The document discusses convex minimization problems under submodular constraints. It begins with an overview of submodular functions and base polytopes. It then defines the convex minimization problem of minimizing a separable convex function g(x) over the base polytope B(f). Several examples of this problem are given, including lexicographically optimal bases and submodular utility allocation markets. It shows these problems are equivalent and discusses the α-curve representation of their optimal solutions.
i give some indeas on how to use asymptotic series and expansion to prove Riemann Hypothesis, solve integral equations and even define a regularized integral of powers
i give some indeas on how to use asymptotic series and expansion to prove Riemann Hypothesis, solve integral equations and even define a regularized integral of powers
The derivative of a composition of functions is the product of the derivatives of those functions. This rule is important because compositions are so powerful.
On the Application of a Classical Fixed Point Method in the Optimization of a...BRNSS Publication Hub
This work on classical optimization reveals the Newton’s fixed point iterative method as involved in
the computation of extrema of convex functions. Such functions must be differentiable in the Banach
space such that their solution exists in the space on application of the Newton’s optimization algorithm
and convergence to the unique point is realized. These results analytically were carried as application
into the optimization of a multieffect evaporator which reveals the feasibility of theoretical and practical
optimization of the multieffect evaporator.
The derivative of a composition of functions is the product of the derivatives of those functions. This rule is important because compositions are so powerful.
On the Application of a Classical Fixed Point Method in the Optimization of a...BRNSS Publication Hub
This work on classical optimization reveals the Newton’s fixed point iterative method as involved in
the computation of extrema of convex functions. Such functions must be differentiable in the Banach
space such that their solution exists in the space on application of the Newton’s optimization algorithm
and convergence to the unique point is realized. These results analytically were carried as application
into the optimization of a multieffect evaporator which reveals the feasibility of theoretical and practical
optimization of the multieffect evaporator.
Here's a toy problem: What is the SMALLEST number of unit balls you can fit in a box such that no more will fit?
In this talk, I will show how just thinking about a naive greedy approach to this problem leads to a simple derivation of several of the most important theoretical results in the field of mesh generation.
We'll prove classic upper and lower bounds on both the number of balls and the complexity of their interrelationships.
Then, we'll relate this problem to a similar one called the Fat Voronoi Problem, in which we try to find point sets such that every Voronoi cell is fat
(the ratio of the radii of the largest contained to smallest containing ball is bounded).
This problem has tremendous promise in the future of mesh generation as it can circumvent the classic lowerbounds presented in the first half of the talk.
Unfortunately the simple approach no longer works.
In the end we will show that the number of neighbors of any cell in a Fat Voronoi Diagram in the plane is bounded by a constant
(if you think that's obvious, spend a minute to try to prove it).
We'll also talk a little about the higher dimensional version of the problem and its wide range of applications.
The first report of Machine Learning Seminar organized by Computational Linguistics Laboratory at Kazan Federal University. See http://cll.niimm.ksu.ru/cms/lang/en_US/main/seminars/mlseminar
Image sciences, image processing, image restoration, photo manipulation. Image and videos representation. Digital versus analog imagery. Quantization and sampling. Sources and models of noises in digital CCD imagery: photon, thermal and readout noises. Sources and models of blurs. Convolutions and point spread functions. Overview of other standard models, problems and tasks: salt-and-pepper and impulse noises, half toning, inpainting, super-resolution, compressed sensing, high dynamic range imagery, demosaicing. Short introduction to other types of imagery: SAR, Sonar, ultrasound, CT and MRI. Linear and ill-posed restoration problems.
Transcript: Selling digital books in 2024: Insights from industry leaders - T...BookNet Canada
The publishing industry has been selling digital audiobooks and ebooks for over a decade and has found its groove. What’s changed? What has stayed the same? Where do we go from here? Join a group of leading sales peers from across the industry for a conversation about the lessons learned since the popularization of digital books, best practices, digital book supply chain management, and more.
Link to video recording: https://bnctechforum.ca/sessions/selling-digital-books-in-2024-insights-from-industry-leaders/
Presented by BookNet Canada on May 28, 2024, with support from the Department of Canadian Heritage.
