PaperlayerView exact revision on arXiv

Source-linked AI summary

Popularity Prediction on Social Platforms with Coupled Graph Neural Networks

Qi Cao, Huawei Shen, Jinhua Gao, Bingzheng Wei, Xueqi Cheng

arXiv:1906.09032v2cs.SIcs.LG

TL;DR

Popularity prediction is difficult because existing methods inadequately capture cascading diffusion across social networks. CoupledGNN models this cascade with two coupled graph neural networks and significantly outperforms state-of-the-art methods on synthetic and Sina Weibo data.

  • Problem

    Existing popularity-prediction methods are less effective at capturing cascading information diffusion through underlying social networks, despite the value of accurate prediction.

  • Method

    CoupledGNN uses two coupled graph neural networks to model the interplay between user activation states and interpersonal influence along the network.

  • Results

    CoupledGNN significantly outperforms state-of-the-art methods on synthetic and real-world Sina Weibo datasets, exceeding DeepCas by more than 10% under MRSE on Sina Weibo.

  • Takeaways & Limitations

    Modeling interactions among early adopters, potential active users, and network-connected users improves future popularity prediction.

  • Takeaways & Limitations

    The study targets macro popularity prediction on global diffusion networks rather than micro action-status prediction on fixed-sized local networks.

Abstract

from arXiv · show

Predicting the popularity of online content on social platforms is an important task for both researchers and practitioners. Previous methods mainly leverage demographics, temporal and structural patterns of early adopters for popularity prediction. However, most existing methods are less effective to precisely capture the cascading effect in information diffusion, in which early adopters try to activate potential users along the underlying network. In this paper, we consider the problem of network-aware popularity prediction, leveraging both early adopters and social networks for popularity prediction. We propose to capture the cascading effect explicitly, modeling the activation state of a target user given the activation state and influence of his/her neighbors. To achieve this goal, we propose a novel method, namely CoupledGNN, which uses two coupled graph neural networks to capture the interplay between node activation states and the spread of influence. By stacking graph neural network layers, our proposed method naturally captures the cascading effect along the network in a successive manner. Experiments conducted on both synthetic and real-world Sina Weibo datasets demonstrate that our method significantly outperforms the state-of-the-art methods for popularity prediction.

1 INTRODUCTION

Popularity prediction seeks to anticipate future content popularity from early adopters, but cascading diffusion makes accurate prediction challenging. CoupledGNN addresses this by combining network structure with coupled modeling of activation and influence.

  • Popularity prediction uses early adopters observed at a specific time to estimate future popularity, supporting early discovery of hot information amid information overload.
  • Existing approaches use regression or classification and extract demographic, activity, degree, density, substructure, and community features from early adopters.
  • Statistics such as average user fan counts approximate cascading impact but ignore the explicit network structure governing diffusion, reducing prediction effectiveness.
  • CoupledGNN models the interplay between user activation states and interpersonal influence with two coupled graph neural networks.
  • The method is evaluated on synthetic data and real-world Sina Weibo data, significantly outperforming state-of-the-art methods.

2 RELATED WORK

Prior popularity-prediction and diffusion approaches use content, temporal, adopter, or network information, but often inadequately represent explicit user interactions during cascades. CoupledGNN instead uses neighborhood aggregation and coupled GNNs to model network-wide cascading effects.

  • 2.1 Popularity Prediction: Popularity prediction covers future popularity from observed early adopters, with prior work treating it as regression or classification across several content types.
  • 2.1 Popularity Prediction: Existing methods incorporate content, temporal information, early adopters, and network structure through handcrafted features or representation learning.
  • 2.1 Popularity Prediction: Handcrafted early-adopter and network features depend heavily on heuristic feature quality, while prior methods neglect explicit user interactions governing cascading diffusion.
  • 2.2 Diffusion Models: Classic diffusion models iteratively simulate propagation, whereas CoupledGNN directly models cascading influence through graph-neighborhood aggregation without specifying a diffusion process.
  • 2.3 Graph Neural Networks: Graph neural networks recursively aggregate neighborhood representations, making them suitable for graph-based representation learning and social influence prediction.
  • 2.3 Graph Neural Networks: DeepInf targets micro-action prediction on fixed-size local networks, unlike this paper’s macro popularity prediction on global diffusion networks.
  • 2.3 Graph Neural Networks: The paper adapts GNNs to information diffusion through CoupledGNN, designed to characterize cascading effects in popularity prediction.

