I/O-EFFICIENT SIMILARITY JOIN R. Pagh, N. Pham, F. Silvestri, M. - - PowerPoint PPT Presentation

i o efficient similarity join
SMART_READER_LITE
LIVE PREVIEW

I/O-EFFICIENT SIMILARITY JOIN R. Pagh, N. Pham, F. Silvestri, M. - - PowerPoint PPT Presentation

I/O-EFFICIENT SIMILARITY JOIN R. Pagh, N. Pham, F. Silvestri, M. Stckel Similarity Join R = Q . / r S = { ( q, x ) Q S | d ( q, x ) r } Similarity join example 1 (Record linkage) Country Name Token ID USA IBM Mircosoft 1 USA


slide-1
SLIDE 1

I/O-EFFICIENT SIMILARITY JOIN

  • R. Pagh, N. Pham, F. Silvestri, M. Stöckel
slide-2
SLIDE 2

Similarity Join

R = Q . /r S = {(q, x) ∈ Q × S|d(q, x) ≤ r}

slide-3
SLIDE 3

Similarity join example 1 (Record linkage)

Country Name USA IBM USA Microsoft Germany SAP China Baidu Token ID Mircosoft 1 SAP SE 2 I.B.M. 3 baidu.com 4

slide-4
SLIDE 4

Similarity join example 2 (Classification)

Class Name Features Mammalia Cat 1100101 Mammalia Dog 1101101 Reptilia Snake 1011000 Aves Parrot 1010111 Features Image ID 0101101 1 1001000 2 1101101 3 1101110 4

slide-5
SLIDE 5

Memory hierarchy

  • Input sets do not fit in memory
  • Use external memory model [Vitter 2008]
  • I/O complexity of an algorithm: number of I/Os
  • Cache-aware algorithm: algorithm that knows M, B
  • Cache-oblivious algorithm: algorithm that doesn’t

know M, B

CPU Memory M Block transfer size B

slide-6
SLIDE 6

The N2 barrier

  • Let
  • We cannot avoid N2/(BM) I/Os in the worst case
  • Consider N points into a ball of radius r/2: N2 near pairs
  • No more than M2 pairs can be created with M/B I/Os
  • Naïve approach is worst-case optimal!
  • Split Q, S into chunks of size M
  • Load in memory each combination of chunks from Q and S
  • Check all pair in memory
  • Cache-oblivious version by [He and Luo, 2006]
  • We are interested in output-sensitive algorithms

N = |Q| = |S|

slide-7
SLIDE 7

Locality Sensitive Hashing

sensitive family H of functions (r, cr, p1, p2)− if d(x, y) ≤ r then Prh∈H[h(x) = h(y)] ≥ p1 if d(x, y) > cr then Prh∈H[h(x) = h(y)] ≤ p2

Quality of an LSH is given by (the smaller, the better)

ρ = log p1 log p2

slide-8
SLIDE 8

Similarity join with LSH

False positives pairs Filtered away in time

|R . /

>r,≤cr S|

slide-9
SLIDE 9

Similarity join with LSH + I/O model

  • Simple implementation
  • Construct a data structure for c-approximate r-near neighbor using

set Q

  • Query I/O
  • Space
  • For each point in R search the r-near neighbor in the above data

structure

  • Total I/O (with batched search)

˜ O ✓✓N B ◆ρ◆ O

  • N 1+ρ

˜ O ✓N B ◆1+ρ!

slide-10
SLIDE 10

Our results

  • Cache-aware and cache-oblivious algorithms for similarity join

that find all near pairs w.h.p., use linear space and require

  • Algorithms work for any distance measure with a monotonic

LSH

˜ O @ ✓ N M ◆ρ @N B + |R . /

≤r S|

MB 1 A + |R . /

≤cr S|

MB 1 A I/O Overhead Cost of generating output Cost of reading input Nasty term Probability of collision decreases with distance

slide-11
SLIDE 11

Is so bad?

|R . /

>r,≤cr S|

slide-12
SLIDE 12

Cache-aware LSH

  • Suppose we have this LSH

collisions with near points (≤ r) collisions with far points (> cr)

  • At most M far points collide with a given point (in

expectation)

p1 ∼ ✓M N ◆ρ p2 ∼ M N

slide-13
SLIDE 13

Cache-aware algorithm

Repeat 1/p1 times

1.