Generating a custom Ruby SDK for your web service or Rails API using Smithyg2nightmarescribd
Have you ever wanted a Ruby client API to communicate with your web service? Smithy is a protocol-agnostic language for defining services and SDKs. Smithy Ruby is an implementation of Smithy that generates a Ruby SDK using a Smithy model. In this talk, we will explore Smithy and Smithy Ruby to learn how to generate custom feature-rich SDKs that can communicate with any web service, such as a Rails JSON API.
Software Delivery At the Speed of AI: Inflectra Invests In AI-Powered QualityInflectra
In this insightful webinar, Inflectra explores how artificial intelligence (AI) is transforming software development and testing. Discover how AI-powered tools are revolutionizing every stage of the software development lifecycle (SDLC), from design and prototyping to testing, deployment, and monitoring.
Learn about:
• The Future of Testing: How AI is shifting testing towards verification, analysis, and higher-level skills, while reducing repetitive tasks.
• Test Automation: How AI-powered test case generation, optimization, and self-healing tests are making testing more efficient and effective.
• Visual Testing: Explore the emerging capabilities of AI in visual testing and how it's set to revolutionize UI verification.
• Inflectra's AI Solutions: See demonstrations of Inflectra's cutting-edge AI tools like the ChatGPT plugin and Azure Open AI platform, designed to streamline your testing process.
Whether you're a developer, tester, or QA professional, this webinar will give you valuable insights into how AI is shaping the future of software delivery.
Builder.ai Founder Sachin Dev Duggal's Strategic Approach to Create an Innova...Ramesh Iyer
In today's fast-changing business world, Companies that adapt and embrace new ideas often need help to keep up with the competition. However, fostering a culture of innovation takes much work. It takes vision, leadership and willingness to take risks in the right proportion. Sachin Dev Duggal, co-founder of Builder.ai, has perfected the art of this balance, creating a company culture where creativity and growth are nurtured at each stage.
GraphRAG is All You need? LLM & Knowledge GraphGuy Korland
Guy Korland, CEO and Co-founder of FalkorDB, will review two articles on the integration of language models with knowledge graphs.
1. Unifying Large Language Models and Knowledge Graphs: A Roadmap.
https://arxiv.org/abs/2306.08302
2. Microsoft Research's GraphRAG paper and a review paper on various uses of knowledge graphs:
https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/
State of ICS and IoT Cyber Threat Landscape Report 2024 previewPrayukth K V
The IoT and OT threat landscape report has been prepared by the Threat Research Team at Sectrio using data from Sectrio, cyber threat intelligence farming facilities spread across over 85 cities around the world. In addition, Sectrio also runs AI-based advanced threat and payload engagement facilities that serve as sinks to attract and engage sophisticated threat actors, and newer malware including new variants and latent threats that are at an earlier stage of development.
The latest edition of the OT/ICS and IoT security Threat Landscape Report 2024 also covers:
State of global ICS asset and network exposure
Sectoral targets and attacks as well as the cost of ransom
Global APT activity, AI usage, actor and tactic profiles, and implications
Rise in volumes of AI-powered cyberattacks
Major cyber events in 2024
Malware and malicious payload trends
Cyberattack types and targets
Vulnerability exploit attempts on CVEs
Attacks on counties – USA
Expansion of bot farms – how, where, and why
In-depth analysis of the cyber threat landscape across North America, South America, Europe, APAC, and the Middle East
Why are attacks on smart factories rising?
Cyber risk predictions
Axis of attacks – Europe
Systemic attacks in the Middle East
Download the full report from here:
https://sectrio.com/resources/ot-threat-landscape-reports/sectrio-releases-ot-ics-and-iot-security-threat-landscape-report-2024/
Encryption in Microsoft 365 - ExpertsLive Netherlands 2024Albert Hoitingh
In this session I delve into the encryption technology used in Microsoft 365 and Microsoft Purview. Including the concepts of Customer Key and Double Key Encryption.
Key Trends Shaping the Future of Infrastructure.pdfCheryl Hung
Keynote at DIGIT West Expo, Glasgow on 29 May 2024.
Cheryl Hung, ochery.com
Sr Director, Infrastructure Ecosystem, Arm.
The key trends across hardware, cloud and open-source; exploring how these areas are likely to mature and develop over the short and long-term, and then considering how organisations can position themselves to adapt and thrive.
Securing your Kubernetes cluster_ a step-by-step guide to success !KatiaHIMEUR1
Today, after several years of existence, an extremely active community and an ultra-dynamic ecosystem, Kubernetes has established itself as the de facto standard in container orchestration. Thanks to a wide range of managed services, it has never been so easy to set up a ready-to-use Kubernetes cluster.
However, this ease of use means that the subject of security in Kubernetes is often left for later, or even neglected. This exposes companies to significant risks.
In this talk, I'll show you step-by-step how to secure your Kubernetes cluster for greater peace of mind and reliability.
Elevating Tactical DDD Patterns Through Object CalisthenicsDorra BARTAGUIZ
After immersing yourself in the blue book and its red counterpart, attending DDD-focused conferences, and applying tactical patterns, you're left with a crucial question: How do I ensure my design is effective? Tactical patterns within Domain-Driven Design (DDD) serve as guiding principles for creating clear and manageable domain models. However, achieving success with these patterns requires additional guidance. Interestingly, we've observed that a set of constraints initially designed for training purposes remarkably aligns with effective pattern implementation, offering a more ‘mechanical’ approach. Let's explore together how Object Calisthenics can elevate the design of your tactical DDD patterns, offering concrete help for those venturing into DDD for the first time!
2. 本発表の構成
劣モジュラ関数 f に付随して定義される基多面体 B(f) 上
での凸関数最小化問題に関する理論と応用のはなし
プロローグ Spanning tree のはなし
Part 1 劣モジュラ制約下の凸最適化の理論のはなし
Nagano (IPCO 2007) など
Part 2 劣モジュラ制約下の凸最適化の応用のはなし
Nagano, Kawahara, Aihara, ICML 2011
3. グラフの spanning tree
Spanning tree :
グラフ G のすべての node をつなげ,
かつサイクルなしの edge 部分集合
Spanning tree はたくさんある
プロローグ
4. 最小重み spanning tree
5 8 各 edge に重みがあるとし, 重み和
7 最小のspanning tree を求めたい
9 3
spanning tree を全列挙する必要なし
最小重み 貪欲アルゴリズムで効率的に計算可能
spanning tree Kruskal 1956
プロローグ
5. Spanning tree が作る多面体
edgeの有無を1と0で表した
Spanning tree の特性ベクトル 0-1ベクトル
a b
c (1,1,1,0,0) (1,0,1,0,1) (0,1,1,1,0) (0,0,1,1,1)
d e
(1,1,0,1,0) (1,1,0,0,1) (0,1,0,1,1) (1,0,0,1,1)
多面体 PG = (これらの凸包)
凸包
実はこの多面体は基多面体B(f)の例になっている
プロローグ
7. 本発表の構成
劣モジュラ関数 f に付随して定義される基多面体 B(f) 上
での凸関数最小化問題に関する理論と応用のはなし
プロローグ Spanning tree のはなし
Part 1 劣モジュラ制約下の凸最適化の理論のはなし
Nagano (IPCO 2007) など
Part 2 劣モジュラ制約下の凸最適化の応用のはなし
Nagano, Kawahara, Aihara, ICML 2011
8. Part 1 劣モジュラ制約下の凸最適化の理論のはなし
On Convex Minimization
over Base Polytopes
Kiyohito Nagano
Notations:
• V = {1, ..., n}
• x ∈RV は v (∈V) 番目の成分が
x(v) であるような n 次元ベクトル
• x(S) =∑ x(v) (S⊆V)
v∈S
9. Submodular Functions
V = {1, ... , n } is a finite set
Set function f : 2V→ R is
monotone if ∀S ,T s. t. S ⊆T , f (S ) ≤ f (T )
submodular if
V
∀S ,T ⊆V, S
T
f (S ) + f (T ) ≥ f (S ∪T ) + f (S ∩T )
• Graphs, Networks
• Game Theory, Auction Theory, Machine Learning
Part 1
10. Submodular Functions
V = {1, ... , n } is a finite set
Set function f : 2V→ R is
monotone if ∀S ,T s. t. S ⊆T , f (S ) ≤ f (T )
∀S ,T s. t. S ⊆T, and ∀v ∉ T
(equivalently) submodular if
V
v
T
S
f (S +v ) – f (S ) ≥ f (T +v ) – f (T )
Diminishing reterns
• Graphs, Networks
• Game Theory, Auction Theory, Machine Learning
Part 1
11. An Interpretation of function f
In this part, we assume
f : 2V→ R is monotone & submodular, f (φ ) = 0
Such f can be seen as a natural cost function
For S ⊆V, we incur cost f (S ) S
One extreme example Another example
∑a(v) Monotone & β >0
a ∈RV
2(S ) = 0 if S =φ
>0 v∈S submodular
f 1(S ) = a(S ) functions f
β o. w.
