Sequential Recommendation: Session Embedding & Next-Item Prediction (2D)

Session Model

Recency weight λ 1.4

Softmax temperature T 0.35

Current Session

—

Session length

Top confidence

–

Prediction entropy

–

Next-Item Prediction

How it works

Each of the 48 catalogue items has a fixed embedding vector e_i ∈ ℝ² (shown as its 2D position, colour-coded by genre cluster). A session-based recommender doesn't just average what a user has clicked — it weighs recent interactions more heavily, the way self-attention / RNN models (GRU4Rec, SASRec) do:

session vector  s = Σᵢ wᵢ · e_{item(i)}  /  Σᵢ wᵢ
recency weight  wᵢ = exp(λ · (i − n))      i = 1..n, most recent = n
next-item score score_j = exp( −‖s − e_j‖² / T )
prediction      P(j)   = score_j / Σₖ score_k    (softmax)

λ (recency weight) — at λ = 0 every past click counts equally, so the session vector is just the mean embedding (pure "bag of items", like classic collaborative filtering). Raise λ and the vector snaps toward the last item clicked — the defining behaviour of sequential recommendation, where order matters and a session can pivot direction after one strong signal.
T (temperature) — a low temperature makes the softmax sharp (one confident winner); a high temperature spreads probability across many plausible next items.
Click a dot — appends that item to the session; a yellow line traces the visit order, and the white cross marks the live session vector s.
Auto-Play — samples the next click from the model's own predicted distribution, showing how a session can self-reinforce toward one embedding neighbourhood, the same feedback loop real recommenders exhibit.

Real systems (YouTube, Spotify, TikTok's "For You" feed) use exactly this shape — a learned sequence encoder producing a session vector, scored against a catalogue of item embeddings — just with hundreds of dimensions and billions of items instead of 2 and 48.

Click a dot to add it to the session

Item embeddings Session path Predicted next item