Lightweight GCN Predicts Classroom Grades with High Accuracy

Shenzhen, China, September 1, 2025

News Summary

A lightweight two-layer Graph Convolutional Network (GCN) can predict four levels of classroom performance with strong accuracy by combining student attributes and social interaction data. Tested on a cleaned dataset of 732 students and a social graph of 5,184 edges, the model uses a 16-feature input matrix and achieves AUC scores near 0.91–0.92 and an F1 around 87%. The approach outperforms GAT and GraphSAGE, and ablation shows social ties are critical. The study highlights interpretability via GNNExplainer, notes limits in scale and multimodality, and recommends ethical adaptation before wider deployment.

Lightweight GCN predicts classroom grades with AUC ≈0.91–0.92, study finds

Researchers report a compact graph convolutional network that predicts four-class classroom performance with high accuracy and an area under the curve near 0.91–0.92. The study, titled Application of artificial intelligence graph convolutional network in classroom grade evaluation, appears in Scientific Reports (Sci Rep), volume 15, Article 32044 (2025) and is identified by DOI 10.1038/s41598-025-17903-4. The paper was received by the journal on 12 June 2025, accepted 28 August 2025 and published 01 September 2025.

What the team did and why it matters

The study proposes a classroom performance evaluation model that treats each student as a node in a graph and uses student interactions as weighted edges. By combining individual student data and social-interaction signals from multiple sources, the model aims to make classroom grade assessment more objective and accurate than traditional teacher‑centric methods. The corresponding author and data contact is Shuying Wu (email: wushuying1234@126.com); datasets are available on reasonable request.

Quick results

• Final clean dataset: 732 student records (initially 802, records with >30% missing removed).
• Graph size: 732 nodes and 5,184 edges; average node degree ≈ 14.16.
• Feature matrix: dimension 732 × 16 (16 input features combining personal, classroom and online behaviors).
• Model architecture: lightweight GCN with two hidden layers ([128, 64]), ReLU activations and 0.5 dropout.
• Performance: GCN AUC ≈ 0.92 in some reports and averages near 0.91; precision ≈ 88.5%, recall ≈ 86.5%, F1 ≈ 87.3%.

Data sources and labels

The model uses multi-source data: school teaching management systems, classroom observation logs, and online learning platforms, including the Smart Education Platform for Primary and Secondary Schools of Shenzhen and the xueleyun Teaching Platform. Data were gathered from 12 classes across four grade levels in two schools in Shenzhen over two semesters.

The supervised label fused mid-term and final exam scores with weights (mid-term α = 0.4, final β = 0.6) to build a 0–100 score. That fused score was discretized into four classes: Excellent (≥90), Good (80–89), Qualified (70–79) and To be improved (<70). Teacher, self and peer evaluations were further combined by a fusion framework described in the study.

Features and preprocessing

Each student node had a 16‑dimensional feature vector that included age, gender (one‑hot), class, historical achievements, attendance, self‑rating, classroom speech frequency (standardized by class), group cooperation activity, teacher rating, peer rating, video learning time, homework timeliness, forum posts, platform access frequency, online questioning, and click path length. Numeric features were standardized (mean 0, SD 1), categorical variables were one‑hot encoded, and missing values were handled with multiple interpolation.

How the social graph was built

The core idea is to build weighted social graphs. One primary construction combined three normalized interaction indicators: class cooperation frequency f_ij⁽¹⁾, online interaction frequency f_ij⁽²⁾, and peer rating f_ij⁽³⁾. Teacher judgments set weights λ1 = 0.4, λ2 = 0.3, λ3 = 0.3 for the combined edge weight w_ij^comb. Alternatives included cosine similarity of interaction vectors, peer evaluation graphs, Pearson correlation graphs and fully connected graphs. Experiments showed peer evaluation graphs often produced the best AUC (≈0.91) while fully connected graphs performed worst (≈0.81), showing that carefully constructed social edges matter.

Modeling and training

The lightweight GCN used a self‑looped normalized adjacency and degree matrices for propagation. Training used cross‑entropy loss with L2 weight decay (0.0005), Adam optimizer (initial LR 0.01 with decay), dropout 0.5, and was trained on stratified splits (70% train, 15% val, 15% test) with 5‑fold cross‑validation. Hardware included dual Intel Xeon Gold CPUs, 256 GB RAM and four NVIDIA A100 GPUs; software stack used PyTorch 2.1, PyG 2.3, CUDA 12.1, cuDNN 8.9, NetworkX, scikit‑learn and other common tools.

Key comparisons and analysis

Against graph baselines (GAT, GraphSAGE) and traditional methods (SVM, linear regression, decision trees, rule‑based scoring), the GCN led in AUC and F1. GCN AUC was 0.92 vs GAT 0.88 and GraphSAGE 0.85. Ablation studies highlighted the critical role of social graph structure: removing edges dropped AUC sharply to 0.68 and accuracy to 71%.

Interpretability, limits and ethics

The team used GNNExplainer and visualization (t‑SNE, PCA) to show which neighbors and features influenced classifications; participation frequency, teacher ratings and assignment timeliness often mattered most. The authors note limits: current graphs rely on logs and questionnaires and lack multimodal signals such as audio or video; scaling to much larger networks will need more efficient methods. Ethics review and informed consent were obtained (Liyuan Foreign Language Primary School Ethics Committee, Approval Number 2023.39498000). The authors declare no competing interests and the article is open access under CC BY‑NC‑ND 4.0.

Funding and affiliations

Funding was provided by local and provincial education projects and the Futian District Primary School Chinese Master Teacher Studio in Shenzhen. Institutional work is linked to Liyuan Foreign Language Primary School, Futian District, Shenzhen.

Where to get data and code

Datasets are available from the corresponding author Shuying Wu on reasonable request (wushuying1234@126.com). The paper specifies software versions and environment details to support reproducibility; code was managed under Git with Conda environments, though no public repository link is listed in the article.

---

Frequently Asked Questions

---

Key features at a glance

Feature	Detail
Model	Lightweight Graph Convolutional Network (2 layers, [128,64])
Data sources	Teaching management systems, classroom observation records, Smart Education Platform, xueleyun Teaching Platform
Sample size	732 valid student records; 732 nodes, 5,184 edges
Input features	16 features: personal, classroom, online behaviors (e.g., attendance, peer/teacher ratings, forum posts)
Graph construction	Weighted social graph combining cooperation, online interaction, peer rating (λ1=0.4, λ2=0.3, λ3=0.3); alternatives tested
Performance	AUC ≈ 0.91–0.92; precision ≈ 88.5%; F1 ≈ 87.3%
Ethics & license	Approved by local ethics committee; CC BY‑NC‑ND 4.0 open access
Contact	Shuying Wu — wushuying1234@126.com (data requests)

Deeper Dive: News & Info About This Topic

Additional Resources

• Scientific Reports: Application of artificial intelligence graph convolutional network in classroom grade evaluation (Article)
• Wikipedia: Graph convolutional network
• Scientific Reports: DOI 10.1038/s41598-025-17903-4
• Google Search: Application of artificial intelligence graph convolutional network classroom grade evaluation
• Scientific Reports: Article PDF (s41598-025-17903-4.pdf)
• Google Scholar: graph convolutional network student performance
• Scientific Reports: Figures (s41598-025-17903-4/figures)
• Encyclopedia Britannica: Graph neural network (search)
• Scientific Reports: Article metrics (s41598-025-17903-4/metrics)
• Google News: graph neural network education