Privacy-Preserving Distributed Link Predictions Among Peers in Online Classrooms Using Federated Learning

Seyyedali Hosseinalipour (Ali Alipour), Anurata Prabha Hridi, Muntasir Hoq, Zhikai Gao, Collin Lynch, Rajeev Sahay, Seyyedali Hosseinalipour, Bita AkramEdm 2025

TL;DR

This paper introduces a novel approach to predict future interactions between students in online classrooms using Federated Learning (FL). By training models collaboratively across different classrooms without sharing raw data, it addresses privacy concerns while improving prediction accuracy through personalized FL models.

What Problem Does This Paper Address?

Online learning platforms generate rich data about student interactions, which can be modeled as Social Learning Networks (SLNs). Analyzing these SLNs can reveal valuable insights into learning patterns and help personalize educational strategies. However, traditional methods for analyzing SLNs rely on centralized training, where all student data is aggregated on a single server. This approach raises significant privacy concerns, especially given regulations like FERPA, and can be impractical in real-world educational settings where data sharing across different classrooms is restricted. Furthermore, training on isolated classroom data limits the model's ability to identify common interaction patterns across multiple learning environments, thus hindering overall performance. This paper tackles the challenge of performing privacy-preserving link prediction in SLNs by leveraging Federated Learning.

Methods

The core of this work lies in applying Federated Learning (FL) to the problem of link prediction in SLNs. FL allows multiple clients (in this case, different classrooms) to collaboratively train a machine learning model without sharing their raw data. Instead, only model updates (parameters) are shared with a central server for aggregation.

The authors propose several key methodological contributions:

  • Federated Learning (FL) for SLNs: They adapt the standard FedAvg FL architecture for the specific task of predicting future student interactions (link formation) in SLNs.
  • Personalized FL (pFL): Recognizing that each classroom has unique interaction dynamics (non-IID data), they implement and evaluate three personalized FL methods:
  • FedAvg+FT: A post-training fine-tuning approach where the global model is further trained on local data.
  • PerFedAvg-HF: A meta-learning approach that aims for a global model adaptable to local datasets.
  • FedALA: An adaptive local aggregation method that uses learnable weight matrices to balance global and local model information.
  • Feature Engineering: They extract six graph-theoretic features (Jaccard Coefficient, Adamic-Adar Index, Resource Allocation Index, Preferential Attachment Score, Cosine Similarity, and Dice Similarity) to represent the similarity between student pairs.
  • Explainable AI (XAI): They employ SHapley Additive exPlanations (SHAP) to understand which features are most influential in predicting link formation and how this importance varies across different classrooms and personalization strategies.

Datasets

The study utilizes five datasets:

  • Four MOOC datasets: 'algo' (Algorithms: Design and Analysis, Part 1), 'comp' (English Composition I), 'shake' (Shakespeare in Community), and 'ml' (Machine Learning).
  • One real-world classroom dataset from a North American university: 'csc' (Software Development Fundamentals).

These datasets cover a mix of STEM and humanities subjects, with 'csc' being a non-MOOC dataset. The data consists of student interaction logs from online forums. Importantly, the datasets are anonymized and do not contain student demographic information.

Key Insights and Ideas

  • Privacy is Paramount: The paper strongly emphasizes that centralized data aggregation for SLN analysis is a significant privacy hurdle in educational settings, making FL a crucial alternative.
  • Personalization is Key for Non-IID Data: Vanilla FL (FedAvg) struggles with the inherent differences between classroom interaction patterns (non-IID data). Personalized FL methods are essential for achieving superior performance.
  • FedALA Shines: The FedALA method consistently outperformed other approaches, demonstrating the effectiveness of adaptive local aggregation in balancing global knowledge with classroom-specific nuances.
  • Feature Importance Varies by Context: The study reveals that the importance of different graph-theoretic features for predicting link formation is not uniform. For instance, Resource Allocation Index is crucial for MOOCs, while Preferential Attachment Score is more dominant in non-MOOC STEM courses. This suggests that educational interventions might need to be tailored based on course type and platform.
  • XAI Unlocks Understanding: SHAP analysis provides deep insights into why certain predictions are made, showing how feature contributions shift with personalization, which is vital for building trust and informing educators.

Key Results

  • Performance Gains: Personalized FL models, particularly FedALA, significantly outperformed the centralized baseline and vanilla FedAvg across all datasets. FedALA achieved an average improvement of 1.6% in accuracy and 4.2% in mean AUC compared to the centralized model.
  • Fairness Improvements: FedALA showed better consistency in True Positive Rates (TPR) across classrooms, indicating improved adherence to the "equal opportunity" fairness measure. While FPR consistency was less conclusive, the TPR improvement is a significant step towards fairer models.
  • Feature Impact: SHAP analysis revealed that features like Jaccard Coefficient, Cosine Similarity, and Resource Allocation Index positively influenced link prediction in the 'algo' classroom, while low Preferential Attachment Score and Adamic-Adar Index had a negative impact. The relative importance of features shifted notably between the centralized and FedALA models.

Limitations and Future Work

  • Lack of Demographic Data: The study acknowledges that the absence of student demographic information prevented them from exploring potential biases against underrepresented groups and fine-tuning models for specific demographic segments.
  • Future Directions:
    • Investigating the direct correlation between predicted links and actual student learning gains.
  • Exploring few-shot learning techniques to adapt models to smaller or underrepresented classrooms.
  • Leveraging Large Language Models (LLMs) for generating human-readable explanations, extracting richer semantic features, and suggesting personalized interventions.
  • Addressing potential biases and improving fairness, especially for underrepresented student groups.

Summary updated

BibTeX
@inproceedings{alipour2025privacy,
  author = {Seyyedali Hosseinalipour (Ali Alipour) and Anurata Prabha Hridi and Muntasir Hoq and Zhikai Gao and Collin Lynch and Rajeev Sahay and Seyyedali Hosseinalipour and Bita Akram},
  title = {{Privacy-Preserving Distributed Link Predictions Among Peers in Online Classrooms Using Federated Learning}},
  booktitle = {Edm},
  year = {2025},
}
Figure