3 PRELIMINARIES

This section defines network-aware popularity prediction from observed cascades and an underlying user network, then outlines the general neighborhood-aggregation framework of GNNs.

  • 3.1 Problem Definition: An observed information cascade records the users who adopt or activate the information within the observation time window.The framework considers M pieces of information and represents each observed cascade as its early adopters during that window.
  • 3.1 Problem Definition: Network-aware popularity prediction uses observed cascades and an underlying network G = (V, E) to predict an information item's final popularity.The network captures interactions between early adopters and potential active users, including interactions among potential active users.
  • 3.1 Problem Definition: The underlying network consists of users as nodes and relationships between users as edges.Formally, V is the set of users and E ⊆ V × V is the set of user relationships.
  • 3.2 General Framework of GNNs: A GNN updates each node representation by recursively aggregating representations from neighboring nodes.The choices of AGGREGATE and COMBINE functions determine how neighborhood information is incorporated.
  • 3.2 General Framework of GNNs: A READOUT function converts node representations into a representation of the entire graph.READOUT may use summation or a more sophisticated graph-level pooling function.

4 METHODS

CoupledGNN models network-aware popularity prediction by coupling activation-state and influence representations across social-network neighborhoods. After stacked graph-neural-network layers, predicted user activation probabilities are sum-pooled into content popularity and trained with MRSE plus regularization.

  • CoupledGNN framework: CoupledGNN captures cascading effects by modeling the iterative interplay between node activation states and influence spread.The model uses neighborhood aggregation to propagate information through the social network.
  • Coupling mechanisms: The two graph neural networks are coupled through influence and state gating mechanisms to jointly update activation and influence representations.The state network uses influence gating, while the influence network uses state gating.
  • State Graph Neural Network: The state graph neural network models each user’s activation state using neighborhood information and heterogeneous influence weights.Each user has a one-dimensional activation value, while influence gating represents variation in interpersonal influence.
  • Influence Graph Neural Network: The influence graph neural network diffuses interpersonal influence through neighborhood aggregation and state gating.Active users’ influence representations propagate along network structure, with state gating reflecting nonlinear state effects.
  • Output layer: After K layers, final user activation probabilities are sum-pooled across users to obtain predicted popularity.The output activation probability is produced by the final state-network layer.
  • Optimization: Training uses mean relative square error with L2 regularization and user-level cross-entropy regularization.MRSE is described as robust to outliers, smooth, and differentiable.
  • Complexity: The coupled graph neural network has O(p|V| + q|E|) computational complexity, with mini-batching offering O(R) complexity.Here p and q are small constants, while R denotes the number of samples in a mini-batch.

5 EXPERIMENTAL SETUP

The experiments compare CoupledGNN with several state-of-the-art methods on different datasets and evaluation metrics.

  • Experimental design: CoupledGNN is evaluated against several state-of-the-art methods under various evaluation metrics.The paper introduces detailed experimental settings for these comparisons.

5.1 Data Sets

The study evaluates CoupledGNN on synthetic cascades and real-world Sina Weibo retweet cascades, using network structures and early diffusion observations.

  • Datasets: Experiments use both synthetic data and a real-world dataset from Sina Weibo.The two settings provide synthetic and platform-derived information cascades.
  • Synthetic Data Set: The synthetic network contains 1,086 nodes and 4,038 edges after retaining the largest connected component of a Kronecker-generated network.The generator produces heavy-tailed degree distributions and small diameters.
  • Synthetic Data Set: Synthetic cascades use power-law seed-set sizes and the IC model, with activation probability 1/d_v and observation at t = 0 and t = 1.The observation window is T = 2 time steps.
  • Synthetic Data Set: 27,218 synthetic cascades remain after filtering cascades with fewer than 3 active users, with 80% for training, 10% validation, and 10% testing.The initial generation produced 108,600 cascades.
  • Sina Weibo Data Set: The Sina Weibo following network contains 1.78 million users and 308 million following relationships, alongside 300 thousand retweet cascades.Following relationships are relevant because followed users’ messages appear in a user’s feed.
  • Sina Weibo Data Set: The selected Sina Weibo network contains 23,681 users and 1,802,146 edges, with 3,228 cascades and observation windows of 1, 2, and 3 hours.Cascades with fewer than 5 active users are filtered out.

