Source-linked AI summary
DeepInf: Social Influence Prediction with Deep Learning
Jiezhong Qiu, Jian Tang, Hao Ma, Yuxiao Dong, Kuansan Wang, Jie Tang
TL;DR
Social influence prediction requires estimating each user’s future action, but hand-crafted features depend on domain expertise and generalize poorly across domains. DeepInf learns latent representations from a user’s local network using network embedding, graph convolution, and attention, and its experiments across four networks report significantly better performance than hand-crafted baselines.
Problem
User-level social influence prediction is difficult because hand-crafted user- and network-specific features require domain knowledge and are difficult to generalize across domains.
Method
DeepInf is an end-to-end framework that samples a user’s local neighbors and combines network embedding, graph convolution, and graph attention to learn predictive latent representations.
Results
DeepInf significantly outperforms baselines with hand-crafted features in predicting social influence locality across OAG, Digg, Twitter, and Weibo.
Takeaways & Limitations
The results support representation learning as a promising approach for social influence analysis and social and information network mining.
Takeaways & Limitations
The neighbor-sampling procedure uses random walks with restart without side information and is loosely coupled with the neural network model.
Abstract
from arXiv · showhide
Social and information networking activities such as on Facebook, Twitter, WeChat, and Weibo have become an indispensable part of our everyday life, where we can easily access friends' behaviors and are in turn influenced by them. Consequently, an effective social influence prediction for each user is critical for a variety of applications such as online recommendation and advertising. Conventional social influence prediction approaches typically design various hand-crafted rules to extract user- and network-specific features. However, their effectiveness heavily relies on the knowledge of domain experts. As a result, it is usually difficult to generalize them into different domains. Inspired by the recent success of deep neural networks in a wide range of computing applications, we design an end-to-end framework, DeepInf, to learn users' latent feature representation for predicting social influence. In general, DeepInf takes a user's local network as the input to a graph neural network for learning her latent social representation. We design strategies to incorporate both network structures and user-specific features into convolutional neural and attention networks. Extensive experiments on Open Academic Graph, Twitter, Weibo, and Digg, representing different types of social and information networks, demonstrate that the proposed end-to-end model, DeepInf, significantly outperforms traditional feature engineering-based approaches, suggesting the effectiveness of representation learning for social applications.
1 INTRODUCTION
The paper targets user-level social influence prediction, where a user’s future action is inferred from nearby users’ actions and local network structure. DeepInf replaces domain-specific feature engineering with an end-to-end representation-learning framework evaluated across four social and information networks.
- Prior work: Existing approaches mainly predict global or aggregated influence patterns, such as cascade size within a time frame.
- Problem motivation: User-level prediction estimates whether a user will act from her near neighbors’ observed action statuses and local structural information.The motivating example asks whether a user will buy a product after some friends have bought it.
- Problem motivation: Hand-crafted user- and network-specific features require extensive domain knowledge and are difficult to generalize across domains.
- Approach: DeepInf integrates network embedding, graph convolution, and graph attention to discover predictive signals from a sampled local network.The local network is obtained by sampling neighbors through random walks with restart.
- Evaluation: Experiments on Open Academic Graph, Digg, Twitter, and Weibo compare DeepInf with conventional hand-crafted-feature methods and a graph classification model.The reported results suggest improved prediction performance and the promise of representation learning for social and information network mining.
2 PROBLEM FORMULATION
The paper defines social influence locality as predicting a user’s future action from her near-neighbor actions and induced local network. It formalizes this task as binary graph classification under a closed-world assumption about influence.
- Definitions: An r-neighbor is a user within shortest-path distance r, and the r-ego network is the subnetwork induced by those r-neighbors.
- Definitions: A social action is represented by a binary status, where 1 means the user has performed the action by time t and 0 means she has not.Examples include retweets and citation actions.
- Problem formulation: Social influence locality assumes users’ decisions are influenced only by near neighbors, excluding external sources from the model.
- Problem formulation: The task quantifies the probability of a user’s future action conditioned on her r-ego network and near-neighbor action states.The prediction concerns activation after a specified interval Δt.
- Problem formulation: Each instance contains a user, action, timestamp, local network, observed neighbor statuses, and the user’s later status, forming a binary graph classification problem.The model parameters are learned by minimizing a negative log-likelihood objective.
- Assumptions: The interval Δt is assumed sufficiently large to predict the ego user’s status at the end of the observation window.
3 MODEL FRAMEWORK
DeepInf samples fixed-size local subnetworks and processes them with an end-to-end neural model that combines network embeddings, user features, and graph convolution or attention. The model predicts the ego user's future action status using a hidden representation learned from local structure and action states.
- 3.1 Sampling Near Neighbors: DeepInf samples a fixed-size sub-network from each user’s r-ego network because ego networks vary in size and can be very large.Random walk with restart starts from the ego or a randomly selected active neighbor and collects n vertices.
- 3.2 Neural Network Model: The sampled network and sampled neighbors’ action statuses are fed to a neural model whose output predicts the ego user’s action status after a time interval.Training compares the model output with ground truth using negative log-likelihood loss.
