SlideShare a Scribd company logo
1 of 34
Download to read offline
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Low-rank tensor completion:
a Riemannian manifold preconditioning approach
Hiroyuki Kasai † and Bamdev Mishra††
†The University of Electro-Communications, Japan
††Amazon Development Centre India, India
ICML2016, New York, USA, June 20, 2016
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 1
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Tensor completion problem
▶ Goal: estimate entries of unknown or missing values.
Matrix completion problem Tensor completion problem
▶ Applications: collaborative filtering, signal recovery, etc.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 2
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Our contributions (Summary)
▶ Address a Riemannian optimization framework.
▶ Propose a novel Riemannian metric by exploiting the
symmetry of constraints and the structure of the cost.
(Riemannian manifold preconditioning)
▶ Develop concrete matrix expressions and algorithms with this
novel metric.
▶ Show superior performances of our proposed algorithms for
low-sampling and ill-conditioned large-scale data (1012).
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 3
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Mathematical problem definition
▶ 3-order tensors X ∈ Rn1×n2×n3 is addressed hereafter.
Definition (Tensor completion problem with fixed-rank)
min
X∈Rn1×n2×n3
1
|Ω|
∥PΩ(X) − PΩ(X⋆
)∥2
F
subject to rank(X) = r,
(1)
▶ PΩ(·) extracts only elements in the observed set Ω.
▶ rank(X) = r = (r1, r2, r3) is multilinear rank of X.
▶ Higher scalability to huge-data due to NO expensive SVD
than nuclear norm regularization approaches.
▶ Assume a low-rank structure by Tucker decomposition.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 4
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Tucker decomposition and Tucker manifold
▶ Tucker decomposition of rank r (=(r1, r2, r3)) is
[Kolda and Bader, 2009]
X = G×1U1×2U2×3U3 (∈ Rn1×n2×n3 )
▶ Ud ∈ St(rd, nd), i.e., Stiefel manifold of matrices of size
nd × rd with orthogonal columns, and G ∈ Rr1×r2×r3
.
▶ Tucker manifold M is defined as
M := St(r1, n1) × St(r2, n2) × St(r3, n3) × Rr1×r2×r3 .
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 5
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
The quotient structure of Tucker decomposition
▶ Solutions are not isolated, but equivalence classes due to
invariant property for Od ∈ O(rd),
[(U1, U2, U3, G)] :=
{(U1O1, U2O2, U3O3, G×1OT
1 ×2OT
2 ×3OT
3 ) : Od ∈ O(rd)}.
▶ Quotient manifold M/∼
M/∼ := M/(O(r1) × O(r2) × O(r3)),
⇓
▶ Solve (1) as an unconstrained optimization problem on a
Riemannian quotient manifold by endowing M/∼.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 6
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
The least-squares structure of the cost function
▶ Problem (1) is convex and quadratic in X and
(U1, U2, U3, G) individually.
⇓
▶ Address the block diagonal approximation of ∥X − X⋆∥2
F for
a new metric.
((G1GT
1 ) ⊗ In1 , (G2GT
2 ) ⊗ In2 , (G3GT
3 ) ⊗ In3 , Ir1r2r3 ).
▶ As a result, exploit second-order information in first-order
algorithms.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 7
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Propose a novel Riemannian metric
▶ Propose a metric from symmetry and least-squares structure.
Definition (New Riemannian metric)
gx : TxM × TxM → R is defined as
gx(ξx, ηx) = ⟨ξU1 , ηU1 (G1GT
1 )⟩ + ⟨ξU2 , ηU2 (G2GT
2 )⟩
+⟨ξU3 , ηU3 (G3GT
3 )⟩ + ⟨ξG, ηG⟩,
ξx = (ξU1
, ξU2
, ξU3
, ξG) and ηx = (ηU1
, ηU2
, ηU3
, ηG) ∈
TxM are tangent vectors.
▶ gx(ξx, ηx) is invariant to [x] (= [(U1, U2, U3, G)]).
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 8
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Geometry of quotient manifold
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 9
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Geometry of Tucker manifold
▶ Show concrete mathematical expressions.
▶ Propose a preconditioned nonlinear conjugate gradient.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 10
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Numerical comparisons
▶ We compare with a number of state-of-the-art algorithms under
various scenarios.
▶ Tucker decomposition based algorithms;
▶ TOpt [Filipovi´c and Juki´c, 2013], geomCG
[Kressner et al., 2014]
▶ Nuclear norm minimization algorithms;
▶ HaLRTC [Liu et al., 2013], Latent [Tomioka et al., 2011],
Hard [Signoretto et al., 2014]
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 11
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Results on synthetic and real datasets (1)
10 20 30 10 20 30 10 20 30
10
0
10
1
10
2
10
3
10
4
100 × 100 × 100
150 × 150 × 150
200 × 200 × 200
Timeinseconds
OS
Proposed
geomCG
HaLRTC
TOpt
Latent
Hard
(a) Small scale
10 20 30
10
−15
10
−10
10
−5
10
0
10
5
OS
MeansquareerroronΓ
Proposed(100x100x100)
geomCG(100x100x100)
HalLRTC(100x100x100)
TOpt(100x100x100)
Latent(100x100x100)
Hard(100x100x100)
Proposed(150x150x150)
geomCG(150x150x150)
HalLRTC(150x150x150)
TOpt(150x150x150)
Latent(150x150x150)
Hard(150x150x150)
Proposed(200x200x200)
geomCG(200x200x200)
HalLRTC(200x200x200)
TOpt(200x200x200)
Latent(200x200x200)
Hard(200x200x200)
(b) Small scale
3000 5000 10000 3000 5000 10000
10
0
10
1
10
2
10
3
10
4
Tensor Size (per dimension)
Timeinseconds
r=(5,5,5)
r=(10,10,10)Proposed
geomCG
(c) Large scale.
0 50 100 150 200
10
−4
10
−2
10
0
10
2
Time in seconds
MeansquareerroronΓ
Proposed (CN=5)
geomCG (CN=5)
Proposed (CN=50)
geomCG (CN=50)
Proposed (CN=100)
geomCG (CN=100)
(f) Ill-cond. & low sample
(5,5,5) (5,5,5) (5,5,5) (7,6,6) (10,5,5)(15,4,4)
0
100
200
300
400
500
600
700
rank
Timeinseconds
(a)
20000
× 7000
× 7000
30000
× 6000
× 6000
40000
× 5000
× 5000
(b) 10000 × 10000
× 10000
Proposed
geomCG
(h) Rectangular instance
0 100 200 300
10
−4
10
−2
10
0
10
2
10
4
Time in seconds
MeansquareerroronΓ
Proposed
geomCG
HaLRTC
TOpt
Latent
Hard
(i) hyperspectral image.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 12
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Results on synthetic and real datasets (2)
Hyperspectral image
Ribeira OS = 11 OS = 22
Algorithm Time MSE on Γ Time MSE on Γ
Proposed 33 ± 13 8.2095 · 10−4
± 1.7 · 10−5
67 ± 43 6.9516 · 10−4
± 1.1 · 10−5
geomCG 36 ± 14 3.8342 · 10−1
± 4.2 · 10−2
150 ± 48 6.2590 · 10−3
± 4.5 · 10−3
HaLRTC 46 ± 0 2.2671 · 10−3
± 3.6 · 10−5
48 ± 0 1.3880 · 10−3
± 2.7 · 10−5
TOpt 80 ± 32 1.7854 · 10−3
± 3.8 · 10−4
27 ± 21 2.1259 · 10−3
± 3.8 · 10−4
Latent 553 ± 3 2.9296 · 10−3
± 6.4 · 10−5
558 ± 3 1.6339 · 10−3
± 2.3 · 10−5
Hard 400 ± 5 6.5090 · 102
± 6.1 · 101
402 ± 4 6.5989 · 102
± 9.8 · 101
MovieLens-10M
MovieLens Proposed geomCG
r Time MSE on Γ Time MSE on Γ
(4, 4, 4) 1748 ± 441 0.6762 ± 1.5 · 10−3
2981 ± 40 0.6956±?2.8 · 10−3
(6, 6, 6) 6058 ± 47 0.6913 ± 3.3 · 10−3
6554 ± 655 0.7398±?7.1 · 10−3
(8, 8, 8) 11370 ± 103 0.7589 ± 7.1 · 10−3
13853 ± 118 0.8955±?3.3 · 10−2
(10, 10, 10) 32802 ± 52 1.0107 ± 2.7 · 10−2
38145 ± 36 1.6550±?8.7 · 10−2
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 13
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Conclusions
▶ Tackled a low-rank tensor completion problem in a
Riemannian manifold preconditioning approach.
▶ Proposed a novel Riemannian metric that exploits the
fundamental structures of symmetry, due to non-uniqueness of
Tucker decomposition, and least-squares of the cost function.
