Statistic estimation over data stream Slides modified from Minos Garofalakis ( yahoo! research) and S. Muthukrishnan (Rutgers University)

2 Outline  Introduction  Frequent moment estimation  Element Frequency estimation

3 Data Stream Processing Algorithms  Generally, algorithms compute approximate answers –Provably difficult to compute answers accurately with limited memory  Approximate answers - Deterministic bounds –Algorithms only compute an approximate answer, but bounds on error  Approximate answers - Probabilistic bounds –Algorithms compute an approximate answer with high probability With probability at least, the computed answer is within a factor of the actual answer

4 Sampling: Basics  Idea: A small random sample S of the data often well-represents all the data –For a fast approximate answer, apply “modified” query to S –Example: select agg from R (n=12) –If agg is avg, return average of the elements in S –Number of odd elements ? Data stream: Sample S: answer: 11.5

5 Probabilistic Guarantees  Example: Actual answer is within 11.5 ± 1 with prob  0.9  Randomized algorithms: Answer returned is a specially- built random variable  Use Tail Inequalities to give probabilistic bounds on returned answer –Markov Inequality –Chebyshev’s Inequality –Chernoff/Hoeffding Bound

6 Basic Tools: Tail Inequalities  General bounds on tail probability of a random variable (that is, 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 Probability distribution Tail probability Markov: Chebyshev:

7 Tail Inequalities for Sums  Possible to derive even stronger bounds on tail probabilities for the sum of independent Bernoulli trials  Chernoff Bound: Let X1,..., Xm be independent Bernoulli trials such that Pr[Xi=1] = p (Pr[Xi=0] = 1-p). Let and be the expectation of. Then, for any,  Application to count queries: –m is size of sample S (4 in example) –p is fraction of odd elements in stream (2/3 in example) Do not need to compute Var(X), but need the independent assumption!

8 The Streaming Model  Underlying signal: One-dimensional array A[1…N] with values A[i] all initially zero –Multi-dimensional arrays as well (e.g., row-major)  Signal is implicitly represented via a stream of updates –j-th update is implying A[k] := A[k] + c[j] (c[j] can be >=0, <0)  Goal: Compute functions on A[] subject to –Small space –Fast processing of updates –Fast function computation –…

9 Streaming Model: Special Cases  Time-Series Model –Only j-th update updates A[j] (i.e., A[j] := c[j])  Cash-Register Model – c[j] is always >= 0 (i.e., increment-only) –Typically, c[j]=1, so we see a multi-set of items in one pass  Turnstile Model –Most general streaming model – c[j] can be >=0 or <0 (i.e., increment or decrement)  Problem difficulty varies depending on the model –E.g., MIN/MAX in Time-Series vs. Turnstile!

10 Frequent moment computation Problem Data arrives online ( a 1,a 2,a 3 …..a m ) Let f(i)=|{ j | a j = i }| ( represented by ||A[i]|| ) Example Data stream: 3, 1, 2, 4, 2, 3, 5,... f(1) f(2) f(3) f(4) f(5) a ∈ {1,2,..., n } i n k F  f k ∑ i i  1 F 0 = 5, F 1 = 7, F 2 = 11 ( 1*1+2*2+2*2+1*1+1*1) ( surprise index) What is F∞?

11 Frequent moment computation  Easy for F 1  How about others ? - focus on the F 2 and F 0 - Estimation of F k

12 Linear-Projection (AMS) Sketch Synopses  Goal:  Goal: Build small-space summary for distribution vector f(i) (i=1,..., N) seen as a stream of i-values  Basic Construct:  Basic Construct: Randomized Linear Projection of f() = project onto inner/dot product of f-vector –Simple to compute over the stream: Add whenever the i-th value is seen –Generate ‘s in small O(logN) space using pseudo-random generators –Tunable probabilistic guarantees on approximation error –Delete-Proof: Just subtract to delete an i-th value occurrence Data stream: 3, 1, 2, 4, 2, 3, 5,... f(1) f(2) f(3) f(4) f(5) where = vector of random values from an appropriate distribution

13 AMS ( sketch ) cont.  Key Intuition: Use randomized linear projections of f() to define random variable X such that –X is easily computed over the stream (in small space) –E[X] = F 2 –Var[X] is small  Basic Idea: –Define a family of 4-wise independent {-1, +1} random variables –Pr[ = +1] = Pr[ = -1] = 1/2 Expected value of each, E[ ] = ? E[ ] = ? –Variables are 4-wise independent Expected value of product of 4 distinct = 0 E( ) = 0 –Variables can be generated using pseudo-random generator using only O(log N) space (for seeding)! Probabilistic error guarantees (e.g., actual answer is 10 ± 1 with probability 0.9)

