CS 591, Lecture 9 Data Analytics: Theory and Applications Boston University

Transcription

1 CS 591, Lecture 9 Data Analytics: Theory and Applications Boston University Babis Tsourakakis February 22nd, 2017

2 Announcement We will cover the Monday s 2/20 lecture (President s day) by appending half an hour to next Monday s and Wednesday s lectures. Next week s office hours for next week: pm Babis Tsourakakis CS 591 Data Analytics, Lecture 9 2 / 25

3 Reminder: Reservoir Sampling Input: Stream of n items that does not fit in memory x 1,..., x n Goal: Keep a uniform random sample Simple: Draw a random integer {1,..., n}, and select that item. Issue: We don t know the length n of the stream a priori Problem: Find a uniform sample s from a stream of unknown length Babis Tsourakakis CS 591 Data Analytics, Lecture 9 3 / 25

4 Algorithm: Reservoir Sampling 1 Initially we set s x 1 2 When i-th element arrives, we set s x i with probability 1 i Intuition: Correctness: What is the probability that s = x i at some time t i? Babis Tsourakakis CS 591 Data Analytics, Lecture 9 4 / 25

5 Algorithm: Reservoir Sampling Correctness: What is the probability that s = x i at some time t i? Pr [s = x i ] = 1 i ( 1 1 ) ( ) = 1 i + 1 t t. Question: What about the space complexity? Babis Tsourakakis CS 591 Data Analytics, Lecture 9 5 / 25

6 Extension: Weighted Reservoir Sampling Input: Stream of w 1, w 2,... Goal: Select index i with probability proportional to w i Again, we don t know the length n of the stream a priori Define for each index i W i = w w i. Any ideas? Babis Tsourakakis CS 591 Data Analytics, Lecture 9 6 / 25

7 Extension: Weighted Reservoir Sampling Intuition: Babis Tsourakakis CS 591 Data Analytics, Lecture 9 7 / 25

8 Algorithm: Weighted Reservoir Sampling 1 Initially we set s x 1 2 When i-th element arrives, we set s x i with probability w i W i 1 +w i Correctness: On seeing w i+1 : We switch to w i+1 with probability w i+1 W i+1 If we don t switch the winner j i remains with probability Pr [s = w j ] = (1 w i+1 W i+1 ) w j W i = w j W i+1. Question: What about the space complexity? Babis Tsourakakis CS 591 Data Analytics, Lecture 9 8 / 25

9 Reservoir Sampling, k items Input: Stream of items x 1, x 2,..., Parameter k 1 Goal: Keep k uniform random samples Ideas? Babis Tsourakakis CS 591 Data Analytics, Lecture 9 9 / 25

10 Extension: Weighted Reservoir Sampling Intuition: # k-sets that do not contain i = ( i 1 k ) ( i k) = 1 k i. Babis Tsourakakis CS 591 Data Analytics, Lecture 9 10 / 25

11 Python implementation import random def reservoir_sampling ( stream, k): S =[] counter = 1 for x in stream : if(counter <=k): S. append (x) else : c = random. randint (0, counter -1); if( c<k): S[c] = x counter +=1 return S Babis Tsourakakis CS 591 Data Analytics, Lecture 9 11 / 25

12 Probabilistic Counting Researchers at Bell Labs needed to count overlapping substrings of three letters. Why? To develop statistic-based spell-checker. This spellchecker (typo) was included in early Unix distributions Robert Morris came up with an elegant way of counting each item approximately [Morris, 1978] Why approximately? 8-bit counters can hold counts up to 255 (too small range)... Babis Tsourakakis CS 591 Data Analytics, Lecture 9 12 / 25

13 Probabilistic Counting Increments performed in a probabilistic manner Suppose we use constant probabilities, i.e., we increase the counter from i to i + 1 with probability p Why is this problematic? Morris idea: probability Pr [i i + 1] = f (i) depends on the actual counter value Outline: Morris Morris+ Morris We will follow Jelani Nelson s exposition (M,M+,M++). Babis Tsourakakis CS 591 Data Analytics, Lecture 9 13 / 25

14 Morris algorithm 1 Initialize X 0. 2 For each event, increment X with probability 1 2 X. 3 Output ñ = 2 X 1. Notice that our data structure just maintains an integer! We will see that this integer grows as log n, hence we need O(log log n) bits. Babis Tsourakakis CS 591 Data Analytics, Lecture 9 14 / 25

