Breaking Information Cocoons: A Hyperbolic Graph-LLM Framework for Exploration and Exploitation in Recommender Systems

📝 Paper Summary

Hyperbolic Graph Neural Networks LLM-enhanced Recommender Systems

HERec combines hyperbolic graph embeddings with LLM-derived semantic profiles to model hierarchical user preferences and allows users to explicitly adjust the trade-off between exploring new content and exploiting known interests.

Core Problem

Modern recommender systems create 'information cocoons' by over-exploiting similar content in Euclidean space, failing to capture hierarchical user interests or effectively balance exploration with exploitation.

Why it matters:

Users get trapped in feedback loops of familiar content (exploitation), missing diverse items (exploration) that might align with broader interests
Existing hyperbolic models capture hierarchy but lack semantic understanding, while Euclidean LLM methods miss the power-law structure of user-item networks
Current systems lack a principled mechanism for users to control how much they want to explore vs. exploit

Concrete Example: A jazz enthusiast might want to explore rock music, but standard models keep recommending jazz (exploitation). Without a hierarchical structure, the system cannot easily pivot to a sibling genre like 'rock' at the appropriate level of granularity.

Key Novelty

Hyperbolic Graph-LLM Framework (HERec)

Aligns LLM-processed semantic descriptions with collaborative filtering signals directly in hyperbolic space, using adaptive gradient magnitudes to preserve hierarchy
Constructs a binary hierarchy tree from learned hyperbolic embeddings by optimizing Dasgupta’s cost, enabling a hyperparameter-free discovery of structure
Introduces a user-controllable exploration mechanism where recommendations are sampled from ancestor clusters in the hierarchy tree based on a temperature parameter

Architecture

The overall HERec framework, illustrating the parallel processing of collaborative information (Hyperbolic GCN) and semantic information (LLM + Encoder), their alignment in hyperbolic space, and the subsequent hierarchy tree construction for exploration.

Evaluation Highlights

Outperforms best baseline (HICF) by up to 5.49% in Recall@20 on Amazon-Book dataset
Achieves 11.39% improvement in Diversity metrics compared to baselines on Amazon-Book, effectively mitigating information cocoons
Successfully balances utility and diversity simultaneously, whereas baselines typically trade one for the other

Breakthrough Assessment

7/10

Strong integration of semantic (LLM) and structural (hyperbolic) signals with a novel user-controllable exploration mechanism. Significant improvements in both diversity and utility.

⚙️ Technical Details

Problem Definition

Setting: Top-K Recommendation with implicit feedback

Inputs: User-item interaction graph and raw text descriptions of users/items

Outputs: Ordered list of K items for each user

Pipeline Flow

Semantic Processing: Raw Text → LLM → Profiles → Text Encoder → Semantic Embeddings
Hyperbolic GCF: Interactions → Gaussian Init → Tangent Projection → Message Passing → Hyperbolic Embeddings
Alignment: Semantic Embeddings + Hyperbolic Embeddings → Alignment Loss
Hierarchy Construction: Hyperbolic Embeddings → Dasgupta Cost Optimization → Hierarchy Tree
Prediction: Hierarchy Tree + User Embedding → Exploration-Exploitation Sampling → Top-K Items

System Modules

Profile Generator (Semantic Processing)

Generate rich text profiles from noisy raw data

Model or implementation: LLM (Specific model not detailed, likely GPT-series or open source equivalent)

Text Encoder (Semantic Processing)

Convert text profiles into dense Euclidean vectors

Model or implementation: Pre-trained Text Encoder (e.g., BERT-based, though specific variant not specified)

Hyperbolic GCN

Learn user/item embeddings reflecting interaction history in hyperbolic space

Model or implementation: Lorentz model-based GNN

Alignment Module

Project semantic embeddings to hyperbolic space and align with collaborative embeddings

Model or implementation: MLP projector + Hyperbolic projection

Hierarchy Builder

Construct binary tree of items for exploration

Model or implementation: Hierarchical clustering minimizing Dasgupta's cost

Novel Architectural Elements

Semantic-enhanced hierarchical mechanism aligning LLM text embeddings with hyperbolic collaborative embeddings
Automatic hierarchical representation builder optimizing Dasgupta's cost to create a tree structure for exploration control

Modeling

Base Model: Hyperbolic Graph Neural Network (Lorentz model)

Training Method: Joint optimization of Ranking Loss and Alignment Loss

Objective Functions:

Purpose: Optimize recommendation ranking with geometric awareness.

