Research map / Probabilistic and scientific modeling / Causal discovery

Unveiling Environmental Sensitivity of Individual Gains in Influence Maximization

NeurIPS 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Causal approaches to fairness · 2 citations (OpenAlex)

Read the paper Check your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CauIM, a causal-inference-based influence maximization framework on hypergraphs that maximizes the sum of dynamic individual treatment effects (ITE), with greedy (G-CauIM) and accelerated gradient-based (A-CauIM) variants.

Problem

Influence maximization where each node's individual gain (ITE) is unobservable for the counterfactual state and depends dynamically on neighbors' activation states, so standard constant-weight IM assumptions fail.

Approach

Model individual gains as individual treatment effects estimated from observational data with a hypergraph-convolution interference representation, and optimize the expected sum of activation-probability-weighted ITE under an SICP hypergraph diffusion, with a greedy approximation guarantee generalized to possibly negative ITE and a derivative-based fast seed selection.</808>

Key results

On GoodReads with K=15, G-CauIM reached 330.25 total ITE versus 297.56 for the traditional greedy baseline and 45.86 for random selection.
A-CauIM ran in 28s on GoodReads versus 1day for the baseline greedy, while reaching 302.17 total ITE.
On SD-100, A-CauIM achieved the highest total ITE (160.49) versus 138.91 for the baseline greedy.

Measured effects reported in the paper

What was changed	Where	Measured	Effect
Use G-CauIM instead of the traditional greedy Baseline	Influence maximization · GoodReads	sum of ITE: 297.56 -> 330.25	Helped
Use G-CauIM instead of the traditional greedy Baseline	Influence maximization · Contact	sum of ITE: 68.12 -> 69.53	Helped
Use G-CauIM instead of the traditional greedy Baseline	Influence maximization · Email-Eu	sum of ITE: 735.28 -> 804.28	Helped
Use G-CauIM instead of Random selection	Influence maximization · SD-100	sum of ITE: 145.97 -> 151.59	Helped
Use A-CauIM instead of the traditional greedy Baseline	Influence maximization · GoodReads	sum of ITE: 297.56 -> 302.17	Helped
Use A-CauIM instead of the traditional greedy Baseline	Influence maximization · Contact	sum of ITE: 68.12 -> 66.78	Hurt
Use A-CauIM instead of the traditional greedy Baseline	Influence maximization · Email-Eu	sum of ITE: 735.28 -> 802.41	Helped
Use A-CauIM instead of Random selection	Influence maximization · SD-100	sum of ITE: 145.97 -> 160.49	Helped
Use A-CauIM instead of G-CauIM	Influence maximization · GoodReads	runtime: 1day02h -> 28s	Helped
Use A-CauIM instead of G-CauIM	Influence maximization · Contact	runtime: 7h -> 115s	Helped
Use A-CauIM instead of G-CauIM	Influence maximization · Email-Eu	runtime: 23h -> 550s	Helped
Use A-CauIM instead of G-CauIM	Influence maximization · SD-100	runtime: 3h -> 53s	Helped

Surprising findings

Traditional greedy IM degenerates to near-random performance in most settings once dynamic ITE is considered.
On Contact, G-CauIM and A-CauIM gains over baseline are small (69.53 and 66.78 vs 68.12).

Limitations

Theoretical guarantee relies on a broad, model-free boundedness condition on reachable probabilities that is not verified in general.
Experiments are on only one synthetic and three real-world datasets, with simulated ITE and treatments for some settings.
Assumes a static hypergraph topology; dynamic graphs are deferred to future work.

Experimental setup

Datasets: GoodReads Contact Email-Eu SD-100 (synthetic)

Benchmarks: GoodReads Contact Email-Eu SD-100

Metrics: sum of ITE spread by selected seeds running time

Baselines: Greedy (traditional IM without parameter adjustment) Random seed selection

Builds on: Kempe influence maximization CELF hypergraph independent cascade (SICP) representation-based ITE estimation

Closest papers by meaning

Better Bounds on the Adaptivity Gap of Influence Maximization under Full-adoption Feedback (AAAI 2021 · similarity 0.88)
Graph Bayesian Optimization for Multiplex Influence Maximization (AAAI 2024 · similarity 0.87)
REM: A Scalable Reinforced Multi-Expert Framework for Multiplex Influence Maximization (AAAI 2025 · similarity 0.86)
Hyperparametric Robust and Dynamic Influence Maximization (AAAI 2025 · similarity 0.85)
DiffIM: Differentiable Influence Minimization with Surrogate Modeling and Continuous Relaxation (AAAI 2025 · similarity 0.85)
Fairness in Social Influence Maximization via Optimal Transport (NeurIPS 2024 · similarity 0.85)

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft