TL;DR
This paper introduces "AugInf," a novel method for predicting social influence on social networks using Graph Neural Networks (GNNs). AugInf uniquely employs data augmentation not only during training but also during testing, significantly improving prediction accuracy, especially on smaller, sparser graphs.
What Problem Does This Paper Address?
Predicting social influence – understanding who is likely to sway others' opinions or actions – is a crucial task in social network analysis. While Graph Neural Networks (GNNs) have shown promise in learning from graph-structured data, their application to social influence prediction often relies on manual feature engineering and lacks robust data augmentation strategies. Existing methods primarily focus on training-time augmentation, leaving a gap in leveraging data augmentation at test-time to enhance GNN performance for this specific task. This paper aims to bridge this gap by proposing a method that integrates both train- and test-time augmentation within a GNN framework for social influence prediction.
Methods
The core of AugInf lies in its innovative use of data augmentation for GNNs. The method utilizes a Variational Graph Autoencoder (VGAE) to learn edge probabilities within a graph. These probabilities are then used to stochastically add new edges, creating augmented graphs. This augmentation process is applied in two phases:
- Train-time Augmentation: Multiple augmented versions of the training graphs are generated and used alongside the original graphs for joint training.
- Test-time Augmentation: For each test graph, several augmented versions are created. Predictions are made on each augmented version, and the final prediction is an average of these individual predictions.
AugInf employs a joint training approach that combines a Graph Autoencoder (GAE) for learning latent graph representations and a GNN prediction module (either a Graph Convolutional Network - GCN, or a Graph Attention Network - GAT) for social influence prediction. This joint training optimizes both the quality of latent representations and the downstream prediction task. Additionally, it leverages pretrained DeepWalk embeddings and influence features as inputs to the GNN.
Datasets
The authors evaluated AugInf on four real-world social network datasets:
- OAG (Open Academic Graph): Represents co-author networks, with citation behaviors as the influence action.
- Digg: A social news aggregation website, where voting behaviors are the influence actions.
- Twitter: A friendship network based on Higgs boson discovery tweets, with retweeting behaviors as the influence action.
- Weibo: A friendship network, with retweeting behaviors as the influence action.
These datasets were preprocessed into sub-networks using random walks, similar to previous work.
Key Insights and Ideas
- Test-time Augmentation is Beneficial for GNNs: The paper demonstrates that augmenting data not only during training but also during inference (test-time) can significantly improve GNN performance for social influence prediction.
- Joint Training Enhances Representations: Combining the GAE's objective of learning latent representations with the GNN's prediction objective leads to more effective features for influence prediction.
- GAT Outperforms GCN for Social Influence: The study highlights that Graph Attention Networks (GATs), which can weigh neighbor importance, perform better than Graph Convolutional Networks (GCNs) for social influence prediction, suggesting that the ability to capture nuanced relationships is crucial.
- Augmentation is Most Effective on Smaller Graphs: The benefits of AugInf are particularly pronounced on datasets with fewer edges, indicating that augmentation helps to compensate for limited structural information.
Key Results
AugInf-GAT achieved state-of-the-art performance on several datasets, outperforming baselines like DeepInf and PSCN in terms of AUC and F1 scores. For instance, on the Digg dataset, AugInf-GAT achieved an AUC of 0.9067 (±0.011) and an F1 score of 0.7385 (±0.017), showing a notable improvement over DeepInf-GAT (AUC: 0.8882, F1: 0.7052). The ablation studies confirmed that both train- and test-time augmentation, along with the jointly trained model, contribute to the overall performance gains. The paper also found that using four augmentations provided optimal performance on the Digg dataset, with diminishing returns beyond that.
Limitations and Future Work
- Dataset Size Dependency: The authors note that the performance improvement from augmentation is less significant on larger, denser graphs (like Twitter and Weibo) compared to sparser ones (like Digg). They suggest future work could explore removing edges as part of the augmentation process to further investigate this.
- GCN Limitations: The paper points out that GCN's unweighted aggregation mechanism is less suitable for social influence prediction, which benefits from considering neighbor importance. This reinforces the advantage of GATs.
- Hyperparameter Sensitivity: While the paper explores hyperparameters like the number of augmentations and the edge addition threshold, further research could delve deeper into their optimal settings across diverse graph structures.
Summary updated