Real-Time Personalization

Real-Time Recommendations · Contextual Personalization · Session-Based Personalization

Definition

Real-time personalization is the adaptation of recommendations and content to a user's actions in the current session, with a delay of under 100–200 ms.

What real-time personalization is

Most recommender systems run in batch mode: an ML model builds a user profile from accumulated
history and recomputes recommendations periodically. That works for repeat customers with a rich
history behind them.

Real-time personalization solves a different problem — adapting to the shopper’s intent right now,
inside the current session, before the next batch recompute.

User → viewed 3 products in Running Shoes
→ the event reaches the engine (< 10 ms)
→ the next recommendation widget shows running socks,
  GPS watches and running shorts — not a cold bestseller list

Why it matters in e-commerce

For new users — there is no history for a batch model to use. Real time makes it possible to
start personalizing from the first click. Three to five actions in a session give enough signal for
a contextual selection.

For occasion-driven behaviour — someone who normally buys electronics has come in to find a
Mother’s Day gift. The batch profile will keep showing gadgets. Real time picks up the gift context
from session clicks within a few pages.

For high intent — an add to cart, a search for a specific SKU, the “in stock” filter are strong
intent signals that have to be acted on immediately.

Technical architecture

Component	Function	Latency requirement
Event API	Receiving events from the client	Async, does not block the UI
Session store (Redis)	Holding the current session context	< 5 ms read/write
Recommendation engine	Serving a request with session context applied	< 50–100 ms
CDN / edge cache	Caching responses for popular contexts	< 10 ms

Important: real-time personalization does not mean recomputing a heavy ML model on every
click. It means reading the current session context quickly and applying lightweight ranking rules
on top of candidates that are already prepared.

Combining it with the long-term profile

The best result comes from the combination: a long-term preference profile (purchase history,
category affinity) plus session context (what the shopper is doing now). When the two contradict each
other, the session signal wins — explicit intent outranks a historical norm.

Common implementation mistakes

Adapting too aggressively — if the recommendations change completely after every click, the
user loses their bearings. The balance: take context into account, but keep some anchor positions
in place.
Personalization with no fallback — if the real-time engine fails, the widget should show
popular products, not an empty space.
Not measuring session lift — an A/B test of real time against batch has to be set up as a
comparison of two strategies, with attribution scoped to the session.

Frequently asked questions

How is real-time personalization different from batch personalization?

Batch personalization recomputes recommendations periodically — hourly, daily — from accumulated history. Real time reacts to every action inside the session: a shopper views three products in Running and the recommendations on the next page are already adjusted to that intent. Batch is good for long-term preferences, real time for current intent.

Why does real-time personalization matter most for new users?

A new user has no historical data for a batch model to work with. Real time makes it possible to adapt to the signals of the first session — which categories they browse, which products they linger on, what they add to the cart. Even two or three clicks give enough context to improve recommendations within that same session.

What latency is acceptable for real-time personalization?

For on-page recommendations, under 200 ms, and ideally under 100 ms. The total is made up of event processing (under 10 ms), the recommendation API call (50–100 ms) and widget rendering. Above 300 ms the user sees stale content, or a flicker as it is replaced — which is worse than no personalization at all.

How is real-time personalization architected?

The event — click, view, add-to-cart — is sent to the server asynchronously through an event API. The server updates the user's session profile in an in-memory store such as Redis. On the next request the recommendation engine reads the current session context and assembles the selection. The critical component is the low-latency session store, not a recompute of a heavy ML model.

Gravity Field