Publications

2026

(#corresponding author, *equal contribution)

• [Mobihoc '26] Continuous Semantic Caching for Low-Cost LLM Serving
  Baran Atalar, Xutong Liu, Jinhang Zuo, Siwei Wang, Wei Chen, Carlee Joe-Wong.
  Accepted by the 27th International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing (Mobihoc), 2026 (Acceptance rate 23%)
  [PDF] [arXiv] TL;DR: The first theoretical framework for semantic LLM caching over a continuous query embedding space, combining dynamic \(\epsilon\)-net discretization with kernel ridge regression to give offline and low-switching online algorithms with sublinear regret and up to 73% lower suboptimality than discrete baselines.
  

• [EMNLP '26] LLM Cache Bandits with Query Heterogeneity for Cost-Efficient Inference
  Hantao Yang, Hong Xie, Xutong Liu, Defu Lian.
  Accepted by the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP Main Conference), 2026 (Acceptance rate 15.4%)
  [arXiv] TL;DR: Revisits the LLM cache bandit problem with heterogeneous query sizes, casting cache selection as a knapsack problem and giving an accumulation-based online algorithm with improved (and the first problem-dependent) regret bounds that cuts inference cost by about 12%.
  

• [UAI '26] FedSteer: Taming Extreme Gradient Staleness in Federated Learning with Corrective Projections and Caching
  Haoran Zhang, Cainã Figueiredo Pereira, Marie Siew, Xutong Liu, Carlee Joe-Wong, Rachid El-Azouzi.
  Accepted by the 42nd Conference on Uncertainty in Artificial Intelligence (UAI).
  [arXiv] TL;DR: FedSteer corrects the direction of stale updates from inactive federated-learning clients by re-applying their cached projection coordinates to an evolving low-dimensional gradient subspace, yielding an unbiased update that provably minimizes aggregation variance under skewed participation.
  

• [DEMO @ ICML ‘26] Offline Preference Learning with Clustering and Active Data-Augmentation
  Jingyuan Liu, Fatemeh Ghaffari, Xuchuang Wang, Xutong Liu#, Mohammad Hajiesmaili, Carlee Joe-Wong.
  Accepted by ICML 2026 Workshop on Decision-Making from Offline Datasets to Online Adaptation: Black-Box Optimization to Reinforcement Learning (DEMO @ ICML ‘26).
  [arXiv] TL;DR: Studies offline preference learning from pairwise feedback across users with different tastes, where Off-C²PL clusters users that can safely pool data and an active-data-augmentation extension targets underrepresented preference dimensions, both with suboptimality guarantees.
  

• [SIGKDD '26] Online Learning to Rank under Corruption: A Robust Cascading Bandits Approach
  Fatemeh Ghaffari, Siddarth Sitaraman, Xutong Liu#, Xuchuang Wang, Mohammad Hajiesmaili.
  Accepted by ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2026. (Acceptance rate 18.5%).
  [arXiv] TL;DR: M²UCB-V makes cascading bandits for online learning to rank robust to click fraud and other corrupted feedback by combining a calibrated mean-of-medians estimator with a variance-aware UCB and model selection over unknown corruption levels, attaining \(O(KC + K\log T/\Delta)\) regret that is optimal without corruption and grows only additively with the total corruption \(C\).
  

• [INFOCOM '26, 🏆Top 10 papers in INFOCOM '26] Constraint-Aware Combinatorial Bandits: Theoretical Foundations and Network Applications
  Xiangxiang Dai, Jin Li, Xutong Liu#, Anqi Yu, John C.S. Lui.
  IEEE International Conference on Computer Communications (INFOCOM), 2026 (10/1740 = 0.6%).
  Top 10 papers in INFOCOM 2026 and selected as fast-track for IEEE Transactions on Networking.
  [arXiv version] TL;DR: A general framework for combinatorial bandits under cost or budget constraints with versatile reward models (e.g., any-win and all-win), solving the NP-hard constrained selection via relaxation and discretization-rounding with regret and violation guarantees, applied to networked settings such as cost-constrained multi-LLM serving.
  

