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.

New study shows a lightweight Graph Convolutional Network can predict classroom grades with AUC ≈0.91–0.92

Researchers report a practical method that uses a Graph Convolutional Network (GCN) to predict four levels of classroom performance with strong accuracy. The work, published in Scientific Reports (volume 15, Article 32044, 2025; DOI: 10.1038/s41598-025-17903-4), tested a lightweight, two-layer GCN that merges students’ personal attributes and social interaction data into a single model and achieved area under the curve scores near 0.91–0.92.

Key facts and publication details

The paper, titled Application of artificial intelligence graph convolutional network in classroom grade evaluation, was received 12 June 2025, accepted 28 August 2025 and published 01 September 2025. The corresponding author for data requests is Shuying Wu (email: wushuying1234@126.com). Authors declare no competing interests. The work was approved by the Liyuan Foreign Language Primary School in Futian District Ethics Committee (Approval Number: 2023.39498000). The article is open access under a CC BY‑NC‑ND 4.0 license.

What the study set out to do

The research aimed to build an objective classroom performance evaluation model by treating students as nodes in a graph and their interactions as weighted edges. The goal was to combine multiple data sources — including school management systems, classroom observations and online learning logs — to improve the fairness and accuracy of grade assessment.

Data and scope

• Data collected from 12 classes across 4 grade levels in two Shenzhen schools over two semesters.
• Initial records: 802 students; final valid sample after cleaning: 732 records.
• Final graph: 732 nodes and 5,184 edges; average node degree ≈ 14.16.
• Feature matrix X dimension: 732 × 16. Each node has a 16‑dimensional feature vector.

Input features and preprocessing

The 16 input features fall into three categories — individual attributes, classroom behavior and online behavior:

• age (normalized), gender (one‑hot), class, historical achievements, attendance rate, self‑rating
• classroom speech frequency (class‑relative standardized), group cooperation activity, teacher rating, peer rating
• video learning duration, homework timeliness rate, forum posts, platform access frequency, online questioning frequency, click path length

Numerical features were standardized (mean=0, SD=1). Categorical variables were one‑hot encoded. Missing values were handled by multiple interpolation. Classroom speech frequency was standardized by class size for fair comparison.

Labeling and class thresholds

Labels were built from a fusion of mid‑term and final scores (mid‑term weight 0.4, final weight 0.6) and combined teacher/self/peer evaluations via a fusion framework. Scores were turned into four classes: Excellent (≥90), Good (80–89), Qualified (70–79), and To be improved (<70).

Graph construction

Students are nodes in graph G = (V, E) with adjacency matrix A and degree matrix D. The team proposed a weighted social graph method combining multiple behavioral indicators. One key strategy used three normalized indicators for combined edge weight w_ij^comb:

• f_ij^(1): frequency of cooperation in class discussion
• f_ij^(2): interaction frequency on the online platform
• f_ij^(3): peer rating

Weighting coefficients were λ1 = 0.4, λ2 = 0.3, λ3 = 0.3. All f_ij terms were normalized to [0,1] before summing. An alternative used cosine similarity between high‑dimensional interaction vectors. The study also compared peer evaluation graphs, Pearson similarity graphs, fully connected graphs and other variants.

Model and training

The chosen architecture is a lightweight two‑layer GCN with hidden sizes [128, 64], ReLU activation, dropout 0.5 after each layer, and L2 weight decay 0.0005. Adam optimizer with initial learning rate 0.01 and decay was used. Output embeddings fed a fully connected layer to predict four‑class probabilities. Experiments used PyTorch 2.1, PyG 2.3 and common Python tools; hardware included Intel Xeon processors, 256 GB RAM and four NVIDIA A100 GPUs.

Evaluation and main results

Data were split stratified into 70% train (512), 15% validation (110) and 15% test (110) with 5‑fold cross‑validation and averages over five runs. Key results:

• GCN AUC ≈ 0.92 (Fig.3); GAT 0.88; GraphSAGE 0.85.
• Average precision 88.52%, recall 86.47%, F1‑score 87.32%.
• With 80% training data, accuracy 87.6%, F1 87.3%, AUC 0.91.
• GCN best identifies Excellent (91% accuracy) and To be improved (86% accuracy); the Qualified class had more misclassifications.

Ablation and graph comparisons

Removing social graph structure dropped AUC to 0.68 and accuracy to 71%, showing social ties are critical. Peer evaluation graphs produced the best AUC (~0.91). Fully connected graphs performed worst (AUC ~0.81), indicating excess noisy edges harm learning.

Interpretability, limits and next steps

The team used GNNExplainer and embedding visualizations (t‑SNE, PCA) to highlight influential neighbors and features (class participation, teacher ratings, homework timeliness were top factors for high performers). Limitations include reliance on questionnaires and logs rather than multimodal signals, potential scale issues for much larger networks, and the need for deeper interpretability. Future work aims to add multimodal data, more efficient GNNs for scale, and richer temporal or multiview fusion models.

Funding, ethics and access

Support came from local Shenzhen education projects and Guangdong provincial talent programs. Ethical approval and informed consent were obtained. Data are available from the corresponding author on reasonable request (wushuying1234@126.com). The authors report no competing interests.

---

Frequently Asked Questions

---

Table: Key features at a glance

Feature	Details
Publication	Scientific Reports, vol.15, Article 32044 (2025); DOI 10.1038/s41598-025-17903-4
Model	Lightweight GCN, 2 hidden layers [128,64], ReLU, dropout 0.5
Data	732 students; 732×16 feature matrix; 5,184 edges in social graph
Input features	16 features including attendance, teacher/peer ratings, online behavior metrics
Graph strategy	Weighted social graph combining cooperation, online interactions, peer ratings (λ: 0.4, 0.3, 0.3)
Performance	AUC ≈ 0.91–0.92; F1 ≈ 87%; strong identification of Excellent and To be improved
Ethics & access	Ethics approved; data available on reasonable request (wushuying1234@126.com)

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