5.2 Baselines

The baselines cover feature-based and deep representation-learning approaches for network-aware popularity prediction, including a method that approximates cascading effects.

  • Baseline categories: Baseline methods are selected from feature-based and deep representation-learning categories, with an additional cascading-effect method.The comparison targets methods using early adopters and network structure without temporal information.
  • Feature-based baseline: Hand-crafted features are extracted from global, cascade, and frontier graphs containing early adopters, their edges, and one-hop neighbors.These features are designed to generalize across datasets.
  • Deep representation-learning baseline: DeepCas learns cascade-graph representations from random-walk sequences using node embeddings, recurrent neural networks, and attention mechanisms.The method assembles a representation of the cascade graph from sequence representations.
  • Cascading-effect baseline: SEISMIC implements a Hawkes self-exciting point process and estimates cascading impact using users’ average number of fans.It represents a direct attempt to capture cascading effects.

5.3 Implementation Details.

The experiments tune regularization and learning rates separately for different parameter groups and models.

  • The L2 coefficient is selected from 10^-8 through 0.1.
  • Feature-based methods use separate learning-rate search spaces for node-id features and other features.
  • DeepCas uses different learning-rate ranges for user embeddings and other parameters, with DeepWalk initialization for its embeddings.
  • CoupledGNN user embeddings are also obtained through a separately described initialization procedure.

5.4 Evaluation Metrics

The evaluation uses multiple error metrics to assess popularity-prediction performance, with lower values indicating better results.

  • The study adopts several evaluation metrics to comprehensively compare methods.
  • Mean relative square error loss is used as an evaluation metric for popularity prediction.
  • Median RSE is the 50th percentile of test-data RSE and is included because SEISMIC is sensitive to outlier error.
  • Wrong prediction error measures the percentage of online contents incorrectly predicted under a given error tolerance ϵ.
  • For all three metrics, smaller values indicate better performance.

6 EXPERIMENTAL RESULTS

CoupledGNN outperforms the baselines across synthetic and Sina Weibo data, while analyses examine observation time, architecture, network completeness, loss weighting, and layer depth.

  • 6.1 Overall Performance: CoupledGNN outperforms all baselines on synthetic and real-world datasets, improving more than 10% over DeepCas on Sina Weibo under MRSE.
  • 6.1 Overall Performance: Longer observation time reduces MRSE, MAPE, and WrongPerc errors for all methods.More observation time provides more information for prediction.
  • 6.2 Compare CoupledGNN with Single-GNN: Single-GCN and Single-GAT perform almost similarly, indicating that GCN normalization already reflects linked-user correlation effectively.
  • 6.2 Compare CoupledGNN with Single-GNN: CoupledGNN significantly improves performance under all evaluation metrics compared with Single-GNN methods.The result supports modeling activation state and influence with separate GNNs coupled by gating mechanisms.
  • 6.3 Parameter Analysis: Adding user-level cross-entropy initially improves macro-popularity prediction, but excessive loss weight shifts attention away from the macro task.The MRSE values for λ = 0.0, 0.5, 1.0, 10.0, and 20.0 are 0.1109, 0.1101, 0.1111, 0.1111, and 0.1141.
  • 6.3.2 The influence of a partial lack of network: The study evaluates partial network information by randomly dropping edges while preserving network connectivity, affecting CoupledGNN and baselines.
  • 6.3.3 The number of layers in CoupledGNN: With a one-hour observation window, K = 3 gives the best validation performance, matching the finding that 99.76% of later activated users lie within three hops of early adopters.This correspondence is presented as guidance for setting K in other situations.

7 CONCLUSION

The paper addresses network-aware popularity prediction by modeling information diffusion with coupled graph neural networks. Experiments on synthetic and real-world data validate the method's effectiveness.

  • CoupledGNN models network-based cascading through two coupled GNNs for node activation states and interpersonal influence.
  • Gated interactions and iterative neighborhood aggregation represent the interplay between activation states and influence along the underlying network.
  • Experiments on synthetic and real-world data validate CoupledGNN for popularity prediction.
Loading 1906.09032v2…