![Page 1: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/1.jpg)
Ranking from Crowdsourced Pairwise
Comparisons
via Matrix Manifold Optimization
Jialin Dong
ShanghaiTech University
1
![Page 2: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/2.jpg)
Outline
Introduction
FourVignettes:
➢ System Model and Problem Formulation
➢ Problem Analysis
❖ Convex Methods
Disadvantage: Why Not Convex Optimization?
❖ Scalable Nonconvex Optimization
Motivation: Why Nonconvex Optimization?
➢ Matrix Manifold Optimization
❖ Regularized Smoothed Maximum-likelihood Estimation
➢ Simulation Results
Summary2
![Page 3: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/3.jpg)
Part I : Introduction
3
![Page 4: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/4.jpg)
Ranking
Given n items, infer an ordering on the items based on partial sampled
data
Applications
4⋮ ⋮ ⋮ ⋮ ⋮
⋯
⋯⋯⋯
⋱⋮ ⋮⋯
?
?
? ? ?
? ? ?
? ?
? ? ?
?
?
![Page 5: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/5.jpg)
Crowdsourcing
Crowdsourcing: Harness the power of human computation to solve
tasks
Applications: Machine learning models, clustering data, scencing
recognition
5
![Page 6: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/6.jpg)
Pairwise Measurements
Pairwise relations over a few object pairs: difficult to directly measure
each individual object
Applications localization, Alignment, registration and synchronization,
community detection... [Chen & Goldsmith, 14]
6
![Page 7: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/7.jpg)
Part II : Four Vignettes
7
![Page 8: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/8.jpg)
Vignettes A: System Model and Problem Formulation
8
![Page 9: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/9.jpg)
9
System Model
m crowd users, n items to be ranked
Underlying weight matrix X: low-rank, unknown
Pairwise comparisons
Sampling set
9
⋮ ⋮ ⋮ ⋮ ⋮
⋯
⋯⋯⋯
⋱⋮ ⋮⋯
?
?
? ? ?
? ? ?
? ?
? ? ?
?
?
![Page 10: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/10.jpg)
10
System Model
BTL model [Bradley & Terry, 1952]:
Associated score
Individual ranking list for user i:10
![Page 11: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/11.jpg)
11
Problem Formulation
Maximum-likelihood estimation (MLE): the negative log-likelihood
function is given by
Low-rank matrix optimization problem
11
How to solve this low-rank matrix optimization problem ?
![Page 12: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/12.jpg)
Vignettes B: Problem Analysis
12
![Page 13: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/13.jpg)
Vignettes B.1: Convex Methods
13
![Page 14: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/14.jpg)
Generic low-rank matrix optimization
Rank-constrained matrix optimization problem
➢ is a real-linear map on matrices
➢ is convex and differentiable
Challenge 1: Reliably solve the low-rank matrix problem at scale
Challenge II: Develop optimization algorithms with optimal storage
14
![Page 15: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/15.jpg)
A brief biased history of convex methods
1990s: Interior-point methods (computationally expensive)
➢ Storage cost for Hessian
2000s: Convex first-order methods
➢ (Accelerated) proximal gradient, spectral bundle methods, and others
➢ Store matrix variable
2008-Present: Storage-efficient convex first-order methods
➢ Conditional gradient method(CGM) and extensions
➢ Store matrix in low-rank form after iterations: no storage guarantees
15Interior-point: Nemirovski & Nesterov 1994; ... First-order: Rockafellar 1976; Helmberg & Rendl
1997; Auslender & Teboulle 2006; ... CGM: Frank & Wolfe 1956; Levitin & Poljak 1967; Jaggi 2013; ...
![Page 16: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/16.jpg)
Convex Relaxation Approach
Nuclear norm relaxation to solve the original problem:
Theoretical foundations: Beautiful, nearly complete theory
Effective algorithms: Spectral projected-gradient (SPG) [Davenport et
al., 14], Newton-ADMM method [Ali et al.,17]…
➢ Use generic methods for not huge problems: high level language support
(CVX/CVXPY/Convex.jl) makes prototyping easy
16
![Page 17: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/17.jpg)
Why not
Convex methods have slow memory hogs, high computational
complexity
➢ Computationally expensive:
➢ Storage issue:
17
![Page 18: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/18.jpg)
Vignettes B.2: Scalable Nonconvex Optimization
18
![Page 19: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/19.jpg)
Recent advances in nonconvex optimization
2009–Present: Nonconvex heuristics
➢ Burer–Monteiro factorization idea + various nonlinear programming methods
➢ Store low-rank matrix factors
Guaranteed solutions: Global optimality with statistical assumptions
➢ Matrix completion/recovery: [Sun-Luo’14], [Chen-Wainwright’15], [Ge-Lee-
Ma’16],…
➢ Phase retrieval: [Candes et al., 15], [Chen-Candes’ 15], [Sun-Qu-Wright’16]
➢ Community detection/phase synchronization [Bandeira-Boumal-
Voroninski’16], [Montanari et al., 17],…
19
![Page 20: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/20.jpg)
Why Nonconvex Optimization
20
Convex methods:
➢ Slow memory hogs:
➢ High computational complexity, e.g., singular value decomposition
Nonconvex methods: fast, lightweight
➢ Under certain statistical models with benign global geometry
➢ Store low-rank matrix factors
![Page 21: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/21.jpg)
Vignettes C: Matrix Manifold Optimization
21
![Page 22: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/22.jpg)
What is matrix manifold optimization?