• [INFOCOM '26] Faster, Smaller, and Smarter: Task-Aware Expert Merging for Online MoE Inference
  Ziyi Han, Xutong Liu#, Ruiting Zhou#, Xiangxiang Dai, John C.S. Lui.
  IEEE International Conference on Computer Communications (INFOCOM), 2026 (Acceptance rate: 18.9%).
  [arXiv] [slides] TL;DR: Tanbr merges the experts of a pre-trained MoE into a single task-aware expert using a binary-tree search over continuous merging weights plus a neural bandit, achieving \(O(\sqrt{T}\log T)\) regret while cutting inference latency by at least 45% and memory by up to 25% with minimal accuracy loss.
  

• [INFOCOM '26] Semantic Caching for Low-Cost LLM Serving: From Offline Learning to Online Adaptation
  Xutong Liu*, Baran Atalar*, Xiangxiang Dai, Jinhang Zuo, Siwei Wang, John C.S. Lui, Wei Chen, Carlee Joe-Wong.
  IEEE International Conference on Computer Communications (INFOCOM), 2026 (Acceptance rate: 18.9%).
  [arXiv] [slides] TL;DR: A principled learning framework for semantic cache eviction in LLM serving under unknown query arrivals and serving costs, with approximation, suboptimality, and regret guarantees (including low cache switching) across the oracle, offline, and online-adaptive settings.
  

• [Preprint ‘26] Steering Frozen LLMs: Adaptive Social Alignment via Online Prompt Routing
  Zeyu Zhang, Xiangxiang Dai, Ziyi Han, Xutong Liu, John C.S. Lui.
  [arXiv] TL;DR: CCLUB steers a frozen LLM at inference time by routing queries to system prompts with a consensus-clustering LinUCB bandit that pools data only where utility and safety similarity agree, achieving sublinear regret and about 11% higher cumulative reward than strong baselines.
  

• [Preprint ‘26] Unlearning Offline Stochastic Multi-Armed Bandits
  Zichun Ye, Runqi Wang, Xuchuang Wang, Xutong Liu, Shuai Li, Mohammad Hajiesmaili.
  [arXiv] TL;DR: The first study of machine unlearning for offline multi-armed bandits, formalizing the privacy constraint and giving adaptive algorithms that switch between Gaussian-mechanism and rollback baselines across single- and multi-source deletion settings, with matching lower bounds.
  

• [TON '26] Corruption-Resilient Combinatorial Bandit Learning for Heterogeneous Network Systems
  Junkai Wang, Xutong Liu#, Jinhang Zuo, Yuedong Xu#, John C.S. Lui.
  IEEE/ACM Transactions on Networking (TON), 2026.
  TL;DR: Journal extension of our INFOCOM ‘25 paper on contextual combinatorial bandits with adversarially corrupted base-arm feedback, giving tight regret bounds under \(L_1\) and \(L_\infty\) corruption measures, lower bounds, and probabilistically triggered arms for heterogeneous network systems.

• [TMC '26] Online Outcome-Driven Configuration Selection with Adaptive Grouping for Live Video Streams
  Xiangxiang Dai, Zeyu Zhang, Ziyi Han, Xutong Liu, Peng Yang, John C.S. Lui.
  IEEE Transactions on Mobile Computing (TMC), 2026.
  TL;DR: Selects configurations for live video streams online by directly optimizing the downstream task outcome rather than proxy quality metrics, adaptively grouping streams with similar behavior so their feedback is shared to accelerate learning.

• [AIJ '26] Exploring Multi-Layered Networks through Random Walks: Bridging Offline Optimization and Online Learning
  Xiangxiang Dai, Xutong Liu#, Jinhang Zuo, Xiaowei Chen, Wei Chen, John C.S. Lui.
  The Journal of Artificial Intelligence (AIJ), 2026.
  TL;DR: Journal version of our ICML ‘21 long-oral MuLaNE paper on allocating random-walk budgets across network layers to maximize distinct nodes visited, with approximation guarantees offline and CMAB-based regret guarantees online.

• [TON '26] Combinatorial Logistic Online Learning and Its Applications in Nonlinear Networked Systems
  Xutong Liu, Xiangxiang Dai, Xuchuang Wang, Carlee Joe-Wong, Mohammad Hajiesmaili, John C.S. Lui.
  IEEE/ACM Transactions on Networking (TON), 2026.
  TL;DR: Journal version of our SIGMETRICS ‘25 Combinatorial Logistic Bandits paper for nonlinear networked systems, with variance-adaptive algorithms that remove the potentially exponentially large nonlinearity factor \(\kappa\) from the leading regret term.

