Table of Contents

Weeks 1-7

Min Points Max Points
Total 10 25


To pass the semestral project you need to get at least 10 points.
Deadline 30.04.2026 23:59
Download the project template. mlm-semestral-project-weeks-1-7.zip
Please, submit a zip file to Brute including a filled-in Jupyter Notebook with your code/solution. Make sure that:
  1. the notebook is runnable
  2. the results can be replicated
  3. the final cell outputs (that you want to present to us) are saved/visible
Issues regarding the project can be discussed on the forum thread.

Environment

  1. Install uv ⇒ https://docs.astral.sh/uv/getting-started/installation/
  2. Run uv venv in the folder with the notebook
  3. Run uv sync in the folder with the notebook
  4. Activate the env with source .venv/bin/activate
  5. Run uv pip install pyg_lib torch_scatter torch_sparse torch_cluster -f https://data.pyg.org/whl/torch-2.10.0+cpu.html

Part 1: Tabular Models (Week 1, 2, 6 & 7) - Total Max 15 points


Task 1.1. - Max 8 points

  • Compare DecisionTree, RandomForest, and XGBoost models on:
    1. Only the target entity table.
    2. The target entity table extended by automatically generated features through propositionalization methods.
  • Choose appropriate evaluation metric(s)
  • Briefly discuss your setup, the results of the experiments.


Task 1.2. - Hyperparameter optimization - Max 2 points

  • Perform hyperparameter optimization and provide brief analysis.
  • To get maximum points optimize at least two parameters on one model, e.g. max_depth and learning_rate on XGBoost.


Task 1.3. - Explanations - Max 3 points

  • Use explainability methods to uncover how the model makes its predictions.
  • To get maximum points use at least two methods, interpret/discuss the meaning of their outputs, and compare them.


Task 1.4. - TabPFN - Max 2 points

  • Extend comparison from Task 1 with TabPFN and discuss the results (also any issues you encounter).

Part 2: Graph Neural Networks (Week 3, 4 & 7) - Total Max 10 points


Task 2.1. - Max 5 points

  • Implement the RDL pipeline (most of the tools were provided in the lab):
    1. Load the database data
    2. Process the data into tensor representation
    3. Prepare an RDL model
    4. Prepare the task target data and loaders/samplers
    5. Train the model
    6. Evaluate the model's performance over a range of values of at least one hyperparameter
  • Provide brief description of the pipeline, setup and results.
If you struggle with the training loop check the examples from RelBench.


Task 2.2. - Max 3 points

  • Compare the performance of a homogenous GNN on the same data represented by:
    1. A homogeneous graph without features (structure only).
    2. A homogeneous graph with one-hot node type features.
  • Discuss the results in comparison to Task 1 (RDL).
HeteroData object in PyG has to_homogeneous() method that can convert the heterogeneous graph to homogeneous graph.


Task 2.3. - GNN Explainer - Max 2 points

  • Use a graph-specific explainer (GNNExplainer, PGExplainer, etc.) on either the RDL Heterogeneous GNN (Task 1) or the Homomogeneous GNN (Task 2).
  • Interpret/discuss the meaning of the explainer's output.
courses/becm36mlm/projects/weeks1-7.txt · Last modified: 2026/03/27 14:03 by pelesjak