User Modeling: Long-Term Taste and Short-Term Intent

The choice is about how many things one user is allowed to be at once. A single profile embedding is cheap to serve and hard to edit. The middle column is not a hypothetical.

Tmall built a recommender on exactly the complaint that column names. MIND — Multi-Interest Network with Dynamic Routing — makes it the opening premise: “Most of the existing deep learning-based models represent one user as a single vector which is insufficient to capture the varying nature of user's interests.” Chao Li and colleagues published it in 2019. Their answer is to give one user several vectors, produced by a capsule-routing multi-interest extractor and combined by label-aware attention. That is the price of “needs routing among interests”: a routing mechanism and an attention layer, not a footnote.

The method was deployed to serve major online traffic on the Mobile Tmall App homepage. The trade-off has been paid at production scale, not argued about.

An event memory plus gate takes the other route. It keeps recent events separate from durable ones and chooses between them. It can also overreact to noise.

A professional-learning platform kept recommending accounting courses after the user moved into software engineering. Historical completions were abundant. The new goal appeared in only a few recent searches. Nothing in the system was broken. The side with more evidence simply won, which is what a single undifferentiated taste signal is built to let happen.

Long-term taste, current session intent, lifecycle stage, explicit preferences, constraints, and negative feedback should not be collapsed into one permanent vector. Keep them in separate representations, or gate between them. That is what lets the policy adapt while the stable needs stay where they are.

Modeling a user also puts the team under obligations it has to be able to name. Several of them are written law rather than good practice. The General Data Protection Regulation gives the person whose data it is three rights over a profile. Article 15: access to what is held about them. Article 16: rectification without undue delay. Article 17: erasure, including where the data are no longer necessary for the purposes for which they were collected.

Article 16 is the one a personalization team meets first. It turns “let the user fix their profile” into a duty with a clock on it: “The data subject shall have the right to obtain from the controller without undue delay the rectification of inaccurate personal data concerning him or her. Taking into account the purposes of the processing, the data subject shall have the right to have incomplete personal data completed, including by means of providing a supplementary statement.” Article 17's erasure ground is a retention question. Data no longer necessary for the purposes they were collected for — that is aimed straight at a stale taste vector.

On top of the article numbers sit the questions no statute answers for you. How long the profile is retained. Whether a household is kept separate. What stops the system inferring sensitive traits beyond the intended purpose.

The layers differ in how fast they are allowed to move. A stated dietary or accessibility need is a stable constraint and should survive any session. A session intent should not harden into durable taste. The bottom layer is where conflicts between them are recorded instead of being resolved in silence.

A shipped system shows what enforcing that separation actually looks like. YouTube's candidate-generation model does not receive an unbounded history. It is fed a maximum bag size of 50 recent watches and 50 recent searches, over a 1M-video and 1M-search-token vocabulary. On top of that sits an explicit “example age” feature, so the model can tell how old an example is instead of treating all history as equally present. Google published the design in 2016, and named the failure the layers exist to prevent: “Machine learning systems often exhibit an implicit bias towards the past because they are trained to predict future behavior from historical examples.”

The team also changed what the model is asked to predict. Not a randomly held-out watch but the user's next watch, because the held-out design “leaks future information and ignores any asymmetric consumption patterns.”

Bounded history, an age feature and a next-item target. Three separate mechanisms, doing what the diagram draws as three separate layers.

Preference fossilization is what the system does; no contestability is what it prevents the user from doing about it. Old behavior keeps dominating after a life or task change, and no control exists to inspect, reset or correct the state that holds it. The two risks in the middle are the ones teams most often treat as speculative. Both have published measurements attached.

Measure how quickly the system responds to a real change without discarding stable preferences. Synthetic change-point tests, user studies, and profile-reset experiments can reveal inertia and overreaction. Monitor whether user controls have the intended effect. That measurement has already been made in public, at scale.

Mozilla audited YouTube's feedback buttons in September 2022, in a report by Becca Ricks and Jesse McCrosky. The evidence was 567,880,195 video recommendations, donated by 22,722 RegretsReporter participants between December 2021 and June 2022, plus a survey of 2,757 people. “Don't recommend channel” prevented about 43% of unwanted recommendations. “Remove from watch history” about 29%, “Dislike” about 12%, “Not interested” about 11%. Of the surveyed users who used the controls, 39.3% did not feel it had changed their recommendations at all.

The report's own summary is flat: “We determined that YouTube's user controls influence what is recommended, but this effect is negligible and most unwanted videos still slip through.” A peer-reviewed version appeared at the ACM Web Conference in 2026.

Those percentages are what a negative signal actually bought. They are also why the general rule holds: a “show less” action that barely changes exposure undermines both trust and the quality of future feedback.

Two results frame how adaptation itself is measured. GRU4Rec came first, at ICLR 2016 — Hidasi and colleagues modelled a whole session with a recurrent network. Ludewig and Jannach then set that family against nearest-neighbour schemes, in User Modeling and User-Adapted Interaction in 2018. The simpler methods, they found, often performed equally well or significantly better. Short-term intent is easy to model badly, and easy to overmodel.

Four horizons are modeled apart or they are modeled by accident. Session, recent, durable, and lifecycle. Each step below has a counterpart in a system somebody shipped or a rule somebody wrote. Bounded recent history and an example-age feature for the first, Article 16 rectification for the fourth, a shared-account dataset for the fifth. The sequence ends on the change events that break a profile: moving home, changing role, sharing a device, returning after months away.