Traditional recommendation systems rely on a "digital paper trail." By analyzing years of a user’s purchase history, models like Matrix Factorization can predict that a user who bought a camera last year might want a lens today. But in the modern web, this trail is often missing. Due to privacy regulations, guest browsing, or transient interests, a significant portion of users are anonymous.
In these scenarios, the system faces a "cold start" not just for a new item, but for every single session. The challenge of Session-based Recommendation (SBR) is to predict a user's next move using only the breadcrumbs left in the current, short-lived interaction sequence.
From History to Sequence
The fundamental shift in SBR is moving from "who is this user?" to "what is this user doing right now?" Early breakthroughs, such as the work by Balázs Hidasi in Session-based recommendations with recurrent neural networks, treated a clickstream like a sentence. Just as a Recurrent Neural Network (RNN) predicts the next word in a phrase, SBR models use Gated Recurrent Units (GRUs) to capture the evolving intent within a session.
The intuition is simple: the order of clicks matters. A user who looks at a laptop, then a mouse, then a keyboard is likely "office shopping." If they looked at those same items in a different context—perhaps starting with a gaming headset—their intent might be "gaming setup." Lin Zhu demonstrated this sequential power in Session-based Sequential Skip Prediction via Recurrent Neural Networks, showing that even music skipping behavior follows predictable patterns that RNNs can decode.
Overcoming Tunnel Vision with Global Graphs
A major limitation of early SBR models was "tunnel vision." If a session only contains two clicks, there isn't enough internal signal to make a high-quality prediction. To solve this, researchers began looking outside the current session to the "wisdom of the crowd."
Wei Wei and colleagues introduced the SRGI framework in Exploring Global Information for Session-based Recommendation, which builds a "Global Graph" of item transitions across all users. If the current anonymous user clicks on Item A, the model doesn't just look at the current session; it looks at the millions of times other users moved from A to B.
By using Graph Neural Networks (GNNs), models like the GC-SAN proposed by Chengfeng Xu in Graph Contextualized Self-Attention Network for Session-based Recommendation can blend this global transition logic with the specific sequence of the current session. This allows the model to "fill in the blanks" for sparse sessions by leveraging the structural patterns of the entire platform.
Disentangling Why Users Click
Not all clicks are created equal. A user might click on a red dress because they like red, because they like dresses, or simply because that dress is the most popular item on the site. Xiaokun Zhang argues in Disentangling ID and Modality Effects for Session-based Recommendation that we must separate these "ID-based" effects (popularity and co-occurrence) from "modality-based" effects (visual style and text descriptions).
By disentangling these factors, models can better generalize. If a user clicks on several minimalist, black items, the system should infer a "minimalist" intent (modality) rather than just looking for items frequently bought with those specific IDs. This is particularly useful for the "long tail" of items that have very few clicks but share clear visual or textual attributes with popular products.
Reasoning with Large Language Models
The latest frontier in SBR involves moving from pattern matching to semantic reasoning. While a GNN can see that Item A often follows Item B, it doesn't "understand" why. Zhu Sun and the team behind PO4ISR (Large Language Models for Intent-Driven Session Recommendations) leverage the reasoning capabilities of Large Language Models (LLMs) to identify diverse user intents.
Instead of just calculating embedding distances, these systems use LLMs to "reflect" on the session. In the Re2LLM framework by Ziyan Wang (Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation), the model learns from its own mistakes by generating "corrective hints." For example, if the model incorrectly predicted a sports car after a user viewed a minivan, the LLM might generate a hint: "The user is looking for family-oriented vehicles; prioritize safety and space over speed."
This semantic layer allows the system to handle the extreme cold-start problem of anonymous browsing by treating the clickstream as a narrative of human intent rather than just a sequence of numbers.