面向课堂教学的计算机课程学习路径优化研究

The rapid development of computer science has posed great challenges on arranging its classroom studies.　However, there are only very few researches focused on optimizing the learning pathes of classroom studies by now. This project uses the courses of computer science as examples, and for the first time studies the following three relevant problems jointly: teaching contents sorting, teaching contents allocation,and course timetabling.The innovations of the project are as follows: 1.In domain knowledge modeling, a) to overcome the disadvantage of current machine learning methods of ignoring the deep semantic informations in learning materials, it presents the CNN+HCRF model to deeply and flexibly represent the contextual characteristics of the learning materials; b) to improve the robustness of the model, it uses ensemble learning to combine the results of different experts and machine learning together; 2.in optimization research, c) different from current works which regard the problem as a path optimization problem, it classifies the problem as a hierarchical graph-cut problem with multi constraints based on the modularity, hierarchy of domain knowledge and teaching units, and the graph structure of domain knowledge; d) it studies on designing targeted local search algorithm by developing nested mechanism for neighborhoods combination; e) it studies on constructing new function for evaluating the similarities among solutions based on ensembling different factors in the constraints of the problem, whereupon designes cooperative evolutionary algorithm.

计算机学科的飞速发展对科学、动态的安排其课堂教学内容提出了挑战。但目前学习路径优化主要针对网络环境而非课堂环境。本项目以计算机课程为案例，首次将教学内容的排序与教学内容在学时中的分配、课程表调度三个紧密关联的问题结合研究。其创新在于：一、领域知识建模上，１、针对目前的机器学习方法没有表征学习资料的语义的不足，提出CNN+HCRF的两层机器学习模型，以柔性、深度的建模变长学习资料的上下文特征；２、通过集成学习实现对不同专家评估结果、机器学习结果的自适应加权，以提高领域知识评估的鲁棒性和可靠性。二、问题的优化研究上，3、不同于将问题归为路径优化问题的主流思路，针对知识、学时的模块性和层次性，知识的图状拓扑，将问题分类为多目标约束下的多粒度图分割问题；４、通过建立新的嵌套邻域组合机制为问题设计针对性的局部搜索算法；５、针对问题约束特性构建集成多侧面要素的格局相似度评估函数，及多种群协同演化算法。

基于计算机领域在大国博弈、国民经济、产业革命中的重要地位，研究面向终身学习的计算机领域的学习路径优化方法。研究包括两方面的内容。第一，面向学习路径规划的优化算法研究；第二，计算机领域高度动态的知识体系的自动建模方法研究。针对第一个问题，项目主持人完成了学习路径优化紧密相关、可直接用于课程路径优化的图最大分割（max-cut）方面的研究，在运筹学中科院1区期刊上发表论文一篇，相关演化算法改进了经典的NP难优化问题maxcut的最常用算例集的6个最难算例的世界记录结果，并首次提出了以边为核心的图分割问题的演化算法设计思路。针对第二个问题，研究了基于多标签文本分类的层级正则化知识点自动识别方法，显著改进了直接基于BERT的知识点标注的准确性和题目难度评估的准确性；同时，提出的潜式标签的思路，在多标签文本分类问题上也改进了两个常用算例集的当时最好结果。此外，针对关键词作为知识点表征存在的大量冗余问题，研究从词义、相关性、形态三方面提取特征，通过迁移学习进行去重，取得了良好的效果。