▶ Use the versatile Riemannian optimization framework.
▶ Proposed a preconditioned nonlinear conjugate gradient
algorithm.
▶ Also proposed a stochastic gradient algorithm for online setup.
▶ Concrete matrix expressions are worked out.
▶ Numerical comparisons suggest that our proposed algorithm
has a superior performance on different benchmarks.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 14
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Thank you for your attention
▶ Paper
▶ ICML paper:
http://jmlr.org/proceedings/papers/v48/kasai16.pdf
▶ ICML supplementary paper:
http://jmlr.org/proceedings/papers/v48/kasai16-supp.pdf
▶ Software (Matlab codes)
▶ Independent project:
http://bamdevmishra.com/codes/tensorcompletion/
▶ Built-in project in Manopt:
http://www.manopt.org/fixedrankfactory_tucker_
preconditioned
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 15
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
References I
▶ Absil, P.-A., Mahony, R., and Sepulchre, R. (2008).
Optimization Algorithms on Matrix Manifolds.
Princeton University Press.
▶ Boumal, N., Mishra, B., Absil, P.-A., and Sepulchre, R. (2014).
Manopt: a Matlab toolbox for optimization on manifolds.
J. Mach. Learn. Res., 15(1):1455–1459.
▶ Filipovi´c, M. and Juki´c, A. (2013).
Tucker factorization with missing data with application to low-n-rank tensor
completion.
Multidim. Syst. Sign. P.
Doi: 10.1007/s11045-013-0269-9.
▶ Foster, D. H., Nascimento, S. M. C., and Amano, K. (2007).
Information limits on neural identification of colored surfaces in natural scenes.
Visual Neurosci., 21(3):331–336.
▶ Kolda, T. G. and Bader, B. W. (2009).
Tensor decompositions and applications.
SIAM Review, 51(3):455–500.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 16
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
References II
▶ Kressner, D., Steinlechner, M., and Vandereycken, B. (2014).
Low-rank tensor completion by Riemannian optimization.
BIT Numer. Math., 54(2):447–468.
▶ Liu, J., Musialski, P., Wonka, P., and Ye, J. (2013).
Tensor completion for estimating missing values in visual data.
IEEE Trans. Pattern Anal. Mach. Intell., 35(1):208–220.
▶ Ring, W. and Wirth, B. (2012).
Optimization methods on Riemannian manifolds and their application to shape
space.
SIAM J. Optim., 22(2):596–627.
▶ Sato, H. and Iwai, T. (2015).
A new, globally convergent Riemannian conjugate gradient method.
Optimization, 64(4):1011–1031.
▶ Signoretto, M., Dinh, Q. T., Lathauwer, L. D., and Suykens, J. A. K. (2014).
Learning with tensors: a framework based on convex optimization and spectral
regularization.
Mach. Learn., 94(3):303–351.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 17
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
References III
▶ Tomioka, R., Hayashi, K., and Kashima, H. (2011).
Estimation of low-rank tensors via convex optimization.
Technical report, arXiv preprint arXiv:1010.0789.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 18
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Supplemental slides
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 19
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Proposal: Invariant property of new Riemannian metric
Proposition (Invariant property of new Riemannian metric)
Let (ξU1 , ξU2 , ξU3 , ξG) and (ηU1 , ηU2 , ηU3 , ηG) be tangent
vectors to the quotient manifold at (U1, U2, U3, G),
and (ξU1O1 , ξU2O2 , ξU3O3 , ξG×1OT
1 ×2OT
2 ×3OT
3 )
and (ηU1O1 , ηU2O2 , ηU3O3 , ηG×1OT
1 ×2OT
2 ×3OT
3
) be
tangent vectors to the quotient manifold at
(U1O1, U2O2, U3O3, G×1OT
1 ×2OT
2 ×3OT
3 ). The new
metric (2) are invariant along the equivalence class (6) as
g(U1,U2,U3,G)((ξU1 , ξU2 , ξU3 , ξG), (ηU1 , ηU2 , ηU3 , ηG))
= g(U1O1,U2O2,U3O3,G×1OT
1 ×2OT
2 ×3OT
3 )
((ξU1O1 , ξU2O2 , ξU3O3 , ξG×1OT
1 ×2OT
2 ×3OT
3
),
(ηU1O1 , ηU2O2 , ηU3O3 , ηG×1OT
1 ×2OT
2 ×3OT
3
)).
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 20
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Quotient manifold and horizontal lift
▶ x and y belong to the same equivalence class and they
represent a single point [x] := {y ∈ M : y ∼ x} on M/∼.
▶ Tangent space is decomposable into TxM = Vx ⊕ Hx, where
▶ the vertical space Vx is the tangent space of the equivalence
class [x], and
▶ the horizontal space Hx is the orthogonal subspace to Vx.
▶ Vx does NOT induce a displacement along the equivalence
class [x].
▶ horizontal lift
▶ a unique element ξx ∈ Hx that ξ[x] ∈ T[x](M/∼) at [x] has.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 21
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Tangent space projection operator (1)
Normal space NxM
▶ Obtain from the extracted component of normal space to
TxM in the ambient space must locate on the tangent space.
▶ The following lemma can be shown for the normal space Nx;
Lemma (Normal space)
The quotient manifold endowed with the Riemannian metric
(2) has the matrix characterization of the normal space NxM
defined as
{(U1SU1 (G1GT
1 )−1
, U2SU2 (G2GT
2 )−1
, U3SU3 (G3GT
3 )−1
, 0)
: SUd
∈ Rrd×rd
, ST
Ud
= SUd
, for d ∈ {1, 2, 3}}.
where SUd
for all d ∈ {1, 2, 3} are symmetric matrices.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 22
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Tangent space projection operator (2)
Tangent space projector Ψx
Proposition (Tangent space projection operator)
The quotient manifold endowed with the Riemannian met-
ric (2) admits a tangent projector Ψx : Rn1×r1 × Rn2×r2 ×
Rn3×r3 × Rr1×r2×r3 → TxM defined as
Ψx(YU1, YU2, YU3, YG) = (YU1−U1SU1 (G1GT
1 )−1,
YU2−U2SU2 (G2GT
2 )−1,
YU3−U3SU3 (G3GT
3 )−1, YG),
where SUd
is the solution to the Lyapunov equation below;
SUd
GdGT
d + GdGT
d SUd
= GdGT
d (YT
Ud
Ud + UT
d YUd
)GdGT
d
▶ They are solved efficiently with the Matlab’s lyap routine.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 23
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Horizontal space projection operator (1)
Vertical space Vx and Horizontal space Hx
▶ This is obtained by
▶ Calculate the condition (Lemma below) of the horizontal space
from the orthogonal relationship with the vertical space.
▶ Remove the component along the vertical space, and the
remaining components must satisfy the condition above.
▶ The vertical space Vx has the matrix characterization
{(U1Ω1, U2Ω2, U3Ω3, −(G×1Ω1 + G×2Ω2 + G×3Ω3))
: Ωd ∈ Rrd×rd
, ΩT
d = −Ωd for d ∈ {1, 2, 3}}.
▶ where Ωd for all d ∈ {1, 2, 3} are skew symmetric matrices.
Lemma (Horizontal space)
Horizontal space ξx = (ξU1 , ξU2 , ξU3 , ξG) ∈ Hx must satisfy
(GdGT
d )ξT
Ud
Ud + ξGd
GT
d is symmetric for d ∈ {1, 2, 3}.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 24
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Horizontal space projection operator (2)
Vertical space Vx and Horizontal space projector Πx
Proposition (Horizontal space projection operator)
The quotient manifold endowed with the Riemannian metric
(2) admits a horizontal projector Πx : TxM :→ Hx : ηx →
Πx(ηx) defined as
Πx(ηx) = (ηU1− U1Ω1, ηU2− U2Ω2, ηU3− U3Ω3,
ηG −(−(G×1Ω1+G×2Ω2+G×3Ω3))),
where ηx = (ηU1 , ηU2 , ηU3 , ηG) ∈ TxM and Ωd is a skew-
symmetric matrix of size rd × rd that is the solution to the
coupled Lyapunov equations below;
(continue to the next page.)
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 25
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Horizontal space projection operator (3)
Horizontal space projector Πx
Proposition (Horizontal space projection operator)
(continued from the previous page.)



