To address the cold-start problem in recommendation tasks, a cold-start mitigation method integrating text semantics and network structure (CMMITSNS) is proposed to simultaneously learn semantic representations for cold-start items and structural representations within the heterogeneous information network (HIN). Firstly, a multi-task document representation model SPECTER2 is used to obtain the text semantics vector representation of cold-start items. Secondly, the scalable feature learning for networks (node2vec) is used to capture the structural information of nodes in the HIN, and the learned representation results are used as the initial embedding of the graph convolutional network (GCN). Finally, the inner product between the target item and the candidate item is calculated, and the item with a higher score is recommended. Experimental results on AAN and ACM datasets show that CMMITSNS can effectively alleviate the data sparsity and cold-start problems, and outperform the existing methods in terms of the Recall and normalized discounted cumulative gain (NDCG). In ablation experiments, CMMITSNS performs better than single GCN and node2vec methods on the two datasets, and its NDCG@20 on the AAN dataset is 8.11% and 10.26% higher than those of w/o node2vec and w/o GCN, respectively, indicating the effectiveness of multi-dimensional information complementation.
针对上述问题,本文提出一种融合文本语义与网络结构的冷启动缓解方法(Cold-Start Mitigation Method Integrating Text Semantics and Network Structure, CMMITSNS)。在文本语义的学习方面,使用SPECTER2[6]获取文本的语义向量表示,通过文本语义的相似性增强冷启动物品与其他物品之间的联系。在网络结构的获取方面,利用可扩展的网络特征学习方法(Scalable Feature Learning for Networks, node2vec)[7]捕获网络中节点的结构信息,并将学习到的表示结果输入给GCN作为初始化嵌入。最后计算目标物品与候选物品之间的内积,并选取得分较高的物品形成推荐列表。
1 相关工作
近年来,图与图表示学习方法[8-9]取得很大进展,并且这类方法用低维稠密向量来表达图上任意两个节点间的语义或结构联系,有些学者已经在推荐任务中尝试使用该类方法以缓解数据稀疏和冷启动问题。在推荐系统中,数据稀疏性(Data Sparsity)表现为用户与项目间的交互数据量有限。如在电影推荐场景下,用户仅对数十部电影进行评分,而平台电影总数可达上万部,致使用户—项目交互矩阵中大部分元素为零。这种数据分布特征极易引发模型过拟合问题。冷启动问题(Cold-Start Problem)是指当推荐系统引入新实体(包括新用户或新项目)时,由于缺乏历史交互数据,传统基于历史行为的推荐算法无法为这些新实体生成有效推荐。文献[3]提出将SPECTER与记忆网络相结合的科学引文推荐模型(Scientifc Paper Recommendation Employing SPECTER with Memory Network, SPR-SMN)。文献[10]基于包含语义链接的加权异质信息网络(Weighted Heterogeneous Information Network Containing Semantic Linking Algorithm, WHIN-CSL),利用node2vec挖掘异质书目网络中异质书目的同类型及不同类型的节点之间的语义联系。文献[11]提出一种基于生成对抗网络的图嵌入方法(Global Citation Recommendation Employing Generative Adversarial Network, GCR-GAN),采用SPECTER学习语义表示,利用去噪自动编码器网络捕捉网络结构的邻近性。文献[12]通过合作关系、引用关系和研究领域的归属关系构建异质信息网络,使用随机游走方法和自然语言模型学习网络中节点的表示进行推荐。上述模型部分缓解了数据稀疏和冷启动问题,但仍存在不足:多数推荐模型将异质网络视为同质网络处理,难以显式地区分不同类型的节点和节点间的语义关系,导致高阶语义信息缺失,进而严重影响推荐性能。
近年来,GCN被应用于节点分类、链路预测、图聚类、知识图补全和引文预测等任务[13]。针对多模态推荐中的噪声干扰问题,文献[14]提出一个具有双级融合的复合图卷积网络(Composite Graph Convolutional Network with Dual-Stage Fusion for Multimodal Recommendation, COHESION),引入一种双阶段融合策略以过滤无关信息,并构建一种复合图卷积架构,深度挖掘用户与物品内部潜在的同质与异质关联。文献[15]在顺序推荐任务中提出基于位置提示的解码器,将绝对物品位置信息看作外部提示,引入一个外部关注图卷积网络,通过两个外部存储单元线性测量节点之间的全局关联。上述研究为基于图的推荐任务提供了发展方向。
使用node2vec和GCN学习网络中的节点表示。node2vec旨在通过生成节点的低维向量表示来捕捉网络的结构信息。为了灵活采样,深度优先采样(Depth First Sampling, DFS)策略和广度优先采样(Breadth First Sampling, BFS)策略被设计生成个节点的邻域集,并在随机游动过程中同时执行。优化目标是最大化节点的邻域的对数概率。每个顶点将获得一个特征表示。
5)知识图嵌入与IRGAN模型相结合的改进引文推荐方法(Improved Paper Recommendation Model Incorporating Knowledge Graph Embedding with IRGAN Model, KGE-GAN)[16]。该方法采用MRotatE算法与SPECTER文档嵌入模型,分别将图结构特征与基于内容的文本语义映射为查询论文及候选论文的初始向量表示,再将此两个表示作为输入分别投入到生成器、鉴别器中进行对抗训练,以此来得到更好的推荐效果。
为验证CMMITSNS能够有效缓解冷启动问题,选取冷启动目标论文“TERZI E,TSAPARAS P. Efficient algorithms for sequence segmentation”作为典型案例进行对比分析。CMMITSNS和GCN为冷启动目标论文推荐前10篇论文,如表5所示。测试集中目标论文对应的引文有8篇,采用(√)标记正确的推荐结果。
