S
Surya Kalipattapu
Guest
How intent, context, relationships, and changing tastes redefine personalization
TL;DR: A static representation of “what this user likes” is not enough because relevance depends on what the person is trying to do now. The next generation of personalization will combine durable taste, session intent, social context, and direct user steering—while knowing when old behavior should be forgotten.
Recommendation systems love stable identities. People do not cooperate.
We change our minds. We share devices. We enter new life stages. We binge a topic for a week and never return. We watch something for work that we would never choose for pleasure. We put on children's music for a toddler, research a medical question, revisit a nostalgic show, or suddenly become interested in tennis because a tournament has started.
Yet many personalization systems begin with a compressed question: What does this user like?
A better question is: What is this person likely trying to accomplish in this context, and which parts of their history are relevant to that intent?
Taste has a half-life
Some preferences are durable. Others decay quickly.
A useful user model separates at least four kinds of memory:
- Identity memory: relatively durable affinities, relationships, languages, and accessibility needs
- seasonal memory: interests that recur around events, releases, holidays, or routines
- session memory: the intent expressed by the current sequence of actions
- ephemeral memory: accidental, one-off, or context-specific behavior that should decay fast
Flattening all four into one embedding creates familiar errors. A single curiosity becomes a permanent theme. A child's viewing changes a parent's feed. A research session is mistaken for a new identity. A stale relationship continues to dominate a social tray.
Recency weighting helps, but time decay alone is crude. The system also needs to infer why an action occurred.
Intent is not the same as preference
Netflix's FM-Intent work makes this distinction concrete. Its model predicts latent session dimensions such as whether a member wants to continue something or discover something new, whether the current preference is for a movie or show, and how genre or release recency matters in the session. Those intent predictions then inform next-item prediction.[^11]
This is a meaningful shift. A user's globally favorite show may be irrelevant when the current job is “find a 25-minute comedy before bed.” A beloved artist may be wrong during a focus session. A close friend's old post may matter less than a timely update from a weaker tie.
Intent is the routing layer between history and the current decision.
For product teams, this means defining a small, interpretable intent taxonomy before reaching for a universal representation. On an entertainment home surface, intents might include continue, browse, choose together, revisit, or sample. On a social surface, they might include catch up, discover, converse, create, or share. The taxonomy does not need to be exposed to the user or perfectly labeled. It provides structure for training, evaluation, and product reasoning.
Social relevance is relational
Entertainment recommendation can often treat relevance as a user-content relationship. Social products add another variable: the relationship between people.
In Stories, the identity of the creator may be more important than production quality. In comments, the relationship between commenter, author, viewer, and conversation can change the value of the same text. In sharing, the system is not merely predicting who likes the content; it is predicting who is meaningful as a recipient from this sender.
This creates a richer unit of modeling:
It also changes negative feedback. Skipping a friend's Story may say little about the friendship. Not sharing a post with a frequently contacted person does not mean the relationship weakened. Social signals are sparse, asymmetric, and sometimes deliberately private.
The best social rankers therefore balance content understanding with relationship recency, interaction type, reciprocity, and surface-specific norms. A global user embedding is useful—but it should not erase the graph.
Context is not a feature at the edge
Time of day, device, location, session depth, connectivity, and available time are often treated as extra features appended to a user representation. In many experiences, context should instead determine which version of the user's preferences matters.
Spotify's 2025 work on homepage calibration illustrates this. Historical consumption may suggest a stable mix of music and podcasts, while actual preferences shift by time, day, and device. Spotify used a contextual-bandit approach to predict the appropriate content-type distribution for the current request rather than simply reproducing a long-term average, and reported online gains after deployment.[^12]
The broader principle is:
Cold start is an invitation to ask
New users and new items expose the limits of behavioral personalization. With little interaction history, systems commonly fall back to popularity, geography, coarse cohorts, or explicit onboarding choices.
TikTok has publicly described asking new users for interest categories and otherwise starting with a generalized feed, then adapting quickly from early interactions. It also describes using video information such as captions, sounds, and hashtags in addition to behavioral signals.[^13]
Modern multimodal and language models improve cold start by understanding the semantics of a new item before it has accumulated engagement. But product design remains important. A system can ask directly through:
- Interest selection
- Natural-language requests
- “More like this” and “less like this” controls
- Topic sliders
- Feed reset or history management
- Surface modes such as Following, Discover, Focus, or Kids
Explicit input is not a concession that the model failed. It is high-bandwidth preference data—and often more respectful than forcing the user to train the system through repeated rejection.
Forgetting is a personalization capability
Recommendation teams invest heavily in memory. They should invest equally in forgetting.
TikTok's feed-refresh feature lets a user restart For You recommendations without changing followed accounts, after which new interactions rebuild the feed. The company also offers “not interested” and topic or keyword controls.[^14] Meta has described chronological and Favorites alternatives alongside feedback controls for Facebook and Instagram.[^15]
These features solve more than occasional relevance problems. They acknowledge that the system's inference about a person is provisional.
A strong personalization architecture needs:
- Decay: old signals lose influence at appropriate rates.
- segmentation: work, family, shared-device, and other contexts do not automatically contaminate one another.
- correction: users can negate an inference.
- reset: users can escape a locally stable but unwanted state.
- explanation: the product can communicate the broad reasons for a recommendation.
The right to correct the model is part of relevance.
Build a living hypothesis, not a permanent profile
The language of “knowing the user” can make product teams overconfident. A user representation is not the person. It is a lossy, policy-shaped hypothesis built from partial observations.
The ranking system itself chose many of the items the user could react to. The absence of an interaction may mean absence of interest—or absence of exposure. A recent cluster may be a genuine new passion—or a temporary task.
So the system should carry uncertainty, competing intent hypotheses, and multiple timescales. It should be willing to ask, explore, and forget.
The goal of personalization is not to build the most detailed permanent portrait of a person. It is to make the best respectful decision for the moment while leaving room for the person to become someone else.
Sources and further reading
[^11]: Netflix Technology Blog, “FM-Intent: Predicting User Session Intent with Hierarchical Multi-Task Learning”, 2025.
[^12]: Spotify Research,“Calibrated Recommendations with Contextual Bandits on Spotify Homepage”, 2025.
[^13]: TikTok,“How TikTok recommends videos #ForYou”, 2020.
[^14]: TikTok,“Introducing a way to refresh your For You feed”, 2023.
[^15]: Meta,“How AI Influences What You See on Facebook and Instagram”, 2023.