Submodular Optimization in Social Computing DingZhu Du UT
Submodular Optimization in Social Computing Ding-Zhu Du UT Dallas
What is a submodular function? Consider a function f on all subsets of a set E. f is submodular if Consider as Discrete Convex Function.
Discrete Convex Analysis Kazuo Murota
Similar to Convex • Unconstrained submodular minimization can be done in polynomial-time. • Unconstrained submodular maximization is NP -hard.
Similar to Concave,too! 1 2
Table of Contents 1. Submodular Maximization 2. Submodular Minimization 3. Submodular Partition.
Table of Contents 1. Submodular Maximization Influence Maximization Multi-products: Knapsack Constraint Profit Maximization: Nonmonotone Continuous Greedy 2. Submodular Minimization 3. Submodular Partition.
Influence Maximization
Kate Middleton effect “Kate Middleton effect The trend effect that Kate, Duchess of Cambridge has on others, from cosmetic surgery for brides, to sales of coral-colored jeans. ” 9
Hike in Sales of Special Products n According to Newsweek, "The Kate Effect may be worth £ 1 billion to the UK fashion industry. " n Tony Di. Masso, L. K. Bennett’s US president, stated in 2012, ". . . when she does wear something, it always seems to go on a waiting list. " 10
Influence Maximization • Given a digraph and k>0, • Find k seeds (Kates) to maximize the number of influenced persons. 11
How to Find Kate? • Influential persons often have many friends. • Kate is one of the persons that have many influenced persons in this social network. For more Kates, it’s not as easy as you might think! 12
Theorem Proof 13
Modularity of Influence 14
Submadular Function Max 15
Greedy Algorithm 16
Performance Ratio Theorem (Nemhauser et al. 1978) Corollary 17
Proof Monotone nondecreasing Submodular! Why? 18
Lower Bound 19
Multi-product: Knapsack Constraint
Knapsack Constraint
Knapsack Constraint
Submadular Function Max 23
Submadular Function Max 24
Profit Maximization
Influence and Profit: Two Sides of the Coin 13 th IEEE International Conference on Data Mining (ICDM 2013) Yuqing Zhu 5/22/2021 26
Influence vs. Profit · Classical models do not fully capture monetary aspects of product adoptions ·Being influenced Being willing to purchase · Classical models do not consider the willingness the active nodes on spreading the influence ·Being influenced Being willing to spread the influence 5/22/2021 27
Influence vs. Profit · Influence: · Profit: · In market, a famous company does not always make generous profit. E. g. Twitter, SONY, Weibo 5/22/2021 28
Product Adoption · Product adoption is a two-stage process (Kalish 85) · 1 st stage: Awareness ·Get exposed to the product ·Become familiar with its features · 2 nd stage: Actual adoption ·Only if valuation outweighs price ·Only in this case the company gains real profit · The 2 nd stage is not captured in existing work 5/22/2021 29
Price Related (PR) Frame Neutral Rules in IC or LT Influenced Active · Three node states: Neutral, Influenced, and Active · Neutral Influenced: same as in LT or IC · Influenced Active: Only if the valuation is at least the quoted price · Only active nodes will propagate influence to inactive neighbors 5/22/2021 30
Pricing Strategies ·Binar. Y pri. Cing (BYC) ·PAnoramic Pricing (PAP) 5/22/2021 31
BIPMax Problem Definition Input Problem Select a set of seeds & determine a vector of quoted price, such that the is maximized under the PR Frame Output A directed graph representing a social network, with influence weights on edges 5/22/2021 32
A Restricted Special Case · : The uniform price for every non-seed · : production cost · max: Not monotone nondecreasing submodular! 5/22/2021 33
Nonmonotone Submodular Max 1 2 3
Feige et all, FOCS 2007 For nonnegative submodular maximization, • 1/3 -approximation (deterministic local search) • 2/5 -approximation (randomized) • ½-approximation (NP-hard) for certain submodular function
Lee et al. STOC 2009 For nonnegative submodular maximization, • With k matroid constraints, • With k knapsack constraints, • With k>2 knapsack constraints,
|S|<k • a knapsack constraint • Also a matroid constraint because {S | |S|<k } form a matroid.
Feldman et al. FOCS 2011 For nonnegative submodular maximization, • (1/e-o(1))-approximation for O(1) knapsack constraints.
Can “nonnegative” be removed? • No! • With possible negative, no good approximation exists in general. • This means that the profit maximization needs some new work.
Table of Contents 1. Submodular Maximization 2. Submodular Minimization Seed Selection Positive Influence Weighted Positive Influence Active Friending 3. Submodular Partition.
Active Friending
Linked. In • Do you receive invitations from Linked. In everyday? • Does Linked. In have the following function: Linked. In may suggest a list of invitations when you want to include a target person into friend list. 42
Problem Formulation s S R t Theorem 43
Size-constrained submodular Min
Approximation Z. Svitkina and L. Fleischer, Submodular approximation: Sampling-based algorithms and lower bounds, FOCS 2008.
Problem Formulation s S R t Theorem 46
Table of Contents 1. Submodular Maximization 2. Submodular Minimization 3. Submodular Partition. Community Partition
Influence-based Community Detection
Model-Based Detection Community Detection Accurate or not? Formulation (Model) Solve formulated problem 49
Idea • People in a same community share common interests in - clothes, music, beliefs, movies, food, etc. • Influence each other strongly. 50
Definition 51
5 2 3 4 1 52
5 2 3 4 1 53
Model 54
Lemma 1 Proof 55
56
< 57
Thanks end
- Slides: 58