Formally: Hyperbolic margin ranking loss with Hyperbolic-aware Margin Learning (HAML) and Hyperbolic Informative Negative Sampling (HINS).
Purpose: Align semantic and collaborative spaces.

Formally: Hyperbolic distance minimization between projected semantic embeddings and collaborative embeddings.

Key Hyperparameters:

curvature_kappa: 1.0
cluster_proportion_k: 2 (binary tree)

Compute: Not reported in the paper

Comparison to Prior Work

vs. HICF/HGCF: HERec integrates semantic (LLM) signals and provides explicit exploration control via hierarchy trees
vs. LightGCN: HERec operates in hyperbolic space to capture power-law distributions and hierarchy
vs. LLM-based Recs (e.g. ChatRec [not cited in paper]): HERec uses LLMs for representation alignment rather than direct inference, reducing latency

Limitations

Construction of the hierarchy tree adds computational complexity (O(N K^2) or O(N log N) depending on implementation details)
Relies on the quality of LLM-generated profiles; poor text data could introduce noise
Hyperbolic operations are generally more computationally expensive than Euclidean equivalents

Reproducibility

Code: https://github.com/Martin-qyma/HERec

Code is publicly available at https://github.com/Martin-qyma/HERec. The specific LLM used for profile generation and the text encoder architecture are not explicitly named in the text provided.

📊 Experiments & Results

Evaluation Setup

Top-K recommendation on three public datasets

Benchmarks:

Amazon-Book (Product Recommendation)
Yelp2018 (POI Recommendation)
Gowalla (Location Recommendation)

Metrics:

Recall@20
NDCG@20
Diversity (unspecified specific metric definition in excerpt, likely ILD or coverage)
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
HERec consistently outperforms baseline methods on utility metrics (Recall, NDCG) across multiple datasets.
Amazon-Book	Recall@20	0.0510	0.0538	+0.0028
Amazon-Book	NDCG@20	0.0409	0.0428	+0.0019
Yelp2018	Recall@20	0.0696	0.0725	+0.0029
Gowalla	Recall@20	0.1837	0.1912	+0.0075
Diversity analysis shows HERec improves diversity metrics alongside utility, unlike baselines.
Amazon-Book	Diversity	0.079	0.088	+0.009

Main Takeaways

HERec achieves state-of-the-art performance on both utility and diversity metrics simultaneously.
The model effectively boosts recommendations for tail items, as evidenced by the analysis on head and tail item distributions.
Ablation studies confirm that both the semantic alignment and the hierarchical structure contribute to the performance gains.

📚 Prerequisite Knowledge

Prerequisites

Hyperbolic geometry (Poincaré ball or Lorentz model)
Graph Neural Networks (GNNs)
Collaborative Filtering
Basics of Large Language Models

Key Terms

Information Cocoons: A phenomenon where users are exposed only to content similar to what they have already interacted with, limiting diversity

Hyperbolic Space: A geometric space with negative curvature where volume grows exponentially, ideal for modeling hierarchical tree-like structures

Lorentz Model: A specific mathematical representation of hyperbolic space used for stable optimization in neural networks

Dasgupta's Cost: A cost function for hierarchical clustering that measures how well a tree structure preserves the similarity between data points

Tangent Space: A Euclidean vector space that locally approximates a point on a curved manifold, used for performing standard neural network operations

Exponential Map: A function that maps vectors from the tangent space onto the hyperbolic manifold

Logarithmic Map: The inverse of the exponential map, mapping points from the manifold back to the tangent space

Exploration-Exploitation: The dilemma of choosing between the best-known option (exploitation) and testing new options (exploration) to potentially find something better