Source-linked AI summary
Jointly embedding the local and global relations of heterogeneous graph for rumor detection
Chunyuan Yuan, Qianwen Ma, Wei Zhou, Jizhong Han, Songlin Hu
TL;DR
Rumor detection methods have not fully used local semantic relations and global structure in heterogeneous propagation graphs. GLAN combines retweet-aware local attention with global heterogeneous-graph modeling, and experiments report superior performance on rumor detection and early detection across three real-world datasets.
Problem
Existing rumor-detection methods underuse local semantic relations and global structural information in message propagation graphs.
Method
GLAN fuses source-tweet and retweet semantics with attention and models relationships among source tweets, retweets, and users as a heterogeneous graph.
Results
GLAN significantly outperforms state-of-the-art models on rumor classification and early detection across three real-world datasets.
Takeaways & Limitations
Jointly encoding local semantic and global structural information provides an effective representation for rumor detection.
Abstract
from arXiv · showhide
The development of social media has revolutionized the way people communicate, share information and make decisions, but it also provides an ideal platform for publishing and spreading rumors. Existing rumor detection methods focus on finding clues from text content, user profiles, and propagation patterns. However, the local semantic relation and global structural information in the message propagation graph have not been well utilized by previous works. In this paper, we present a novel global-local attention network (GLAN) for rumor detection, which jointly encodes the local semantic and global structural information. We first generate a better integrated representation for each source tweet by fusing the semantic information of related retweets with the attention mechanism. Then, we model the global relationships among all source tweets, retweets, and users as a heterogeneous graph to capture the rich structural information for rumor detection. We conduct experiments on three real-world datasets, and the results demonstrate that GLAN significantly outperforms the state-of-the-art models in both rumor detection and early detection scenarios.
I. INTRODUCTION
Rumor detection is difficult and existing approaches underuse relationships across heterogeneous propagation graphs. GLAN addresses this by combining local semantic attention with global structural modeling, achieving strong results on rumor classification and early detection.
- Motivation: Social media rumors spread rapidly, making automatic early-stage detection necessary because ordinary users may struggle to distinguish rumors from other information.
- Existing limitations: Feature-based methods use text, publisher, and propagation information but can be time-consuming, biased, labor-intensive, and sensitive to unavailable or manipulated features.
- Existing limitations: Neural rumor-detection approaches often ignore global structural information among different microblogs and users.
- Proposed approach: The model constructs a global heterogeneous network linking source tweets, retweets, and users to capture structural and semantic relationships across microblogs.
- Evaluation: Experiments on three real-world datasets show superior performance over strong baselines for both rumor classification and early detection.
- Proposed approach: GLAN fuses source-tweet and retweet context with local multi-head attention to produce an integrated representation for each source tweet.
II. RELATED WORK
Prior rumor-detection research uses handcrafted features, neural models, and propagation-tree methods, but these approaches leave important information underused. In particular, existing methods may overlook global relationships among heterogeneous social-media entities and across propagation trees.
- Feature-based Classification Methods: Rumor detection research has used text, user profiles, propagation patterns, and other related information to distinguish rumors from non-rumors.
- Deep Learning Methods: Deep learning methods automatically learn features, but some ignore propagation patterns or represent propagation paths only as sequences.
- Propagation Tree Related Methods: Propagation-tree methods model information-transmission structures, but they do not consider relationships among different propagation trees.
- Propagation Tree Related Methods: Social-media propagation is naturally heterogeneous, involving users, posts, locations, hashtags, and relations such as retweets and friendships.
IV. THE PROPOSED MODEL
The proposed model builds microblog representations with word embeddings and CNNs, then prepares source tweets and retweets for local-relation encoding. Its architecture includes microblog representation, local relation encoding, global relation encoding, and rumor detection modules.
- The model has four components: microblog representation, local relation encoding, global relation encoding, and rumor detection with local and global relations.
- Microblog Representation: Each microblog is represented using word embeddings, with shorter texts zero-padded and longer texts truncated to a fixed length.
- Microblog Representation: A CNN applies convolutional filters and max-over-time pooling to extract semantic features from microblog word-embedding matrices.
- Microblog Representation: Filters with receptive-field sizes 3, 4, and 5 are concatenated to form each microblog representation, and the same process represents retweets.
- Local Relation Encoding: The resulting retweet representations are stacked into a retweet matrix for encoding local relations between source tweets and retweet comments.
B. Local Relation Encoding
The local relation encoder uses multi-head attention to model contextual dependencies between source tweets and retweets. It combines attention-weighted representations while preserving the query dimension.
- The model avoids recurrent networks because they are difficult to parallelize and costly for capturing multi-grained semantic representations.
- Multi-head attention receives query, key, and value sentences and lets each query word attend to words in the key sentence.
- Attention weights are applied to the value sentence to capture dependencies and compose representations from query and value elements.
- Multiple attention heads model relationships from different aspects, concatenate their outputs, and apply a linear transformation.
- The attention output retains the same dimension as the query representation, enabling different query, key, and value configurations.
2) Local Context Representation:
GLAN encodes local semantic relations by refining retweet representations, linking source microblogs with retweets, and fusing the resulting context. It then uses attention over heterogeneous-graph neighbors and multiple relation types to produce global node representations.
