Lightweight GCN Predicts Classroom Grades with High Accuracy

Publication details

The article is titled Application of artificial intelligence graph convolutional network in classroom grade evaluation. It appears in Scientific Reports, volume 15, Article number 32044 (2025). The DOI is 10.1038/s41598-025-17903-4. The manuscript was received by the journal on 12 June 2025, accepted on 28 August 2025, and published on 01 September 2025. The corresponding author for data is Shuying Wu, with the dataset available on reasonable request via email at wushuying1234@126.com. Authorship highlights include Shuying Wu being credited with conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing (original draft and review/editing), visualization, supervision, project administration, and funding acquisition.

Data, scope, and ethics

The study draws on multi-source data from 12 classes across 4 grade levels in two Shenzhen schools. The data span classroom management systems, classroom observation records, and online learning platforms. The online platforms named include a Smart Education Platform for Primary and Secondary Schools of Shenzhen and the xueleyun Teaching Platform. All data collection was conducted with authorization from the schools and educational authorities, with ethics approval from the Liyuan Foreign Language Primary School in Futian District (Approval Number: 2023.39498000). Participants provided written informed consent, and the study followed relevant guidelines and regulations. The article is open access under the Creative Commons Attribution‑NonCommercial‑NoDerivatives 4.0 International (CC BY‑NC‑ND 4.0) license.

Dataset and graph basics

Dataset scope includes data from 12 classes across four grade levels in Shenzhen, collected over two academic semesters. The initial sample contained 802 individual student records, with 732 records retained after cleaning (records with more than 30% missing data excluded). The resulting graph contains 732 nodes and 5,184 edges, yielding an average degree of 14.16 where, on average, each student connects to about 14 peers. The feature matrix is a 732 × 16 array, meaning each node carries a 16‑dimensional feature vector.

Features and preprocessing

The 16 input features fall into three categories: individual attributes, classroom behavior, and online behavior. They include age (normalized), gender (one‑hot), class, historical achievements, attendance rate, self‑rating, classroom speech frequency, 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, standard deviation 1), categorical features were one‑hot encoded, and missing values were filled via multiple interpolation. Classroom speech frequency was standardized by class size to ensure comparability across classes.

Graph construction and model design

The core idea is to treat students as graph nodes and to connect them through weighted edges that encode social interactions and collaboration. The study proposes a specific edge weight combination for a primary method, w_ij^comb, using three indicators: frequency of cooperation in class discussions, interaction frequency on online platforms, and peer ratings of learning communication. Weights are set as λ1 = 0.4, λ2 = 0.3, and λ3 = 0.3, with all indicators normalized to the range [0, 1] before they are combined into A, the adjacency matrix. In addition, an auxiliary approach uses cosine similarity between high‑dimensional behavior vectors to compute edge weights (w_ij^cos). Other graph variants tested include peer‑evaluation graphs, Pearson‑correlation graphs, fully connected graphs, and graphs built from different strategies for comparison.

GCN architecture and training

The researchers used a two‑layer lightweight GCN, with hidden layer sizes of 128 and 64 neurons, respectively. ReLU activation was employed, and dropout probability was set to 0.5 after each layer. The loss function combined cross‑entropy for multi‑class classification with L2 regularization (weight decay of 0.0005). The optimizer was Adam with an initial learning rate of 0.01 and a learning rate decay schedule. Training followed standard gradient descent procedures.

Outputs, evaluation, and baselines

The model outputs per‑node probability vectors ŷ_i ∈ [0, 1]^4 for the four classes. An array of baselines was tested, including Graph Attention Network (GAT), GraphSAGE, Support Vector Machine (SVM), linear regression, decision trees, and a rule‑based method. A random classifier served as a minimal baseline. The dataset was split into training (70% ≈ 512 samples), validation (15% ≈ 110), and test (15% ≈ 110), with stratified sampling to preserve label distribution. Five‑fold cross‑validation was conducted, and averages across five runs were reported to reduce randomness.

Key results

Across the board, the GCN outperformed the baselines. In one figure, the GCN achieved an AUC of about 0.92, while the GAT reached approximately 0.88 and GraphSAGE about 0.85. In cross‑validation (Table 4 in the study), the GCN delivered precision around 88.52%, recall about 86.47%, and F1‑score near 87.32% (all rounded). Compared with traditional methods such as linear regression and decision trees, GCN showed superior performance, with the F1‑score advantage exceeding 13% relative to those baselines. When the training set ratio increased to 80% (Figure 5), the model reached an accuracy of 87.6%, an F1‑score of 87.3%, and an AUC of 0.91, with gains leveling off beyond this point.