• [AAAI '26] Online Multi-LLM Selection via Contextual Bandits under Unstructured Context Evolution
  Manhin Poon, XiangXiang Dai, Xutong Liu, Fang Kong, John C.S. Lui, Jinhang Zuo.
  40th Annual AAAI Conference on Artificial Intelligence (AAAI), 2026 (Acceptance rate: 17.6%).
  [arXiv] TL;DR: The first contextual-bandit formulation of sequential LLM selection under unstructured, black-box prompt evolution, with a LinUCB-based algorithm (plus budget-aware and position-aware variants) that attains sublinear myopic regret without forecasting future contexts.
  

• [Preprint] HiLoRA: Adaptive Hierarchical LoRA Routing for Training-Free Domain Generalization
  Ziyi Han, Huanyu Wang, Zeyu Zhang, Xiangxiang Dai, Xutong Liu, John C.S. Lui.
  [arXiv] TL;DR: HiLoRA is a training-free hierarchical router over LoRA pools that selects LoRAs and their rank-one components at the sequence and token levels via Gaussian likelihoods, provably identifying relevant LoRAs and improving domain-generalization accuracy by up to 55%.
  

• [SIGMETRICS '26] Heterogeneous Multi-agent Multi-armed Bandits on Stochastic Block Models
  Mengfan Xu, Liren Shan, Fatemeh Ghaffari, Xuchuang Wang, Xutong Liu, and Mohammad Hajiesmaili.
  ACM International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS), 2026 (Acceptance rate 20.7%).
  [PDF] TL;DR: Studies multi-agent bandits where a stochastic-block-model cluster structure governs both the communication topology and the reward heterogeneity across agents, giving algorithms with system-wide regret guarantees for both known and unknown clusters.
  

2025

• [NeurIPS '25] Learning Across the Gap: Hybrid Multi-armed Bandits with Heterogeneous Offline and Online Data
  Qijia He, Minghan Wang, Xutong Liu, Zhiyong Wang, Fang Kong.
  The Thirty-ninth Conference on Neural Information Processing Systems (NeurIPS), 2025. (Acceptance rate 24.5%).
  [Openreview] TL;DR: A general hybrid bandit framework that exploits offline data with a different feedback structure (e.g., absolute rewards offline vs. dueling feedback online) and distribution shift, with regret guarantees quantifying how much heterogeneous, possibly biased, offline data helps.
  

• [Preprint] Offline Clustering of Linear Bandits: Unlocking the Power of Clusters in Data-Limited Environments
  Jingyuan Liu, Zeyu Zhang, Xuchuang Wang, Xutong Liu, John C.S. Lui, Mohammad Hajiesmaili, Carlee Joe-Wong.
  [arXiv] TL;DR: Brings clustering of bandits to the offline setting by deciding from a fixed dataset whether users can be confidently pooled, where Off-C²LUB works well for any amount of user data and Off-CLUB nearly matches a new lower bound when data are sufficient.
  

• [SIGMETRICS '25] Asynchronous Multi-Agent Bandits: Fully Distributed vs. Leader-Coordinated Algorithms
  Xuchuang Wang, Yu-Zhen Janice Chen, Lin Yang, Xutong Liu, Mohammad Hajiesmaili, Don Towsley, and John C.S. Lui.
  ACM International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS), 2025 (Acceptance rate 15.7%).
  TL;DR: For cooperative bandits where an unknown, time-varying subset of agents is active each round, both a fully distributed algorithm (successive elimination with accuracy-adaptive, on-demand communication) and a leader-coordinated one (a low-switching Tsallis-INF leader-selection scheme) attain near-optimal regret and communication, almost matching the synchronous case.

• [SIGMETRICS '25, 🏆Best Paper Runner-Up] Combinatorial Logistic Bandits
  Xutong Liu, Xiangxiang Dai, Xuchuang Wang, Mohammad Hajiesmaili, John C.S. Lui.
  ACM International Conference on Measurement and Modeling of Computer Systems (SIGMETRICS), 2025.
  Selected as Best Paper Runner-Up at SIGMETRICS 2025.
  [arXiv] [code] [slide] TL;DR: Introduces combinatorial logistic bandits (CLogB), where base-arm outcomes are binary with a logistic link to contexts, and gives variance-adaptive algorithms that remove the potentially exponentially large nonlinearity factor \(\kappa\) from the leading regret term.
  