$5
Intermediate
• Customer 1 eats functions of f 1 & f 2 The case where
• Customer 2 eats we can make a
• Customer 3 eats copy for free
Part 1
12. Base Polytope B (f )
f : 2V→ R is monotone & submodular, f (φ ) = 0
x(S) ≤ f (S ) ( φ ≠S ⊊ V ) 2n − 2 ineq
The base polytope B(f ) ⊆RV : Notations:
• x = (x(v):v ∈V )∈RV
x(V ) = f (V ) 1 equality • x(S) =∑ x(v)
v∈S
x(3)
f 1(S ) = a(S ) f 2(S ) = 0 if S =φ
B(f )⊆RV ≥0 β o.w.
B(f 1) B(f 2)
x(1)
6 inequalities
x(2) n=3 1 equality
Part 1
13. Our Problem
Let g : RV → R is an n-variable, separable convex function
We mainly consider
x(3)
minx g (x) = ∑ gv (x(v)) 1-variable,
v∈V convex
sub. to x(S) ≤ f (S ) ∀S
x ∊ B (f )
x(V ) = f (V ) x(1)
x(2)
Assumptions :
f : 2V→ R is monotone & submodular, f (φ ) = 0
f is given by a value-giving oracle
Function gv (x(v)) is differentiable & strictly convex
Part 1
14. Outline of Part 1
Definitions
Examples & Our Results
Equivalence of Problems
Remark: Efficient algorithm
minx g (x ) over B (f )
Part 1
15. Examples of minx g (x ) over B (f )
Lexicographically optimal bases Fujishige 80
• weight vector w ∈ RV>0
• For x ∊ B (f ) ,
x(v ) x(vn) x(v ) x(v )
Inc.Orderw (x) = ( 1 , ... ,
w(v1) w(vn) )
with w(v1 ) ≤ ≤ w(vn )
1 n
the sequence of weighted components of x in nondecreasing order
x = (1,3,4), w = (10,10, 20) Inc.Orderw (x )
x(3)
x(1) 1 3 4 =( 1 , 4 , 3 )
10 10 20 10 20 10
x(2)
• Lexicographically optimal base :
min ∑ x(v)2 over
lex. max Inc.Orderw(x) over B(f ) B(f )
x x v ∈V w(v)
equivalent
Part 1
16. Examples of minx g (x ) over B (f )
Lexicographically optimal bases Fujishige 80
• weight vector w ∈ RV>0
• Lexicographically optimal base :
min ∑ x(v)2 over
lex. max Inc.Orderw(x) over B(f ) B(f )
x x v ∈V w(v)
equivalent
Related problems:
Generalization of lex opt flow Megiddo 74
Egalitarian solution in convex game Dutta & Ray 89
Lex opt base x* minimum ratio problem min f (S )
X ⊆V w(S )
Min norm point x* submodular function minimization (SFM)
(w = 1) Fujishige 84 (Now f is not necessarily monotone)
Size-constrained SFM (See Part 2)
Part 1
17. Examples of minx g (x ) over B (f ) (continued)
Lex opt bases Fujishige 80 2
min ∑ x(v) over B(f )
• weight vector w ∈ RV >0 x v ∈V w(v)
Submodular utility allocation (SUA) markets Jain & Vazirani 07
• money vector m ∈ RV >0 max ∑ m (v ) ln x(v ) over B(f )
x v ∈V
Surprisingly, we will see
Corollary (Our Result)
The lex opt base problem and the SUA market problem
have the same solution if w = m