14 Suppose { } : 1)1,2  {1}, 3,4  {-1} then Z = ? 2)4  {1}, 1,3,4  {-1} then Z = ? AMS ( sketch ) cont. Example 2 Data stream R :

15 AMS ( sketch ) cont.  Expected value of X = F 2  Using 4-wise independence, possible to show that 0 1

16 Boosting Accuracy  Chebyshev’s Inequality:  Boost accuracy to by averaging over several independent copies of X (reduces variance)  By Chebyshev: x x x Average y

17 Boosting Confidence  Boost confidence to by taking median of 2log(1/ ) independent copies of Y  Each Y = Bernoulli Trial Pr[|median(Y)-F 2 | F 2 ] (By Chernoff Bound) = Pr[ # failures in 2log(1/ ) trials >= log(1/ ) ] y y y copiesmedian 2log(1/ )“FAILURE”:

18  Step 1: Compute random variables:  Step 2: Define X= Z 2  Steps 3 & 4: Average independent copies of X; Return median of averages  Main Theorem : Sketching approximates F 2 to within a relative error of with probability using space –Remember: O(log N) space for “seeding” the construction of each X Summary of AMS Sketching for F 2 x x x Average y x x x y x x x y copies median 2log(1/ )

19 Binary-Join COUNT Query  Problem: Compute answer for the query COUNT(R A S)  Example:  Exact solution: too expensive, requires O(N) space! –N = sizeof(domain(A)) Data stream R.A: Data stream S.A: = 10 ( )

20 Basic AMS Sketching Technique [AMS96]  Key Intuition: Use randomized linear projections of f() to define random variable X such that –X is easily computed over the stream (in small space) –E[X] = COUNT(R A S) –Var[X] is small  Basic Idea: –Define a family of 4-wise independent {-1, +1} random variables

21 AMS Sketch Construction  Compute random variables: and –Simply add to X R (X S ) whenever the i-th value is observed in the R.A (S.A) stream  Define X = X R X S to be estimate of COUNT query  Example: Data stream R.A: Data stream S.A:

22 Binary-Join AMS Sketching Analysis  Expected value of X = COUNT(R A S)  Using 4-wise independence, possible to show that  is self-join size of R (second/L2 moment) 0 1

23 Boosting Accuracy  Chebyshev’s Inequality:  Boost accuracy to by averaging over several independent copies of X (reduces variance)  By Chebyshev: x x x Average y

24 Boosting Confidence  Boost confidence to by taking median of 2log(1/ ) independent copies of Y  Each Y = Bernoulli Trial Pr[|median(Y)-COUNT| COUNT] (By Chernoff Bound) = Pr[ # failures in 2log(1/ ) trials >= log(1/ ) ] y y y copiesmedian 2log(1/ )“FAILURE”:

25  Step 1: Compute random variables: and  Step 2: Define X= X R X S  Steps 3 & 4: Average independent copies of X; Return median of averages  Main Theorem (AGMS99): Sketching approximates COUNT to within a relative error of with probability using space –Remember: O(log N) space for “seeding” the construction of each X Summary of Binary-Join AMS Sketching x x x Average y x x x y x x x y copies median 2log(1/ )

26 Distinct Value Estimation ( F 0 )  Problem: Find the number of distinct values in a stream of values with domain [0,...,N-1] –Zeroth frequency moment –Statistics: number of species or classes in a population –Important for query optimizers –Network monitoring: distinct destination IP addresses, source/destination pairs, requested URLs, etc.  Example (N=64)  Hard problem for random sampling! –Must sample almost the entire table to guarantee the estimate is within a factor of 10 with probability > 1/2, regardless of the estimator used! Data stream: Number of distinct values: 5

27  Assume a hash function h(x) that maps incoming values x in [0,…, N-1] uniformly across [0,…, 2^L-1], where L = O(logN)  Let lsb(y) denote the position of the least-significant 1 bit in the binary representation of y –A value x is mapped to lsb(h(x))  Maintain Hash Sketch = BITMAP array of L bits, initialized to 0 –For each incoming value x, set BITMAP[ lsb(h(x)) ] = 1  Prob[ lsb(h(x) = i ] = ? Hash (aka FM) Sketches for Distinct Value Estimation [FM85] x = 5 h(x) = lsb(h(x)) = BITMAP

