1. Carnegie Mellon AISTATS 2009 Jonathan Huang Carlos Guestrin Carnegie Mellon University Xiaoye Jiang Leonidas Guibas Stanford University

2. Identity management [Shin et al., ‘03] Identity Mixing @Tracks 1,2 Track 1 Track 2 Where is Donald Duck?

3. Identity management Mixing @Tracks 1,2 Mixing @Tracks 1,3 Mixing @Tracks 1,4 Track 1 Track 2 Track 3 Track 4 Where is ? Where is ?

4. Reasoning with permutations Represent uncertainty in identity management with distributions over permutations Identities Track Permutations Probability of each track permutation Probability of each track permutation A B C D P(σ) 1 2 3 40 2 1 3 40 1 3 2 41/10 3 1 2 40 2 3 1 41/20 3 2 1 41/5 1 2 4 30 2 1 4 30 [1 3 2 4] means: “Alice is at Track 1, and Bob is at Track 3, and Cathy is at Track 2, and David is at Track 4 with probability 1/10” [1 3 2 4] means: “Alice is at Track 1, and Bob is at Track 3, and Cathy is at Track 2, and David is at Track 4 with probability 1/10”

5. Storage complexity There are n! permutations in S n (set of permutations of n objects) Need to compress… x 1,800,000

6. Fourier theoretic approaches Approximate distributions over permutations with low frequency basis functions [Kondor2007, Huang2007] low frequency high frequency Fourier analysis on the real line sinusoidal basis Fourier analysis on S n (Permutations of n objects)

7. 1 st order summary matrix Store marginals of the form P(identity i is at track j) Requires storing only n 2 numbers! 0 1/2 1/5 3/100 1/51/209/20 ABCD 1 2 3 4 Identities Tracks “Bob is at Track 2 with zero probability” “Bob is at Track 2 with zero probability” “Cathy is at Track 3 with probability 1/20” “Cathy is at Track 3 with probability 1/20” 1 st order marginals reconstructible from O(n 2 ) lowest frequency Fourier coefficients!

8. Uncertainty principle on a line Uniform distribution “time domain” hard to represent “Fourier domain” easy to represent Peaked distribution “time domain” easy to represent “Fourier domain” hard to represent Uncertainty Principle: a signal f cannot be sparsely represented in both the time and Fourier domains Power spectrumSignal f

9. Uncertainty principle on permutations 246810121416182022 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 frequencyfraction of energy stored in frequency harder to represent Uniform on 1 group of 8 2 subgroups of 4 4 subgroups of 2 8 subgroups of 1 (peaked dist.) Power spectrums C C B B D D E E F F G G A A H H 3 3 2 2 4 4 5 5 6 6 7 7 1 1 8 8 tracks identities confusion between tracks 1,2 confusion between tracks 3,4 No uncertainty Keep full joint distribution over S 8 Keep 8 independent distributions over S 1 Uncertainty within group, no mixing between groups Decomposed distributions can be captured more accurately, with fewer coefficients

10. Adaptive decompositions Our approach: adaptively factor problem into subgroups allowing for higher order representations for smaller subgroups “This is Bob” (and Bob was originally in the Blue group) “This is Bob” (and Bob was originally in the Blue group) Claim: Adaptive Identity Management can be highly scalable, more accurate for sharp distributions

11. Contributions Characterization of constraints on Fourier coefficients on permutations implied by probabilistic independence Two algorithms: for factoring a distribution (Split) and combining independent factors in the Fourier domain (Join) Adaptive algorithm for scalable identity management (handles up to n~100 tracks)

12. First-order independence condition Mutual Exclusivity “Alice and Bob not both at Track k” Independence = = 0 Alice Bob Track k Alice Bob Track k or Product of first-order marginals

13. First-order independence condition Tracks Identities Tracks Identities Not independentIndependent vs. Can verify condition using first-order marginals

14. First-order independence First-order condition is insufficient: “Alice is in red team” “Bob is in blue team” “Alice guards Bob” First-order independence ignores the fact that Alice and Bob are always next to each other! image from [sullivan06]

15. The problem with first-order First-order marginals look like: P(Alice is at Track 1) = 3/5 P(Bob is at Track 2) = 1/2 Now suppose Alice guards Bob, and… Tracks 1,2 very far apart Can write as second-order marginal: P({Alice,Bob} occupy Tracks  {1,2}) = 0

