Portrait of Wei Chen

Wei Chen

Principal Researcher

Selected Projects

Influence diffusion dynamics and influence maximization in social networks

– Monograph summarizing research results on modeling, optimization, and learning of information and influence diffusions: “Information and Influence Propagation in Social Networks, Morgan and Claypool, 2013“.

– Multi-round influence maximization [KDD’18]: We study the multi-round influence model where propagation from potentially different seeds occurs independently in different rounds, with the goal to collectively active nodes in social networks. This corresponds to multiple stages in a marketing campaign. We develop scalable algorithm for both non-adaptive and adaptive settings.

– Influence-based centrality [WWW’17]: We propose the study of influence-based network centrality and provide a comparative study on two centrality measures: single node influence (SNI) centrality and Shapley centrality. We provide axiomatic characterizations for both centralities as well as scalable centrality computation algorithms based on the reverse influence sampling approach.

– Robust influence maximization [KDD’16, arXiv:1601.06551]: We address the issue of robust optimization when parameters for the influence maximization are not accurate. We propose a new algorithm to achieve robust influence maximization with a guarantee, and  investigate sampling methods to effectively improve parameter accuracy for the robust influence maximization task.

– Influence diffusion model covering the spectrum from competition to complementarity [VLDB’16, arXiv:1507.00317]: We propose an extension to the classical independent cascade model to cover the entire spectrum from competition to complementarity for two propagating items in a social network. We study the properties of the model and two variants of optimization problems for the mutually complementary regime. The paper also proposes a general sandwich approximation and a general condition for applying reverse reachable set approach to any diffusion model.

– Amphibious influence maximization [EC’15, arXiv:1507.03328]: we propose the  model of combining traditional media marketing with viral marketing by selecting seed content providers to initiate cascades and selecting seed consumers to propagate the cascades in the social network. We prove hard-to-approximate result of the problem for general graphs, and propose an approximation algorithm for a certain class of restricted bipartite graph between content providers and consumers.

– Seed minimization with probabilistic coverage guarantee [KDD’14, arXiv: 1402.5516]: We look into the optimization problem of minimizing the seed set to ensure a probabilistic influence coverage guarantee. The main difficulty is that the objective function is not submodular. We provide theoretical analysis on the approximation guarantee of this problem and experimental study on the effectiveness of our algorithm.

– Active friending in social networks [KDD’13, arXiv: 1302.7025]: Active friending provides a new perspective on using social influence in social networks. Instead of finding seed nodes to maximize influence, it needs to find influence pathways to maximize influence. We provide efficient algorithms for finding intermediate nodes so as to increase the probability of a target node accepting the friending request.

– Scalable influence maximization: We study new algorithms and heuristics to improve the speed of finding influential nodes in a social network for viral marketing.

    • MixedGreedy and DegreeDiscount [KDD’09]: improve the greedy algorithm by edge sampling, and propose DegreeDiscount for uniform independent cascade model.
    • PMIA [KDD’10, DAMI’12]: scalable heuristic for the independent cascade model, using local tree structures to speed up influence computation for 3 orders of magnitude.
    • LDAG [ICDM’10, MSR-TR-2010-133]: scalable heuristic for the linear threshold model, using local directed acyclic graph structures to speed up influence computation for 3 orders of magnitude.
    • IRIE [ICDM’12, arXiv:1111.4795]: even faster scalable heuristic combining efficient influence ranking with influence estimation method, achieving up to 2 orders of magintude faster than PMIA in the independent cascade model.
    • Dataset used: [data for two collaboration graphs NetHEPT and NetPHY][DBLP graph data]
    • Code is available upon request by email

– Influence diffusion modeling and maximization for complex social interactions: We model complex social interactions including negative opinions due to product defects, competitive influence diffusion, and friend/foe relationships, and study influence maximization problem in these contexts.

  • IC-N model and MIA-N algorithm [SDM’11, MSR-TR-2010-137]: We extend the independent cascade model to incorporate the emergence and propagation of negative opinions due to product defects, and study the influence maximization problem in this new model.
  • Competitive influence diffusion CLT model and influence blocking maximization algorithm CLDAG in [SDM’12, arXiv:1110.4723]: We study the competitive linear threshold model, and propose efficient heuristic algorithm CLDAG for selecting the positive seed set that most effectively block the diffusion of negative influence.
  • Time-critical influence maximization with time-delayed diffusion process [AAAI’12, arXiv:1204.3074]: We study influence maximization in which diffusion on each step may be delayed, and the objective is to maximize influence spread within a certain deadline. Both IC and LT models are extended, and efficient algorithms are proposed and evaluated.
  • Influence diffusion dynamics and maximization in networks with friend and foe relationships [WSDM’13, arXiv:1111.4729]: We extend the voter model to signed networks representing both friend and foe relationships, and study the influence diffusion dynamics and influence maximization problem in this context.

