Session-Based Recommendations

Session-based Recsys · Real-time Session Recommendations

Definition

Session-based recommendations are a class of algorithms that predict a shopper's next interaction from the sequence of actions in the current session, without using a long-term profile.

Why session-based recommendations matter

Most visitors to an online store are anonymous. Signed-in shoppers with accumulated history may be
20–40% of traffic; the other 60–80% arrive without an account or for the first time. Personalized
recommendations are unavailable to them — there is no profile.

Session-based recommendations close that gap: they work from the first action, using only the
current session as a signal source.

The algorithmic stack

How session recommendation algorithms evolved:

Algorithm	Basis	Characteristic
Item-based KNN	Co-view patterns	Simple, interpretable
GRU4Rec	Recurrent networks (RNN)	Accounts for event order
BERT4Rec	Transformer (self-attention)	Understands the whole session context
NARM	RNN plus attention	Balances local and global context

In practice most production systems run simpler algorithms with fast inference — BERT4Rec demands
significant compute on every refresh.

How a session algorithm builds a prediction

Shopper session:
  t=0: opened the card for Nike Air Max 270
  t=1: viewed Adidas Ultraboost
  t=2: added Nike Air Max 270 to the cart

Algorithm:
  → current interest: men's sneakers, urban style, mid-to-high price band
  → the cart signal reinforces the intent to buy Nike
  → recommendations: comparable sneakers, socks, shoe care

Every new event recomputes the intent vector for the session.

Combining with a long-term profile

For signed-in shoppers the optimal architecture is hybrid:

Session signal (high weight): what they are looking at right now — the immediate intent
Long-term profile (medium weight): category preferences, purchase history
Popularity (low weight): the fallback when personal signals are weak

The component weights are configurable and can be A/B tested.

Tip: session recommendations are particularly effective in product page blocks — Similar items,
People also viewed — where the shopper has already declared an intent the algorithm can extend in
a specific direction.

Frequently asked questions

What is the difference between session-based and personalized recommendations?

Personalized recommendations use the whole history: past purchases, category preferences, the affinity profile. Session-based use only the current session. Personalized are more precise for repeat customers; session-based are irreplaceable for anonymous visitors, who are frequently more than 70% of e-commerce traffic.

How do session-based recommendations handle cold start?

That is their main advantage: they need no history. The first action — opening a product page — is already a signal. The algorithm predicts from what people view next patterns aggregated across all users rather than from an individual history.

How fast should session-based recommendations refresh?

Ideally after every event (a product view, an add to cart). Real-time refresh under 100 ms latency is what makes the response feel live. Session models are easier to serve at low latency than heavy collaborative models.

Gravity Field

Session Recommendations in Gravity Field

Gravity Field serves session-based recommendations in real time — the algorithm reads the current actions of an anonymous visitor and recommends accurately from the first interaction.

✦Recommendations from the current session with no sign-in

✦The prediction refreshed in real time on every new event

✦Session and long-term signals combined for signed-in shoppers

✦Recently viewed, Similar items and Frequently bought together blocks

✦A/B testing of recommendation strategies

Explore the platform →