Matrix manifold (or Riemannian) optimization problem
➢ is a smooth function
➢ is a Riemannian manifold: spheres, orthonormal bases (Stiefel), rotations,
positive definite matrices, fixed-rank matrices, Euclidean distance matrices,
semidefinite fixed-rank matrices, linear subspaces (Grassmann), phases,
essential matrices, fixed-rank tensors, Euclidean spaces...
22
How to reformulate the original problem to Riemannian optimization problem?
![Page 23: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/23.jpg)
Reformulate to the Riemannian optimization problem
The original problem
The reformulated matrix manifold optimization problem
❖ Problem with respect to Fixed-rank matrices manifold
➢ F removing the origin
❖ Challenge: nonsmooth elementwise infinity norm constraint23
![Page 24: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/24.jpg)
Smoothing
24
Motivation: derive smooth objective to implement Riemannian optimization
![Page 25: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/25.jpg)
Regularization
25
Motivation: Address the constraint of smoothed surrogate of the
element-wise infinity norm
Convex regularized function:
How to develop the algorithm on the manifold?
![Page 26: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/26.jpg)
Manifold optimization paradigms
Generalize Euclidean gradient (Hessian) to Riemannian gradient (Hessian)
We need Riemannian geometry: 1) linearize search space into atangent space ; 2) pick a metric, i.e., Riemannian metric, on togive intrinsic notions of gradient and Hessian
26
Riemannian Gradient Euclidean Gradient
Retraction Operator
![Page 27: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/27.jpg)
Optimization on the manifold: main idea
27
![Page 28: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/28.jpg)
Optimization on the manifold: main idea
28
![Page 29: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/29.jpg)
Optimization on the manifold: main idea
29
![Page 30: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/30.jpg)
Optimization on the manifold: main idea
30
![Page 31: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/31.jpg)
Example: Rayleigh quotient
Optimization over (sphere) manifold
➢ The cost function is smooth on , symmetric matrix
Step 1: Compute the Euclidean gradient in
Step 2: Compute the Riemannian gradient on via projecting to
the tangent space using the orthogonal projector
31
![Page 32: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/32.jpg)
Ranking problem in this paper
Low-rank optimization for ranking via Riemannian optimization
32
How to efficiently compute the descent direction on the manifold ?
![Page 33: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/33.jpg)
Riemannian trust-region method
Sub-optimal problem
Update the iterate
33
![Page 34: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/34.jpg)
Vignettes D: Simulation Result
34
![Page 35: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/35.jpg)
Algorithms & Metric
Algorithms
➢ Proposed Riemannian trust-region algorithm solving log-sum-exp regularized
problem ( PRTRS)
➢ Bi-factor gradient descent solving log-barrier regularized problem (BFGDB)
[Park et al., 16]
➢ Spectral projected-gradient (SPG) [Davenport et al., 14]
Metric
➢ Sampling size: , rescaled sample size : d
➢ Relative mean square error (RMSE):
35
![Page 36: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/36.jpg)
Convergence rate
36
Objective
➢ PRTRS
➢ BFGDB [Park et al., 16]
➢ SPG [Davenport et al., 14]
PRTRS:
➢ Faster rate of convergence than BFGDB algorithm
➢ Comparable with SPG algorithm
![Page 37: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/37.jpg)
Performance
37
Conclusion
➢ Rescaled sample size increases, relative
MSE reduces
➢ PRTRS: better performance in terms of
MSE than both SPG and BFGDB
![Page 38: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/38.jpg)
GLOBECOM 2017 TUTORIAL
Computational Cost
38
Conclusion
➢ Computational time with different sizes Krespectively
➢ PRTRS: dramatical advantage in computational
time of both SPG and BFGDB
![Page 39: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/39.jpg)
Part III: Summary
39
![Page 40: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/40.jpg)
Concluding remarks
Ranking from pairwise comparisons in the crowdsourcing system
Scalable nonconvex optimization algorithms
➢ Store low-rank matrix factors
➢ Global optimality with statistical assumptions
Matrix manifold optimization
➢ Smoothed regularized MLE
➢ Riemannian trust-region algorithm: Outperform the state the state-of-art algorithms
❖ performance (i.e., relative MSE)
❖ computational cost
❖ convergence rate
40
![Page 41: Ranking from Crowdsourced Pairwise Comparisons via Matrix ...dml.cs.byu.edu/icdm17ws/Jialin.pdf · m crowd users,n items to be ranked Underlying weight matrix X:low-rank,unknown Pairwise](https://reader036.vdocuments.us/reader036/viewer/2022090608/605e7b9a5099bd30205f39b1/html5/thumbnails/41.jpg)
Thanks
41