基于层次观测认知诊断的考生成绩预测方法  

Hierarchical observation-based cognitive diagnosis for predicting examinee performance

作　　者：方寒月 曹浪财 FANG Hanyue;CAO Langcai(Xiamen Key Laboratory of Big Data Intelligent Analysis and Decision-making,School of Aerospace Engineering,Xiamen University,Xiamen 361005,China)

机构地区：[1]厦门大学航空航天学院,厦门市大数据智能分析与决策重点实验室,福建厦门361005

出　　处：《厦门大学学报（自然科学版）》2022年第6期1044-1052,共9页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金(62176225,62171391)。

摘　　要：认知诊断旨在发现考生对特定知识点的掌握程度,从而实现对考生成绩的预测.以嵌入向量的形式表示考生和试题,提出了一个由知识点掌握程度、试题掌握程度、试题观测分数构成的基于层次观测的认知诊断框架(HoCDF)模拟考生的分数生成过程.该框架基于神经网络搭建,通过试题观测分数的差异区分不同的试题类型,从而探索不同试题类型之间的关系并实现联合建模.此外,通过增加不含激活函数的单层网络作为额外通路,HoCDF提高了对考生嵌入向量的学习效果,更加充分地诊断出不同考生对同一知识点的掌握程度差异,从而更加清楚地区分不同考生对同一知识点的掌握程度差异.实验结果表明:HoCDF的性能远优于其他考生建模方法,不仅能成功诊断出考生对知识点的掌握程度,而且在相同训练集比例下,可以更加有效地预测考生成绩.Cognitive diagnoses aim to discover the examinee′s proficiency on specific knowledge concepts so that examinees′performance can be predicted.In this paper,examinees and problems are represented in the form of embedded vectors,and a hierarchical observational-based cognitive diagnosis framework(HoCDF)is proposed to simulate the examinee′s score-generation process that includes Examinees′proficiencies,problem mastery,and problem-observation scores.The framework is built on the basis of neural networks,and distinguishes different problem types through the difference in the problem-observation score,so that the relationship between different problem types and realize joint modeling can be explored.Furthermore,by adding a single-layer network without activation function as an extra pass,HoCDF improves the learning effect of the examinees′embedding,and fully diagnoses the difference in proficiency on a knowledge concept among different examinees,thereby clearly distinguishing the difference in proficiency on a knowledge concept among different examinees.Results show that,far outperforming other examinee modeling methods,HoCDF not only successfully diagnoses the examinee′s proficiency of knowledge concepts,but also can effectively predict the examinee′s performance under the same training set ratio.

关 键 词：认知诊断 层次观测 掌握程度 知识点 成绩预测 神经网络 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]