AI Recommenders Learn When Your Tastes Shift

A recommender that always treats yesterday’s clicks as today’s priorities can be confidently wrong. People browse for a new interest, settle back into an old habit or change their mind within a week, yet many artificial-intelligence systems compress their entire interaction history into a single representation. A new open-access study proposes a more responsive design that keeps two versions of a user in view: a recent stream for immediate intentions and a longer stream for durable preferences. The architecture then decides, separately for each recommendation, how strongly each signal should influence the result. On three public benchmarks, this dual-stream approach generally ranked the correct next item more effectively than earlier recurrent and transformer-based recommenders, although detailed statistical analysis shows that the advantages depend strongly on the density and length of the available interaction history.

The work, presented in Discover Artificial Intelligence, introduces the Dual-Stream Adaptive Sequential Recommendation system, or DSASR. Its central premise is that a user’s behavior contains several timescales at once. The most recent interactions may capture a temporary mood, a current search theme or an impulse purchase, while a longer sequence can reveal a stable preference for a genre, product category or style. Conventional sequential models such as SASRec and BERT4Rec can learn complex dependencies through attention, but they generally process the history through one encoder and must infer the balance between immediate and enduring taste implicitly. DSASR instead gives each timescale its own causal transformer, preventing the short stream from being diluted by older events and allowing the long stream to preserve a broader sense of the user’s established habits.

Technically, the two encoders do not share weights. A short-term encoder reads a compact window of recent item embeddings, while a long-term encoder reads a larger window. Each uses causal self-attention, so a position can attend only to earlier interactions, and both stack two transformer blocks with multi-head attention, feed-forward layers, residual connections and normalization. Item embeddings are shared, but separate positional-embedding tables allow each stream to learn its own temporal structure. The default windows were 10 recent and 50 historical interactions for the MovieLens datasets, while shorter windows of 5 and 20 were used for Amazon Beauty because its user histories average only 8.9 interactions.

The system’s most distinctive component is a “Consistency Gate,” which measures disagreement between the two streams and translates that signal into a personalized blending weight. Preference drift is calculated as one minus the cosine similarity between the short- and long-term representations: a value near zero indicates agreement, whereas a larger value signals a stronger departure from historical behavior. The gate concatenates both representations, the drift value and a user-context vector, then passes the result through a two-layer multilayer perceptron with a sigmoid output. That output lies between zero and one and determines the final user representation. A value close to one gives more weight to recent behavior; a value close to zero privileges the long-term stream. Because the gate is conditioned on individual history rather than using a fixed average, it can respond differently to users whose tastes are stable and those undergoing rapid change.

Demographic information supplies additional context to that decision. The model embeds user identity, gender and occupation, alongside a normalized age value, and projects the combined vector into the gate’s input space. This allows demographic attributes to influence how strongly recent behavior is trusted, rather than serving only as an auxiliary label. The architecture remains operational when such information is absent: Amazon Beauty has no comparable demographic metadata, so dummy values replace them and the gate relies primarily on preference drift. That makes the method more flexible, but it also exposes a central tension. Explicit demographic features can improve personalization, while potentially reinforcing stereotypes embedded in historical behavior or limiting the model when users do not disclose such attributes.

DSASR also changes how the model is trained. It combines binary cross-entropy, which rewards accurate scores for known positive interactions, with Bayesian Personalized Ranking loss, which pushes a true next item above sampled non-interactions. The second term is designed for the top-N nature of recommendation, where getting the correct item near the top matters more than merely classifying it as plausible. During training, each positive item was accompanied by five randomly selected negatives that the user had not previously interacted with. Evaluation used a temporal leave-one-out protocol: the last interaction was reserved for testing, the penultimate for validation, and earlier interactions for training. At test time, the correct item was ranked against 99 sampled unseen items, with performance summarized through Hit Rate at 10, Normalized Discounted Cumulative Gain at 10 and Mean Reciprocal Rank at 10.

On MovieLens 100K, DSASR reached an HR@10 of 0.7253, an NDCG@10 of 0.4650 and an MRR@10 of 0.3627. The corresponding figures on MovieLens 1M were 0.8043, 0.5832 and 0.5130, while the sparse Amazon Beauty dataset produced 0.4353, 0.2686 and 0.2172. The paper reports improvements of 9.08%, 19.32% and 19.19% over a demographic-aware single-stream comparator on MovieLens 100K, and of 4.0%, 6.93% and 8.4% on MovieLens 1M. On Amazon Beauty, the gains over BERT4Rec were 43.9% for HR@10, 44.2% for NDCG@10 and 27.7% for MRR@10. A separate comparison against the LLM-based SEALR system also favored DSASR, but the authors appropriately caution that the candidate pools were constructed differently, making that comparison approximate rather than exact.

Ablation experiments helped isolate what actually mattered. Replacing a single-stream model with two streams and a naïve 50/50 average reduced HR@10 by 1.8% on MovieLens 100K and 0.7% on Amazon Beauty, showing that merely adding a short-term channel is not enough. Substituting the Consistency Gate for fixed averaging recovered the loss and produced larger gains in the initial ablation. More rigorous paired analysis, however, painted a more nuanced picture. On MovieLens 1M, the gate significantly improved all three ranking metrics over fixed fusion after statistical correction, including a 5.54% relative gain in NDCG@10 and a 7.56% gain in MRR@10. The positive differences on Amazon Beauty and MovieLens 100K were smaller and not statistically significant. The gate therefore appears most reliable when users have long, information-rich histories, while its behavior in sparse settings remains promising rather than conclusive.

Sensitivity experiments likewise showed that DSASR is not a single rigid formula. The best temporal windows varied by dataset, a generic learned gate sometimes beat the proposed gate, and fixed fusion remained competitive on some MovieLens metrics. What the proposed gate offers is an interpretable compromise: it combines an explicit measure of preference change with user-specific context instead of learning an opaque fusion coefficient from representations alone. Its clearest advantage over fixed averaging appeared on Amazon Beauty, where it produced the best NDCG@10 and MRR@10 among four tested fusion mechanisms. Adding demographic context improved all three metrics on both MovieLens datasets, with the largest changes on MovieLens 1M. Collectively, these findings frame DSASR less as an uncontested new champion than as a useful blueprint for recommendation systems that must decide when a user’s latest behavior deserves to override their established history.

There are clear reasons not to treat the benchmark results as a forecast of effortless industrial deployment. DSASR was tested on public datasets containing at most about one million ratings, rather than proprietary systems with millions of users and items. Running two transformers adds computational overhead, and the study did not measure distributed training, online inference latency or continuously changing catalogues. The method also remains vulnerable to the cold-start problem, because a new user with too few interactions offers little to either stream. It does not yet use interface signals such as dwell time or skipped items, which could sharpen the distinction between genuine interest and accidental clicks. The researchers point to richer content features, multi-domain recommendation and industrial-scale testing as future directions. The wider lesson is that a recommender need not choose between remembering a user and reacting to them; it must learn when each kind of memory deserves a vote.

Subject of Research: Adaptive AI recommendation systems that model short-term intent and long-term user preferences separately.

Article Title: Dual stream adaptive sequential recommendation

Article References: Ghosh, S., Mandal, S., Basak, R., Prasad, D. K., & Sekh, A. A. (2026). Dual stream adaptive sequential recommendation. Discover Artificial Intelligence, 6(1), Article 1201. https://doi.org/10.1007/s44163-026-02282-0

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02282-0

Keywords: recommendation systems, artificial intelligence, transformers, preference drift, personalization, causal attention, user behavior, ranking metrics, demographic data, machine learning, Dual, stream