- Multi-head self-attention refines each retweet by allowing it to attend to similar retweets, encoding semantic relations among them.
- Cross attention treats the source microblog as a key and aggregates attention-weighted retweets into a new text representation.
- A fusion gate combines the original microblog representation with the retweet-enhanced representation using sigmoid activation and learnable parameters.
- The global heterogeneous graph contains combined text nodes and user nodes, with user-centric and microblog-centric relations encoded through neighborhood attention.
- Multi-head attention applies K independent transformations whose features are concatenated, with ELU activation producing the output representation.
- At each iteration, nodes collect embeddings from local neighbors, weight them with attention, and use their weighted average as the current representation.
D. Rumor Classification
The model concatenates local and global microblog representations as final features, then projects them into class probabilities and optimizes rumor classification with cross-entropy loss.
- Local and global microblog representations are concatenated as final features for rumor classification.
- Fully connected layers project the concatenated representation into the target space of class probabilities.
- Cross-entropy loss is used as the optimization objective, with gold rumor probabilities and all model parameters included in training.
A. Data sets
The evaluation uses three real-world datasets with heterogeneous graphs constructed from source tweets, responsive tweets, and related users, and compares GLAN with established rumor-detection baselines.
- A. Data sets: The evaluation covers Weibo, Twitter15, and Twitter16, collected from popular social media websites in China and the United States.
- A. Data sets: Weibo has binary false-rumor and non-rumor labels, whereas Twitter15 and Twitter16 each use four labels: FR, NR, UR, and TR.
- A. Data sets: Each dataset’s heterogeneous graph is constructed from source tweets, responsive tweets, and related users.
- A. Data sets: Because the original datasets lacked user information, the authors crawled related user profiles through the Twitter API.
- B. Baseline Models: The comparison includes decision-tree, SVM, random-forest, recurrent, propagation-tree, recursive-neural-network, and PPC baseline models.
- B. Baseline Models: PPC is described as the state-of-the-art rumor-detection method when the paper was submitted and combines recurrent and convolutional networks for propagation-path classification.
C. Data Preprocessing
Experiments used three real-world datasets with standard evaluation metrics and a fixed training setup. GLAN outperformed the compared baselines, achieving strong accuracy across Weibo, Twitter15, and Twitter16.
- Data preprocessing: The model used trainable 300-dimensional word vectors and convolutional kernels of sizes 3, 4, and 5.Each kernel size used 100 kernels, with Adam optimization and a learning rate initialized at 1e-3.
- Data preprocessing: The experiments used three datasets and evaluated accuracy, precision, recall, and F1 score.The datasets were split into development, training, and testing subsets, with a 10% development set and a 3:1 training-to-testing ratio.
- Results and analysis: 94.6% accuracy was achieved on Weibo, alongside 90.5% and 90.2% on the two Twitter datasets.These results indicate performance across different dataset types.
- Results and analysis: 92.1% to 94.6% accuracy improvement was obtained on Weibo against the best propagation-path baseline.On Twitter15 and Twitter16, accuracy increased from 84.2% to 90.5% and from 86.3% to 90.2%, respectively.
F. Ablation Study
Ablation studies show that both local and global relation encoding contribute substantially to GLAN. Combining them provides complementary information and improves performance across the evaluated datasets.
- Ablation design: The ablation study compared variants without local encoding, without global encoding, and using only source-tweet text.The study was reported in Table V for Weibo, Twitter15, and Twitter16.
- Local relation encoding: Removing local relation encoding reduced accuracy by 7.8%, 7.8%, and 3.8% on Weibo, Twitter15, and Twitter16.The local relation encoding module captures semantic relations between each source tweet and its corresponding retweets.
- Global relation encoding: Removing global relation encoding caused significant performance declines on all three datasets.The global relation encoding module models structural relations among users and microblogs.
- Combined relation encoding: Combining local and global relation information produced significant performance improvements.The results characterize the two information sources as complementary from local and global perspectives.
G. Early Detection
GLAN was evaluated for rumor detection under varying propagation delays and showed strong early performance. Its accuracy remained better than state-of-the-art models as additional propagation information introduced both signal and noise.
- Evaluation protocol: Early detection performance was evaluated by accuracy after incrementally adding test data up to specified time-delay checkpoints.Only tweets posted before each selected delay were used for evaluation.
- Early detection: Within four hours, GLAN achieved 94% accuracy on Weibo, 88% on Twitter15, and 90% on Twitter16.Using less than four-hour data, GLAN outperformed tree-based classification methods using all-time data.
- Early detection: From four to twelve hours, GLAN remained better than state-of-the-art models despite a slight performance drop.The authors attribute the setting to increasing structural and semantic information alongside increased noise.
- Parameter sensitivity: Tweet length significantly influenced classification performance, with longer tweets generally improving results when short tweets lacked sufficient information.This analysis varied tweet length while controlling other parameters.
- Parameter sensitivity: Kernel-size experiments showed performance increasing with kernel size, peaking at 3 or 4, while size 1 lost word-location information.A size-1 kernel captures unigram features similar to a unigram bag of words.
- Model overview: The proposed heterogeneous graph combines local semantic and global structural information for rumor detection.Local context is fused from source tweets and retweets with multi-head attention, while global structure is modeled with global attention.