G1GT
1 Ω1 + Ω1G1GT
1 − G1(Ir3 ⊗ Ω2)GT
1 − G1(Ω3 ⊗ Ir2 )GT
1
= Skew(UT
1 ηU1 G1GT
1 ) + Skew(G1ηT
G1
),
G2GT
2 Ω2 + Ω2G2GT
2 − G2(Ir3 ⊗ Ω1)GT
2 − G2(Ω3 ⊗ Ir1 )GT
2
= Skew(UT
2 ηU2 G2GT
2 ) + Skew(G2ηT
G2
),
G3GT
3 Ω3 + Ω3G3GT
3 − G3(Ir2 ⊗ Ω1)GT
3 − G3(Ω2 ⊗ Ir1 )GT
3
= Skew(UT
3 ηU3 G3GT
3 ) + Skew(G3ηT
G3
).
▶ Skew(·) extracts the skew-symmetric part of a square matrix, i.e.,
Skew(D) = (D − DT
)/2.
▶ The coupled Lyapunov equations are solved efficiently with the
Matlab’s pcg routine that is combined with a specific preconditioner
resulting from the Gauss-Seidel approximation.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 26
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Retraction Rx(ξx)
▶ Map vectors in the horizontal space to points on the search
space M and satisfies the local rigidity condition
[Absil et al., 2008, Definition 4.1].
▶ Provides a natural way to move on the manifold along a
search direction.
▶ Due to the product nature of M, we can choose a retraction
by combining retractions on the individual manifolds, i.e.,
Rx(ξx) = (uf(U1 + ξU1 ), uf(U2 + ξU2 ), uf(U3 + ξU3 ), G + ξG),
where ξx ∈ Hx and uf(·) extracts the orthogonal factor of a
full column rank matrix, i.e., uf(A) = A(AT
A)−1/2.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 27
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Vector transport Tηx
ξx
▶ Smooth mapping that transports a tangent vector ξx ∈ TxM
at x ∈ M to a vector in the tangent space at Rx(ηx)
[Absil et al., 2008, Section 8.1.4].
▶ Generalize the classical concept of translation of vectors in the
Euclidean space to manifolds.
▶ The horizontal lift of the abstract vector transport Tη[x]
ξ[x] on
M/∼ has the matrix characterization
ΠRx(ηx)(Tηx ξx) = ΠRx(ηx)(ΨRx(ηx)(ξx)),
where ξx and ηx are the horizontal lifts in Hx of ξ[x] and η[x]
that belong to T[x](M/∼).
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 28
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Riemannian gradient computations (1)
Horizontal lift (1)
▶ Finally, from the Riemannian submersion theory
[Absil et al., 2008, Section 3.6.2], the horizontal lift of
grad[x]f is equal to gradxf = Ψ(egradxf).
▶ Subsequently, we obtain below;
the horizontal lift of grad[x]f =
(S1(U3 ⊗ U2)GT
1 (G1GT
1 )−1
− U1BU1 (G1GT
1 )−1
,
S2(U3 ⊗ U1)GT
2 (G2GT
2 )−1
− U2BU2 (G2GT
2 )−1
,
S3(U2 ⊗ U1)GT
3 (G3GT
3 )−1
− U3BU3 (G3GT
3 )−1
,
S ×1 UT
1 ×2 UT
2 ×3 UT
3 ),
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 29
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Riemannian gradient computations (2)
Horizontal lift (2)
▶ where BUd
for d ∈ {1, 2, 3} are the solutions to the Lyapunov
equations



