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

Session-based recommenders (GRU4Rec, SASRec, and the models behind YouTube's and TikTok's next-video queues) don't treat a user's history as an unordered bag of likes — they encode the sequence, weighting recent interactions more heavily so a session can pivot direction after a single strong signal. This simulator places 48 catalogue items as embedding points in a 2D latent space, lets you click items to build a real viewing session, and computes a recency-weighted session vector live. A softmax over the distance from that vector to every unclicked item's embedding produces a genuine next-item probability distribution, visualised as glowing rings sized by confidence — with a recency-weight slider that morphs the model continuously between plain averaging (bag-of-items, like classic collaborative filtering) and sharp last-item-dominant sequential prediction.