长尾分布下基于层内相似关系的认知诊断模型
2025-07-01王冕张玉红刘菲卜晨阳胡学钢



中图分类号:TP391 文献标志码:ADOI:10. 13705/j. issn. 1671-6841. 2023115
文章编号:1671-6841(2025)03-0035-07
Cognitive Diagnosis Model Based on Intra-layer Similarity with Long Tail Distribution
WANG Mian, ZHANG Yuhong,LIU Fei,BU Chenyang,HU Xuegang (School of Computer Science and Information Engineering,Hefei University of Technology,Hefei 230601,China)
Abstract:Most of currnt cognitive diagnostic models that existed in the past predominantly relied on abundant student response records for diagnosis. However, in reality, the interconnections among students'response records,items,and knowledge concepts exhibited a long-tail distribution. That meant that some students had a limited number of response records,and some items were covered by only a few knowledge concepts. This challenge was posed for model training. To address this issue,a cognitive diagnostic model based on intra-layer similarity relationships was proposed. Using a simple matching coefficient,the similarity coefficients of students,items,and knowledge concepts were calculated based on their response records.This process established intra-layer similarity relationships for students,items, and knowledge concepts. These intra-layer relationships were then utilized by the model,and a relational graph convolutional network was employed to propagate information from head nodes to tail nodes.This approach aimed to improve the sparsity of inter-layer relationships in the tail nodes.A diagnostic function that incorporated knowledge point representations was used for cognitive diagnosis.
Key words:cognitive diagnosis;long-tailed distribution;similarity; intra-layer relationships;graph convolutional networks
0 引言
近年来,人工智能技术应用于教育领域引起了广泛关注。认知诊断是其中的重要任务之一,旨在根据学生做题记录来评估学生对知识点的掌握情况,在教学评价和资源推荐等任务中具有重要作用[1-3] 。
近年来,认知诊断的相关研究取得了丰富的成果,其方法主要分为两类。一类基于传统概率模型如 IRT[4] 、MIRT[5]等方法,使用习题难度、习题区分度和学生的熟练度模拟学生与习题的交互过程。确定性输入、噪声“与\"门模型(DINA)[6结合Q 矩阵和学生的答题记录,对学生的认知状态进行建模。然而,上述模型依赖于人工设计的交互函数,难以捕捉学生和习题之间的复杂交互关系。另一类基于深度学习方法,该类方法基于神经网络,或建模学生和习题之间的交互关系或对学生习题知识点的图结构进行表示,如CDGK[8]、RCD[9]。上述方法表明基于图结构的认知诊断方法具有良好性能。
然而,上述模型的良好性能依赖于充分的数据,包括学生的做题记录、习题和知识点的关联关系等。在实际应用中,这种数据可能是不充分的,在图结构中往往呈长尾分布。例如ASSIST2009—2010(简称ASSIST)数据集中,做题记录多于平均数的学生仅占19% ,而做题记录小于40题的学生却占 45% 。此外,在ASSIST数据集中,关联多个知识点的题目仅占17% ,而关联单知识点的题目占 83% 。这种学生做题记录少或知识点关联信息少的节点被称作学生尾节点或习题尾节点。显然,长尾分布中的尾节点由于具有较少信息难以进行有效建模。……