BU1 G1GT
1 + G1GT
1 BU1 = 2Sym(G1GT
1 UT
1 (S1(U3 ⊗ U2)GT
1 ),
BU2 G2GT
2 + G2GT
2 BU2 = 2Sym(G2GT
2 UT
2 (S2(U3 ⊗ U1)GT
2 ),
BU3 G3GT
3 + G3GT
3 BU3 = 2Sym(G3GT
3 UT
3 (S3(U2 ⊗ U1)GT
3 ),
which are solved efficiently with the Matlab’s lyap routine.
Sym(·) extracts the symmetric part of a square matrix, i.e.,
Sym(D) = (D + DT
)/2.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 30
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Preconditioned Conjugate Gradient (CG) Algorithm
▶ Off-the-shelf conjugate gradient implementation of Manopt 1
[Boumal et al., 2014] for any smooth cost function.
▶ The convergence analysis of the Riemannian conjugate
gradient method follows from
[Sato and Iwai, 2015, Ring and Wirth, 2012].
▶ Cost function specific ingredients are needed;
▶ Compute the Riemannian gradient.
▶ Compute an initial guess for the step-size for CG.
▶ The total computational cost per iteration of our algorithm is
O(|Ω|r1r2r3), where |Ω| is the number of known entries.
1
http://www.manopt.org/Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 31
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Experiment conditions
▶ Comparisons with state-of-the-art algorithms that include
▶ Tucker decomposition based algorithms;
▶ TOpt [Filipovi´c and Juki´c, 2013], geomCG
[Kressner et al., 2014]
▶ Nuclear norm minimization algorithms;
▶ HaLRTC [Liu et al., 2013], Latent [Tomioka et al., 2011],
Hard [Signoretto et al., 2014]
▶ Problem instances
▶ Case S2: small-scale instances.
▶ Case S3: large-scale instances.
▶ Case S5: influence of ill-conditioning and low sampling.
▶ Case S7: asymmetric instances.
▶ Case R1: hyperspectral image “Ribeira” [Foster et al., 2007].
▶ Case R2: MovieLens-10M.
▶ Note that, compared with ONLY geomCG for large-scale
instances, i.e., except Case S2.
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 32
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Synthetic dataset cases
▶ Case S2: small-scale instances.
▶ Size 100 × 100 × 100, 150 × 150 × 150, and 200 × 200 × 200
and rank r = (10, 10, 10) are considered. OS is {10, 20, 30}.
▶ Case S3: large-scale instances.
▶ Tensors of size 3000 × 3000 × 3000, 5000 × 5000 × 5000, and
10000 × 10000 × 10000 and ranks r = (5, 5, 5) and
(10, 10, 10). OS is 10.
▶ Case S5: influence of ill-conditioning and low sampling.
▶ Case S4 with OS = 5. We impose a diagonal core G with
exponentially decaying positive values of condition numbers
(CN) 5, 50, and 100.
▶ Case S7: asymmetric instances.
▶ (a): tensors size 20000 × 7000 × 7000, 30000 × 6000 × 6000,
and 40000 × 5000 × 5000 with rank r = (5, 5, 5).
▶ (b): tensor size 10000 × 10000 × 10000 with ranks (7, 6, 6),
(10, 5, 5), and (15, 4, 4).
Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 33
Introduction Metric proposal Tucker manifold geometry Numerical Comparisons
Real-world dataset cases
▶ Case R1: hyperspectral image “Ribeira”.
[Foster et al., 2007]
▶ The tensor size is 203 × 268 × 33.
▶ Compare all the algorithms, and perform five random samplings
of the pixels based on the OS values 11 and 22, corresponding
to the rank r=(15, 15, 6) adopted in [Kressner et al., 2014].
▶ While OS = 22 corresponds to the observation ratio of 10%
studied in [Kressner et al., 2014], OS = 11 considers a
challenging scenario with the observation ratio of 5%.
▶ Case R2: MovieLens-10M2.
▶ This dataset contains 10000054 ratings corresponding to
71567 users and 10681 movies.
▶ Split the time into 7-days wide bins results, and finally, get a
tensor of size 71567 × 10681 × 731.
▶ The fraction of known entries is less than 0.002%.
2
http://grouplens.org/datasets/movielens/.Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 34

