stream
– Correct in expectation, bound variance O(mF32)
[Indyk, Woodruff ‘05, Bhuvangiri et al. ‘06]: Optimal solutions by extracting different frequencies– Use hashing to sample subsets of items and fi’s
– Combine these to build the correct estimator
– Cost is O(m1-2/k poly-log(m,n,1/)) space
Data Stream Algorithms69
Combined Frequency Moments
Want to focus on number of distinct communication pairs, not size of communication
So want to compute moments of F0 values...
Consider network traffic data: defines a communication graph
eg edge: (source, destination)
or edge: (source:port, dest:port)
Defines a (directed) multigraph
We are interested in the underlying (support) graph on n nodes
Data Stream Algorithms70
Multigraph Problems
Let G[i,j] = 1 if (i,j) appears in stream:edge from i to j. Total of m distinct edges
Let di = j=1n G[i,j] : degree of node i
Find aggregates of di’s:
– Estimate heavy di’s (people who talk to many)
– Estimate frequency moments:number of distinct di values, sum of squares
– Range sums of di’s (subnet traffic)
Data Stream Algorithms71
F (F0) using CM-FM
Find i’s such that di > i di
Finds the people that talk to many others Count-Min sketch only uses additions, so can apply:
Data Stream Algorithms72
Accuracy for F(F0)
Focus on point query accuracy: estimate di. Can prove estimate has only small bias in expectation
– Analysis is similar to original CM sketch analysis, but now have to take account of F0 estimation of counts
Gives an bound of O(1/3 poly-log(n)) space:– The product of the size of the sketches
Remains to fully understand other combinations of frequency moments, eg. F2(F0), F2(F2) etc.
Data Stream Algorithms73
Exercises / Problems
1. (Research problem) What can be computed for other combinations of frequency moments, e.g. F2 of F2 values, etc.?
2. The F2 algorithm uses the fact that +1/-1 values square to preserve F2 but are 0 in expectation. Why won’t it work to estimate F4 with h {-1, +1, -i, +i}?
3. (Research problem) Read, understand and simplify analysis for optimal Fk estimation algorithms
4. Take the sampling Fk algorithm and combine it with F0 estimators to approximate Fk of node degrees
5. Why can’t we use the sketch approach for F2 of node degrees? Show there the analysis breaks down
Data Stream Algorithms74
Frequency Moments
Introduction to Frequency Moments and Sketches Count-Min sketch for Fand frequent items
AMS Sketch for F2
Estimating F0
Extensions: – Higher frequency moments– Combined frequency moments
Data Stream Algorithms Lower Bounds
Graham Cormode [email protected]
Data Stream Algorithms76
Streaming Lower Bounds
Lower bounds for data streams– Communication complexity bounds– Simple reductions– Hardness of Gap-Hamming problem– Reductions to Gap-Hamming
1 0 1 1 1 0 1 0 1 …
Alice
Bob
Data Stream Algorithms77
This Time: Lower Bounds
So far, have seen many examples of things we can do with a streaming algorithm
What about things we can’t do? What’s the best we could achieve for things we can do? Will show some simple lower bounds for data streams
based on communication complexity
Data Stream Algorithms78
Streaming As Communication
Imagine Alice processing a stream Then take the whole working memory, and send to Bob Bob continues processing the remainder of the stream
1 0 1 1 1 0 1 0 1 …
Alice
Bob
Data Stream Algorithms79
Streaming As Communication
Suppose Alice’s part of the stream corresponds to string x, and Bob’s part corresponds to string y...
...and that computing the function on the stream corresponds to computing f(x,y)...