Subproblems in approximation algorithms
• Facility location with submodular penalty (Chudak & Nagano, 07)
• Submodular cost set cover problem (Iwata & Nagano, 09)
Part 1
18. Outline of Part 1
Definitions
Examples & Our Results
Equivalence of Problems
Remark: Efficient algorithm
minx g (x ) over B (f )
Part 1
19. α -curve C = { xα∈ RV : α ∈ J }
Let C⊆RV be the set of optimal solutions to
minx { Σ gv (x(v )) : x(V ) = β } ∀β > 0
v ∈V
minimization under a budget constraint
β3
β2
β1
β1 β2 β3 Curve C
β1 β2
β3
parameter
‐1
C = { xα ∈ RV : α ∈ J ⊆ R } where xα (v ) = (gv) (α ),
α -curve J is an interval⊆ R
Since gv is convex, α1 < α2 ⇒ xα1< xα2
component-wise
Part 1
20. Examples of α - curve C
Positive vectors w, m ∈ RV >0
gv (x(v )) = x(v )2 / w (v ) gv (x(v )) = - m (v ) ln x(v )
Lex opt base, Fujishige 80 SUA market, Jain-Vazirani 07
α -curve C α -curve C
w m
half line half line
C = { 2α w : α ∈ (0, +∞) } C = { - 1 m : α ∈(- ∞, 0) }
α
xα xα
If w = m, two curves coincide
Part 1
21. Rationality of optimal solutions
From the correctness of Decomposition Algorithm, we obtain
sufficient conditions for the rationality of optimal solutions :
Theorem (Our Result) y*
∀β ∈Q>0,∀U⊆V , the optimal solution y* to β
min U { Σ gv (y(v )) : y(U ) = β } is rational β
y ∈R v ∈U
then the optimal solution x* to
minx g (x ) over B (f ) is rational
Part 1
22. Equivalence of problems
Suppose that we are given
α - curve C(1)
(1) (1)
g (x ) = ∑ gv (x(v ))
v ∈V
(2)
α - curve C(2)
(2)
g (x ) = ∑ gv (x(v ))
v ∈V
Theorem (Our Result)
(2)
Curves C(1) and C coincide
minx g(1)(x ) over B (f ) and minx g(2)(x ) over B (f )
have the same optimal solution
Part 1
23. Equivalence of problems (examples)
Corollary
The lex opt base problem (Fujishige 80) and the SUA market
problem (Jain&Vazirani 07) have the same solution if w = m
gv (x(v )) = x(v )2 / w (v ) gv (x(v )) = - m(v ) ln x(v )
w α -curve C m α -curve C
2 curves coincide
Corollary
Suppose all functions gv are the same (gv = g0)
x ∈ B(f ) minimizes Euclidean norm ||x|| if and only if
x ∈ B(f ) minimizes g(x ) = ∑ g0 (x(v))
v ∈V
Part 1
24. Outline of Part 1
Definitions
Examples & Our Results
Equivalence of Problems
Remark: Efficient algorithm
minx g (x ) over B (f )
Part 1
25. Remark: Efficient Algorithm
In this work, we generalized the discussion of
Fleischer & Iwata 03 for lex. opt. base problem
An efficient algorithm for minx g(x ) over B (f )
via parametric submodular function minimization
The method is similar to the parametric cut
