Lightweight GCN Predicts Classroom Grades with High Accuracy

Lightweight AI graph model predicts classroom grades with high accuracy

Key finding: A lightweight Graph Convolutional Network (GCN) that combines individual student attributes and social interaction data can predict four-class classroom performance with an area under the curve (AUC) of about 0.91–0.92, offering a practical route for more objective classroom assessment.

Most important details up front

The study, titled Application of artificial intelligence graph convolutional network in classroom grade evaluation. was published in Scientific Reports (volume 15, Article 32044) with DOI 10.1038/s41598-025-17903-4. The article was received on 12 June 2025, accepted 28 August 2025 and published 01 September 2025. The authors report a GCN model that treats students as graph nodes and uses weighted edges derived from classroom cooperation, online interaction and peer ratings to predict student performance divided into four classes: Excellent (≥90), Good (80–89), Qualified (70–79) and To be improved (<70).

Why it matters

The approach addresses a longstanding problem in classroom assessment: teacher‑centred grading can miss peer influence and interaction patterns that shape learning. By combining logs from teaching systems, observation records and online platform behavior into a social graph, the model captures both individual and social signals. The authors claim the GCN improves objectivity and classification performance compared with traditional rule‑based and common machine learning methods.

Data and ethics

Data came from 12 classes across four grade levels in two primary/secondary schools in Shenzhen over two semesters. An initial 802 student records were collected; after cleaning (excluding records with >30% missing data), 732 valid records remained. The final graph contains 732 nodes and 5,184 edges with an average node degree of 14.16. All online behavior data were collected with authorization from schools and education authorities. The study received ethics approval from Liyuan Foreign Language Primary School in Futian District Ethics Committee (Approval Number: 2023.39498000). Participants provided written informed consent and privacy protections were applied.

Features and labels

Each student is represented by a 16‑dimensional feature vector (feature matrix X of size 732 × 16). The features span three categories: individual attributes, classroom behavior and online behavior. Notable inputs include normalized age, one‑hot encoded gender, historical achievements, attendance rate, self‑rating, classroom speech frequency standardized by class size, group cooperation activity, teacher rating, peer rating, video learning duration, homework timeliness rate, forum posts, platform access frequency, online questioning frequency and click path length. Numerical features were standardized (mean 0, SD 1); categorical variables were one‑hot encoded and missing values filled by multiple interpolation. Labels were formed by fusing mid‑term and final scores (weights α = 0.4 and β = 0.6) and further combined teacher/self/peer evaluations per the paper’s fusion framework, then discretized into the four classes above.

Graph construction

The study proposes a weighted social graph generation method. One main strategy builds edge weights as a weighted sum of three normalized indicators: cooperation frequency in class discussion, interaction frequency on the online platform, and peer rating. Weighting coefficients were set by pre‑investigation teacher judgments as λ1 = 0.4, λ2 = 0.3 and λ3 = 0.3. An alternative method used cosine similarity between high‑dimensional interaction vectors. Other graph variants tested included peer evaluation graphs, Pearson correlation graphs and fully connected graphs for comparison.

Model and training

A lightweight two‑layer GCN was used with hidden sizes [128, 64], ReLU activations and dropout = 0.5 after each layer. Training used cross‑entropy loss with L2 weight decay (0.0005) and the Adam optimizer (initial LR 0.01) with learning‑rate decay. Outputs were node embeddings fed to a fully connected layer producing four‑class probabilities. Experiments used stratified splits (70% train, 15% validation, 15% test) and 5‑fold cross‑validation; reported metrics are averages across five runs.

Performance and comparisons

The GCN achieved an AUC reported as 0.92 in one figure and cross‑validation AUC around 0.91. Averaged classification metrics include precision 88.52%, recall 86.47% and F1‑score 87.32%. Compared baselines included Graph Attention Network, GraphSAGE, SVM, linear regression, decision tree and a rule‑based method. GCN outperformed these, improving F1 by over 13% relative to traditional methods. The model identified the Excellent and To be improved groups most accurately (91% and 86% respectively), while the Qualified group had higher confusion with the lower category.

Ablation and sensitivity

Ablation showed the social graph structure is critical: removing edges cut AUC to 0.68 and accuracy to 71%. Using only individual attributes yielded AUC ≈ 0.74. Graph construction strategy mattered: peer evaluation graphs gave the best AUC (≈0.91), Pearson graphs ≈0.87, and fully connected graphs performed worst (≈0.81). Hidden size [128,64] was chosen as a compromise between performance and computation.

Interpretability, reproducibility and resources

GNNExplainer identified key neighbors and features for sample predictions; for example, frequent group collaborators and features such as class participation, teacher rating and assignment timeliness strongly influenced an “excellent” prediction. Embedding visualizations via t‑SNE showed clustering by performance category. The authors provide implementation details (PyTorch 2.1, PyG 2.3, CUDA 12.1, cuDNN 8.9, Ubuntu 22.04) and hardware specs (dual Intel Xeon Gold 6330, 256 GB RAM, four NVIDIA A100 GPUs). Code management used Git, Conda and SLURM. Datasets are available from corresponding author Shuying Wu on reasonable request via e‑mail: wushuying1234@126.com.

Funding, license and limitations

Funding came from local and regional education projects in Shenzhen and Guangdong. The open‑access article is published under a Creative Commons Attribution‑NonCommercial‑NoDerivatives 4.0 International (CC BY‑NC‑ND 4.0) licence. Authors declare no competing interests. Limitations include reliance on questionnaires and behavioral logs (future work should add multimodal signals), potential scaling challenges on much larger school networks, and the need for extended interpretability measures to support deployment in educational settings.

Bottom line

The paper presents a practical, interpretable GCN framework that fuses multi‑source student and social‑interaction data to produce high‑accuracy four‑class performance predictions. The results suggest social structure matters strongly for classroom evaluation and that a lightweight GCN can be a viable tool to support more objective, dynamic classroom assessment.

Key features at a glance