针对推荐任务中存在的冷启动问题,提出一种融合文本语义与网络结构的冷启动缓解方法(CMMITSNS)。CMMITSNS通过预训练语言模型SPECTER2获取冷启动物品的文本语义的向量表示,利用node2vec捕获异质信息网络中节点的结构信息,并将学习到的表示结果输入给GCN作为初始化嵌入。文本语义和网络结构的协同作用有助于模型挖掘隐含的信息,缓解冷启动问题,提升推荐方法的性能。CMMITSNS在具有复杂语义结构的异质信息网络场景以及中小规模图谱上能通过引入强先验知识显著提升冷启动推荐效果,但两阶段训练增加了该方法整体的时间成本和工程复杂度。未来工作中,将考虑使用可微分的随机游走(Differentiable Random Walks)或将node2vec的损失函数联合到GCN的总损失中,实现联合训练。
ABBASM A, AJAYIS, BILALM, et al. A deep learning approach for context-aware citation recommendation using rhetorical zone classification and similarity to overcome cold-start problem[J]. Journal of Ambient Intelligence and Humanized Computing, 2024,15(1):419-433.
[2]
WANGH, LIW J. Relational collaborative topic regression for recommender systems[J]. IEEE Transactions on Knowledge and Data Engineering, 2015,27(5):1343-1355.
[3]
ALI Z, QIG L, KEFALASP, et al. SPR-SMN: scientific paper recommendation employing SPECTER with memory network[J]. Scientometrics, 2022,127(11):6763-6785.
[4]
CAIX Y, HANJ W, YANGL B. Generative adversarial network based heterogeneous bibliographic network representation for personalized citation recommendation[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2018:5747-5754.
SINGHA, D’ARCYM, COHANA, et al. SciRepEval: a multi-format benchmark for scientific document representations[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Singapore. Stroudsburg, USA: ACL, 2023:5548-5566.
[7]
GROVERA, LESKOVECJ. node2vec: scalable feature learning for networks[C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York, USA: ACM, 2016:855-864.
[8]
SHIB S, WANGY Q, GUOF D, et al. Domain adaptation for graph representation learning: challenges, progress, and prospects[J]. Journal of Computer Science and Technology, 2025,40(2):283-300.
[9]
SUNY D, ZHUD J, WANGY S, et al. GTC: GNN-Transformer co-contrastive learning for self-supervised heterogeneous graph representation[J]. Neural Networks, 2025,181:No.106645.
[10]
CHENJ, LIUY, ZHAOS, et al. Citation recommendation based on weighted heterogeneous information network containing semantic linking[C]//Proceedings of the 2019 IEEE International Conference on Multimedia and Expo. Piscataway, USA: IEEE, 2019:31-36.
[11]
ALI Z, QIG L, MUHAMMADK, et al. Global citation recommendation employing generative adversarial network[J]. Expert Systems with Applications, 2021,180:No.114888.
[12]
DUN N, GUOJ, WUC Q, et al. Recommendation of academic papers based on heterogeneous information networks[C]//Proceedings of the 2020 IEEE/ACS 17th International Conference on Computer Systems and Applications. Piscataway, USA: IEEE, 2020. DOI: 10.1109/AICCSA50499.2020.9316516 .
[13]
SADASIVANA, GANANATHANK, NESAMONY ROSE MARY J DPAL, et al. A systematic survey of graph convolutional networks for artificial intelligence applications[J]. WIREs Data Mining and Knowledge Discovery, 2025,15(2):e70012.
[14]
XUJ F, CHENZ Y, WANGW, et al. COHESION: composite graph convolutional network with dual-stage fusion for multimodal recommendation[C]//Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York, USA: ACM, 2025:1830-1839.
[15]
ZHANGJ Y, LIC, ZHAOZ Y. Lightweight yet efficient: an external attentive graph convolutional network with positional prompts for sequential recommendation[J]. ACM Transactions on Information Systems, 2025,43(3):No.79.
[16]
THIERRYN, CHRIST CHATELAINI B, ALI Z, et al. Improved paper recommendation model incorporating knowledge graph embedding with IRGAN model[J]. ACM Transactions on Recommender Systems, 2025,3(4):No.50.