PersonalizedFLEducation

Assessing Academic Performance: Personalized Federated Learning for Identifying Overperforming and Underperforming Schools and Individual-Level Implications

Research conducted in Computer Science Department @ Florida State University

Young Scholars Program: https://ysp.osta.fsu.edu

Model Details/Structure

Dataset: Florida Department of Education - School Grades 2022 Dataset

Client Model:

• Type: Deep Neural Network (DNN)
• Layers: 1 input, 4 hidden, 1 output | 16, 32, 64, 32, 16, 1
• Inputs: 16 features (e.g. school grade, % minority/economically disadvantaged, ELA/science Achievement, school type, etc)
• Output: Mathematics Achievement (value from 0-100, used as proxy for school performance)
• Loss: Mean Squared Error, Optimizer: Adam, Learning Rate: 0.001
• Local Epochs: 3, Batch Size: 128

Global Model:

• Type: Personalized Federated Learning
• Aggregating Epochs: 100, Batch Size: 128
• Model Weightage: 70% global, 30% local

Model Results

Coefficient of Determination (r^2) Convergence

• Personalized FL: ~0.86
• FL: ~0.86
• Centralized: ~0.89

Mean Absolute Error (MAE) Convergence

• Personalized FL: ~5
• FL: ~5
• Centralized: ~4.5