Choose h in H uniformly at random

2.

Partition Q and S into buckets with h

3.

For each value v where Qv and Sv are not empty

a.

Split Qv and Sv into chunks of size M/2

b.

Load Qv,i and Rv,i for every i,j and emit all near pairs in Qv,i x Rv,i

slide-14
SLIDE 14

Example

S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ S ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ Q ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡

slide-15
SLIDE 15

Example

S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ S ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ Q ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡

slide-16
SLIDE 16

Example

S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ S ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ S ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ S ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ Q ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ Q ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ S ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ S ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ Q ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 1 ¡ 1 ¡ 0 ¡ Q ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡ 1 ¡ 1 ¡ 0 ¡ 0 ¡

slide-17
SLIDE 17

I/O cost

  • If Qv or Rv fit in memory à #reading complexity
  • If Qv and Rv don’t fit in memory à Cost |Qv||Rv|/(MB)
  • 1/(MB) amortized I/O per collided pair
  • Total cost follows by bounding the number of expected

collisions

slide-18
SLIDE 18
  • Total expected cost:

#rounds * (#far pairs + #near pairs + #c-near pairs)/(BM)

I/O cost (2)

Each near pair collides with probability at most 1 Each c-near pair collides with probability at most p1 At most M far points per point

˜ O @ ✓ N M ◆ρ @N B + |R . /

≤r S|

MB 1 A + |R . /

≤cr S|

MB 1 A I/O

1/p1 rounds

slide-19
SLIDE 19

Getting the obliviousness

  • In previous approach
  • p1 and p2 are aware of M
  • Buckets are divided into chunks of size M/2
  • Suppose the collision probabilities p1 and p2 are given and

we cannot tune them.

  • No guarantee on subproblem size
  • We use recursion!

98% of recursive c.

  • blivious

algorithms are optimal!*

* Some non-constant factors may apply. Read papers before coding.

slide-20
SLIDE 20

Cache-oblivious algorithm

1.

If or depth = compute the join with the cache-oblivious “nested loop” in [He and Luo, 2006] and return.

2.

Find all points in Q that are near to a constant fraction of points in S (use sampling)

3.

Repeat 1/p1 times

1.

Choose h in H uniformly at random

2.

Partition Q and S into buckets with h

3.

For each value v where Qv and Sv are not empty

a. Compute recursively

min{|Q|, |S|} = O(1) log1/p2 N Technicalities required by the proof |Qv . /

≤r Rv|

slide-21
SLIDE 21

Recursion and repetition

  • Recursive calls increase the probability to break far pairs
  • Repetition guarantees that near pairs are found
  • No space blow up if repetitions are performed sequentially

(within the same node)

slide-22
SLIDE 22

Branching processes

  • The I/O follows by computing the total number of

collisions

  • expected cost per pair: 1/(MB)
  • What is the probability that a pair (q,x) in

Q x S survives up to level i?

slide-23
SLIDE 23

I/O cost

  • We model the algorithm with a Galton-Watson process
  • Expected #far pairs at level i:
  • Expected #c-near pairs at level i:
  • Expected #near pairs at level i:

|Q . /

>cr S|(p2/p1)i

|Q . /

>r,≤cr S|

|Q . /

≤r S|(1/p1)i

slide-24
SLIDE 24

I/O complexity, sketch

log1/p2(n/M)

slide-25
SLIDE 25

Too many near pairs!

  • LSH guarantees that near pairs collide with probability at least

p1

  • In the “worst case”, pairs can collide with probability 1 and thus

each near pair can be emitted more than 1/p1 times

  • Naïve solution: store all near pairs, sort them and remove

duplicates.

  • Assume that the collision probability can be

explicitly computed for any pair

  • It applies to some LSH for Hamming, l1, l2, Jaccard, and angular

distances

  • Solution: at level i, emit a pair with probability

✓ p1 Pr[h(q) = h(x)] ◆i Pr[h(q) = h(x)]

slide-26
SLIDE 26

Conclusion and open questions

  • Our upper bound:
  • Is our approach practical?
  • Can we improve the upper bound, even for specific

distance measures?

˜ O @ ✓ N M ◆ρ @N B + |R . /

≤r S|

MB 1 A + |R . /

≤cr S|

MB 1 A I/O

slide-27
SLIDE 27

Thank you!