Abstract
Based on the construction of the knowledge map of "Data Structure and Programming" course, the data of PTA platform, rain class data, MOOC and other platforms are integrated, and the comparative analysis is carried out according to the different granularity of knowledge points, trying to find out students' learning difficulties and blind spots, so as to provide more specific and effective help for students' learning. At the same time, it provides the basis for the adjustment of the follow-up curriculum exercises, the targeted adjustment of the textbook and the concrete implementation of the teaching work.
Introduction
In the era of "Internet +" education, students interact with a fragmented landscape of MOOCs, programming assistants, and classroom tools. While these platforms generate massive amounts of data, it is often difficult for educators to see the "big picture." Traditional analysis treats course content as a linear list, ignoring the complex, non-linear relationships between concepts—for instance, how understanding a "Binary Tree" is a prerequisite for mastering "Heap Sort."
Researchers at the Army Engineering University of PLA have addressed this challenge by developing a knowledge graph-based framework to identify specific "blind spots" in student learning. By moving beyond broad performance metrics, they aim to understand exactly which nodes in a knowledge network cause friction and how different teaching models (supervised vs. unsupervised) impact the mastery of complex topics.
Methodology
The core of the study is the construction of a Curriculum Knowledge Graph for the "Data Structure and Programming" course. The process involved three primary stages:
- Graph Construction: Entities such as Course, Chapter, Knowledge Point, and Resources were defined. Three relationship types were established: Inclusion (hierarchical structure), Association (related concepts), and Subsequent (prerequisite dependencies like Complete Binary Tree $\rightarrow$ Heap $\rightarrow$ Heap Sort).
- Data Integration: Learning data from 556 students (2017–2021) was aggregated from the PTA (Programming Teaching Assistant) platform, Rain Classroom, and MOOC platforms.
- Comparative Analysis: Performance was mapped onto specific nodes in the knowledge graph to compare three demographics: Military Students (strict mixed teaching), Civil Air Defense Students (primarily offline), and MOOC Social Learners (unsupervised online).
Key Insights: The "Difficulty Divergence"
A critical finding of the research is that granularity matters. Broad knowledge points are often too general to be actionable. By refining the graph into smaller sub-concepts, researchers could pinpoint exactly where students were getting lost.
Data suggests that at low difficulty levels, there is negligible difference between supervised and unsupervised learners. However, as conceptual complexity rises, the performance gap widens significantly. For complex topics like Graph Structures and Hash Tables, the "degree of differentiation" between military and social learners reached as high as 0.45 on a 1.0 scale, highlighting that teacher guidance becomes exponentially more important as difficulty increases.
Results and Performance Hierarchy
The analysis revealed a consistent performance hierarchy: Military Students > Civil Air Defense Students > Social Learners.
Specific findings include:
- Blind Spots: The largest gaps in mastery were found in "Graph Structure," "Hash Tables," and "Sorting," where the performance gap between groups exceeded 0.4.
- Quantitative Differentiation: For difficulty level 3 (e.g., Huffman Trees), the differentiation between military and social learners was 0.255. At difficulty level 4 (Complete Binary Trees), it jumped to 0.45.
- Long-term Success: Since the implementation of this KG-based analysis and subsequent curriculum adjustment in 2018, the rate of "good" student performance has improved by approximately 6% per year.
Conclusion and Future Work
While the current framework effectively maps group performance, the authors acknowledge the need for a "micro-portrait" of individual learners. Future work will focus on using the knowledge graph to provide personalized learning recommendations and deeper data mining of individual behaviors to better reflect the cognitive load of different concepts.
Summary updated