salmon1802/unirank

UniRank is a research toolkit for comparing different AI model designs that power recommendation systems, the kind of technology that decides which videos, products, or posts to show a user next. Companies like TikTok, YouTube, and e-commerce platforms use ranking models to sort through millions of candidates and pick what to show in a given moment. This project gives researchers a common testing environment so they can compare different model architectures fairly, using the same data splits and the same metrics. The specific problem UniRank focuses on is that modern ranking models need to juggle two things at once: they need to understand the relationships between many different features about a user and an item (age, category, context), and they need to track what a user has been doing over time (their history of clicks, likes, shares, and comments). Different research groups have proposed different ways to combine those two concerns. UniRank organizes these approaches into two camps: ones that first summarize the history into a compact form before mixing it with other features, and ones that keep history tokens and feature tokens together and let them interact throughout the whole model. The benchmark includes implementations of more than a dozen published models from industrial research papers. It supports training setups where the model learns from multiple types of user feedback at once, not just click-through rate but also follows, long views, and conversions. For large datasets, the project includes tools to load data in chunks and to run training across multiple GPUs. Three preprocessed datasets are available on Hugging Face for anyone who wants to reproduce experiments without hunting down raw data. The README describes this as a work in progress at version 0.1.0. Setup follows standard Python conventions with a requirements file. Training is launched from a single script that accepts experiment configuration files, and a separate script handles hyperparameter search.