16. Second-order summaries Store summaries for ordered pairs: 2 nd order summary requires O(n 4 ) storage (A,B)(B,A)(A,C) (1,2) (2,1) (C,A) (1,3) (3,1) 1/61/121/8 1/12 1/81/24 1/8 1/61/12 1/81/12 Identities Tracks “Bob is at Track 1 and Alice is at Track 3 with probability 1/12” “Bob is at Track 1 and Alice is at Track 3 with probability 1/12” store marginal probability that identities (k,l) map to tracks (i,j)

17. Trade-off – capture higher frequencies by storing more numbers Remark: high-order marginals contain low-order information Higher orders and connections to Fourier Sum over entire distribution (always equal to 1) Sum over entire distribution (always equal to 1) Recovers original distribution, requires storing n! numbers Recovers original distribution, requires storing n! numbers

18. Fourier coefficient matrices Fourier coefficients on permutations are a collection of square matrices ordered by “frequency”: Bandlimiting - keep a truncated set of coefficients Fourier domain inference – prediction/conditioning in the Fourier domain [Kondor et al,AISTATS07] [Huang et al,NIPS07] 0 th order1 st order2 nd ordern th order

19. Back to independence Need to consider two operations Groups join when tracks from two groups mix Groups split when an observation allows us to reason over smaller groups independently “This is Bob” (and Bob was originally in the Blue group) “This is Bob” (and Bob was originally in the Blue group)

20. Problems If the joint distribution h factors as a product of distributions f and g: Distribution over tracks {1,…,p} Distribution over tracks {p+1,…,n} (Join problem) Find Fourier coefficients of the joint h given Fourier coefficients of factors f and g? (Split problem) Find Fourier coefficients of factors f and g given Fourier coefficients of the joint h?

21. First-order join Given first-order marginals of f and g, what does the matrix of first-order marginals of h look like? first-order marginals f h g zeroes

22. Higher-order joining Given Fourier coefficients of the factors f and g at each frequency level: Compute Fourier coefficients of the joint distribution h at each frequency level factors joint frequencies

23. Higher-order joining Joining for higher-order coefficients gives similar block- diagonal structure Also get Kronecker product structure for each block Blocks appear multiple times (multiplicities related to Littlewood-Richardson coefficients) same block

24. Problems If the joint distribution h factors as a product of distributions f and g: Distribution over tracks {1,…,p} Distribution over tracks {p+1,…,n} (Join problem) Find Fourier coefficients of the joint h given Fourier coefficients of factors f and g? (Split problem) Find Fourier coefficients of factors f and g given Fourier coefficients of the joint h?

25. Splitting Want to “invert” the Join process: Consider recovering 2 nd Fourier block Need to recover A, B from – only possible to do up to scaling factor Our approach: search for blocks of the form: or Theorem: these blocks always exist! (and are efficient to find)

26. Marginal preservation Problem: In practice, never have entire set of Fourier coefficients! Marginal preservation guarantee: Conversely, get a similar guarantee for splitting (Usually get some higher order information too) Theorem: Given m th -order marginals for independent factors, we exactly recover m th - order marginals for the joint. bandlimited representation

27. Detecting independence To adaptively split large distributions, need to detect independent subsets Recall first-order independence condition: Can use (bi)clustering on matrix of marginals to discover an appropriate ordering! matrix of marginals with appropriate ordering on identities and tracks matrix of marginals with appropriate ordering on identities and tracks In practice, get unordered identities, tracks… In practice, get unordered identities, tracks… Balance constraint: force nonzero blocks to be square Balance constraint: force nonzero blocks to be square p p

28. First-order independence First-order condition is insufficient: Can check for higher order independence after detecting at first-order What if we call Split when only the first-order condition is satisfied? “Alice is in red team” “Bob is in blue team” “Alice guards Bob” Even when higher-order independence does not hold: Theorem: Whenever first-order independence holds, Split returns exact marginals of each subset of tracks.

29. Experiments - accuracy 00.050.10.150.20.250.3 0 0.2 0.4 0.6 0.8 1 ratio of observations label accuracy better nonadaptive adaptive dataset from [Khan et al. 2006] 20 ants

30. Experiments – running time 00.10.20.3 0 500 1000 1500 2000 ratio of observations elapsed time per run (seconds) 20406080100 0 500 1000 1500 Number of ants Running time (seconds) better nonadaptive adaptive nonadaptive

31. Conclusions Completely Fourier-theoretic characterization of probabilistic independence marginalization, conditioning, join, split Two new algorithms Scalable and adaptive identity management algorithm to track up to n=100 objects Thank you