15 Example: Morris algorithm Source: [Flajolet, 1985] Pr [ñ = 1] = 8 38, Pr [ñ = 2] =, Pr [ñ = 3] = 17 1, Pr [ñ = 4] = Babis Tsourakakis CS 591 Data Analytics, Lecture 9 15 / 25

16 Morris algorithm: Analysis X n := Morris counter after n updates Let s compute E [ 2 Xn ]. Idea 1: Let s apply the formula E [ 2 Xn] = n 2 l p n,l. l=0 We have to compute Pr [X n = l] = p n,l Babis Tsourakakis CS 591 Data Analytics, Lecture 9 16 / 25

17 Morris algorithm: Analysis Suppose we reached state l via n 1 transitions from state 1 to state 1, n 2 from state 2 to state 2,..., n l from state l to l (1 2 1 ) n (1 2 2 ) n (1 2 (l 1) ) n l 1 2 (l 1) (1 2 l ) n l. with the condition that n n l + l 1 = n. Tedious approach... Let s try an inductive approach. Babis Tsourakakis CS 591 Data Analytics, Lecture 9 17 / 25

18 Morris algorithm : Analysis 1 Base case. It s obviously true for n = 0. 2 Induction step. E [ 2 ] X n+1 = = = Pr [X n = j] E [ 2 X n+1 X n = j ] j=0 Pr [X n = j] (2 j (1 1 2 ) j+1 ) j 2 j j=0 Pr [X n = j] 2 j + j=0 j = E [ 2 Xn] + 1 = (n + 1) + 1 Pr [X n = j] Babis Tsourakakis CS 591 Data Analytics, Lecture 9 18 / 25

19 Morris algorithm : Analysis In a similar way, we can show that E [ 2 2Xn] = 3 2 n n + 1. Therefore Var [ 2 Xn] < 1 2 n2. Chebyshev s inequality: Pr [ ñ n ɛn] < 1 2ɛ 2. Babis Tsourakakis CS 591 Data Analytics, Lecture 9 19 / 25

20 Details For the second moment, again we use induction: E [ 2 2Xn] = j 0 = j 0 2 2j Pr [X n = j] 2 2j ( Pr [X n 1 = j] ( 1 2 j) + Pr [X n 1 = j 1]2 (j 1) ) =... = E [ 2 2X n 1 ] + 3E [ 2 X n 1 ] = 3 2 (n 1)2 + 3 (n 1) n 2 = 3 2 n n + 1 Babis Tsourakakis CS 591 Data Analytics, Lecture 9 20 / 25

21 Morris algorithm Source: [Flajolet, 1985] Exact probability distribution for Morris counter for n = 1024 Variance is not bad, but we want to do better! Morris+ Babis Tsourakakis CS 591 Data Analytics, Lecture 9 21 / 25

22 Morris+ algorithm We instantiate s independent copies of Morris and average their outputs. s = 1 2ɛ 2 δ Our estimate becomes ñ = 1 s s ñ i. i=1 Now Chebyshev s inequality gives Next step is Morris++! Pr [ ñ n ɛn] < 1 2sɛ 2 < δ. Babis Tsourakakis CS 591 Data Analytics, Lecture 9 22 / 25

23 Morris++ algorithm Run t = 36 lg 1 δ copies of Morris+ with failure probability δ = 1 3 Output the median estimate from the t copies of Morris+ In other words, we have t coin tosses, each results in 1 with probability 1 3, and in 0 with probability 2 3. Failure if median is 0, otherwise success Chernoff! Babis Tsourakakis CS 591 Data Analytics, Lecture 9 23 / 25

24 Morris++ algorithm analysis Y i = By Chernoff bound, { 1, if i-th Morris+ fails. 0, otherwise. [ ] [ [ ] ] Pr Y i > t Pr Y i E Y i > t < δ. 2 6 i i i Morris++ estimate ñ (1 ± ɛ)-approximates n with probability at least (1 δ). Question: What is the space complexity that we achieved? Babis Tsourakakis CS 591 Data Analytics, Lecture 9 24 / 25

25 references I Flajolet, P. (1985). Approximate counting: a detailed analysis. BIT Numerical Mathematics, 25(1): Morris, R. (1978). Counting large numbers of events in small registers. Communications of the ACM, 21(10): Babis Tsourakakis CS 591 Data Analytics, Lecture 9 25 / 25