Original title: Dual stream adaptive sequential recommendation
Swastika Ghosh, Sourav Mandal, Rohini Basak, Dilip K. Prasad +1 · September 2026
Recommendation systems try to predict what you will want next based on items you clicked, rated, or bought before. However, people often balance long-standing habits with sudden new interests. Researchers built a system named DSASR to solve this. It processes your choices through two separate streams using transformers, which are AI models that track patterns in sequences. One stream learns your recent, short-term interests. The other stream learns your stable, long-term history. Then, a tool called a Consistency Gate measures preference drift, which is how much your recent taste has shifted away from your past habits. It combines this shift with details like your age, gender, occupation, and user identity to blend the two streams together for you. The researchers found that DSASR beat existing systems across three datasets. It worked better because its personalized blend weight customizes recommendations to each user.
Imagine you have a friend who loves both action movies and quiet dramas. One weekend they binge five action films in a row. You need to recommend a movie for Saturday night. If you only look at their last five picks, you'd suggest another action film. But if you only look at their all-time favorites, you'd suggest a drama. The smart move is to check how much their recent mood has shifted from their usual taste, then blend both signals. In this paper, the system does exactly that—it tracks recent clicks and long-term history, measures the drift, and mixes them differently for each user, also factoring in their age and gender.
When you shop or stream, most recommendation systems treat your last few clicks as your only signal, ignoring that you still love old favorites. This system keeps both—your recent mood and your stable identity—and blends them per person. That means fewer irrelevant suggestions and more finds that actually fit you.
The abstract does not discuss computational cost, training time, or scalability to very large datasets. It also does not test on non-consumer domains or compare against more recent transformer-based recommenders beyond those mentioned. The demographic attributes used (age, gender, occupation) may not capture all relevant user identity factors.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Dual stream adaptive sequential recommendation — Swastika Ghosh, Sourav Mandal, Rohini Basak et al. — Discover Artificial Intelligence — doi:10.1007/s44163-026-02282-0
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.