• [SIGKDD '25] A Unified Online-Offline Framework for Co-Branding Campaign Recommendations
  Xiangxiang Dai, Xiaowei Sun, Jinhang Zuo, Xutong Liu#, John C.S. Lui.
  ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2025. (Acceptance rate 18.4%).
  [arXiv] [code] TL;DR: The first systematic study of co-branding partner recommendation, modeling brands as a bipartite graph and combining CUCB-style online learning that avoids costly redundant exploration with offline submodular budget allocation, improving over baselines by up to 39%.
  

• [ICML '25] Fusing Reward and Dueling Feedback in Stochastic Bandits
  Xuchuang Wang, Qirun Zeng, Jinhang Zuo, Xutong Liu#, Mohammad Hajiesmaili, John C.S. Lui, Adam Wierman
  Forty-second International Conference on Machine Learning (ICML), 2025. (Acceptance rate 26.3%).
  [arXiv] TL;DR: For bandits receiving both absolute rewards and pairwise comparisons each round, a lower bound shows an efficient learner pays only the smaller per-arm regret of the two feedback types, and the proposed decomposition-fusion algorithm nearly matches it.
  

• [ICML '25] Offline Learning for Combinatorial Multi-armed Bandits
  Xutong Liu, Xiangxiang Dai, Jinhang Zuo, Siwei Wang, Xuchuang Wang, Carlee Joe-Wong, John C.S. Lui, and Wei Chen.
  Forty-second International Conference on Machine Learning (ICML), 2025. (Acceptance rate 26.3%)
  [arXiv] [slide] TL;DR: Off-CMAB is the first offline learning framework for combinatorial bandits, with the pessimism-based CLCB algorithm achieving a near-optimal suboptimality gap under two new data-coverage conditions and applications to learning to rank, LLM caching, and influence maximization.
  

• [ACM JCSS] Dynamic Incentive Allocation for City-scale Deep Decarbonization
  Anupama Sitaraman, Adam Lechowicz, Noman Bashir, Xutong Liu, Mohammad Hajiesmaili, and Prashant Shenoy.
  ACM Journal on Computing and Sustainable Societies (JCSS), 2025.
  [arXiv] TL;DR: Dynamically allocates an incentive budget across households to directly maximize city-wide carbon reduction, learning unknown household adoption willingness with an offline contextual bandit and delivering up to 32% more carbon reduction than status-quo incentives on real data.
  

• [ICLR '25] Stochastic Bandits Robust to Adversarial Attacks
  Xuchuang Wang, Maoli Liu, Jinhang Zuo, Xutong Liu#, John C.S. Lui, and Mohammad Hajiesmaili.
  The 13th International Conference on Learning Representations (ICLR), 2025.
  [arXiv] TL;DR: For bandits where an attacker observes the pulled arm and alters the reward under a known or unknown attack budget \(C\), gives algorithms with regret depending on \(C\) either additively, \(O((K/\Delta)\log T + KC)\), or multiplicatively, \(\tilde{O}(\sqrt{KTC})\).
  

• [INFOCOM '25] Learning Best Paths in Quantum Networks
  Xuchuang Wang, Maoli Liu#, Xutong Liu#, Zhuohua Li, Mohammad Hajiesmaili, John C.S. Lui, Don Towsley.
  IEEE International Conference on Computer Communications (INFOCOM), 2025. (Acceptance rate 18.7%).
  [PDF] TL;DR: Casts best-path selection in quantum networks as an online learning problem, with BeQuP-Link and BeQuP-Path identifying the best path under link-level and path-level benchmarking feedback, respectively, with theoretical guarantees.
  

• [INFOCOM '25] Robust Contextual Combinatorial Multi-Armed Bandits for Unreliable Network Systems
  Junkai Wang, Xutong Liu#, Jinhang Zuo, Yuedong Xu#.
  IEEE International Conference on Computer Communications (INFOCOM), 2025. (Acceptance rate 18.7%).
  [PDF] TL;DR: Studies contextual combinatorial bandits with adversarially corrupted base-arm feedback, giving tight regret upper bounds under \(L_1\) and \(L_\infty\) corruption measures, the first lower bounds, and an extension to probabilistically triggered arms, applied to federated learning, content delivery, and 360° VR streaming.
  

