What it is
This project turns a recommender model experiment into a small, repeatable research loop. The setup is intentionally narrow: Amazon Books, ComiRec-SA, fixed data preparation, and a strict short runtime budget.
Problem
Many recommendation experiments become hard to compare because the data preparation, training budget, and evaluation setup keep drifting between runs. That makes it difficult to know whether a result came from a better idea or a changed protocol.
Approach
The repository splits stable infrastructure from mutable experiment code. prepare.py defines the fixed protocol, while train.py stays as the main place for iteration, model edits, and short-run comparisons.
Why it matters
The project is useful as a minimal testbed for autoresearch-style workflows in recommender systems, where the goal is not a full platform but a clean loop for trying one idea at a time.