[Purpose/Significance] Accurately predicting the popularity of social information is critically important for government regulation, platform governance, and public decision-making, as viral misinformation can trigger social panic and disrupt public order. While deep learning methods have advanced this task, most operate under a homogeneous graph assumption that fails to capture the heterogeneous nature of real-world social networks. Even existing heterogeneous graph approaches typically encode high-level “topic” semantics as implicit edge attributes rather than explicit structural entities, severely limiting their capacity for fine-grained semantic reasoning across the complex, non-linear dependencies among users, texts, and topics. To address these fundamental gaps, this paper proposes a novel meta-path enhanced feature learning framework that explicitly materializes topics as first-class graph nodes and introduces an adaptive meta-path selection mechanism to fully exploit heterogeneous semantic associations. [Method/Process] The proposed method proceeds in three stages. First, a heterogeneous social graph is constructed with three node types, namely users (U), texts (V), and topics (T), where topics are extracted via LDA and modeled as independent nodes. Raw features from BERT, user influence metrics, and topic distributions are projected into a unified semantic space via type-specific transformations. An adaptive meta-path selection mechanism then automatically identifies predictive semantic patterns: it generates candidate paths through breadth-first search with type-constrained pruning, evaluates their importance using an attention mechanism that fuses structural and semantic path information, and employs a dual-regularization loss with L1 and entropy terms to dynamically reinforce high-contribution paths while eliminating redundant ones during training. Second, a two-stage feature enhancement strategy refines target node representations. Stage One applies a novel type-role-position three-factor attention mechanism that aggregates fine-grained semantics from all nodes along a meta-path, not merely the endpoints, by jointly modeling each node’s type, its path role, and its positional distance from the target. Stage Two introduces a cross-layer residual gating mechanism to combat over-smoothing in deep architectures, dynamically balancing shallow local features with deep global semantics to preserve local distinctiveness while integrating high-order structural information. Finally, enhanced features from multiple meta-paths are adaptively fused via path-level attention and fed into a Multi-Layer Perceptron for popularity prediction. [Result/Conclusion] Extensive experiments on a large-scale Sina Weibo dataset demonstrate the effectiveness of the proposed framework. Ablation studies validate the indispensable contribution of each component. Comparative results against eight representative baselines covering feature engineering, time series analysis, and state-of-the-art deep learning methods show that the proposed model achieves the best performance on both MSE and MAE metrics. Parameter sensitivity analysis is conducted to determine the optimal hyperparameter configuration, and topic perturbation experiments confirm the model’s robustness to suboptimal topic settings. Qualitative case studies further verify the model’s strong generalization ability across different popularity levels. [Innovation/Value] This work makes dual contributions. Theoretically, it provides a new heterogeneous graph learning paradigm by structurally elevating topics to explicit nodes and introducing an adaptive, data-driven meta-path selection mechanism that obviates manual design. The synergistic two-stage feature enhancement constitutes a novel methodological advance for capturing non-linear heterogeneous semantics. Practically, the proposed method can support public opinion monitoring and platform content management by enabling more accurate popularity prediction. [Insufficient/Improvement] Several limitations point to future work. The LDA-based topic modeling neglects deep contextual semantics, which can be improved with neural topic models. The current framework only leverages textual and structural features, omitting multimodal content pervasive in social media. Finally, cross-platform validation is needed to verify generalizability.
[1] 王震宇, 朱学芳. 基于多模态Transformer的虚假新闻检测研究[J]. 情报学报, 2023, 42(12): 1477-1486. (WANG Z Y, ZHU X F. Research on fake news detection based on multimodal Transformer[J]. Journal of the China Society for Scientific and Technical Information, 2023, 42(12): 1477-1486.)
[2] 陈娟, 谢丹, 邓胜利, 等. 面向重大突发公共卫生事件舆情的政府辟谣信息效果评价研究——基于信息覆盖与信息说服双视角的实证[J]. 图书情报工作, 2024, 68(23): 63-75. (CHEN J, XIE D, DENG S L, et al. Evaluating the effectiveness of government rumor refutation in major public health emergencies: an empirical study from the dual perspectives of information coverage and persuasion[J]. Library and information service, 2024, 68(23): 63-75.)
[3] 周潇, 王博, 胡玉琳, 等. 基于时序图神经网络的潜在高价值专利识别研究[J]. 情报学报, 2024, 43(6): 697-711. (ZHOU X, WANG B, HU Y L, et al. Potential high-value patent identify based on a time-series graph neural network[J]. Journal of the China Society for Scientific and Technical Information, 2024, 43(6): 697-711.)
[4] FU X, KING I. MECCH: metapath context convolution-based heterogeneous graph neural networks[J]. Neural networks, 2024, 170: 266-275.
[5] 李亚聪, 刘皓冰, 蒋若冰, 等. 异质图表征学习综述[J]. 软件学报, 2025, 36(6): 2794-2826. (LI Y C, LIU H B, JIANG R B, et al. Survey on heterogeneous graph representation learning[J]. Journal of software, 2025, 36(6): 2794-2826.)
[6] 李妍, 何洪波, 王闰强. 微博热度预测研究综述[J]. 数据与计算发展前沿, 2023, 5(2): 119-135. (LI Y, HE H B, WANG Y Q. A Survey of research on microblog popularity prediction[J]. Frontiers of data & computing, 2023, 5(2): 119-135.)
[7] 黄微, 刘熠, 许烨婧, 等. 网络舆情推文的热度测度模型构建[J]. 图书情报工作, 2019, 63(20): 17-25. (HUANG W, LIU Y, XU Y J, et al. The construction of heat assessment model for tweets of network public opinion[J]. Library and information service, 2019, 63(20): 17-25.)
[8] CHEN X, ZHOU X, CHAN J, et al. Event popularity prediction using influential hashtags from social media[J]. IEEE transactions on knowledge and data engineering, 2022, 34(10): 4797-4811.
[9] 王龙, 黄嘉凯, 逄华, 等. 基于多维度特征融合的信息热度预测研究——以校园信息平台为例[J]. 情报杂志, 2024, 43(12): 133-141. (WANG L, HUANG J K, PANG H, et al. Research on information popularity prediction based on multi-dimensional feature fusion: a case study of campus information platforms[J]. Journal of intelligence, 2024, 43(12): 133-141.)
[10] YANG Y, FAN C, GONG Y, et al. Forwarding in social media: forecasting popularity of public opinion with deep learning[J]. IEEE transactions on computational social systems, 2025, 12(2): 749-763.
[11] 谢柏林, 黎琦, 邝建. 基于隐半马尔可夫模型的微博流行信息检测方法[J]. 计算机科学, 2022, 49(S1): 291-296. (XIE B L, LI Q, KUANG J. Microblog popular information detection based on Hidden Semi-Markov Model [J]. Computer science, 2022, 49(S1): 291-296.)
[12] MA Y, JIAO H, SONG L. Tracking attention of social media event by Hidden Markov Model: cases from Sina Weibo[J]. IEEE access, 2021, 9: 68240-68252.
[13] GAO S, MA J, CHEN Z. Modeling and predicting retweeting dynamics on microblogging platforms[C]//Proceedings of the eighth ACM international conference on web search and data mining. New York: Association for Computing Machinery, 2015: 107-116.
[14] ZENG Y, XIANG K. Persistence augmented graph convolution network for information popularity prediction[J]. IEEE transactions on network science and engineering, 2023, 10(6): 3331-3342.
[15] CHEN G, KONG Q, XU N, et al. NPP: a neural popularity prediction model for social media content[J]. Neurocomputing, 2019, 333: 221-230.
[16] WANG Y, LI D, SUN J, et al. Combining macro and micro: feature-driven dynamic graph learning for social media popularity prediction[J]. World wide web, 2025, 28(1): 1-20.
[17] JIAO P, SONG W, WANG Y, et al. Cascade popularity prediction: a multi-view learning approach with socialized modeling[J]. IEEE transactions on network science and engineering, 2025, 12(2): 1198-1209.
[18] JIA X, SHANG J, LIU D, et al. HeDAN: heterogeneous diffusion attention network for popularity prediction of online content[J]. Knowledge-based systems, 2022, 254: 109659.
[19] WANG D, ZHOU Y. The popularity of contradictory information about COVID-19 vaccine on social media in China[J]. Computers in human behavior, 2022, 134: 107320.
[20] 吴树芳, 王宏彬, 朱杰, 等. 基于两层异质网络的社交短文本扩展研究[J]. 数据分析与知识发现, 2024, 8(10): 77-88. (WU S F, WANG H B, ZHU J, et al. Social short text expansion based on two-layer heterogeneous network[J]. Data analysis and knowledge discovery, 2024, 8(10): 77-88.)
[21] 吴树芳, 王宏彬, 朱杰. 融入情感反馈转变和交互演化的社交用户动态影响力研究[J]. 情报理论与实践, 2025, 48(4): 135-142, 162. (WU S F, WANG H B, ZHU J. Research on the dynamic influence of social users incorporated with emotional feedback transformation and interactive evolution[J]. Information studies: theory & application, 2025, 48(4): 135-142, 162.)
[22] SUN Y, HAN J. Mining heterogeneous information networks: a structural analysis approach[J]. ACM SIGKDD explorations newsletter, 2013, 14(2): 20-28.
[23] SUN Y, HAN J, YAN X, et al. Pathsim: meta path-based top-k similarity search in heterogeneous information networks[J]. Proceedings of the VLDB endowment, 2011, 4(11): 992-1003.
[24] 邱林山, 房子荃, 陈璐, 等. 面向异质信息网络的双通道协同聚类算法[J]. 计算机学报, 2023, 46(11): 2416-2430. (QIU L S, FANG Z Q, CHEN L, et al. A bi-channel co-clustering algorithm for heterogeneous information networks[J]. Chinese journal of computers, 2023, 46(11): 2416-2430.)
[25] WANG X, JI H, SHI C, et al. Heterogeneous graph attention network[C]// World wide web conference. New York: Association for Computing Machinery, 2019: 2022-2032.
[26] 薛欣, 朱天晨, 孙庆赟, 等. 融合图神经网络的高效子图匹配算法[J]. 计算机研究与发展, 2025, 62(3): 694-708. (XUE X, ZHU T C, SUN Q Y, et al. Efficient subgraph matching algorithm with graph neural network[J]. Journal of computer research and development, 2025, 62(3): 694-708.)
[27] 郑作武, 邵斯绮, 高晓沨, 等. 基于社交圈层和注意力机制的信息热度预测[J]. 计算机学报, 2021, 44(5): 921-936. (ZHENG Z W, SHAO S Q, GAO X F, et al. Social circle and attention based information popularity prediction[J]. Chinese journal of computers, 2021, 44(5): 921-936.)