...then if f(x,y) has communication complexity (g(n)), then the streaming computation has a space lower bound of (g(n))
Proof by contradiction: If there was an algorithm with better space usage, we could run it on x, then send the memory contents as a message, and hence solve the communication problem
Data Stream Algorithms80
Deterministic Equality Testing
Alice has string x, Bob has string y, want to test if x=y Consider a deterministic (one-round, one-way) protocol
that sends a message of length m < n There are 2m possible messages, so some strings must
generate the same message: this would cause error So a deterministic message (sketch) must be (n) bits
– In contrast, we saw a randomized sketch of size O(log n)
1 0 1 1 1 0 1 0 1 …
1 0 1 1 0 0 1 0 1 …
Data Stream Algorithms81
Hard Communication Problems
INDEX: x is a binary string of length ny is an index in [n]Goal: output x[y]Result: (one-way) (randomized) communication complexity of INDEX is (n) bits
DISJOINTNESS: x and y are both length n binary strings Goal: Output 1 if i: x[i]=y[i]=1, else 0Result: (multi-round) (randomized) communication complexity of DISJOINTNESS is (n) bits
Data Stream Algorithms82
Simple Reduction to Disjointness
F: output the highest frequency in a stream Input: the two strings x and y from disjointness Stream: if x[i]=1, then put i in stream; then same for y Analysis: if F=2, then intersection; if F1, then disjoint.
Conclusion: Giving exact answer to F requires (N) bits– Even approximating up to 50% error is hard– Even with randomization: DISJ bound allows randomness
x: 1 0 1 1 0 1
y: 0 0 0 1 1 0
1, 3, 4, 6
4, 5
Data Stream Algorithms83
Simple Reduction to Index
F: output the number of items in the stream Input: the strings x and index y from INDEX Stream: if x[i]=1, put i in stream; then put y in stream Analysis: if (1-)F’0(xy)>(1+)F’0(x) then x[y]=1, else it is 0
Conclusion: Approximating F0 for <1/N requires (N) bits– Implies that space to approximate must be (1/)– Bound allows randomization
x: 1 0 1 1 0 1
y: 5
1, 3, 4, 6
5
Data Stream Algorithms84
Hardness Reduction Exercises
Use reductions to DISJ or INDEX to show the hardness of:
1. Frequent items: find all items in the stream whose frequency > N, for some .
2. Sliding window: given a stream of binary (0/1) values, compute the sum of the last N values– Can this be approximated instead?
3. Min-dominance: given a stream of pairs (i,x(i)), approximate i min(i, x(i)) x(i)
4. Rank sum: Given a stream of (x,y) pairs and query (p,q) specified after stream, approximate |{(x,y)| x<p, y<q}|
Data Stream Algorithms85
Streaming Lower Bounds
Lower bounds for data streams– Communication complexity bounds– Simple reductions– Hardness of Gap-Hamming problem– Reductions to Gap-Hamming
1 0 1 1 1 0 1 0 1 …
Alice
Bob
Data Stream Algorithms86
Gap Hamming
GAP-HAMM communication problem: Alice holds x {0,1}N, Bob holds y {0,1}N
Promise: H(x,y) is either N/2 - pN or N/2 + pN Which is the case? Model: one message from Alice to Bob
Requires (N) bits of one-way randomized communication [Indyk, Woodruff’03, Woodruff’04, Jayram, Kumar, Sivakumar ’07]
Data Stream Algorithms87
Hardness of Gap Hamming
Reduction to an instance of INDEX Map string x to u by 1 +1, 0 -1 (i.e. u[i] = 2x[i] -1 ) Assume both Alice and Bob have access to public
random strings rj, where each bit of rj is iid {-1, +1} Assume w.l.o.g. that length of string n is odd (important!) Alice computes aj = sign(rj u)
Bob computes bj = sign(rj[y]) Repeat N times with different random strings, and
consider the Hamming distance of a1... aN with b1 ... bN
Data Stream Algorithms88
Probability of a Hamming Error
Consider the pair aj= sign(rj u), bj = sign(rj[y])