- 3.2 Neural Network Model: The architecture comprises network embedding, instance normalization, a formal input layer, graph convolutional or attention layers, and an output layer.The embedding layer maps each user to a D-dimensional representation, while the input layer adds action, ego, and customized vertex features.
- 3.2 Neural Network Model: Instance normalization gives each instance’s user embeddings zero mean and unit variance, encouraging downstream models to use relative rather than absolute latent positions.The paper reports that instance normalization can help avoid overfitting during training.
- 3.2 Neural Network Model: GCN uses a static normalized adjacency matrix, whereas GAT computes attention only across graph edges and self-loops before applying graph propagation.GCN adds self-loops and uses the adjacency and degree matrices; multi-head attention runs independent attention mechanisms in parallel.
4 EXPERIMENT SETUP
The experiments evaluate DeepInf on four social and information networks from different domains, using constructed positive and negative instances for binary influence prediction. Performance is measured against feature-engineering and graph-classification baselines with AUC, precision, recall, and F1.
- 4.1 Datasets: Experiments use four networks from different domains: Open Academic Graph, Digg, Twitter, and Weibo.The datasets represent co-author citation behavior, story voting, retweeting of Higgs-related tweets, and Weibo retweeting, respectively.
- 4.1 Datasets: The data preparation creates positive instances for observed influence events and negative instances for neighbors never observed as active during the observation window.The target is to distinguish positive from negative instances, but the resulting datasets have imbalance problems.
- 4.2 Evaluation: Predictive performance is evaluated with Area Under Curve, Precision, Recall, and F1-Measure.Additional experiments analyze hyper-parameter sensitivity and use case studies to examine the framework’s effectiveness.
- 4.3 Comparison Methods: DeepInf is compared with logistic regression and linear-kernel SVM using ego-user, network-embedding, and hand-crafted ego-network features.It is also compared with PSCN, a state-of-the-art graph-classification model that constructs ordered vertex-neighbor sequences before one-dimensional convolution.
- 4.3 Comparison Methods: DeepInf-GCN and DeepInf-GAT are evaluated as variants, while both DeepInf and PSCN use vertex-level features only and therefore omit ego-network features.The GCN variant uses normalized graph convolution; the GAT variant uses graph attention.
5 EXPERIMENTAL RESULTS
Across four datasets, DeepInf-GAT generally outperforms conventional baselines and other DeepInf variants, while parameter and component analyses identify locality, attention, and normalization effects.
- Overall comparison: DeepInf-GAT achieves significantly better AUC and F1 than baselines, with a 3.8% relative AUC gain on OAG and Digg.The gain is attributed to discovering hidden mechanisms and dynamics of social influence locality.
- Overall comparison: PSCN outperforms linear predictors such as LR and SVM but does not perform as well as DeepInf-GAT.PSCN selects vertices using a user-defined ranking function, whereas DeepInf-GAT uses graph attention.
- Variant analysis: DeepInf-GCN achieves the worst performance among the evaluated methods.The authors attribute this inferiority to GCN’s homophily assumption, which can mix predictive signals with noise when averaging neighbors.
- Feature analysis: DeepInf-GAT retains comparable performance without hand-crafted vertex features, except for the pre-trained network embedding.Experiments with features otherwise use quantities such as page rank score and clustering coefficient.
- Parameter analysis: Increasing return probability or sampled-network size produces slight or slow prediction improvements, indicating a locality pattern and benefits from additional neighbor information.Return probability is varied from 10% to 90%, while sampled-network size is varied from 10 to 100; both analyses report AUC and F1.
- Parameter analysis: Multi-head attention improves DeepInf, but performance decreases as each head receives fewer hidden units; Instance Normalization significantly reduces overfitting on Digg.The attention analysis fixes 128 total hidden units, while the normalization analysis compares training loss and test AUC with and without IN.
6 RELATED WORK
Prior work studies social influence at both macro and user levels, while graph representation learning develops latent representations for vertices and whole graphs. These lines of research motivate local, representation-based analysis of influence.
- Macro-social studies model influence as cascades and examine global patterns such as rise-and-fall dynamics and cascade size.
- DeepInf's case study examines how different graph attention heads highlight different network areas across selected Digg examples.
- User-level studies model influence through near neighbors, including pairwise, topic-level, group-formation, and structural-diversity mechanisms.
- A Facebook field experiment connected online social influence with offline voting behavior during the 2010 US congressional elections.
- Graph representation learning includes vertex embeddings that learn low-dimensional latent factors for network vertices.
7 CONCLUSION
DeepInf formulates social influence locality as a deep-learning problem and combines network embedding, graph convolution, and self-attention. Experiments across four networks show improved prediction over hand-crafted-feature baselines, while neighbor sampling remains an open direction.
- DeepInf combines network embedding, graph convolution, and self-attention to learn social influence locality.
- Experiments on OAG, Digg, Twitter, and Weibo show DeepInf significantly outperforms baselines with rich hand-crafted features.
- The framework can summarize local network areas for downstream tasks such as link prediction, similarity search, and network alignment.
- Neighbor sampling uses random walk with restart without side information and remains loosely coupled with the neural network model.