Skip to content

XwaeK/ML_with_Graphs_Project

Repository files navigation

📘 Machine Learning with Graphs - Final Project Report

Paper Analysis & Experimentation

Institut Polytechnique de Paris (IPP)
Course: Machine Learning with Graphs
Students: Fabien Lagnieu & Kenneth Browder
Date of presentation: April 1st, 2025, at 8:45 AM
Deadline for submission (slides + report): March 31st, 2025


🔹 Project Description

This repository contains all the material for the final project in the Machine Learning with Graphs course (IPP 2025).
We are analyzing and reproducing experiments from the following papers:

  1. Main paper (Presentation + Report):
    "Interpretable Lightweight Transformer via Unrolling of Learned Graph Smoothness Priors"
    ArXiv Link

  2. Related paper (Comparison Section of the Report):
    "Inpainting-Driven Graph Learning via Explainable Neural Networks"
    ArXiv Link


🔹 Task Distribution

Task Responsible Status
Read and summarize Paper 1 Fabien Lagnieu ✅ Done
Prepare the presentation PDF slides Fabien Lagnieu ✅ Done
Reproduce at least one experiment from Paper 1 Kenneth Browder ✅ Done
+ a "new experiment" (2nd mail) Kenneth Browder ✅ Done
Analyze and summarize Paper 2 Fabien Lagnieu ✅ Done
Write Part 1 of the report (Paper 1 summary) Fabien Lagnieu ✅ Done
Implement experiments on the report Kenneth Browder ✅ Done
Write Part 2 of the report (Paper comparison) Fabien Lagnieu ✅ Done
Final review and corrections Both ✅ Done
Upload PDF slides & final report to the platform Fabien Lagnieu ✅ Done

🔹 Deliverables & Deadlines

  • Presentation PDF slides:

    • Deadline: March 31st, 2025
    • Upload Link: Partage IMT - Slides
    • Naming convention: 08h45_April01_KennethBrowder_FabienLagnieu.pdf
  • Final report (ICML format, 4 pages excluding references):

    • Deadline: March 31st, 2025
    • Upload Link: Partage IMT - Report
    • Naming convention: KennethBrowder_FabienLagnieu.pdf
  • Oral presentation:

    • Date: April 1st, 2025
    • Time: 8:45 AM
    • Duration: 10 minutes presentation + 5 minutes Q&A

🔹 Oral Presentation Guidelines

  • Timing: Strictly 10 minutes for both speakers combined.
  • Questions: Prepare short, direct answers (< 1 min per question).
  • Both team members must be ready to answer any technical questions.
  • No reading scripts, speak freely and naturally.
  • Focus on:
    • Technical contributions.
    • Novelty and relevance to the literature.
    • Experimental results supporting the claims.
    • The additional experiment we reproduced (discussion).

🔹 Report Guidelines

The report is divided into two sections:

  1. Paper 1 summary:

    • Main contributions.
    • Related works (brief).
    • Overview of the methodology (GLR, GTV, optimization).
    • Key experimental results.
  2. Paper 2 comparison:

    • Justify the choice of this paper.
    • Brief summary of Paper 2.
    • Relation to Paper 1 (similarities/differences).
  • Format: ICML 2025 template (Overleaf link)
  • Length: 4 pages max (excluding references).
  • Figures/tables: Only the most relevant.

🔹 Experiment Reproduction

  • Reproduce at least one key experiment from Paper 1:

    • Report the experimental setup: settings, code version, datasets used.
    • Analyze results compared to those reported in the paper.
    • If reproduction fails, provide a detailed explanation (e.g., Out-Of-Memory, hardware limitations, or implementation issues).
  • Propose and perform an additional experiment, as required:

    • This is mandatory for the final grade.
    • Include a discussion of the new experiment in the final report -> Appendix !
    • Possible directions ?:
      • Ablation study (e.g., number of layers, regularizer effect).
      • Applying the model to another dataset.
      • Varying hyperparameters (e.g., learning rate, neighborhood size).
      • Testing model robustness (e.g., noise injection).
      • Etc.

🔹 Useful Links


🔹 Contact

About

Final project for the Machine Learning with Graphs course at Institut Polytechnique de Paris. Analysis and experimental reproduction of the paper: “Interpretable Lightweight Transformer via Unrolling of Learned Graph Smoothness Priors”. Includes a comparative study with a second related paper. Authors: Fabien Lagnieu & Kenneth Browder.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors