TLDR: HIPHOP is a new AI model for session-based recommendations that predicts what a user will click next in an anonymous browsing session. It improves accuracy by using large language models to understand item meanings, analyzing user interests within a session, and learning from both short-term and long-term patterns across different user sessions, while also filtering out irrelevant information. Experiments show it significantly outperforms previous methods.
In today’s digital world, recommendation systems are crucial for helping us find what we’re looking for, whether it’s a new product, a video, or an article. Among these, Session-based Recommendation (SBR) is particularly important. SBR aims to predict what a user will interact with next, based on their current, anonymous browsing session. This is vital for real-time environments like e-commerce, where user identities might not be known, or historical data is limited.
However, existing SBR models face several challenges. Many focus only on the current session, missing out on valuable insights from past interactions or related sessions. When they do try to include data from other sessions, they often struggle with irrelevant information, which can actually make recommendations worse. Furthermore, most models rely on simple item ID connections, overlooking the rich semantic details—like descriptions, categories, or features—that could provide a deeper understanding of items.
To address these issues, researchers have introduced a new approach called HIPHOP: Hierarchical Intent-guided Optimization with Pluggable LLM-Driven Semantics for Session-based Recommendation. This innovative model aims to provide more accurate and relevant recommendations by tackling the limitations of previous systems.
Understanding Items Better with AI
One of HIPHOP’s core innovations is its ‘Pluggable LLM-Driven Semantic Embedding Module’. This module uses large language models (LLMs) to generate high-quality semantic representations for items. Instead of just relying on an item’s ID, HIPHOP can understand what an item is about from its title, description, and attributes. For example, it can recognize that a ‘sports wristband’ and a ‘smartwatch’ are semantically similar, even if users don’t frequently buy them together. This module is designed to be flexible, meaning it can be easily added to or removed from existing SBR models.
Modeling User Behavior Within and Across Sessions
HIPHOP uses Graph Neural Networks (GNNs) to model how users move from one item to another within a single browsing session. This helps it understand the immediate flow of user interest. Beyond individual sessions, the model also incorporates a ‘Dynamic Multi-Intent Capture Module’. This module identifies various user intentions within a session, recognizing that a user might have multiple interests at once. For instance, someone browsing smartphones might also be looking for phone cases or wireless chargers.
Crucially, HIPHOP introduces a ‘Hierarchical Inter-Session Similarity Learning Module’. This part of the model looks at both long-term and short-term user interests across different sessions. It builds two types of graphs: a global graph for overall session similarities and a local graph focusing on recent item interactions. To prevent irrelevant information from degrading recommendations, an ‘intent-guided denoising strategy’ is applied during this inter-session learning process. This helps filter out noise, like recommending ‘running shoes’ to someone browsing ‘smartphones’.
Refining Recommendations with Contrastive Learning
Finally, HIPHOP enhances its ability to distinguish between relevant and irrelevant items through ‘Contrastive Learning’. This technique optimizes the session representations by making similar sessions closer together in a mathematical space and dissimilar sessions further apart. This robust optimization leads to more precise and reliable recommendations.
Also Read:
- SASRecLLM: A Hybrid Approach for Enhanced Recommendation Systems with LLMs
- A Novel Approach to Evaluating LLM Generalization: Predicting User Behavior
Demonstrated Effectiveness
Experiments conducted on various datasets, including public SBR datasets and new Amazon-derived datasets with rich item semantic information, show that HIPHOP significantly outperforms existing state-of-the-art methods. The research demonstrates that each module contributes to the model’s improved performance, highlighting the importance of semantic embeddings, multi-intent capture, and hierarchical inter-session learning. The pluggable nature of the LLM-driven semantic embedding module also proved its value by enhancing the performance of other SBR models when integrated.
For more technical details, you can refer to the full research paper here.
