session-based recommendation
**Session-based recommendation** predicts **what users want next within a browsing session** — analyzing current session behavior (clicks, views, searches) to recommend items in real-time, without requiring user accounts or long-term history, ideal for e-commerce and anonymous browsing.
**What Is Session-Based Recommendation?**
- **Definition**: Recommend based on current session activity.
- **Input**: Sequence of items viewed/clicked in current session.
- **Output**: Next items user likely to interact with.
- **Goal**: Capture short-term intent, immediate needs.
**Why Session-Based?**
- **Anonymous Users**: No login required, works for guests.
- **Immediate Intent**: Capture what user wants right now.
- **E-Commerce**: Shopping sessions have clear goals.
- **Privacy**: No long-term tracking needed.
- **Real-Time**: Adapt to user behavior instantly.
**Techniques**
**Markov Chains**: Predict next item from current item transition probabilities.
**Recurrent Neural Networks**: LSTMs, GRUs learn session sequences.
**Transformers**: Self-attention over session items (BERT4Rec, SASRec).
**Graph Neural Networks**: Model item-to-item transitions as graph.
**Session Features**: Item sequence, dwell time, clicks, add-to-cart, searches, filters applied.
**Applications**: E-commerce (Amazon, eBay), news (Google News), video (YouTube), music (Spotify).
**Challenges**: Short sessions, noisy signals, cold start for first item, session boundaries.
**Tools**: TensorFlow Recommenders, RecBole, GRU4Rec, BERT4Rec implementations.