融合知识点与图卷积的在线编程题目推荐算法  

Online Judge Problem Recommendation Algorithm by Knowledge Point and Graph Convolution Network

作　　者：罗文劼 肖梓良 LUO Wen-jie;XIAO Zi-liang(School of Cyberspace Security and Computer,Hebei University,Baoding 071000,China)

机构地区：[1]河北大学网络空间安全与计算机学院,河北保定071000

出　　处：《小型微型计算机系统》2023年第10期2331-2337,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61375075)资助;河北省自然科学基金项目(F2019201451)资助;河北省高等学校科学技术研究项目(2019GJJG016)资助;2019教育部-百腾产学合作育人项目(201902127001)资助.

摘　　要：针对在线编程场景下题目推荐方法信息利用不足,在高稀疏度下效果较差的问题,提出一种融合知识点与图卷积的题目推荐算法LGCNR.首先通过潜在狄利克雷分布模型提取题目文本中的知识点信息,并结合用户答题记录的统计数据综合计算题目的相似度.同时将用户的答题记录建模为用户答题意愿度,通过隐语义模型分解用户答题意愿度得到题目的潜在特征.最后利用图卷积层与自注意力机制动态融合题目相似度与题目潜在特征,缓解数据稀疏问题.基于一个真实数据集的实验结果表明,所提算法具有比基线方法更好的推荐效果,在高稀疏度下优势更加明显.To solve the problem of insufficient information utilization of question recommendation methods in online judge scenarios and poor results in high sparseness,a problem recommendation algorithm LGCNR by knowledge points and graph convolution network is proposed.Firstly,the knowledge point information in the title text is extracted by the latent dirichlet allocation model,and the similarity of the question is calculated by combining the statistics recorded by the user′s answer.At the same time,the user′s answer record is modeled as the user′s answer willingness,and the potential characteristics of the question are obtained by decomposing the user′s answer willingness through the latent factor model.Finally,using the graph convolution layer and self-attention mechanism dynamically fuses question similarity and potential features to alleviate data sparseness.Experiments based on a real dataset show that the proposed algorithm has a better recommendation than the baseline method,and its advantages are more evident in high sparsity.

关 键 词：在线编程系统 推荐系统 潜在狄利克雷分布 图卷积 自注意力机制 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]