PaperlayerView exact revision on arXiv

Source-linked AI summary

DyHGCN: A Dynamic Heterogeneous Graph Convolutional Network to Learn Users' Dynamic Preferences for Information Diffusion Prediction

Chunyuan Yuan, Jiacheng Li, Wei Zhou, Yijun Lu, Xiaodan Zhang, Songlin Hu

arXiv:2006.05169v1cs.SIcs.LG

TL;DR

Information diffusion prediction methods have not jointly modeled social and diffusion graphs or users’ changing preferences. DyHGCN combines those graphs, encodes temporal information to model dynamic preferences, and uses attention to capture current-path context. It reports improvements over state-of-the-art models on three real-world datasets.

  • Problem

    Prior methods focus on diffusion sequences or social networks separately and do not jointly model both graphs or users’ changing preferences.

    Source: §1 · p. 2

  • Method

    DyHGCN combines social and dynamic diffusion graphs, encodes temporal information to learn dynamic preferences, and applies multi-head attention to current-path context.

    Source: §1 · p. 2

  • Results

    More than 5% absolute improvement in hits@100 and map@100 is reported for DyHGCN-H and DyHGCN-S over state-of-the-art methods.

    Source: §5.4 · p. 11

  • Takeaways & Limitations

    Experiments on three real-world datasets report significant improvements over state-of-the-art models, supporting the model’s effectiveness and feasibility for real-world applications.

    Source: §7 · p. 15

    The work does not use the text content of diffused information, although content may also shape users’ reposting preferences.

    Source: §7 · p. 15

Abstract

from arXiv · show

Information diffusion prediction is a fundamental task for understanding the information propagation process. It has wide applications in such as misinformation spreading prediction and malicious account detection. Previous works either concentrate on utilizing the context of a single diffusion sequence or using the social network among users for information diffusion prediction. However, the diffusion paths of different messages naturally constitute a dynamic diffusion graph. For one thing, previous works cannot jointly utilize both the social network and diffusion graph for prediction, which is insufficient to model the complexity of the diffusion process and results in unsatisfactory prediction performance. For another, they cannot learn users' dynamic preferences. Intuitively, users' preferences are changing as time goes on and users' personal preference determines whether the user will repost the information. Thus, it is beneficial to consider users' dynamic preferences in information diffusion prediction. In this paper, we propose a novel dynamic heterogeneous graph convolutional network (DyHGCN) to jointly learn the structural characteristics of the social graph and dynamic diffusion graph. Then, we encode the temporal information into the heterogeneous graph to learn the users' dynamic preferences. Finally, we apply multi-head attention to capture the context-dependency of the current diffusion path to facilitate the information diffusion prediction task. Experimental results show that DyHGCN significantly outperforms the state-of-the-art models on three public datasets, which shows the effectiveness of the proposed model.

1 Introduction

DyHGCN addresses limitations in prior diffusion prediction by combining social and dynamic diffusion graphs and modeling users’ changing preferences. The paper reports improved performance over state-of-the-art models on three public datasets.

1 Introduction

  • Information diffusion prediction supports applications including social-influence prediction, misinformation-spread analysis, and malicious-account detection.Source: §1 · p. 2
  • Prior methods do not jointly use social and diffusion graphs or learn users’ changing preferences, although preferences can be reflected in dynamic diffusion structure.The authors argue that combining these aspects is beneficial for prediction.Source: §1 · p. 2
  • DyHGCN combines the social network and dynamic diffusion graph, encodes temporal information to learn dynamic preferences, and models current-path context dependency.The paper identifies dynamic preference modeling as, to the authors’ knowledge, a first for information diffusion prediction.Source: §1 · p. 2
  • DyHGCN outperforms state-of-the-art models on three public datasets, with reported effectiveness and efficiency.The introduction presents this as evidence of the model’s real-world feasibility.Source: §1 · p. 2

2 Related Work

Related work is organized around diffusion-path methods and social-network methods. The paper argues that jointly modeling the social graph and dynamic diffusion graph also captures global reposting relations and dynamic individual preferences.

2.1 Diffusion Path based Methods.

  • Neural diffusion-path models learn from past sequences; examples include TopoLSTM, DeepDiffuse, and NDM [26].The cited approaches model topology-aware embeddings, infection timestamps, or long-term dependencies, respectively.Source: §2.1 · p. 3
  • Diffusion-path methods infer interpersonal influence from observed sequences, but sequence-based approaches omit social relations that help identify information-flow direction.Some earlier approaches rely on specifying a prior diffusion model, whose effectiveness depends on a hypothesis that can be hard to specify or verify in practice.Source: §2.1 · p. 3

2.2 Social Graph based Methods.

  • Social-graph methods use user relations for prediction, with prior work studying social roles, sequential diffusion with user connections, and multi-scale prediction.Source: §2.2 · pp. 3–4
  • The paper argues that existing graph-based methods miss other messages’ diffusion paths and global reposting relations; DyHGCN jointly learns social and dynamic diffusion graph structure and models dynamic preferences.Source: §2.2 · p. 4

3 Problem Formulation

The paper represents a message’s diffusion as timestamp-ordered reposting behavior by users. Given observed traces, the task is to predict which users will repost at a future time, using both social and diffusion graphs.

3 Problem Formulation

  • A message’s diffusion process is represented by successive user activations, with reposting behaviors recorded in timestamp order.The paper describes real cascades as often recorded in single-chain sequence form.Source: §3 · p. 4
  • Given observed diffusion traces, the task is to predict diffusion behavior at a future timestamp t′, including which users may repost.Figure 1 illustrates possible reposting behavior and potential activation users.Source: Figure 1 · p. 4, §3 · p. 5
  • The formulation jointly uses social and diffusion graphs; the example contrasts direct social links with repost history when identifying likely future reposts.Users u2 and u6 are more likely than u3 to repost from u5 based on direct follows, while u3 may receive the message through u2 or u6 and has reposted twice from u2.Source: §3 · p. 5

4 Framework

DyHGCN combines social relations and time-varying repost relations to represent users’ preferences, then uses those representations to predict future diffusion. Its framework builds heterogeneous graphs, encodes them across time, and applies attention to the current diffusion path.

4 Framework

  • DyHGCN constructs heterogeneous social and diffusion graphs, learns user representations with dynamic preferences, and combines them to predict future infected users.The framework has three stages: graph construction, dynamic heterogeneous graph encoding, and prediction.Source: §4 · p. 5

4.1 Heterogeneous Graph Construction

  • The model represents social relations as a directed, unweighted graph and repost relations in each time interval as a directed, weighted diffusion graph.It uses repost history to capture changes in users’ preferences.Source: §4.1 · p. 6

4.2 Heterogeneous Graph Convolutional Network (HGCN)

  • At each time interval, a heterogeneous graph uses user nodes and follow and repost relations; GCNs learn representations for each relation and fuse them.The relation representations are combined using their concatenation, element-wise product, and difference, followed by a learned projection.Source: §4.2 · p. 6

4.3 Dynamic Graph Encoding

  • Dynamic graph encoding splits the historical diffusion timeline into intervals, applies HGCN at each interval, and collects the resulting user representations.The collected representations are passed to the next stage.Source: §4.3 · p. 7

Hard Selection Strategy

  • Hard selection uses the user representation from the interval in which that user reposts.For a repost at t′ in [t3, t4), the strategy selects ut3.Source: Hard Selection Strategy · p. 7

Soft Selection Strategy

  • Soft selection uses time-aware attention to fuse historical user representations, unlike hard selection, which uses only the repost-time interval.The attention mask assigns zero weight to future timestamps to avoid label leakage.Source: Soft Selection Strategy · p. 8

4.5 Information Diffusion Prediction

  • The prediction stage applies masked multi-head self-attention to encode dependencies in the current diffusion sequence while masking future information.A two-layer fully connected network then computes diffusion probabilities from user representations across diffusion time steps.Source: §4.5 · pp. 8–9

5 Experiments

DyHGCN is evaluated on three public datasets against diffusion-path and social-network baselines. Both DyHGCN variants consistently outperform state-of-the-art methods by more than 5% on Hits@100 and MAP@100.