Let w = i y u[i] rj[i]
– w is a sum of (n-1) values distributed iid uniform {-1,+1} Case 1: w 0. So |w| 2, since (n-1) is even
– so sign(aj) = sign(w), independent of x[y]
– Then Pr[aj bj] = Pr[sign(w) sign(rj[y]) = ½
Case 2: w = 0. So aj = sign(rju) = sign(w + u[y]rj[y]) = sign(u[y]rj[y])
– Then Pr[aj bj] = Pr[sign(u[y]rj[y]) = sign(rj[y])]
– This probability is 1 is u[y]=+1, 0 if u[y]=-1– Completely biased by the answer to INDEX
Data Stream Algorithms89
Finishing the Reduction
So what is Pr[w=0]?– w is sum of (n-1) iid uniform {-1,+1} values– Textbook: Pr[w=0] = c/n, for some constant c
Do some probability manipulation:– Pr[aj = bj] = ½ + c/2n if x[y]=1
– Pr[aj = bj] = ½ - c/2n if x[y]=0
Amplify this bias by making strings of length N=4n/c2
– Apply Chernoff bound on N instances – With prob>2/3, either H(a,b)>N/2 + N or H(a,b)<N/2 - N
If we could solve GAP-HAMMING, could solve INDEX– Therefore, need (N) = (n) bits for GAP-HAMMING
Data Stream Algorithms90
Streaming Lower Bounds
Lower bounds for data streams– Communication complexity bounds– Simple reductions– Hardness of Gap-Hamming problem– Reductions to Gap-Hamming
1 0 1 1 1 0 1 0 1 …
Alice
Bob
Data Stream Algorithms91
Lower Bound for Entropy
Alice: x {0,1}N, Bob: y {0,1}N
Entropy estimation algorithm A Alice runs A on enc(x) = (1,x1), (2,x2), …, (N,xN) Alice sends over memory contents to Bob Bob continues A on enc(y) = (1,y1), (2,y2), …, (N,yN)
010011
(6,0)(5,1)(4,0)(3,0)(2,1)(1,1)Bob
(6,1)(5,1)(4,0)(3,0)(2,1)(1,0)
110010Alice
Data Stream Algorithms92
Lower Bound for Entropy
Observe: there are– 2H(x,y) tokens with frequency 1 each– N-H(x,y) tokens with frequency 2 each
So, H(S) = log N + H(x,y)/N Thus size of Alice’s memory contents = (N).
Set = 1/(p(N) log N) to show bound of (/log 1/)-2)
010011
(6,0)(5,1)(4,0)(3,0)(2,1)(1,1)Bob
(6,1)(5,1)(4,0)(3,0)(2,1)(1,0)
110010Alice
Data Stream Algorithms93
Lower Bound for F0
Same encoding works for F0 (Distinct Elements)– 2H(x,y) tokens with frequency 1 each– N-H(x,y) tokens with frequency 2 each
F0(S) = N + H(x,y) Either H(x,y)>N/2 + N or H(x,y)<N/2 - N
– If we could approximate F0 with < 1/N, could separate
– But space bound = (N) = (-2) bits
Dependence on for F0 is tight
Similar arguments show (-2) bounds for Fk, – Proof assumes k (and hence 2k) are constants
Data Stream Algorithms94
Lower Bounds Exercises
1. Formally argue the space lower bound for F2 via Gap-Hamming
2. Argue space lower bounds for Fk via Gap-Hamming
3. (Research problem) Extend lower bounds for the case when the order of the stream is random or near-random
4. (Research problem) Kumar conjectures the multi-round communication complexity of Gap-Hamming is (n) – this would give lower bounds for multi-pass streaming
Data Stream Algorithms95
Streaming Lower Bounds
Lower bounds for data streams– Communication complexity bounds– Simple reductions– Hardness of Gap-Hamming problem– Reductions to Gap-Hamming
1 0 1 1 1 0 1 0 1 …
Alice
Bob
Data Stream Algorithms Extensions and Open Problems
Graham Cormode [email protected]
Data Stream Algorithms97
This Time: Extensions
Have given “the basics” of streaming: streams of items, frequency moments, upper and lower bounds
Many variations with many open problems– Streams representing different combinatorial objects– Streams that are distributed, correlated, uncertain– Systems for processing streams– Different models of streams
See also “Open problems in Data Streams” [McGregor ’07]– Result of a workshop held at IIT Kanpur in Dec 2006
Data Stream Algorithms98
Deterministic Streaming Algorithms
Focus so far has been on randomized algorithms Many important problems can be solved deterministically!
