PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Propagation Uncertainty: Edge-enhanced Bayesian Graph Convolutional Networks for Rumor Detection

Lingwei Wei, Dou Hu, Wei Zhou, Zhaojuan Yue, Songlin Hu

arXiv:2107.11934v1cs.AI

TL;DR

Unreliable and incomplete propagation relations make rumor-detection graphs uncertain, a gap that can limit effective feature learning. EBGCN uses Bayesian relation weighting and edge-wise consistency training, and the authors report better performance than baselines on both rumor-detection tasks across three benchmarks.

  • Problem

    Propagation relations can be manipulated or incompletely annotated, but rumor-detection methods often treat graph edges as reliable despite the resulting uncertainty.

    Source: §1 · p. 2, §2.1 · p. 3

  • Method

    EBGCN uses a Bayesian approach to adaptively infer latent-relation weights and adds edge-wise consistency training for unlabeled relations.

    Source: §1 · p. 2

  • Results

    Across three benchmark datasets, EBGCN significantly outperforms baselines on both rumor detection and early rumor detection.

    Source: §1 · p. 2, §7 · p. 9

  • Takeaways & Limitations

    The authors conclude that modeling propagation uncertainty is effective for rumor detection on the evaluated benchmark datasets.

    Source: §7 · p. 9

    Latent relations are modeled independently, with a factorized Gaussian distribution for each relation.

    Source: §4.3 · p. 5

Abstract

from arXiv · show

Detecting rumors on social media is a very critical task with significant implications to the economy, public health, etc. Previous works generally capture effective features from texts and the propagation structure. However, the uncertainty caused by unreliable relations in the propagation structure is common and inevitable due to wily rumor producers and the limited collection of spread data. Most approaches neglect it and may seriously limit the learning of features. Towards this issue, this paper makes the first attempt to explore propagation uncertainty for rumor detection. Specifically, we propose a novel Edge-enhanced Bayesian Graph Convolutional Network (EBGCN) to capture robust structural features. The model adaptively rethinks the reliability of latent relations by adopting a Bayesian approach. Besides, we design a new edge-wise consistency training framework to optimize the model by enforcing consistency on relations. Experiments on three public benchmark datasets demonstrate that the proposed model achieves better performance than baseline methods on both rumor detection and early rumor detection tasks.

1 Introduction

Unreliable or manipulated propagation relations make observed graphs uncertain and can cause error accumulation in message passing. EBGCN models relation uncertainty probabilistically, with edge-wise consistency training for unlabeled relations.

  • Unreliable propagation relations can cause severe error accumulation across message-passing layers and limit effective feature learning.The paper attributes uncertainty to manipulated relations and incomplete or noisy spread-relation annotations.Source: §1 · pp. 1–2
  • EBGCN infers latent-relation weights from observed-graph node features and uses refined edges to guide message passing.The model adaptively controls message passing rather than relying on fixed edge weights.Source: §1 · p. 2
  • Edge-wise consistency training combines unlabeled-relation consistency with supervised learning and aligns latent-edge and observed-node-feature distributions using KL divergence.Training optimizes claim cross-entropy and the Bayes by Backprop loss for latent relations.Source: §1 · p. 2
  • Experiments on three real-world benchmark datasets demonstrate effectiveness on both rumor detection and early rumor detection.Source: §1 · p. 2

2 Related Work

Prior rumor-detection work uses text and propagation structure, while graph methods generally treat propagation edges as reliable. Related graph-neural-network research uses Bayesian approaches to represent graph uncertainty.

2.1 Rumor Detection

  • Prior methods progressed from handcrafted-feature classifiers to temporal deep models and propagation trees or graphs; Wei et al. (2019) jointly modeled structure and temporal evolution.Examples include recurrent networks, attention, CNNs, and Transformers for temporal features.Source: §2.1 · pp. 2–3
  • Most propagation-structure methods assume reliable edges, whereas this paper targets uncertainty from unreliable relations in rumor-detection graphs.Source: §2.1 · p. 3

2.2 Graph Neural Networks

  • Graph neural networks have limited ability to handle graph-structure uncertainty; prior Bayesian approaches motivate this paper's probabilistic treatment of propagation uncertainty.The passage cites Luo et al. (2020) and Zhang et al. (2019b) as approaches to uncertain graph information.Source: §2.2 · p. 3

3 Problem Statement

The problem is framed as classifying claims that contain a source tweet, related retweets, and a propagation structure. That structure is represented as a directed graph rooted at the source tweet.

  • A claim includes a source tweet, related retweets, and a propagation structure, and rumor detection is framed as multiclass classification.Source: §3 · p. 3
  • The propagation structure is a directed graph with the source tweet as root and edges connecting tweets to corresponding retweets.Source: §3 · p. 3
  • Each claim has a ground-truth label, and the objective is to learn a classifier that predicts labels for claims.Source: §3 · p. 3

4 The Proposed Model

EBGCN models rumor propagation uncertainty with graph convolution and an edge-wise consistency training framework. It builds paired propagation and dispersion graphs, updates their node features and edge weights, and combines graph representations for classification.