5.1 Datasets

  • Twitter records URL tweets with follow ties; Douban treats books or movies as information items and discussion co-occurrence as friendship; Memetracker has no underlying social graph.The three public datasets are used to evaluate information diffusion prediction.Source: §5.1 · pp. 9–10
  • Cascades are randomly split 80% for training, 10% for validation, and 10% for testing.Source: §5.1 · p. 10

Based on diffusion path

Based on social network

  • Social-network baselines include SNIDSA and FOREST; DyHGCN is tested with hard selection and soft, time-aware attention strategies.SNIDSA integrates sequential and structural information, while FOREST incorporates macroscopic diffusion size into an RNN-based model.Source: Based on social network · p. 10

5.3 Evaluation Metrics and Parameter Settings.

  • Next-infected-user prediction is evaluated as a retrieval task using MAP@k and Hits@k.Source: §5.3 · p. 11
  • The selected configuration uses 64-dimensional user and temporal-interval embeddings, two GCN layers, 14 attention heads, and eight time intervals.Configuration is selected on the validation set and evaluated on the test set.Source: §5.3 · p. 11

5.4 Experimental Results

  • DyHGCN-H and DyHGCN-S outperform state-of-the-art methods by more than 5% on Hits@100 and MAP@100 across the three datasets.Improvements are reported as statistically significant at p < 0.01 on paired t-tests.Source: Table 2 · p. 11, §5.4 · p. 11, Table 3 · p. 12

6 Further Study

The ablations indicate that both graph encodings, their heterogeneous combination, and time-aware attention contribute to DyHGCN’s performance. Parameter analyses on Twitter examine the effects of snapshot intervals and attention-head count.

6.1 Ablation Study

  • Removing heterogeneous encoding degrades performance more than removing either graph alone, indicating complementary information in the social and repost relations.The ablation study covers Twitter and Douban.Source: Table 5 · p. 13, §6.1 · p. 13
  • Replacing time-aware attention with hard selection slightly reduces performance compared with DyHGCN-S.The authors report that time-aware attention fuses historical user representations to produce a better user representation.Source: §6.1 · p. 13
  • Removing either the social-graph or diffusion-graph encoding substantially degrades performance compared with DyHGCN.The results identify both relation-encoding modules as essential for information diffusion prediction.Source: §6.1 · p. 13

6.2 Parameter Analysis

  • Performance increases as the number of diffusion-graph snapshots grows up to n = 8, while larger n brings limited further improvement.The authors describe larger n as splitting the diffusion graph into more pieces, allowing finer-grained changes to be learned.Source: §6.2 · p. 15
  • Increasing the number of attention heads initially improves performance slightly, but too many heads cause a significant drop; H = 14 is most suitable.The authors attribute the drop with too many heads to overfitting.Source: §6.2 · p. 15

7 Conclusion and Future Work

DyHGCN encodes social and dynamic diffusion graph structure to learn users’ dynamic preferences. Experiments on three real-world datasets show improvements over state-of-the-art models; future work will add diffused-message text content.

7 Conclusion and Future Work

  • DyHGCN encodes social and dynamic diffusion graph structure to learn users’ dynamic preferences for information diffusion prediction.The conclusion describes experiments on three real-world datasets.Source: §7 · p. 15
  • The authors report significant improvements over state-of-the-art models on three real-world datasets.They describe the results as showing the model’s effectiveness and feasibility for real-world applications.Source: §7 · p. 15
  • The model does not use diffused-message text content, which the authors identify as future work.They note that users may repost messages based on preferences for particular topics or content.Source: §7 · p. 15

Related papers

Loading 2006.05169v1…