– Finding frequent items/ heavy hitters– Finding quantiles of a distribution
For many problems, lower bounds show randomization is necessary for sublinear space:– Anything involving equality testing as a special case– Frequency moments
When they are possible, deterministic algorithms are often faster and use less space: more practical to implement
Data Stream Algorithms99
Clustering On Data Streams
Goal: output k cluster centers at end - any point can be classified using these centers.
Use divide and conquer approach [Guha et al. ’00]:– Buffer as many points as possible, then cluster them– Cluster the clusters– Cluster the cluster clusters, etc...– Each level of clustering gives up extra factors in quality
Input: Output:
Data Stream Algorithms100
Geometric Streaming
Stream specifies a sequence of d-dimensional points Answer various geometric problems such as:
– Convex hull– Minimum spanning tree weight– Facility location– Minimum enclosing ball
Gridding approach reduces to Fk or related problems [Indyk ’03]
Core-set: keep a carefully chosen small subset of points and evaluate on them [Har-Peled 02, Chan’06]– Simple example: For minimum enclosing ball, keep
extremal points in evenly-space directions
Data Stream Algorithms101
Sliding Window Computations
In a sliding window, we only consider the last W items– W still very large, so want poly-log(W) solutions
Exponential Histograms [Datar et al.02] and Waves [Gibbons Tirthapura’02]– Deterministic structure tracks counts in a window– Based on doubling bucket sizes to give relative error– Same structure + sketches solves for aggregates
Asynchronous streams: items not in timestamp order– Relative error counts possible [Busch, Tirthapura ’07]– Extend concept to other aggregates [C. et al. ’08]
Data Stream Algorithms102
Time Decay
Assign a weight to each item as a function of its age– E.g. Exponential decay or polynomial decay– Implies “weighted” versions of problems
Cohen and Strauss [2003]: – Can reduce sum and counts to multiple
instances of sliding window queries C., Korn and Tirthapura [2008]:
– Same observations applies to othercomputations (quantiles, frequent items)
age
Data Stream Algorithms103
Multi-Pass Algorithms
Some situations allow multiple passes of the stream– E.g. scanning over slow storage (tape):
random access not possible, but can scan multiple times Earliest work in streaming [Munro, Paterson ’78] studied the
pass/space tradeoff for finding medians Lower bounds can follow from multi-round
communication complexity bounds
1 0 1 1 1 0 1 0 1 …
Data Stream Algorithms104
Other Massive Data Models
Massive Unordered Data (MUD) model [Feldman et al. ‘08] – Abstracts computations in MapReduce/Hadoop settings– Can provably simulate deterministic streaming algs– What about randomized computations, multiple passes?
