Fair Online Influence Maximization
Authors: Xiangqi Wang, Shaokun Zhang, Jose Efraim Aguilar Escamilla, Qingyun Wu, Xiangliang Zhang, Jian Kang, Huazheng Wang
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive empirical evaluations on three real-world networks demonstrate the efficacy of our proposed framework. 6 Empirical Results We assess FOIM with CMAB backbones under multiple fairness metrics, detailing datasets and bandit variants, then report results that confirm superior performance and show the algorithm accurately recovers edge weights key to optimal influence maximization. |
| Researcher Affiliation | Academia | Xiangqi Wang 1, Shaokun Zhang2, Jose Efraim Aguilar Escamilla3, Qingyun Wu2, Xiangliang Zhang1, Jian Kang4, Huazheng Wang 3 1University of Notre Dame 2Pennsylvania State University 3Oregon State University 4University of Rochester |
| Pseudocode | Yes | Algorithm 1: Fair Online Influence Maximization (FOIM) 1: Input Graph G, time budget T, seed nodes size K, reward function rµ(S), edge weight estimator of CMAB A(G, ˆµt, rt), fair oracle Fair ORACLE(G, ˆµt, K). 2: Initialize t 0 3: for t in {1,. . . , T} do 4: ˆµt A(G, ˆµt 1, rt 1) // estimate edge weights 5: St=Fair ORACLE(G, ˆµt, K) // obtain seed nodes 6: rt rµ(St) // obtain rewards 7: end for |
| Open Source Code | Yes | The code can be found at https://github.com/XiangqiWang77/FOIM |
| Open Datasets | Yes | We conduct our experiments on five real-world networks , namely NBA, German, Pokec-z, and bail, which are commonly used for fair graph learning (Dai & Wang, 2020; Ma et al., 2022), as well as You Tube. Detailed dataset statistics and descriptions are deferred in Table 1 and Appendix A.7. The code can be found at https://github.com/yushundong/Graph-Mining-Fairness-Data/tree/main/dataset/ |
| Dataset Splits | No | The paper does not explicitly provide training/test/validation dataset splits. It discusses online learning where models interact with the network over time, rather than using predefined static splits. The 'seed set volume' is mentioned, but this refers to a parameter of the influence maximization problem, not a data split for model evaluation. |
| Hardware Specification | Yes | The Youtube dataset (the largest dataset) is run on Google Cloud with N4 VM and Google Axion Processor including 32 v CPUs and 128 GB of DDR5 memory, while other datasets are tested on an AMD 5800H CPU. |
| Software Dependencies | No | The paper mentions various CMAB algorithms (CUCB, ϵ-greedy, IMLin UCB, IMFB) and offline oracles (multi-objective Frank-Wolfe, Mixed Integer Programming, degree discount). However, it does not provide specific version numbers for these software components or any programming languages/libraries used. |
| Experiment Setup | Yes | In all experiments, the dimensionalities for the susceptibility factor in IMLin UCB and the latent factors in IMFB are both set to 20. The seed set volume is set to K = 300 for Pokec-z and bail dataset, and K = 10 for NBA and K = 50 for German dataset respectively. For all datasets, we select the first 8 features for each dataset with the sensitive attribute included. Then we initialize the activation probability with the heuristics as presented in Table 1 between node features. |