• [TON] Variance-Aware Bandit Framework for Dynamic Probabilistic Maximum Coverage Problem with Triggered or Self-Reliant Arms
  Xiangxiang Dai, Xutong Liu#, Jinhang Zuo, Hong Xie, Carlee Joe-Wong, John C.S. Lui.
  IEEE/ACM Transactions on Networking (TON), 2024.
  [PDF] TL;DR: Journal extension of our INFOCOM ‘23 work that broadens probabilistic maximum coverage bandits to dynamic settings with probabilistically triggered or self-reliant arms, using variance-adaptive algorithms to tighten regret over prior CMAB results.
  

2024

• [Preprint ‘24] Cost-Effective Online Multi-LLM Selection with Versatile Reward Models
  Xiangxiang Dai, Jin Li, Xutong Liu, Anqi Yu, John C.S. Lui.
  [arXiv] TL;DR: C2MAB-V selects combinations of LLMs per task under cost budgets with versatile reward models, solving the NP-hard selection via relaxation and discretization-rounding with regret and violation guarantees (superseded by the INFOCOM ‘26 paper above).
  

• [ACM MM '24] AxiomVision: Accuracy-Guaranteed Adaptive Visual Model Selection for Perspective-Aware Video Analytics
  Xiangxiang Dai, Zeyu Zhang, Peng Yang, Yuedong Xu, Xutong Liu#, John C.S. Lui.
  ACM Multimedia (MM), 2024 (1149/4385 = 26.2%).
  [Link] [arXiv], [Poster], [Code], [ACM showcase on Kudos] TL;DR: An accuracy-guaranteed edge-cloud framework that selects the best visual model per video-analytics task via continual online learning, accounting for the camera’s perspective and using topology-driven grouping to accelerate selection.
  

• [TKDE] Conversational Recommendation with Online Learning and Clustering on Misspecified Users
  Xiangxiang Dai*, Zhiyong Wang*, Jize Xie, Xutong Liu, John C.S. Lui.
  IEEE Transactions on Knowledge and Data Engineering (TKDE), 2024.
  [Link] TL;DR: Combines conversational key-term feedback with robust clustering of users whose linear preference models are only approximately shared, giving regret guarantees despite model misspecification.
  

• [ICML '24] Combinatorial Multivariant Multi-Armed Bandits with Applications to Episodic Reinforcement Learning and Beyond
  Xutong Liu, Siwei Wang, Jinhang Zuo, Han Zhong, Xuchuang Wang, Zhiyong Wang, Shuai Li, Mohammad Hajiesmaili, John C.S. Lui, and Wei Chen.
  Forty-first International Conference on Machine Learning (ICML), 2024 (Acceptance rate 27.5%).
  [arXiv] [poster] [slides] TL;DR: Generalizes combinatorial bandits to multivariate base-arm outcomes with a new 1-norm triggering-probability-modulated smoothness condition, yielding batch-size-independent regret and recovering episodic RL as a special case with new or improved bounds.
  

• [ICML '24] Quantum Algorithm for Online Exp-concave Optimization
  Jianhao He, Chengchang Liu, Xutong Liu, Lvzhou Li, John C.S. Lui.
  Forty-first International Conference on Machine Learning (ICML), 2024 (Acceptance rate 27.5%).
  [arXiv] TL;DR: Quantum online quasi-Newton methods for zeroth-order online exp-concave optimization achieve \(O(n\log T)\) regret with \(O(1)\) queries per round, beating the optimal classical rate by a factor of \(T^{2/3}\).
  

• [AAAI '24] Federated Contextual Cascading Bandits with Asynchronous Communication and Heterogeneous Users
  Hantao Yang, Xutong Liu#, Zhiyong Wang, Hong Xie, John C.S. Lui, Defu Lian, and Enhong Chen.
  The 38th AAAI Conference on Artificial Intelligence (AAAI), 2024. (Acceptance rate 23.75%).
  [arXiv] TL;DR: Federated online learning to rank with asynchronous client-server communication and users in latent preference clusters, where a UCB-type algorithm attains regret on par with the synchronous setting using only logarithmic communication.
  

