Northwestern and Purdue's RecSysSWT Brings Netflix-Style AI Recommendations to a Single GPU

Long user histories are gold for recommender systems, but training on them at full length has been treated as a luxury reserved for companies with warehouse-scale GPUs. A new open-source release from researchers at Northwestern and Purdue pushes back on that assumption, showing that industrial-style sliding window training can be reproduced on a single university cluster while still delivering measurable ranking gains.

The reproducibility gap in long-context recommenders

Long interaction histories sit at the core of modern recommenders, yet training on the full sequence gets dismissed as impractical under realistic memory and latency budgets. This work argues otherwise and backs it up at academic scale. The technique in question, sliding window training, was popularized by a Netflix paper at RecSys 2024 but shipped without code, without runtime numbers, and without a strategy for handling the million-item vocabularies that make real catalogs painful to embed.

The idea itself is simple. Rather than feeding a model a truncated recent slice of a user's history, you slide a fixed-size context window across the entire sequence over multiple epochs. The model still sees a short context per step, respecting inference latency budgets, but is eventually exposed to the full temporal range of behavior. Nobody outside the original team had a working recipe until now.