4.2.1 Graph Construction and Text Embedding

  • EBGCN constructs top-down propagation and bottom-up dispersion graphs, initializing both with the same TF-IDF features from the claim’s top 5,000 words.The source tweet is represented by x0 in the feature matrix.Source: §4.2.1 · pp. 3–4

4.2.2 Node Update

  • Graph convolutional layers iteratively aggregate neighboring-node features, with textual features initializing the node representations.The node update is described as a differentiable message-passing framework.Source: §4.2.2 · p. 4

4.2.3 Edge Inference

  • The edge inference module transforms previous-layer node representations to update adjacency weights, then shares edge-inference parameters across the two graphs.After two transformation layers, the model accumulates normalized neighbor features driven by latent relations.Source: §4.2.3 · p. 4

4.2.4 Classification

  • Mean-pooled representations of the propagation and dispersion graphs are concatenated and passed through a fully connected layer and softmax to predict claim labels.The graphs’ node representations are aggregated into graph representations before classification.Source: §4.2.4 · pp. 4–5

4.3 Edge-wise Consistency Training Framework

  • The framework adaptively adjusts edge weights from the observed graph to support message passing and learning discriminative rumor features.The unsupervised loss uses the predicted distribution of latent relations.Source: §4.3 · p. 5
  • Training minimizes a weighted sum of supervised cross-entropy and an unsupervised Bayes-by-Backprop loss over latent relations.The trade-off coefficient γ weights the two losses; the posterior over each latent relation is modeled with a factorized Gaussian whose means and variances are set by a transformation layer.Source: §4.3 · p. 5

5 Experimental Setup

The experiments use Twitter15, Twitter16, and PHEME, with dataset-specific cross-validation protocols. They compare EBGCN with prior methods using classification metrics and tuned training settings.

5.1 Datasets

  • Twitter15 and Twitter16 use five-fold cross-validation, while PHEME uses leave-one-event-out testing across nine events.Twitter15 and Twitter16 contain 1,490 and 818 claims; PHEME contains 2,402 claims.Source: §5.1 · p. 5

5.2 Baselines

  • The baselines include credibility, time-series, propagation-tree, recurrent, transformer, and graph-convolution approaches, with dataset-specific comparison sets.BiGCN is included for Twitter15, Twitter16, and PHEME; PHEME’s baselines also include NileTMRG, BranchLSTM, and Hierarchical GCN-RNN.Source: §5.2 · pp. 5–6

5.3 Evaluation Metrics

  • Twitter15 and Twitter16 are evaluated with accuracy and per-class F1, while PHEME uses accuracy, macro-F1, and weighted-F1.Weighted-F1 is reported for PHEME because of its imbalanced classes.Source: §5.3 · p. 6

5.4 Parameter Settings

  • The model uses 64-dimensional hidden vectors, trains for up to 200 epochs with early stopping, and selects hyperparameters on designated validation sets.For PHEME, the study also evaluates a skip-gram word-embedding variant with 200-dimensional textual features; baseline results come from a BiGCN implementation or cited original papers.Source: §5.4 · p. 6

6 Results and Analysis

Across three benchmarks, EBGCN leads baseline comparisons for rumor detection. The analysis also examines edge inference and consistency training, early detection, and an example of an unreliable propagation link.

6.1 Performance Comparison with Baselines

  • Reported EBGCN improvements: Twitter15, 2.4% accuracy/3.7% unverified-rumor F1; Twitter16, 3.4% accuracy/6.0% nonrumor F1; PHEME, 40.2% accuracy/34.7% mF1/18.0% wF1.The paper reports EBGCN as the best-performing baseline comparison overall.Source: §6.1 · p. 6, Table 2 · p. 7

6.2 Model Analysis

  • Best edge-inference T differs by dataset: 2 on Twitter15, 3 on Twitter16, and 4 on PHEME.Posterior-based edge refinement adapts relation reliability; omitting consistency training gave the worst performance, with optimal γ values of 0.4, 0.3, and 0.3, respectively.Source: §6.2 · pp. 7–8

6.3 Early Rumor Detection

  • EBGCN has relatively high accuracy at each early-detection deadline or tweet count and outperforms BiGCN in early detection.The paper also reports that structural-feature models outperform temporal-only and handcrafted-feature comparators in early detection.Source: §6.3 · p. 8, Figure 4 · p. 8

6.4 The Case Study

  • In the PHEME case, an irrelevant reply edge from tweet 5 to source tweet 1 is downweighted, reducing its negative effect on node 1's aggregation.Figure 5 depicts the sampled rumor and the constructed top-down and bottom-up graphs.Source: §6.4 · pp. 8–9, Figure 5 · p. 9

7 Conclusion

EBGCN models uncertainty in rumor propagation with Bayesian adjustment of unreliable-relation weights and edge-wise consistency training. Experiments on three benchmark datasets show it significantly outperforms baselines on rumor detection and early rumor detection.

  • EBGCN significantly outperforms baselines on both rumor detection and early rumor detection across experiments on three benchmark datasets.Source: §7 · p. 9
  • EBGCN handles propagation uncertainty with a Bayesian method that adaptively adjusts the weights of unreliable relations.Source: §7 · p. 9
  • An edge-wise consistency training framework incorporates unsupervised relation learning to enforce consistency on latent relations.Source: §7 · p. 9
Loading 2107.11934v1…