algorithm by Gallo, Grigoriadis & Tarjan 89
Parametric version of Orlin's algorithm (IPCO07)
Theorem (Nagano 07, manuscript)
minx g(x ) over B (f ) can be solved in O(n6+ n5EO) time,
where EO is the time for one function evaluation
Part 1
26. 本発表の構成
劣モジュラ関数 f に付随して定義される基多面体 B(f) 上
での凸関数最小化問題に関する理論と応用のはなし
プロローグ Spanning tree のはなし
Part 1 劣モジュラ制約下の凸最適化の理論のはなし
Nagano (IPCO 2007) など
Part 2 劣モジュラ制約下の凸最適化の応用のはなし
Nagano, Kawahara, Aihara, ICML 2011
28. Size-constrained Submodular Minimization
through Minimum Norm Base
Outline of Part 2
The densest subgraph problem
This problem is an important special case of
Introduction
the size-constrained submodular minimization
Size-constrained submodular minimization
The algorithm through minimum norm base
Concluding remarks
Part 2
29. Example: Finding a dense graph
1
G = (V, E) : undirected graph 1 2 1 1 4
node set V = {1, ... , n} 3 2 1 3
edge set E 5 6
2
Edge weights we ≥ 0(e∈E) S={2,3,6}
Integer k (0≤ k ≤ n) I ({2,3,6}) =1+1 =2
Densest k-subgraph problem :
Find a k-subset S ⊆ V (|S|=k ) that maximizes I(S),
where I(S) is sum of weights of edges within S
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
30. Example: Finding a dense graph
1 2
Densest k-subgraph problem : 1 1 1 4
maximize I(S) 3 2 1 3
5 6
subject to S⊆{1, ... , n}, |S|= k 2
If k =3, an optimal solution is {1,5,6}
NP-hard and constant factor approx algorithms are not known
Applications:
Community detection problem in complex networks
Identification of functional modules in protein-protein
interaction networks
size-constrained
submodular minimization
In this talk, we deal with
a more general discrete optimization problem
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
31. Size-constrained Submodular Minimization
through Minimum Norm Base
Outline of Part 2
Introduction
Size-constrained submodular minimization
The algorithm through minimum norm base
Concluding remarks min f (S)
sub. to S ⊆{1,..., n}, |S| = k
f : submodular
This problem generalizes
densestsize constraint
a k-subgraph problem
size-constrained minimum
cut problem
Fundamental NP-hard problems
Part 2
32. Definition: Submodular function
V ={1, ... , n } is a finite set
A real-valued function f defined on 2V = { S : S ⊆ V } is
submodular if V
S T
∀S, T ⊆V 1
2
3 4
7
f (S)+ f (T )≥ f (S ∪T)+ f (S ∩T)
6
5 8
Discrete analogue of convex function (Lovász 1983)
This function arises in many fields, e.g.,
graph theory, game theory, machine learning, etc.
