基于扩散模型的多结构感知知识推荐OA
A Sequential recommendation model with structure-guided diffusion and knowledge tracing
已有基于知识概念的序列推荐方法大多仅依赖简单的ID嵌入或原始文本语义进行推荐,忽略了文本描述中存在的语义噪声以及知识点之间固有的先修后继结构关系,此外,现有方法往往未能充分考量学习者在学习过程中动态变化的知识掌握状态,导致推荐结果难以精准匹配用户的当前能力水平.针对这些问题,提出了一种基于结构引导扩散与知识追踪的序列推荐模型SG-DiffRec.不同于传统的静态表征学习,SG-DiffRec利用扩散模型的生成流形特性,将知识图谱的拓扑结构作为条件引导注入语义生成过程.该机制能够将发散的语义分布强行对齐到符合教学逻辑的结构轨道上,从而生成兼具语义深度与结构精度的增强表征.同时,模型引入深度知识追踪模块实时捕捉学习者的动态认知能力,并通过多模态注意力机制实现静态知识属性与动态能力状态的自适应校准,综合考虑知识图谱的全局拓扑结构,交互数据的局部时序结构,语义空间的深层特征结构.在真实的三个教育数据集Junyi,ASSISTments09,ASSISTments12上进行了大量的实验,实验结果表明SG-DiffRec优于现有的流行序列推荐模型.
Most existing concept-aware sequential recommendation methods mainly rely on ID embeddings or raw textual semantics,while neglecting semantic noise in text,prerequisite-successor relations among knowledge concepts,and learners dynamically changing knowledge mastery.As a result,recommendation results often fail to match learners current cognitive states.To address these limitations,SG-DiffRec,a sequential recommendation model that integrates structure-guided diffusion with knowledge tracing is proposed.Specifically,SG-DiffRec incorporates the topological structure of a knowledge graph into the semantic generation process of diffusion models,guiding representations toward pedagogically consistent structural patterns and producing embeddings with both semantic richness and structural consistency.In addition,a deep knowledge tracing module is introduced to capture learners dynamic mastery states,while a multimodal attention mechanism adaptively fuses static knowledge attributes with dynamic learner states.Extensive experiments on three real-world educational datasets,Junyi,ASSISTments09,and ASSISTments12,demonstrate that SG-DiffRec outperforms several competitive sequential recommendation methods.
宗宇航;罗昌银;李书豪;陈凌飞;张思
华中师范大学 计算机学院,湖北 武汉 430074华中师范大学 计算机学院,湖北 武汉 430074华中师范大学 计算机学院,湖北 武汉 430074华中师范大学 计算机学院,湖北 武汉 430074华中师范大学 人工智能学部,湖北 武汉 430074
信息技术与安全科学
序列推荐扩散模型知识图谱深度知识追踪
sequential recommendationdiffusion modelknowledge graphdeep knowledge tracing
《中南民族大学学报(自然科学版)》 2026 (6)
806-816,11
国家自然科学基金资助项目(62577026)
评论