DQ4FairIM: Fairness-aware Influence Maximization using Deep Reinforcement Learning

TL;DR

<3-5 sentence high-level summary> The paper tackles fairness in Influence Maximization by formulating the problem as a Markov Decision Process and introducing DQ4FairIM, a deep Q-learning framework that leverages Structure2Vec embeddings to learn seed-selection policies under a maximin fairness objective. By incorporating a fairness-aware reward (outreach plus phi times fairness), the method explicitly balances equitable influence across communities while maintaining competitive total outreach. Empirical results on synthetic and real networks show that DQ4FairIM achieves higher fairness compared to both fairness-agnostic and fairness-aware baselines and generalizes well to unseen and evolving graphs. The approach offers a scalable, generalizable solution for fair information diffusion with potential applications in public health, political mobilization, and social campaigns.

Abstract

The Influence Maximization (IM) problem aims to select a set of seed nodes within a given budget to maximize the spread of influence in a social network. However, real-world social networks have several structural inequalities, such as dominant majority groups and underrepresented minority groups. If these inequalities are not considered while designing IM algorithms, the outcomes might be biased, disproportionately benefiting majority groups while marginalizing minorities. In this work, we address this gap by designing a fairness-aware IM method using Reinforcement Learning (RL) that ensures equitable influence outreach across all communities, regardless of protected attributes. Fairness is incorporated using a maximin fairness objective, which prioritizes improving the outreach of the least-influenced group, pushing the solution toward an equitable influence distribution. We propose a novel fairness-aware deep RL method, called DQ4FairIM, that maximizes the expected number of influenced nodes by learning an RL policy. The learnt policy ensures that minority groups formulate the IM problem as a Markov Decision Process (MDP) and use deep Q-learning, combined with the Structure2Vec network embedding, earning together with Structure2Vec network embedding to solve the MDP. We perform extensive experiments on synthetic benchmarks and real-world networks to compare our method with fairness-agnostic and fairness-aware baselines. The results show that our method achieves a higher level of fairness while maintaining a better fairness-performance trade-off than baselines. Additionally, our approach learns effective seeding policies that generalize across problem instances without retraining, such as varying the network size or the number of seed nodes.

Figures (4)

• Figure 1: Graphical overview of the DQ4FairIM algorithm as described in Algorithm \ref{['DQ4FairIM']}. The process starts by randomly selecting a graph from the pool of graphs. This selected graph creates a new environment, and the agent then interacts with the environment and solves the MDP defined in Section \ref{['formulationusingRL']} for this specific graph. It chooses a new node (action) based on the $\epsilon$-greedy policy: it either selects a random node or a node with the highest $Q$-value. The reward it receives is based on both the expected influence outreach and the fairness measure. It picks a new graph at random once the terminal state is reached ($k$ seed nodes are selected) and a new episode begins. Along the way, the parameters of the neural network are updated with the samples stored in the Experience Replay Memory according to the mean squared error loss. First, the current state will be parameterized to an embedding space using the Structure2Vec mechanism. These embeddings will then be fed into a Multi-Layer Perceptron (MLP) to estimate the corresponding $Q$-values.
• Figure 2: Outreach and fairness for varying seed nodes on the HBA10k dataset.
• Figure 3: Generalizability results: Outreach and fairness on evolving graphs where the DQ4FairIM and PIANO models are trained on the HBA1k dataset and tested on bigger size networks, having 1000--3000 nodes. However, other baselines are executed on all new datasets to compute the outreach and fairness.
• Figure 4: (a)--(b): Outreach and fairness, and (c)--(d): Reward while training DQ4FairIM for different levels of $\phi$ on the HBA10k dataset.