• [INFOCOM '24] Learning Context-Aware Probabilistic Maximum Coverage Bandits: A Variance-Adaptive Approach
  Xutong Liu, Jinhang Zuo, Junkai Wang, Zhiyong Wang, Yuedong Xu, and John C.S. Lui.
  IEEE International Conference on Computer Communications (INFOCOM), 2024. (Acceptance rate 19.6%).
  [pdf] [slides] TL;DR: Contextual probabilistic maximum coverage bandits whose activation probabilities depend on contexts, with variance-adaptive confidence bounds giving regret that scales with the actual outcome variance rather than the worst case.
  

2023

• [NeurIPS '23] Online Clustering of Bandits with Misspecified User Models
  Zhiyong Wang, Jize Xie, Xutong Liu, Shuai Li, John C.S. Lui.
  The Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. (Acceptance rate 26.1%).
  [arXiv] TL;DR: Relaxes the assumption that clustered users share identical linear preference models by allowing bounded misspecification, with RCLUMB and RSCLUMB achieving regret that matches the lower bound up to log factors asymptotically in \(T\).
  

• [UAI '23] Exploration for Free: How Does Reward Heterogeneity Improve Regret in Cooperative Multi-agent Bandits?
  Xuchuang Wang, Lin Yang, Yu-zhen Janice Chen, Xutong Liu, Mohammad Hajiesmaili, Don Towsley, John C.S. Lui.
  The 39th Conference on Uncertainty in Artificial Intelligence (UAI), 2023.
  TL;DR: Shows that reward heterogeneity in cooperative multi-agent bandits provides “free exploration”, since agents’ differing locally optimal arms let exploitation double as exploration, and quantifies the resulting regret improvement.

• [ICML '23] Contextual Combinatorial Bandits with Probabilistically Triggered Arms
  Xutong Liu, Jinhang Zuo, Siwei Wang, John C.S. Lui, Mohammad Hajiesmaili, Adam Wierman, Wei Chen.
  The 40th International Conference on Machine Learning (ICML), 2023. (Acceptance rate 27.9%).
  [arXiv] [slides] TL;DR: Contextual combinatorial bandits with probabilistically triggered arms, where C²-UCB-T and the variance-adaptive VAC²-UCB use new triggering-probability-modulated smoothness conditions to remove the batch-size dependence of prior bounds for cascading bandits and influence maximization.
  

• [AISTATS '23] On-Demand Communication for Asynchronous Multi-Agent Bandits
  Yu-Zhen Janice Chen, Lin Yang, Xuchuang Wang, Xutong Liu, Mohammad Hajiesmaili, John C.S. Lui, Don Towsley.
  The 26th International Conference on Artificial Intelligence and Statistics (AISTATS), 2023.
  TL;DR: ODC is an on-demand communication protocol for asynchronous cooperative bandits that shares observations only when enough new information has accumulated, preserving near-optimal regret with near-constant communication.

• [ICLR '23] Achieve Near-Optimal Individual Regret and Low Communications in Multi-Agent Bandits
  Xuchuang Wang, Lin Yang, Yu-Zhen Janice Chen, Xutong Liu, Mohammad Hajiesmaili, Don Towsley, John C.S. Lui.
  The 11th International Conference on Learning Representations (ICLR), 2023.
  TL;DR: UCB-TCOM attains near-optimal regret for both the group and the worst individual agent in cooperative bandits, which matters when one “bad” agent can wreck the system, while using only \(O(\log\log T)\) communication rounds.

• [INFOCOM '23] Variance-Adaptive Algorithm for Probabilistic Maximum Coverage Bandits with General Feedback
  Xutong Liu*, Jinhang Zuo*, Hong Xie, Carlee Joe-Wong, John C.S. Lui.
  IEEE International Conference on Computer Communications (INFOCOM), 2023. (Acceptance rate 19.2%).
  [pdf] [slides] TL;DR: A variance-adaptive algorithm for probabilistic maximum coverage bandits that exploits the low variance of Bernoulli coverage outcomes, improving the leading regret term by a factor of the batch size under feedback more general than semi-bandit.
  

• [AAAI '23] Efficient Explorative Key-term Selection Strategies for Conversational Contextual Bandits
  Zhiyong Wang, Xutong Liu, Shuai Li, John C.S. Lui.
  Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), 2023. (Acceptance rate 19.6%).
  [slides] [poster] TL;DR: Proposes explorative key-term selection for conversational contextual bandits, with ConLinUCB variants that pick key terms to maximally reduce preference uncertainty, giving tighter regret and much faster learning.
  

2022

• [NeurIPS '22] Batch-Size Independent Regret Bounds for Combinatorial Semi-Bandits with Probabilistically Triggered Arms or Independent Arms
  Xutong Liu, Jinhang Zuo, Siwei Wang, Carlee Joe-Wong, John C.S. Lui, Wei Chen.
  Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. (Acceptance rate 25.6%).
  [arXiv] [paper] [slides] [poster] TL;DR: A new triggering-probability and variance modulated (TPVM) smoothness condition with the BCUCB-T algorithm (and SESCB for independent arms) removes the batch-size factor \(K\) from regret bounds for cascading bandits, influence maximization, and probabilistic maximum coverage.
  

• [UAI '22] Federated Online Clustering of Bandits
  Xutong Liu, Haoru Zhao, Tong Yu, Shuai Li, John C.S. Lui.
  The 38th Conference on Uncertainty in Artificial Intelligence (UAI), 2022. (Acceptance rate 32%).
  [paper] [arXiv] [slides] [poster] [code] TL;DR: FCLUB_DC performs federated clustering of bandits across clients without sharing raw data, using a local-then-global clustering scheme and delayed communication to achieve near-optimal regret with low communication.
  

• [AISTATS '22] Online Competitive Influence Maximization
  Jinhang Zuo, Xutong Liu, Carlee Joe-Wong, John C.S. Lui, Wei Chen.
  The 25th International Conference on Artificial Intelligence and Statistics (AISTATS 2022), (Acceptance rate 29%)
  [arXiv] TL;DR: Studies online influence maximization against a competing campaign spreading in the same network, where the reward is neither monotone nor submodular in general, introducing a competitive triggering model and CMAB algorithms with regret guarantees.
  

• [IEEE TMC] Learning with Guarantee via Constrained Multi-armed Bandit: Theory and Network Applications
  Kechao Cai, Xutong Liu, Yuzhen Janice Chen, and John C.S. Lui.
  IEEE Transactions on Mobile Computing (IEEE TMC), 2022.
  DOI: https://doi.org/10.1109/TMC.2022.3173792 TL;DR: Journal extension of our INFOCOM ‘18 work on stochastic constrained bandits with time-varying multi-level rewards, giving algorithms with sublinear regret and bounded constraint violation for network applications.
  

2021 and before

• [ICML '21, 🎤️Long Oral] Multi-layered Network Exploration via Random Walks: From Offline Optimization to Online Learning
  Xutong Liu, Jinhang Zuo, Xiaowei Chen, Wei Chen, John C.S. Lui.
  The 38th International Conference on Machine Learning (ICML), Long Oral, 2021. (Acceptance rate 3%).
  [paper] [arXiv] [slides] [poster] [video] TL;DR: Formulates multi-layered network exploration, allocating random-walk budgets across layers to maximize the number of distinct nodes visited, with offline approximation guarantees and CMAB-based online learning algorithms with regret bounds.
  

• [NSJ] Learning to Count: a Deep Learning Framework for Graphlet Count Estimation
  Xutong Liu, Yu-Zhen Chen, John C.S. Lui., Konstantin Avrachenkov.
  Network Science Journal.
  DOI: https://doi.org/10.1017/nws.2020.35. TL;DR: A deep-learning framework that predicts graphlet counts from graph representations instead of costly per-query sampling, extending our COMPLEX NETWORKS ‘18 CNN approach with fast and accurate estimates.
  

• [INFOCOM '18] An Online Learning Approach to Network Application Optimization with Guarantee
  Kechao Cai, Xutong Liu, Yuzhen Janice Chen, and John C.S. Lui.
  IEEE International Conference on Computer Communications (INFOCOM), 2018.
  TL;DR: Formulates the first stochastic constrained multi-armed bandit with time-varying multi-level rewards for network application optimization, with algorithms achieving sublinear regret and bounded constraint violation.

• [CNA '18] Graphlet Count Estimation via Convolutional Neural Networks
  Xutong Liu, Yu-Zhen Chen, John C.S. Lui, Konstantin Avrachenkov.
  COMPLEX NETWORKS, 2018.
  [PDF] TL;DR: Estimates graphlet counts with convolutional neural networks by turning sampled subgraphs into image-like matrices, trading offline training for fast and accurate online estimation.