Data Stream Algorithms105
Skewed Streams
In practice, not all frequency distributions are worst case– Few items are frequent, then a long tail of infrequent items– Such skew is prevalent in network data, word frequency,
paper citations, city sizes, etc.– “Zipfian” distribution with skew z > 0 (z = [1..2] typical)
Analyze algorithms under assumption of skewed data– Improved F2 space cost = O(-2/(1+z) log 1/), provided z>1
items sorted by frequency
freq
uen
cy
log rank
log
freq
uenc
y
Data Stream Algorithms106
Graph Streaming
Stream specifies a massive graph edge by edge– Most natural problems have (|V|) space lower bounds– Semi-streaming model: allow (|V|) but o(|E|) space
Therefore also o(|V|2) space also Allow one (or few) passes to approximate:
– Minimum Spanning Tree Weight– Graph Distances (based on spanners)– Maximum weight matching– Counting Triangles
(4,5) (2,3) (1,3) (3,5) (1,2) (2,4) (1,5) (3,4)1
2
3
4
5
Data Stream Algorithms107
Matrix Streaming
Stream specifies a massive n n matrix– Either by giving entries in some
order, or updates to entries In one (or few) passes, find:
– CUR Decomposition– Page Rank Vector– Approximate Matrix product– Singular Vector Decomposition
Current methods take small constant number of passes, sample constant number of rows and columns by weight– Sketching methods don’t seem so useful here
( )( )( )=( )A C U R
O(1) Columns O(1) Rows
Carefully chosen U
Data Stream Algorithms108
Permutation Streaming
Stream presents a permutation of items– Abstraction of several settings, more of theoretic interest
Approximate number of inversions in the stream– Locations where i > j but i appears before j in stream– Can be reduced to a variation of quantiles [Gupta, Zane’03]
Find length of longest increasing subsequence– Reduce (up to factor 2) to simpler function [Ergun, Jowhari ’08]– Approximate this using a different variation of quantiles– Deterministic lower bound (N1/2), randomized bound open
1 3 4 2 3 4 1 2
Data Stream Algorithms109
Random Order Streaming
Lower bounds are sometimes based on carefully creating adversarial orders of streams
Random order streams: order is uniformly permuted– Can sometimes give much better upper bounds– prefix of
stream gives a good sample of dbn. to come– Lower bounds in random order give stronger evidence of
“robust” hardness, e.g. [Chakrabarti et al. ’08]– Hardness via communication complexity of random partitions
GAP-HAMMING still has linear lower bound t2-party DISJOINTNESS has (n/t) lower bound
Data Stream Algorithms110
Probabilistic Streams
Instead of exact values, stream of discrete distributions– Specify exponentially many possible worlds
Adds complexity to previously studied problems– Sum and Count are easy (by linearity of expectation)– Avg=Sum/Count is hard! –because of ratio [McGregor et al. ’07]
Linearity of expectation, summation of variance– Allows estimation of Fk over streams [C, Garofalakis ’07]
Example: S = (x, ½, y, 1/3, y, ¼) Encodes 6 “possible worlds”:
G x y x,y y,y x,y,y
Pr[G] ¼ ¼ 5/24 5/24 1/24 1/24
Data Stream Algorithms111
Distributed Streams
Motivated by Sensor Networks – large wireless nets– Communication drains battery: compute more, send less
Key problem: design stream summary data structures that can be combined to summarize the union of streams– Most sketches (AMS, Count-Min, F0) naturally distribute
– Similar results needed for other problems
base station(root, coordinator…)
http
://ww
w.in
tel.c
om/re
sear
ch/e
xplo
rato
ry/m
otes
.htm
Data Stream Algorithms112
Continuous Distributed Model
Goal:: Continuously track (global) query over streams at the coordinator while bounding the communication– Large-scale network-event monitoring, real-time anomaly/
DDoS attack detection, power grid monitoring, …
Results known for quantiles, Fk, clustering...– Cost not much higher than one time computation [C et al. 08]
Coordinator
m sites
local stream(s) seen at each
site
S1 Sm
Track Q(S1,…,Sm)
Data Stream Algorithms113
Extensions for P2P Networks
Much work focused on specifics of sensor and wired nets P2P and Grid computing present alternate models
– Structure of multi-hop overlay networks– “Controlled failure” model: nodes explicitly leave and join
Allows us to think beyond model of “highly resource constrained” sensors.
Implementations such as OpenDHT over PlanetLab [Rhea et al.’05]
Data Stream Algorithms114
Authenticated Stream Aggregation
Wide-area query processing
– Possible malicious aggregators
– Can suppress or add spurious
information
Authenticate query results at the
querier?
– Perhaps, to within some
approximation error
Initial steps in [Garofalakis et al.’06], Sliding window: [Hadjieleftheriou et al. ’07]
Data Stream Algorithms115
Data Stream Algorithms
Slides are on the web on my website Long list of references also on the web http://dimacs.rutgers.edu/~graham
Top Related