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
33. Examples of submodular functions
Graph G = (V, E) with edge weights we ≥ 0 (e∈E)
V ={1, ... , n }
Ex. 1. Intensity function I : 2V→ R S
1 2 3
both of two endpoints
Σ
6
I (S) = we : 4
5
7
8 9
of e∈E are in S 10 11 12
– I is submodular (i.e. I is supermodular)
Ex. 2. Cut function C : 2V→ R S
1 2 3
e∈E has one endpoint in S
C (S) = Σ we :
and one in V – S
4
10
5
6
7
8 9
11 12
C is submodular
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
34. Submodular optimization (general form)
min or max f (S) • V ={1, ... , n}
• submodular function f : 2V→ R
subject to S ∈F
• feasible region F ⊆2V
We deal with minimization problems solvable
in poly time
Unconstrained Submodular Minimization (USM)
Size-constrained Submodular Minimization (SSM) NP-hard
Assumption: Computation of each value f (S) is a basic operation
(Given value oracle which answers value queries)
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
35. Unconstrained submodular minimization (USM)
Problem (USM) min f (S)
solvable in
s. t. S ⊆V ={1, ... , n} poly time
polynomial time algorithm:
ellipsoid method: Grotschel, Lovasz & Schrijver (1981, 1988)
combinatorial algorithm: Schrijver (2000), Iwata+ (2001)
”practical” algorithm: Fujishige-Wolfe algorithm (2005)
much faster in practice (Fujishige+ 2006)
this algorithm computes the minimum norm base
minimum L2-norm point in the base polyhedron B(f ) ⊆Rn
min norm base x*∈Rn minimizer of f ⊆{1, ... , n}
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
36. Size-constrained submodular minimization (SSM)
Problem (SSM) min f (S) integer k (0≤ k ≤ n)
s. t. S ⊆V , |S| = k
NP-hard
There is no constant factor approximation algorithm
that runs in polynomial time (Svitkina & Fleischer, 2008)
special cases 1 1 2 1 1 4
3 2 1 3
• f =–I Densest k-subgraph problem 5
2 6
• f =C Size-constrained minimum cut problem
Both of these are fundamental NP-hard problems
Even for these special cases, good approx algorithms are not known
In this work, we propose a new method for (SSM) that
utilizes the minimum norm base
computes ”a portion of exact optimal solutions”
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
37. Example of outputs of the proposed method
For all k =0, ... , n, consider the densest k-subgraph problems
max I (S) s. t. S ⊆V, |S| = k
proposed
1 1 2 1 1 4 method 1 2 4
3 2 1 3 3
5 2 6 5 6
k=0 yes Outputs: {},{1,5,6},{1,2,5,6},
k=1 no {1,2,3,4,5,6}
k=2 no
k=3 yes Our method gives optimal
k=4 yes solutions for k = 0, 3, 4, 6
k=5 no
a portion of exact
k=6 yes
optimal solutions
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
38. Problem (SSM) in ML min f(S) s. t. S⊆V, |S|= k
Densest k-subgraph problem (f = – I)
This problem naturally formulates a community
detection problem in complex networks
The identification of functional modules in protein-
protein interaction networks is known as an important
application (Dittrich et al., 2008)
Size-constrained minimum cut problem (f = C)
This problem deals with an explicit size constraint
Contrastingly, spectral clustering, which is one of the
most popular clustering algorithms, can deal with a
minimum cut problem with an implicit size constraint
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
39. Size-constrained Submodular Minimization
through Minimum Norm Base
Outline of Part 2
Introduction
Size-constrained submodular minimization
The algorithm through minimum norm base
Concluding remarks
Our algorithm uses the basic polyhedral theory
associated with submodular functions x3 B( f )
f : 2V→ R B(f ) ⊆ Rn x1
submodular func base polyhedron x2
Part 2
40. Base polyhedron
For a submodular function f : 2V→ R with f ({}) =0,
the base polyhedron B(f ) ⊆ Rn is given by
B(f ) = {x∊Rn : Σ xi ≤ f (S) (∀S ⊆ V), Σ xi = f (V)}
i∈S i∈V
B(f ) is determined by 2n – 2 inequalities and 1 equality
If n = 3, B(f) is determined by x3
B( f )
x1 ≤ f ({1}) , x1 + x2 ≤ f ({1, 2})
x2 ≤ f ({2}) , x1 + x3 ≤ f ({1, 3}) 23 – 2 ineq x1
x3 ≤ f ({3}) , x2 + x3 ≤ f ({2, 3})
x2
x1 + x2 + x3 = f ({1, 2, 3}) 1 equality
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
41. Minimum norm base and algorithms
The minimum norm base x*∈ Rn is an optimal solution to
min Σ i∈V xi2 s. t. x∊B(f )
x* can be computed efficiently (Fleischer & Iwata, 2003; Nagano, 2007)
Fujishige-Wolfe algorithm (2005) finds x* much faster in practice
Algorithms through minimum norm base x*∈ Rn
The Fujishige-Wolfe algorithm for (USM) min f(S) s. t. S⊆V
S*={i ∊ V : x*< 0} minimizes f
i The problem can be solved
In FW algorithm, we use partial information about x*
The proposed algorithm for (SSM) min f(S) s. t. S⊆V, |S|= k
In our algorithm, we use full information about x*
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
42. min f(S) s. t. S⊆V, |S|= k
Algorithm for (SSM)
Consider the following algorithm: surprisingly simple!