Category behavior and ablation findings

Confusion analysis showed the strongest accuracy for identifying “Excellent” (about 91%) and “To be improved” (about 86%). The ablation study revealed that using the full feature set with social structure intact yielded the best AUC around 0.91. Removing the social graph dramatically reduced AUC to roughly 0.68 and lowered overall accuracy to about 71%, underscoring the critical role of social relationships in the model. Features based solely on individual attributes produced an AUC around 0.74, while methods using only interaction features performed in between, illustrating the value of combining data sources.

Graph construction comparison and interpretation

Among the graph variants, the peer evaluation graph (mutual student ratings of each other’s learning communication) yielded the best performance with an AUC around 0.91. The Pearson similarity graph produced about 0.87, and a fully connected graph performed worst at roughly 0.81 due to noise from many edges. This points to the importance of a well‑chosen graph structure that respects social feedback in educational settings.

Model interpretability and practical implications

Interpretability analyses used GNNExplainer to identify influential neighbors and input features for predictions on test samples. In an example where a student was predicted as “excellent,” frequent group collaboration and high teacher ratings were among the most influential factors. Visualization of the learned embeddings with t‑SNE showed distinct clustering by performance category, supporting the model’s discriminative power. Practically, the researchers argue that a two‑layer GCN offers a balance of performance and efficiency, making it more feasible for primary and secondary school settings than heavier GNNs.

Limitations, reproducibility, and future work

The authors acknowledge that the current graph relies mainly on questionnaires and behavioral logs; richer multimodal data (voice, video, facial cues) could enhance future graphs. They also note that training on much larger networks could require more efficient architectures or distributed training. While the article provides detailed experimental settings (software versions, hardware, and data processing steps) to aid reproducibility, full public release of code is not stated. The work also emphasizes that further interpretability enhancements, such as additional attention mechanisms or explainers, would strengthen transparency for educational use.

Funding, licensing, and data sharing

The project is supported by multiple sources including Futian District Primary School Chinese Master Teacher Studio in Shenzhen, Guangdong Province’s 2024 “Hundred, Thousand and Ten Thousand Talents Project” Special Research Project, Shenzhen Achievement Cultivation Project, and local school support from Liyuan Foreign Language Primary School. The article is open access under CC BY‑NC‑ND 4.0, and data can be requested from the corresponding author. The study situates itself within the broader goal of advancing objectivity and personal‑level insight in classroom assessment through graph‑based learning models.

Bottom line

Overall, the study demonstrates that a lightweight GCN approach can meaningfully improve the objectivity and accuracy of classroom grade evaluation by integrating multi‑source data and focusing on the dynamics of social interaction. It provides a practical blueprint for schools considering data‑driven assessment tools and points toward future work in multimodal data fusion and scalable training for larger educational networks.

Key takeaways

• GCN effectiveness: AUC around 0.91–0.92 for four‑class prediction. Variants without social graphs underperform.
• Data scope: 12 classes, 4 grade levels, 2 schools, 732 valid samples.
• Graph design: Edge weights combining cooperation, online interaction, and peer ratings work best when weighted properly.
• Interpretability: GNNExplainer and embedding visualizations help explain predictions.
• Accessibility: Open access with data sharing on request; reproducibility details are provided.

Summary table of key features

Frequently Asked Questions

What is the main finding of the study?

The lightweight Graph Convolutional Network achieves an AUC of about 0.91–0.92 in predicting four-class classroom performance by integrating student attributes and social interaction data.

What data were used?

Data from 12 classes across four grade levels in two Shenzhen schools, with 732 valid student records after cleaning, using features spanning individual attributes, classroom behavior, and online behavior.

How was the graph built?

Edges reflect social interactions: cooperation frequency, online interaction frequency, and peer ratings, combined with weights to form a weighted adjacency matrix; alternative graphs based on cosine similarity and peer evaluation were also tested.

What are the key numbers?

Graph has 732 nodes and 5,184 edges; average degree about 14.16; feature matrix 732×16; final four-class thresholds for Excellent/Good/Qualified/To be improved as provided.

Open access and data sharing?

The article is open access under CC BY‑NC‑ND 4.0; the data can be requested from the corresponding author.

What are limitations and future directions?

Future work may include multimodal data like voice or video, larger scales, more efficient GNNs, and enhanced interpretability.