28 Hash (FM) Sketches for Distinct Value Estimation [FM85]  By uniformity through h(x): Prob[ BITMAP[k]=1 ] = –Assuming d distinct values: expect d/2 to map to BITMAP[0], d/4 to map to BITMAP[1],...  Let R = position of rightmost zero in BITMAP –Use as indicator of log(d)  [FM85] prove that E[R] =, where –Estimate d = –Average several iid instances (different hash functions) to reduce estimator variance fringe of 0/1s around log(d) BITMAP position << log(d) position >> log(d) 0L-1

29 Accuracy of FM BITMAP BITMAP m 0 Approximation with probability at least 1-

30  [FM85] assume “ideal” hash functions h(x) (N-wise independence) –In practice h(x) =, where a, b are random binary vectors in [0,…,2^L-1]  Composable: Component-wise OR/add distributed sketches together –Estimate |S1 S2 … Sk| = set-union cardinality Hash (FM) Sketches for Distinct Value Estimation

31 Cash Register Sketch (AMS) Choose random p from 1..n and let Stream sampling | r  |{ q : q  p, a  a } qp Estimator kk X  m ( r− ( r− 1)) Using F 2 ( k=2 ) as example Data stream: 3, 1, 2, 4, 2, 3, 5,... If we choose the first element a 1 r = 2 and X = 7*(2*2-1*1) = 21 And for a 2 r = ?, X= ? a 5 r = ?, X= ? A more general algorithm for F k

32 Cash Register Sketch (AMS)  Y=Average A copies of X and Z = median of B copies Of Y’s x x x Average y x x x y x x x y B copies median Claim: This is a 1 +ε approx to F 2, and space used is O(AB) = words of size O(logn+logm) with probability at least 1-δ. A copies

33 Analysis: Cash Register Sketch E(X) = F 2 V(X) = E(X) 2 - (E(X)) 2. Using (a 2 - b 2 ) ≤ 2(a-b)a, we have V(X) ≤ 2 F 1 F 3. Also, V(X) ≤ 2 F 1 F 3 ≤. Hence, E ( Y )  E ( X )  F ii 2 V ( Y )  V ( X )/ A  ii

34 Analysis Contd. Applying Chebyshev ’ s inequality Hence, by Chernoff bounds, probability that more than B/2 Y i ’ s deviate by far is at most δ, if we take log (1/δ) of Y i ’ s. Hence, median gives the correct approximation.

35 Computation of F k  E(X) = F k  When A = B = Get approximation with probability at least 1 - r  |{ q : q  p, a  a }| qp kk X  m ( r− ( r− 1))

36 Estimate the element frequency  Ask for f(1) = ? f(4) = ? - AMS based algorithm - Count Min sketch. Data stream: 3, 1, 2, 4, 2, 3, 5,... f(1) f(2) f(3) f(4) f(5)

37 AMS ( sketch ) based algorithm.  Key Intuition: Use randomized linear projections of f() to define random variable Z such that For given element A[i] E( Z ) = ||A[i]|| = f i Similar, we have E( Z ) = f j Basic Idea: –Define a family of 4-wise independent {-1, +1} random variables ( same as before ) Pr[ = +1] = Pr[ = -1] = ½ Let Z = So E( Z ) 1 0

38 AMS cont.  Keep an array of w ╳ d counters for Z ij  Use d hash functions to map element x to [1..w] W d a h 1 (a) h d (a) Z[i, h i (a)] += Est(f a ) = median i (Z[i,h i (a)] )

39 The Count Min (CM) Sketch  Simple sketch idea, can be used for point queries ( f i ), range queries, quantiles, join size estimation  Creates a small summary as an array of w ╳ d counters C  Use d hash functions to map element to [1..w] W = d =

40 CM Sketch Structure  Each element x i is mapped to one counter per row  C[ k,h k (x i )] = C[k, h k (x i )]+1 ( -1 if deletion ) or +c[j] if income is  Estimate A[j] by taking min k C[k,h k (j)] +1 h 1 (x i ) h d (x i ) xixi d w

41 CM Sketch Summary  CM sketch guarantees approximation error on point queries less than  in size O(1/  log 1/  ) –Probability of more error is less than 1-   Hints –Counts are biased! Can you limit the expected amount of extra “mass” at each bucket? (Use Markov)