More Related Content

What's hot

Reinforcement Learning, Application and Q-Learning
Reinforcement Learning, Application and Q-LearningReinforcement Learning, Application and Q-Learning
Reinforcement Learning, Application and Q-LearningAbdullah al Mamun
 
SLAM-베이즈필터와 칼만필터
SLAM-베이즈필터와 칼만필터SLAM-베이즈필터와 칼만필터
SLAM-베이즈필터와 칼만필터jdo
 
Mappingasphere
MappingasphereMappingasphere
Mappingaspheresstoffre
 
BRDF Model (English version)
BRDF Model (English version)BRDF Model (English version)
BRDF Model (English version)Teppei Kurita
 

What's hot (6)

Reinforcement Learning, Application and Q-Learning
Reinforcement Learning, Application and Q-LearningReinforcement Learning, Application and Q-Learning
Reinforcement Learning, Application and Q-Learning
 
SLAM-베이즈필터와 칼만필터
SLAM-베이즈필터와 칼만필터SLAM-베이즈필터와 칼만필터
SLAM-베이즈필터와 칼만필터
 
Graph Partitioning and Spectral Methods
Graph Partitioning and Spectral MethodsGraph Partitioning and Spectral Methods
Graph Partitioning and Spectral Methods
 
Mappingasphere
MappingasphereMappingasphere
Mappingasphere
 
BRDF Model (English version)
BRDF Model (English version)BRDF Model (English version)
BRDF Model (English version)
 
Lec14 eigenface and fisherface
Lec14 eigenface and fisherfaceLec14 eigenface and fisherface
Lec14 eigenface and fisherface
 

Similar to ICML2016: Low-rank tensor completion: a Riemannian manifold preconditioning approach

New data structures and algorithms for \\post-processing large data sets and ...
New data structures and algorithms for \\post-processing large data sets and ...New data structures and algorithms for \\post-processing large data sets and ...
New data structures and algorithms for \\post-processing large data sets and ...Alexander Litvinenko
 
Free vibration analysis of composite plates with uncertain properties
Free vibration analysis of composite plates  with uncertain propertiesFree vibration analysis of composite plates  with uncertain properties
Free vibration analysis of composite plates with uncertain propertiesUniversity of Glasgow
 
Efficient Computation of Regret-ratio Minimizing Set: A Compact Maxima Repres...
Efficient Computation ofRegret-ratio Minimizing Set:A Compact Maxima Repres...Efficient Computation ofRegret-ratio Minimizing Set:A Compact Maxima Repres...
Efficient Computation of Regret-ratio Minimizing Set: A Compact Maxima Repres...Abolfazl Asudeh
 
Platoon Control of Nonholonomic Robots using Quintic Bezier Splines
Platoon Control of Nonholonomic Robots using Quintic Bezier SplinesPlatoon Control of Nonholonomic Robots using Quintic Bezier Splines
Platoon Control of Nonholonomic Robots using Quintic Bezier SplinesKaustav Mondal
 
Unbiased Markov chain Monte Carlo
Unbiased Markov chain Monte CarloUnbiased Markov chain Monte Carlo
Unbiased Markov chain Monte CarloJeremyHeng10
 
Introduction of Quantum Annealing and D-Wave Machines
Introduction of Quantum Annealing and D-Wave MachinesIntroduction of Quantum Annealing and D-Wave Machines
Introduction of Quantum Annealing and D-Wave MachinesArithmer Inc.
 
AAE550_Final_LuQi
AAE550_Final_LuQiAAE550_Final_LuQi
AAE550_Final_LuQiLU QI
 
Estimating Future Initial Margin with Machine Learning
Estimating Future Initial Margin with Machine LearningEstimating Future Initial Margin with Machine Learning
Estimating Future Initial Margin with Machine LearningAndres Hernandez
 
MVPA with SpaceNet: sparse structured priors
MVPA with SpaceNet: sparse structured priorsMVPA with SpaceNet: sparse structured priors
MVPA with SpaceNet: sparse structured priorsElvis DOHMATOB
 
Julio Bravo's Master Graduation Project
Julio Bravo's Master Graduation ProjectJulio Bravo's Master Graduation Project
Julio Bravo's Master Graduation ProjectJulio Bravo
 
Introduction to hmc
Introduction to hmcIntroduction to hmc
Introduction to hmcKota Takeda
 
MCQMC_talk_Chiheb_Ben_hammouda.pdf
MCQMC_talk_Chiheb_Ben_hammouda.pdfMCQMC_talk_Chiheb_Ben_hammouda.pdf
MCQMC_talk_Chiheb_Ben_hammouda.pdfChiheb Ben Hammouda
 
[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...
[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...
[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...Yuko Kuroki (黒木祐子)
 

Similar to ICML2016: Low-rank tensor completion: a Riemannian manifold preconditioning approach (20)

New data structures and algorithms for \\post-processing large data sets and ...
New data structures and algorithms for \\post-processing large data sets and ...New data structures and algorithms for \\post-processing large data sets and ...
New data structures and algorithms for \\post-processing large data sets and ...
 
Presentation.pdf
Presentation.pdfPresentation.pdf
Presentation.pdf
 
Free vibration analysis of composite plates with uncertain properties
Free vibration analysis of composite plates  with uncertain propertiesFree vibration analysis of composite plates  with uncertain properties
Free vibration analysis of composite plates with uncertain properties
 
Subquad multi ff
Subquad multi ffSubquad multi ff
Subquad multi ff
 
Efficient Computation of Regret-ratio Minimizing Set: A Compact Maxima Repres...
Efficient Computation ofRegret-ratio Minimizing Set:A Compact Maxima Repres...Efficient Computation ofRegret-ratio Minimizing Set:A Compact Maxima Repres...
Efficient Computation of Regret-ratio Minimizing Set: A Compact Maxima Repres...
 