Algorithm SSM
Step1: Compute the minimum norm base x*∈B(f )⊆ Rn
Step2: Let ξ1 < ξ2 < ⋯ < ξ d be distinct values of x*
Return T0 :={} and Tj := {i ∊ V : xi * ξ j }, ∀j = 1,..., d
≤
Example 1 2
1 1 1 4
The densest k-subgraph problems on 3 2 1 3 B(–I ) ⊆ R6
5
2 6
7 7 7
The minimum norm base is x* = (– , –2, –1, –1, – , – ) ∈B(–I )
3 3 3
7
Thus we have ξ1 =– , ξ2 =–2, ξ3 =–1,
3
T0={} , T1={1,5,6}, T2={1,2,5,6}, T3={1,2,3,4,5,6}
optimal solutions for some of size constraints
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
43. min f(S) s. t. S⊆V, |S|= k
Algorithm for (SSM)
Consider the following algorithm: surprisingly simple!
Algorithm SSM
Step1: Compute the minimum norm base x*∈B(f )⊆ Rn
Step2: Let ξ1 < ξ2 < ⋯ < ξ d be distinct values of x*
Return T0 :={} and Tj := {i ∊ V : xi * ξ j }, ∀j = 1,..., d
≤
Theorem [This work] For each j ∈{0, 1,... , d}, Tj ⊆ V is
an optimal solution to Problem (SSM) w. r. t. k = |Tj |.
Algorithm SSM computes ”a portion of exact optimal solutions”
Running time = computation of the minimum norm base
E.g., an implementation of the Fujishige-Wolfe algorithm can be found in
a toolbox for submodular function optimization by Krause (2010, JMLR)
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
44. Proof of the validity of the algorithm SSM (sketch)
h(λ)
Define a function h : R → R as
λ
h(λ)= min{f (S)–|S|λ : S ⊆ V} (λ ∊ R) 0
linear function in λ
h is the minimum of 2n linear functions
h(λ) λ
Each Sj is an exact optimal
0 solution to problem (SSM)
S0 S1 S w.r.t. some size constraint
2 S3
Furthermore, with the aid of the result of Fujishige
(1980), we can show that Sj=Tj for all j, where each
Tj is the subset returned by the algorithm SSM
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
45. Application to artificial data
Networks are randomly generated from the GENRMF generator
For each randomly generated network, we considered
the size-constrained minimum cut problem
the size-constrained minimum s-t cut problem
the densest k-subgraph problem
The number of subsets (= d) found by the algorithm SSM:
dataset 1 (Genrmf-long) dataset 2 (Genrmf-wide)
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
46. Application to real data
We applied the algorithm SSM to the densest k-
subgraph problems on the network with 5,000
nodes and 31,664 edges (n = 5000)
This is a sub-network of social network data cnr-2000
(http://law.dsi.unimi.it/webdata/cnr-2000)
X 104 Intensity
3
The algorithm provides
optimal solutions for 57 I(S) 2
out of 5000 size levels
1
0
0 1000 2000 3000 4000 5000
k
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base
47. Size-constrained Submodular Minimization
through Minimum Norm Base
Outline of Part 2
Introduction
Size-constrained submodular minimization
The algorithm through minimum norm base
Concluding remarks
Part 2
48. Concluding remarks
To the size-constrained submodular minimization
problem, we have proposed a new method that
computes a portion of exact optimal solutions
This result contrasts sharply with the NP-hardness of SSM
(see also the result of Nagano, Kawahara & Iwata, NIPS 2010)
Our method is simple. We just utilize the minimum
norm base to the fullest extent.
The Fujishige-Wolfe algorithm does not have worst
time complexity bounds, so its complexity analysis
should be given in future works.
Part 2 Size-constrained Submodular Minimization through Minimum Norm Base