– Participation maximization based on influence in online discussion forums [ICWSM’11, MSR-TR-2010-142]: We propose and study the participation maximization problem in online discussion forums, in which forum participation is determined by influence propagation through forum users.

Combinatorial online learning

– Combinatorial multi-armed bandit [ICML’13, JMLR’16, NIPS’17]: We provide a general stochastic framework that encompasses a large class of combinatorial multi-armed bandit problems, including many non-linear reward problems not considered before. We provide a CUCB learning algorithm with a tight analysis to show its regret bound. The framework can be applied to solve online learning problems with nonlinear reward functions such as probabilistic maximum coverage for online advertising, social influence maximization, as well as many optimization problems with linear reward functions. We further extend the model to cover probabilistically triggered arms, with applications to online influence maximization and combinatorial cascading bandits. With the latest improvement, we show that our algorithm could achieve good regret bounds for a large class of problems satisfying triggering probability modulated bounded smoothness conditions.

– Thompson sampling for CMAB [ICML’18]: We apply the Thompson sampling approach to combinatorial semi-bandit problems and provide theoretical regret bounds and proper conditions for the bounds to hold.

– CMAB with general reward functions [NIPS’16]: We address the combinatorial semi-bandit cases where the reward function not only depends on the means of the base arms, but on the entire distributions of the base arms. We design stochastic dominant confidence bound (SDCB) algorithm and show its regret bound, and apply it to K-Max problem and expected utility maximization problems.

– Contextual combinatorial cascading bandit [ICML’16]: We incorporate contextual information, together with position discounts and general reward functions to the combinatorial cascade bandit, which has applications in online advertising and online recommendations.

– Online greedy learning [NIPS’15]: We show how to turn an offline greedy algorithm into an online greedy learning algorithm, with competitive regret bound compared with the solution of the offline greedy algorithm.

– Combinatorial pure exploration [NIPS’14]: We study the problem of how to explore stochastic arms efficiently so that in the end we can output a subset of arms satisfying certain combinatorial constraint (k-set, matching, spanning tree, etc.) and maximizing the sum of expected rewards of these arms with high probability. We provide generic fix-confidence and fix-budget algorithms and prove a lower bound on sample complexity, which implies that our fix-confidence algorithm is tight (up to a logarithmic factor) for combinatorial constraints formed by bases of a matroid.

– Combinatorial partial monitoring [ICML’14, supplementary material]: We study the online learning problem where the optimization problem is combinatorial (e.g. matching) and thus the action space is exponentially large, and the feedback is limited. We provide efficient solutions to this problem, and demonstrate how to apply it to problems such as crowd sourcing task allocations.

Properties of small-world networks

– Navigability of small-world networks [WWW’15, arXiv:1411.4097]: We provide a game-theoretic formulation of navigable small-world networks to explain why real networks are often navigable. We provide strong theoretical and empirical results showing that navigable small-world network is the only stable equilibrium in the game. Our payoff function balances distance with relationship reciprocity, and this is the first work providing a surprising connection between relationship reciprocity and network navigability.

– Complex routing in small-world networks [COCOON’16, arXiv:1503.00448]: We define the decentralized routing problem of the complex contagion, which is the diffusion behavior in which a node is activated if and only if the number of its active friends is above a certain threshold greater than 1. We study the routing efficiency of complex contagions, and show that in Kleinberg’s small-world networks, the decentralized routing time is polynomial to the size of the graph in all parameter range, which means it cannot be efficient.

– The hyperbolicity of small-world and tree-like random graphs [ISAAC’12, arXiv:1201.1717]: There are empirical evidence that many real-world networks such as Internet or social networks are hyperbolic, and theoretical studies show that hyperbolic networks allow efficient decentralized routing behavior. We study the hyperbolicity of Kleinberg small-world random graphs and a family of tree-like random graphs. Our results show that Kleinberg small-world graphs are not strongly hyperbolic, which indicates that efficient decentralized routing does not imply graph hyperbolicity.

Networked game theory and economics

– Sybil-proof mechanisms in query incentive networks [EC’13, arXiv:1304.7432]: Query incentive networks investigate how to design incentives to allow queries to be propagated to a large fraction of networks so as to find the answers to the queries. We study sybil-proof mechanisms for query incentive networks, and show that a class of direct referral mechanisms is both cost effective and avoid users creating fake identities (sybils) in the network.

– Pricing and revenue maximization with social influence consideration [WINE’11, arXiv:1007.1501]: We study the revenue maximization problem in selling a digital product in a social network where social influence affects agents’ purchasing decisions.

– Game-theoretic approach to overlapping community detection [DMKD’10 special issue on ECML PKDD’10]: We propose a community formation game and use its equilibrium to detect overlapping communities in social networks.

– Bounded budget betweenness centrality game [ESA’09, MSR-TR-2008-167, MSR-TR-2009-78]: We study a strategic network formation game in which nodes strategically select connections to other nodes to maximize their betweenness centralities in the network.