### slides

```Principal Component Analysis for
Distributed Data
David Woodruff
Based on works with Ken Clarkson, Ravi Kannan, and Santosh Vempala
Outline
1.  What is low rank approximation?
2.  How do we solve it offline?
3.  How do we solve it in a distributed setting?
2
Low rank approximation
§  A is an n x d matrix
§  Think of n points in Rd
§  E.g., A is a customer-product matrix
§  Ai,j = how many times customer i purchased item j
§  A is typically well-approximated by low rank matrix
§  E.g., high rank because of noise
§  Goal: find a low rank matrix approximating A
§  Easy to store, data more interpretable
3
What is a good low rank approximation?
Singular Value Decomposition (SVD)
Any matrixAAk == argmin
U ¢ Σ ¢rank
V k matrices B |A-B|F
§  U has orthonormal columns
§  Σ is diagonal with non-increasing
positive
1/2
2
(|C|
) )
F = (Σ
i,jVC
i,j
The
rows
of
are
k
entries down the diagonal
the orthonormal
top k principal
§  V has
rows
components
Computing
Ak exactly is
expensive
§  Rank-k approximation:
Ak = Uk ¢ Σk ¢ Vk
4
Low rank approximation
§  Goal: output a rank k matrix A’, so that
|A-A’|F · (1+ε) |A-Ak|F
§  Can do this in nnz(A) + (n+d)*poly(k/ε) time [S,CW]
§  nnz(A) is number of non-zero entries of A
5
Solution to low-rank approximation [S]
§  Given n x d input matrix A
§  Compute S*A using a sketching matrix S with k/ε << n
rows. S*A takes random linear combinations of rows of A
A
SA
§  Project rows of A onto SA, then find best rank-k
approximation to points inside of SA.
6
What is the matrix S?
§  S can be a k/ε x n matrix of i.i.d. normal random
variables
§  [S] S can be a k/ε x n Fast Johnson Lindenstrauss
Matrix
§  Uses Fast Fourier Transform
§  [CW] S can be a poly(k/ε) x n CountSketch matrix
[
[
0010 01 00
1000 00 00
0 0 0 -1 1 0 -1 0
0-1 0 0 0 0 0 1
S ¢ A can be
computed in
nnz(A) time!
7
Caveat: projecting the points onto SA is slow
§  Current algorithm:
1. Compute S*A
2. Project each of the rows onto S*A
3. Find best rank-k approximation of projected points
inside of rowspace of S*A
§  Bottleneck is step 2
§  [CW] Approximate the projection
§  Fast algorithm for approximate regression
minrank-k X |X(SA)-A|F2
§  nnz(A) + (n+d)*poly(k/ε) time
8
Distributed low rank approximation
§  We have fast algorithms, but can they be made to work
in a distributed setting?
§  Matrix A distributed among s servers
§  For t = 1, …, s, we get a customer-product matrix from
the t-th shop stored in server t. Server t’s matrix = At
§  Customer-product matrix A = A1 + A2 + … + As
§  More general than row-partition model in which each
customer shops in only one shop
9
Communication cost of low rank approximation
§  Input: n x d matrix A stored on s servers
§  Server t has n x d matrix At
§  A = A1 + A2 + … + As
§  Output: Server t has n x d matrix Ct satisfying
§  C = C1 + C2 + … + Cs has rank at most k
§  |A-C|F · (1+ε)|A-Ak|F
§  Application: distributed clustering
§  Resources: Each server is polynomial time, linear
space, communication is O(1) rounds. Bound the total
number of words communicated
§  [KVW]: O(skd/ε) communication, independent of n
10
Protocol
§  Designate one machine the Central Processor (CP)
§  Let SProblems:
be one of the poly(k/ε) x n random matrices above
§  S can be generated pseudorandomly from small seed
tUUTfor
§  CP
chooses
smallAseed
S and
sends
to all servers
§  Can’t
output
since
rank
too itlarge
§  Could
communicate
AtU to
CP,
§  Server
t computes
SAt and sends
it to
CPthen CP
computes SVD of Σt AtU UT = AUUT
§  CP computes Σi=1s SAt = SA
§  But communicating AtU depends on n
§  CP sends orthonormal basis UT for row space of SA to each
server
§  Server t computes
AtU
11
Approximate SVD lemma
§  Problem reduces to
Communication
independent of n!
§  Server t has n x r matrix Bt = AtU, where r = poly(k/ε)
§  B = Σt Bt
§  CP outputs top k principal components of B
§  Approximate SVD
§  If WT 2 Rk x r is the matrix of top k principal components of PB,
where P is a random r/ε2 x n matrix,
|B-BW WT|F · (1+ε) |B-Bk|F
§  CP sends P to every server
§  Server t sends PBt to CP who computes PB = Σt PBt
§  CP computes W, sends everyone W
12
The protocol
§  Phase 1:
§  Learn an orthonormal basis U for row space of SA
optimal space in U
U
cost · (1+ε)|A-Ak|F
13
The protocol
§  Phase 2:
§  Find an approximately optimal space W inside of U
optimal space in U
approximate
space W in U
U
cost · (1+ε)2|A-Ak|F
14
Conclusion
§  O(sdk/ε) communication protocol for low rank approximation
§  A bit sloppy with words vs. bits but can be dealt with
§  Almost matching Ω(sdk) bit lower bound
§  Can be strengthened to Ω(sdk/ε) in one-way model
§  Can we remove the one-way restriction?
§  Communication cost of other optimization problems?
§  Linear programming
§  Frequency moments
§  Matching
§  etc.
15
```