Most shoppers never log in. To serve them, we must learn to read the 'Digital Breadcrumbs' they leave behind in the current moment.
1Personalization without Identity
In the modern web, privacy and guest browsing mean that a large percentage of users are Anonymous. Standard Collaborative Filtering (which requires a User ID and long-term history) fails here. Session-Based Recommendation treats each visit as a unique, self-contained sequence. By analyzing the transition from Item A to Item B to Item C within seconds, the model can infer the user's current 'Mission' and suggest the most logical Item D to help them complete it.
2The Sequential Engine (GRU4Rec)
GRU4Rec was the first major breakthrough in this field. It uses a Gated Recurrent Unit (GRU)—a type of simplified LSTM—to process the sequence of interactions. Unlike a standard RNN, it is optimized for the 'Sparse' nature of sessions (which might only have 2 or 3 clicks). It learns to maintain a 'Hidden State' that represents the user's current interest, updating it with every click to ensure the recommendations stay perfectly aligned with the user's real-time focus.
3Sessions as Graphs (GNN)
The latest frontier is Session-based Recommendation with Graph Neural Networks (SR-GNN). Instead of a simple line, it treats the session as a path on a global graph of all item-to-item transitions. If thousands of users went from 'Tent' to 'Sleeping Bag', the graph has a strong 'Edge' there. By projecting the current session onto this global graph, the model can find the most statistically likely 'Next Node', providing a level of accuracy that simple sequential models cannot match.
