How haktanpolatt/federated-learning-on-weather-station-data works
Project type and purpose
• This is a research-oriented machine-learning project that implements and evaluates federated learning algorithms on weather-station datasets.
• The goal is to compare classical tree-based models (Decision Tree, Random Forest, XGBoost) under two federated scenarios, using Python scripts and a Jupyter notebook.
File structure
• Top-level folders:
• Data/ – raw CSV data for local “clients” (weather stations)
• Papers/ – reference material (does not participate in runtime)
• Top-level files:
• Scenario-*.py – orchestrate client-side training and server-side aggregation for each model and scenario
• RandomForest.ipynb – exploratory analysis and visualization of federated runs
• README.md, CHANGELOG.md, .gitattributes, .vscode (IDE config)
Components
a. Main components
• Data Sources (weather station CSVs)
• Client-side Training Modules (scripts for DecisionTree, RandomForest, XGBoost)
• Federated Orchestrator / Server (aggregation logic in each Scenario-*.py)
• Central Model Store (holds the global model parameters)
• Evaluation & Visualization (RandomForest.ipynb, logs)
• Configuration / Scripts Repo (the Git repo itself)
b. Relationships & interactions
1. The Federated Orchestrator reads a list of Data Sources.
2. Each Client-side Training Module loads its assigned CSV, trains a local model, and serializes model parameters.
3. The Orchestrator collects parameters from all clients and performs aggregation (e.g., averaging).
4. The new global model is redistributed to clients for the next round.
5. After convergence, the global model is evaluated; results are visualized in the Jupyter Notebook.
c. Architectural patterns / design principles
• Federated Learning pattern: decoupling data storage (clients) from a central parameter server.
• Modular script organization: one script per model and scenario.
• Separation of concerns: data folder vs. code vs. research artifacts (papers).
d. Relevant technologies & libraries
• Python 3.x runtime
• scikit-learn (DecisionTree, RandomForest)
• XGBoost
• pandas / numpy for data handling
• Jupyter for interactive analysis
Detail without overthinking
• Scenario script as a distinct client-model pair if granular detail is desired.
• Data/ folder feeding into clients, and the Notebook pulling from the Central Model Store.