Platoon Control of Nonholonomic Robots using Quintic Bezier Splines
Platoon Control of Nonholonomic Robots using Quintic Bezier SplinesPlatoon Control of Nonholonomic Robots using Quintic Bezier Splines
Platoon Control of Nonholonomic Robots using Quintic Bezier Splines
 
Unbiased Markov chain Monte Carlo
Unbiased Markov chain Monte CarloUnbiased Markov chain Monte Carlo
Unbiased Markov chain Monte Carlo
 
QMC: Undergraduate Workshop, Introduction to Monte Carlo Methods with 'R' Sof...
QMC: Undergraduate Workshop, Introduction to Monte Carlo Methods with 'R' Sof...QMC: Undergraduate Workshop, Introduction to Monte Carlo Methods with 'R' Sof...
QMC: Undergraduate Workshop, Introduction to Monte Carlo Methods with 'R' Sof...
 
presentation
presentationpresentation
presentation
 
Introduction of Quantum Annealing and D-Wave Machines
Introduction of Quantum Annealing and D-Wave MachinesIntroduction of Quantum Annealing and D-Wave Machines
Introduction of Quantum Annealing and D-Wave Machines
 
talk_NASPDE.pdf
talk_NASPDE.pdftalk_NASPDE.pdf
talk_NASPDE.pdf
 
AAE550_Final_LuQi
AAE550_Final_LuQiAAE550_Final_LuQi
AAE550_Final_LuQi
 
Estimating Future Initial Margin with Machine Learning
Estimating Future Initial Margin with Machine LearningEstimating Future Initial Margin with Machine Learning
Estimating Future Initial Margin with Machine Learning
 
Smart Multitask Bregman Clustering
Smart Multitask Bregman ClusteringSmart Multitask Bregman Clustering
Smart Multitask Bregman Clustering
 
MVPA with SpaceNet: sparse structured priors
MVPA with SpaceNet: sparse structured priorsMVPA with SpaceNet: sparse structured priors
MVPA with SpaceNet: sparse structured priors
 
Julio Bravo's Master Graduation Project
Julio Bravo's Master Graduation ProjectJulio Bravo's Master Graduation Project
Julio Bravo's Master Graduation Project
 
Introduction to hmc
Introduction to hmcIntroduction to hmc
Introduction to hmc
 
MCQMC_talk_Chiheb_Ben_hammouda.pdf
MCQMC_talk_Chiheb_Ben_hammouda.pdfMCQMC_talk_Chiheb_Ben_hammouda.pdf
MCQMC_talk_Chiheb_Ben_hammouda.pdf
 
[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...
[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...
[AAAI2021] Combinatorial Pure Exploration with Full-bandit or Partial Linear ...
 
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
 

Recently uploaded

Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...GQ Research
 
Top 5 Best Data Analytics Courses In Queens
Top 5 Best Data Analytics Courses In QueensTop 5 Best Data Analytics Courses In Queens
Top 5 Best Data Analytics Courses In Queensdataanalyticsqueen03
 
Real-Time AI Streaming - AI Max Princeton
Real-Time AI  Streaming - AI Max PrincetonReal-Time AI  Streaming - AI Max Princeton
Real-Time AI Streaming - AI Max PrincetonTimothy Spann
 
LLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGILLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGIThomas Poetter
 
Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Cathrine Wilhelmsen
 
Multiple time frame trading analysis -brianshannon.pdf
Multiple time frame trading analysis -brianshannon.pdfMultiple time frame trading analysis -brianshannon.pdf
Multiple time frame trading analysis -brianshannon.pdfchwongval
 
Heart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis ProjectHeart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis ProjectBoston Institute of Analytics
 
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...Boston Institute of Analytics
 
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档208367051
 
Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...
Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...
Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...Boston Institute of Analytics
 
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝DelhiRS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhijennyeacort
 
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Colleen Farrelly
 
ASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel CanterASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel Cantervoginip
 
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default  Presentation : Data Analysis Project PPTPredictive Analysis for Loan Default  Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPTBoston Institute of Analytics
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一F sss
 
办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degree
办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degree办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degree
办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degreeyuu sss
 
Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...
Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...
Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...Thomas Poetter
 
Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...
Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...
Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...limedy534
 
Defining Constituents, Data Vizzes and Telling a Data Story
Defining Constituents, Data Vizzes and Telling a Data StoryDefining Constituents, Data Vizzes and Telling a Data Story
Defining Constituents, Data Vizzes and Telling a Data StoryJeremy Anderson
 

Recently uploaded (20)

Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
Biometric Authentication: The Evolution, Applications, Benefits and Challenge...
 
Top 5 Best Data Analytics Courses In Queens
Top 5 Best Data Analytics Courses In QueensTop 5 Best Data Analytics Courses In Queens
Top 5 Best Data Analytics Courses In Queens
 
Real-Time AI Streaming - AI Max Princeton
Real-Time AI  Streaming - AI Max PrincetonReal-Time AI  Streaming - AI Max Princeton
Real-Time AI Streaming - AI Max Princeton
 
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
Deep Generative Learning for All - The Gen AI Hype (Spring 2024)
 
LLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGILLMs, LMMs, their Improvement Suggestions and the Path towards AGI
LLMs, LMMs, their Improvement Suggestions and the Path towards AGI
 
Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)Data Factory in Microsoft Fabric (MsBIP #82)
Data Factory in Microsoft Fabric (MsBIP #82)
 
Multiple time frame trading analysis -brianshannon.pdf
Multiple time frame trading analysis -brianshannon.pdfMultiple time frame trading analysis -brianshannon.pdf
Multiple time frame trading analysis -brianshannon.pdf
 
Heart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis ProjectHeart Disease Classification Report: A Data Analysis Project
Heart Disease Classification Report: A Data Analysis Project
 
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
NLP Data Science Project Presentation:Predicting Heart Disease with NLP Data ...
 
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
原版1:1定制南十字星大学毕业证(SCU毕业证)#文凭成绩单#真实留信学历认证永久存档
 
Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...
Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...
Decoding the Heart: Student Presentation on Heart Attack Prediction with Data...
 
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝DelhiRS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
RS 9000 Call In girls Dwarka Mor (DELHI)⇛9711147426🔝Delhi
 
Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024Generative AI for Social Good at Open Data Science East 2024
Generative AI for Social Good at Open Data Science East 2024
 
ASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel CanterASML's Taxonomy Adventure by Daniel Canter
ASML's Taxonomy Adventure by Daniel Canter
 
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default  Presentation : Data Analysis Project PPTPredictive Analysis for Loan Default  Presentation : Data Analysis Project PPT
Predictive Analysis for Loan Default Presentation : Data Analysis Project PPT
 
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
办理学位证中佛罗里达大学毕业证,UCF成绩单原版一比一
 
办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degree
办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degree办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degree
办美国阿肯色大学小石城分校毕业证成绩单pdf电子版制作修改#真实留信入库#永久存档#真实可查#diploma#degree
 
Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...
Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...
Minimizing AI Hallucinations/Confabulations and the Path towards AGI with Exa...
 
Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...
Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...
Effects of Smartphone Addiction on the Academic Performances of Grades 9 to 1...
 
Defining Constituents, Data Vizzes and Telling a Data Story
Defining Constituents, Data Vizzes and Telling a Data StoryDefining Constituents, Data Vizzes and Telling a Data Story
Defining Constituents, Data Vizzes and Telling a Data Story
 

ICML2016: Low-rank tensor completion: a Riemannian manifold preconditioning approach

