Most recommendation engines look at everything you've ever liked and quietly ignore the one thing that matters most: the order you liked it in.
Think about what your last action says versus a random one from a year ago. If you just finished part two of a trilogy, the next thing you want is obvious to a human and invisible to a system that treats your history as an unordered bag of items.
So imagine a model that reads your history as a sequence — that learns which past actions predict the next one, the way a language model learns which words predict the next word. Cold-start gets easier. "You watched these in this order, last night" becomes a real signal instead of noise.
That's what I built with a self-attention sequential recommender — a SASRec transformer in PyTorch, trained across movies and music. It attends over your interaction history and predicts what comes next, and on a proper held-out split it lands well above the matrix-factorization baseline I kept next to it for honesty.
If your domain has any time structure at all — and almost all of them do — switching from a set-based model to a sequence-based one is the biggest single lever you're probably not pulling.
I write up the things I build → www.divyakush.com
Top comments (0)