  • 1. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Low-rank tensor completion: a Riemannian manifold preconditioning approach Hiroyuki Kasai † and Bamdev Mishra†† †The University of Electro-Communications, Japan ††Amazon Development Centre India, India ICML2016, New York, USA, June 20, 2016 Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 1
  • 2. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Tensor completion problem ▶ Goal: estimate entries of unknown or missing values. Matrix completion problem Tensor completion problem ▶ Applications: collaborative filtering, signal recovery, etc. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 2
  • 3. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Our contributions (Summary) ▶ Address a Riemannian optimization framework. ▶ Propose a novel Riemannian metric by exploiting the symmetry of constraints and the structure of the cost. (Riemannian manifold preconditioning) ▶ Develop concrete matrix expressions and algorithms with this novel metric. ▶ Show superior performances of our proposed algorithms for low-sampling and ill-conditioned large-scale data (1012). Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 3
  • 4. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Mathematical problem definition ▶ 3-order tensors X ∈ Rn1×n2×n3 is addressed hereafter. Definition (Tensor completion problem with fixed-rank) min X∈Rn1×n2×n3 1 |Ω| ∥PΩ(X) − PΩ(X⋆ )∥2 F subject to rank(X) = r, (1) ▶ PΩ(·) extracts only elements in the observed set Ω. ▶ rank(X) = r = (r1, r2, r3) is multilinear rank of X. ▶ Higher scalability to huge-data due to NO expensive SVD than nuclear norm regularization approaches. ▶ Assume a low-rank structure by Tucker decomposition. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 4
  • 5. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Tucker decomposition and Tucker manifold ▶ Tucker decomposition of rank r (=(r1, r2, r3)) is [Kolda and Bader, 2009] X = G×1U1×2U2×3U3 (∈ Rn1×n2×n3 ) ▶ Ud ∈ St(rd, nd), i.e., Stiefel manifold of matrices of size nd × rd with orthogonal columns, and G ∈ Rr1×r2×r3 . ▶ Tucker manifold M is defined as M := St(r1, n1) × St(r2, n2) × St(r3, n3) × Rr1×r2×r3 . Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 5
  • 6. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons The quotient structure of Tucker decomposition ▶ Solutions are not isolated, but equivalence classes due to invariant property for Od ∈ O(rd), [(U1, U2, U3, G)] := {(U1O1, U2O2, U3O3, G×1OT 1 ×2OT 2 ×3OT 3 ) : Od ∈ O(rd)}. ▶ Quotient manifold M/∼ M/∼ := M/(O(r1) × O(r2) × O(r3)), ⇓ ▶ Solve (1) as an unconstrained optimization problem on a Riemannian quotient manifold by endowing M/∼. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 6
  • 7. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons The least-squares structure of the cost function ▶ Problem (1) is convex and quadratic in X and (U1, U2, U3, G) individually. ⇓ ▶ Address the block diagonal approximation of ∥X − X⋆∥2 F for a new metric. ((G1GT 1 ) ⊗ In1 , (G2GT 2 ) ⊗ In2 , (G3GT 3 ) ⊗ In3 , Ir1r2r3 ). ▶ As a result, exploit second-order information in first-order algorithms. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 7
  • 8. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Propose a novel Riemannian metric ▶ Propose a metric from symmetry and least-squares structure. Definition (New Riemannian metric) gx : TxM × TxM → R is defined as gx(ξx, ηx) = ⟨ξU1 , ηU1 (G1GT 1 )⟩ + ⟨ξU2 , ηU2 (G2GT 2 )⟩ +⟨ξU3 , ηU3 (G3GT 3 )⟩ + ⟨ξG, ηG⟩, ξx = (ξU1 , ξU2 , ξU3 , ξG) and ηx = (ηU1 , ηU2 , ηU3 , ηG) ∈ TxM are tangent vectors. ▶ gx(ξx, ηx) is invariant to [x] (= [(U1, U2, U3, G)]). Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 8
  • 9. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Geometry of quotient manifold Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 9
  • 10. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Geometry of Tucker manifold ▶ Show concrete mathematical expressions. ▶ Propose a preconditioned nonlinear conjugate gradient. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 10
  • 11. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Numerical comparisons ▶ We compare with a number of state-of-the-art algorithms under various scenarios. ▶ Tucker decomposition based algorithms; ▶ TOpt [Filipovi´c and Juki´c, 2013], geomCG [Kressner et al., 2014] ▶ Nuclear norm minimization algorithms; ▶ HaLRTC [Liu et al., 2013], Latent [Tomioka et al., 2011], Hard [Signoretto et al., 2014] Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 11
  • 12. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Results on synthetic and real datasets (1) 10 20 30 10 20 30 10 20 30 10 0 10 1 10 2 10 3 10 4 100 × 100 × 100 150 × 150 × 150 200 × 200 × 200 Timeinseconds OS Proposed geomCG HaLRTC TOpt Latent Hard (a) Small scale 10 20 30 10 −15 10 −10 10 −5 10 0 10 5 OS MeansquareerroronΓ Proposed(100x100x100) geomCG(100x100x100) HalLRTC(100x100x100) TOpt(100x100x100) Latent(100x100x100) Hard(100x100x100) Proposed(150x150x150) geomCG(150x150x150) HalLRTC(150x150x150) TOpt(150x150x150) Latent(150x150x150) Hard(150x150x150) Proposed(200x200x200) geomCG(200x200x200) HalLRTC(200x200x200) TOpt(200x200x200) Latent(200x200x200) Hard(200x200x200) (b) Small scale 3000 5000 10000 3000 5000 10000 10 0 10 1 10 2 10 3 10 4 Tensor Size (per dimension) Timeinseconds r=(5,5,5) r=(10,10,10)Proposed geomCG (c) Large scale. 0 50 100 150 200 10 −4 10 −2 10 0 10 2 Time in seconds MeansquareerroronΓ Proposed (CN=5) geomCG (CN=5) Proposed (CN=50) geomCG (CN=50) Proposed (CN=100) geomCG (CN=100) (f) Ill-cond. & low sample (5,5,5) (5,5,5) (5,5,5) (7,6,6) (10,5,5)(15,4,4) 0 100 200 300 400 500 600 700 rank Timeinseconds (a) 20000 × 7000 × 7000 30000 × 6000 × 6000 40000 × 5000 × 5000 (b) 10000 × 10000 × 10000 Proposed geomCG (h) Rectangular instance 0 100 200 300 10 −4 10 −2 10 0 10 2 10 4 Time in seconds MeansquareerroronΓ Proposed geomCG HaLRTC TOpt Latent Hard (i) hyperspectral image. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 12
  • 13. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Results on synthetic and real datasets (2) Hyperspectral image Ribeira OS = 11 OS = 22 Algorithm Time MSE on Γ Time MSE on Γ Proposed 33 ± 13 8.2095 · 10−4 ± 1.7 · 10−5 67 ± 43 6.9516 · 10−4 ± 1.1 · 10−5 geomCG 36 ± 14 3.8342 · 10−1 ± 4.2 · 10−2 150 ± 48 6.2590 · 10−3 ± 4.5 · 10−3 HaLRTC 46 ± 0 2.2671 · 10−3 ± 3.6 · 10−5 48 ± 0 1.3880 · 10−3 ± 2.7 · 10−5 TOpt 80 ± 32 1.7854 · 10−3 ± 3.8 · 10−4 27 ± 21 2.1259 · 10−3 ± 3.8 · 10−4 Latent 553 ± 3 2.9296 · 10−3 ± 6.4 · 10−5 558 ± 3 1.6339 · 10−3 ± 2.3 · 10−5 Hard 400 ± 5 6.5090 · 102 ± 6.1 · 101 402 ± 4 6.5989 · 102 ± 9.8 · 101 MovieLens-10M MovieLens Proposed geomCG r Time MSE on Γ Time MSE on Γ (4, 4, 4) 1748 ± 441 0.6762 ± 1.5 · 10−3 2981 ± 40 0.6956±?2.8 · 10−3 (6, 6, 6) 6058 ± 47 0.6913 ± 3.3 · 10−3 6554 ± 655 0.7398±?7.1 · 10−3 (8, 8, 8) 11370 ± 103 0.7589 ± 7.1 · 10−3 13853 ± 118 0.8955±?3.3 · 10−2 (10, 10, 10) 32802 ± 52 1.0107 ± 2.7 · 10−2 38145 ± 36 1.6550±?8.7 · 10−2 Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 13
  • 14. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Conclusions ▶ Tackled a low-rank tensor completion problem in a Riemannian manifold preconditioning approach. ▶ Proposed a novel Riemannian metric that exploits the fundamental structures of symmetry, due to non-uniqueness of Tucker decomposition, and least-squares of the cost function. ▶ Use the versatile Riemannian optimization framework. ▶ Proposed a preconditioned nonlinear conjugate gradient algorithm. ▶ Also proposed a stochastic gradient algorithm for online setup. ▶ Concrete matrix expressions are worked out. ▶ Numerical comparisons suggest that our proposed algorithm has a superior performance on different benchmarks. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 14
  • 15. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Thank you for your attention ▶ Paper ▶ ICML paper: http://jmlr.org/proceedings/papers/v48/kasai16.pdf ▶ ICML supplementary paper: http://jmlr.org/proceedings/papers/v48/kasai16-supp.pdf ▶ Software (Matlab codes) ▶ Independent project: http://bamdevmishra.com/codes/tensorcompletion/ ▶ Built-in project in Manopt: http://www.manopt.org/fixedrankfactory_tucker_ preconditioned Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 15
  • 16. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons References I ▶ Absil, P.-A., Mahony, R., and Sepulchre, R. (2008). Optimization Algorithms on Matrix Manifolds. Princeton University Press. ▶ Boumal, N., Mishra, B., Absil, P.-A., and Sepulchre, R. (2014). Manopt: a Matlab toolbox for optimization on manifolds. J. Mach. Learn. Res., 15(1):1455–1459. ▶ Filipovi´c, M. and Juki´c, A. (2013). Tucker factorization with missing data with application to low-n-rank tensor completion. Multidim. Syst. Sign. P. Doi: 10.1007/s11045-013-0269-9. ▶ Foster, D. H., Nascimento, S. M. C., and Amano, K. (2007). Information limits on neural identification of colored surfaces in natural scenes. Visual Neurosci., 21(3):331–336. ▶ Kolda, T. G. and Bader, B. W. (2009). Tensor decompositions and applications. SIAM Review, 51(3):455–500. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 16
  • 17. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons References II ▶ Kressner, D., Steinlechner, M., and Vandereycken, B. (2014). Low-rank tensor completion by Riemannian optimization. BIT Numer. Math., 54(2):447–468. ▶ Liu, J., Musialski, P., Wonka, P., and Ye, J. (2013). Tensor completion for estimating missing values in visual data. IEEE Trans. Pattern Anal. Mach. Intell., 35(1):208–220. ▶ Ring, W. and Wirth, B. (2012). Optimization methods on Riemannian manifolds and their application to shape space. SIAM J. Optim., 22(2):596–627. ▶ Sato, H. and Iwai, T. (2015). A new, globally convergent Riemannian conjugate gradient method. Optimization, 64(4):1011–1031. ▶ Signoretto, M., Dinh, Q. T., Lathauwer, L. D., and Suykens, J. A. K. (2014). Learning with tensors: a framework based on convex optimization and spectral regularization. Mach. Learn., 94(3):303–351. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 17
  • 18. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons References III ▶ Tomioka, R., Hayashi, K., and Kashima, H. (2011). Estimation of low-rank tensors via convex optimization. Technical report, arXiv preprint arXiv:1010.0789. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 18
  • 19. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Supplemental slides Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 19
  • 20. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Proposal: Invariant property of new Riemannian metric Proposition (Invariant property of new Riemannian metric) Let (ξU1 , ξU2 , ξU3 , ξG) and (ηU1 , ηU2 , ηU3 , ηG) be tangent vectors to the quotient manifold at (U1, U2, U3, G), and (ξU1O1 , ξU2O2 , ξU3O3 , ξG×1OT 1 ×2OT 2 ×3OT 3 ) and (ηU1O1 , ηU2O2 , ηU3O3 , ηG×1OT 1 ×2OT 2 ×3OT 3 ) be tangent vectors to the quotient manifold at (U1O1, U2O2, U3O3, G×1OT 1 ×2OT 2 ×3OT 3 ). The new metric (2) are invariant along the equivalence class (6) as g(U1,U2,U3,G)((ξU1 , ξU2 , ξU3 , ξG), (ηU1 , ηU2 , ηU3 , ηG)) = g(U1O1,U2O2,U3O3,G×1OT 1 ×2OT 2 ×3OT 3 ) ((ξU1O1 , ξU2O2 , ξU3O3 , ξG×1OT 1 ×2OT 2 ×3OT 3 ), (ηU1O1 , ηU2O2 , ηU3O3 , ηG×1OT 1 ×2OT 2 ×3OT 3 )). Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 20
  • 21. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Quotient manifold and horizontal lift ▶ x and y belong to the same equivalence class and they represent a single point [x] := {y ∈ M : y ∼ x} on M/∼. ▶ Tangent space is decomposable into TxM = Vx ⊕ Hx, where ▶ the vertical space Vx is the tangent space of the equivalence class [x], and ▶ the horizontal space Hx is the orthogonal subspace to Vx. ▶ Vx does NOT induce a displacement along the equivalence class [x]. ▶ horizontal lift ▶ a unique element ξx ∈ Hx that ξ[x] ∈ T[x](M/∼) at [x] has. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 21
  • 22. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Tangent space projection operator (1) Normal space NxM ▶ Obtain from the extracted component of normal space to TxM in the ambient space must locate on the tangent space. ▶ The following lemma can be shown for the normal space Nx; Lemma (Normal space) The quotient manifold endowed with the Riemannian metric (2) has the matrix characterization of the normal space NxM defined as {(U1SU1 (G1GT 1 )−1 , U2SU2 (G2GT 2 )−1 , U3SU3 (G3GT 3 )−1 , 0) : SUd ∈ Rrd×rd , ST Ud = SUd , for d ∈ {1, 2, 3}}. where SUd for all d ∈ {1, 2, 3} are symmetric matrices. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 22
  • 23. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Tangent space projection operator (2) Tangent space projector Ψx Proposition (Tangent space projection operator) The quotient manifold endowed with the Riemannian met- ric (2) admits a tangent projector Ψx : Rn1×r1 × Rn2×r2 × Rn3×r3 × Rr1×r2×r3 → TxM defined as Ψx(YU1, YU2, YU3, YG) = (YU1−U1SU1 (G1GT 1 )−1, YU2−U2SU2 (G2GT 2 )−1, YU3−U3SU3 (G3GT 3 )−1, YG), where SUd is the solution to the Lyapunov equation below; SUd GdGT d + GdGT d SUd = GdGT d (YT Ud Ud + UT d YUd )GdGT d ▶ They are solved efficiently with the Matlab’s lyap routine. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 23
  • 24. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Horizontal space projection operator (1) Vertical space Vx and Horizontal space Hx ▶ This is obtained by ▶ Calculate the condition (Lemma below) of the horizontal space from the orthogonal relationship with the vertical space. ▶ Remove the component along the vertical space, and the remaining components must satisfy the condition above. ▶ The vertical space Vx has the matrix characterization {(U1Ω1, U2Ω2, U3Ω3, −(G×1Ω1 + G×2Ω2 + G×3Ω3)) : Ωd ∈ Rrd×rd , ΩT d = −Ωd for d ∈ {1, 2, 3}}. ▶ where Ωd for all d ∈ {1, 2, 3} are skew symmetric matrices. Lemma (Horizontal space) Horizontal space ξx = (ξU1 , ξU2 , ξU3 , ξG) ∈ Hx must satisfy (GdGT d )ξT Ud Ud + ξGd GT d is symmetric for d ∈ {1, 2, 3}. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 24
  • 25. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Horizontal space projection operator (2) Vertical space Vx and Horizontal space projector Πx Proposition (Horizontal space projection operator) The quotient manifold endowed with the Riemannian metric (2) admits a horizontal projector Πx : TxM :→ Hx : ηx → Πx(ηx) defined as Πx(ηx) = (ηU1− U1Ω1, ηU2− U2Ω2, ηU3− U3Ω3, ηG −(−(G×1Ω1+G×2Ω2+G×3Ω3))), where ηx = (ηU1 , ηU2 , ηU3 , ηG) ∈ TxM and Ωd is a skew- symmetric matrix of size rd × rd that is the solution to the coupled Lyapunov equations below; (continue to the next page.) Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 25
  • 26. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Horizontal space projection operator (3) Horizontal space projector Πx Proposition (Horizontal space projection operator) (continued from the previous page.)    G1GT 1 Ω1 + Ω1G1GT 1 − G1(Ir3 ⊗ Ω2)GT 1 − G1(Ω3 ⊗ Ir2 )GT 1 = Skew(UT 1 ηU1 G1GT 1 ) + Skew(G1ηT G1 ), G2GT 2 Ω2 + Ω2G2GT 2 − G2(Ir3 ⊗ Ω1)GT 2 − G2(Ω3 ⊗ Ir1 )GT 2 = Skew(UT 2 ηU2 G2GT 2 ) + Skew(G2ηT G2 ), G3GT 3 Ω3 + Ω3G3GT 3 − G3(Ir2 ⊗ Ω1)GT 3 − G3(Ω2 ⊗ Ir1 )GT 3 = Skew(UT 3 ηU3 G3GT 3 ) + Skew(G3ηT G3 ). ▶ Skew(·) extracts the skew-symmetric part of a square matrix, i.e., Skew(D) = (D − DT )/2. ▶ The coupled Lyapunov equations are solved efficiently with the Matlab’s pcg routine that is combined with a specific preconditioner resulting from the Gauss-Seidel approximation. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 26
  • 27. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Retraction Rx(ξx) ▶ Map vectors in the horizontal space to points on the search space M and satisfies the local rigidity condition [Absil et al., 2008, Definition 4.1]. ▶ Provides a natural way to move on the manifold along a search direction. ▶ Due to the product nature of M, we can choose a retraction by combining retractions on the individual manifolds, i.e., Rx(ξx) = (uf(U1 + ξU1 ), uf(U2 + ξU2 ), uf(U3 + ξU3 ), G + ξG), where ξx ∈ Hx and uf(·) extracts the orthogonal factor of a full column rank matrix, i.e., uf(A) = A(AT A)−1/2. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 27
  • 28. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Vector transport Tηx ξx ▶ Smooth mapping that transports a tangent vector ξx ∈ TxM at x ∈ M to a vector in the tangent space at Rx(ηx) [Absil et al., 2008, Section 8.1.4]. ▶ Generalize the classical concept of translation of vectors in the Euclidean space to manifolds. ▶ The horizontal lift of the abstract vector transport Tη[x] ξ[x] on M/∼ has the matrix characterization ΠRx(ηx)(Tηx ξx) = ΠRx(ηx)(ΨRx(ηx)(ξx)), where ξx and ηx are the horizontal lifts in Hx of ξ[x] and η[x] that belong to T[x](M/∼). Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 28
  • 29. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Riemannian gradient computations (1) Horizontal lift (1) ▶ Finally, from the Riemannian submersion theory [Absil et al., 2008, Section 3.6.2], the horizontal lift of grad[x]f is equal to gradxf = Ψ(egradxf). ▶ Subsequently, we obtain below; the horizontal lift of grad[x]f = (S1(U3 ⊗ U2)GT 1 (G1GT 1 )−1 − U1BU1 (G1GT 1 )−1 , S2(U3 ⊗ U1)GT 2 (G2GT 2 )−1 − U2BU2 (G2GT 2 )−1 , S3(U2 ⊗ U1)GT 3 (G3GT 3 )−1 − U3BU3 (G3GT 3 )−1 , S ×1 UT 1 ×2 UT 2 ×3 UT 3 ), Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 29
  • 30. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Riemannian gradient computations (2) Horizontal lift (2) ▶ where BUd for d ∈ {1, 2, 3} are the solutions to the Lyapunov equations    BU1 G1GT 1 + G1GT 1 BU1 = 2Sym(G1GT 1 UT 1 (S1(U3 ⊗ U2)GT 1 ), BU2 G2GT 2 + G2GT 2 BU2 = 2Sym(G2GT 2 UT 2 (S2(U3 ⊗ U1)GT 2 ), BU3 G3GT 3 + G3GT 3 BU3 = 2Sym(G3GT 3 UT 3 (S3(U2 ⊗ U1)GT 3 ), which are solved efficiently with the Matlab’s lyap routine. Sym(·) extracts the symmetric part of a square matrix, i.e., Sym(D) = (D + DT )/2. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 30
  • 31. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Preconditioned Conjugate Gradient (CG) Algorithm ▶ Off-the-shelf conjugate gradient implementation of Manopt 1 [Boumal et al., 2014] for any smooth cost function. ▶ The convergence analysis of the Riemannian conjugate gradient method follows from [Sato and Iwai, 2015, Ring and Wirth, 2012]. ▶ Cost function specific ingredients are needed; ▶ Compute the Riemannian gradient. ▶ Compute an initial guess for the step-size for CG. ▶ The total computational cost per iteration of our algorithm is O(|Ω|r1r2r3), where |Ω| is the number of known entries. 1 http://www.manopt.org/Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 31
  • 32. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Experiment conditions ▶ Comparisons with state-of-the-art algorithms that include ▶ Tucker decomposition based algorithms; ▶ TOpt [Filipovi´c and Juki´c, 2013], geomCG [Kressner et al., 2014] ▶ Nuclear norm minimization algorithms; ▶ HaLRTC [Liu et al., 2013], Latent [Tomioka et al., 2011], Hard [Signoretto et al., 2014] ▶ Problem instances ▶ Case S2: small-scale instances. ▶ Case S3: large-scale instances. ▶ Case S5: influence of ill-conditioning and low sampling. ▶ Case S7: asymmetric instances. ▶ Case R1: hyperspectral image “Ribeira” [Foster et al., 2007]. ▶ Case R2: MovieLens-10M. ▶ Note that, compared with ONLY geomCG for large-scale instances, i.e., except Case S2. Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 32
  • 33. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Synthetic dataset cases ▶ Case S2: small-scale instances. ▶ Size 100 × 100 × 100, 150 × 150 × 150, and 200 × 200 × 200 and rank r = (10, 10, 10) are considered. OS is {10, 20, 30}. ▶ Case S3: large-scale instances. ▶ Tensors of size 3000 × 3000 × 3000, 5000 × 5000 × 5000, and 10000 × 10000 × 10000 and ranks r = (5, 5, 5) and (10, 10, 10). OS is 10. ▶ Case S5: influence of ill-conditioning and low sampling. ▶ Case S4 with OS = 5. We impose a diagonal core G with exponentially decaying positive values of condition numbers (CN) 5, 50, and 100. ▶ Case S7: asymmetric instances. ▶ (a): tensors size 20000 × 7000 × 7000, 30000 × 6000 × 6000, and 40000 × 5000 × 5000 with rank r = (5, 5, 5). ▶ (b): tensor size 10000 × 10000 × 10000 with ranks (7, 6, 6), (10, 5, 5), and (15, 4, 4). Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 33
  • 34. Introduction Metric proposal Tucker manifold geometry Numerical Comparisons Real-world dataset cases ▶ Case R1: hyperspectral image “Ribeira”. [Foster et al., 2007] ▶ The tensor size is 203 × 268 × 33. ▶ Compare all the algorithms, and perform five random samplings of the pixels based on the OS values 11 and 22, corresponding to the rank r=(15, 15, 6) adopted in [Kressner et al., 2014]. ▶ While OS = 22 corresponds to the observation ratio of 10% studied in [Kressner et al., 2014], OS = 11 considers a challenging scenario with the observation ratio of 5%. ▶ Case R2: MovieLens-10M2. ▶ This dataset contains 10000054 ratings corresponding to 71567 users and 10681 movies. ▶ Split the time into 7-days wide bins results, and finally, get a tensor of size 71567 × 10681 × 731. ▶ The fraction of known entries is less than 0.002%. 2 http://grouplens.org/datasets/movielens/.Low-rank tensor completion: a Riemannian manifold preconditioning approach (ICML2016) (copyrights by Kasai & Mishra) 34