💬 NLP

Breaking Information Cocoons: A Hyperbolic Framework for Balancing Exploration and Exploitation in Recommender Systems

The paper proposes HERec, a hyperbolic framework that combines semantic-enhanced hierarchical modeling with automatic clustering to effectively balance exploration and exploitation in recommender systems, thereby mitigating information cocoons while outperforming existing Euclidean and hyperbolic baselines.

Original authors: Qiyao Ma, Menglin Yang, Mingxuan Ju, Tong Zhao, Neil Shah, Rex Ying

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qiyao Ma, Menglin Yang, Mingxuan Ju, Tong Zhao, Neil Shah, Rex Ying

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a massive, endless library. In a typical recommendation system, the librarian has a habit of only handing you books that are exactly like the ones you've already read. If you love mystery novels, they only give you more mysteries. Eventually, you end up in a "cocoons" of just one genre, never seeing the sci-fi, history, or poetry sections that might also delight you. This is the problem of Information Cocoons.

The paper introduces a new system called HERec (Hyperbolic Recommendation) designed to break these cocoons. It does this by changing the "shape" of the library and giving you a map that lets you choose how much to explore.

Here is how it works, using simple analogies:

1. The Shape of the Library: Euclidean vs. Hyperbolic

Most current recommendation systems use Euclidean geometry. Imagine this as a flat, square grid. In a flat grid, if you try to organize a massive library with a tree-like structure (like "Fiction" -> "Mystery" -> "Detective" -> "Detective in 1920s"), the branches get crowded and squished together. It's hard to keep the hierarchy clear.

The authors use Hyperbolic geometry. Imagine this as a giant, expanding tree or a coral reef.

The Center (The Trunk): This is where the "popular" or "broad" things live. Think of it as the main aisle where "Music" or "Books" is written.
The Edges (The Branches): As you move away from the center, the space expands exponentially. This is perfect for "niche" items. A specific sub-genre like "1920s Detective Fiction" can have its own spacious corner without crowding the main aisle.

Why it matters: In this "tree-shaped" library, the system can naturally understand that "Jazz" is a child of "Music," and "Bebop" is a child of "Jazz," without everything getting messy.

2. The Problem: The Librarian Doesn't Read

Even with this cool tree-shaped library, previous systems had a blind spot. They only looked at what you clicked (collaborative data) but ignored what the items actually are (text descriptions).

Analogy: Imagine a librarian who only knows you bought a book because you clicked "Buy," but they have never read the back cover. They don't know if the book is a horror story or a romance; they just know you bought it.

HERec's Solution: It brings in a Smart Librarian (Large Language Models).

This librarian reads the book descriptions, user reviews, and titles.
It then translates that "meaning" into the same "tree-shaped" language as the click data.
The Magic: The paper proves mathematically that doing this translation inside the "tree-shaped" (hyperbolic) space is much better than doing it on a flat grid. It aligns the "meaning" of the book with the "clicks" perfectly, helping the system understand why you liked something, not just that you liked it.

3. The New Feature: Your Personal "Exploration Dial"

The biggest innovation is how HERec handles the balance between Exploitation (giving you what you know you love) and Exploration (showing you something new).

In old systems, the balance was fixed by the computer. In HERec, the system builds a hierarchical tree of your preferences automatically, without needing a human to set rules.

The Tree: The top of the tree represents broad categories (e.g., "Music"). The bottom represents very specific tastes (e.g., "1980s Synth-pop").
The Dial: The user gets two controls:
1. How deep to go (Layer): Do you want recommendations from your specific niche (bottom of the tree) or a broader category (top of the tree)?
2. How much to swap (Temperature): If you want to explore, you can tell the system, "Keep 50% of my usual recommendations, but swap the other 50% with items from a slightly different branch of the tree."

The Result: You aren't stuck in a cocoon. You can say, "I love Jazz, but today I want to try something from the 'Rock' branch that is still somewhat related," and the system knows exactly how to find that middle ground.

4. The Results: A Better Library Experience

The authors tested this on real data (like Amazon books and Yelp reviews).

Better Accuracy: It recommended things you actually liked better than the old flat-grid systems.
More Diversity: It successfully broke the cocoons, recommending a wider variety of items (including "tail" items—those niche books or restaurants that usually get ignored).
The Win: It achieved a rare "double win": it was more accurate and more diverse at the same time, improving utility by about 5% and diversity by over 11% compared to the best existing methods.

Summary

HERec is a recommendation system that:

Organizes the world of content in a growing tree shape (Hyperbolic space) instead of a flat grid.
Uses AI to read descriptions and match them to your clicks within that tree shape.
Gives you the controls to decide how much to stick to your favorites versus how much to wander into new territory, all without needing complex settings.

It's like upgrading from a librarian who only hands you the same book over and over, to a guide who knows the entire library's layout, understands the stories inside the books, and lets you choose how far off the beaten path you want to go.

Technical Summary: Breaking Information Cocoons with HERec

Problem Statement

Modern recommender systems frequently create "information cocoons," restricting user exposure to diverse content by over-optimizing for immediate preferences. The core challenge lies in balancing exploration (introducing novel, diverse content) and exploitation (recommending familiar, high-utility items) while allowing users to adjust this trade-off.

Existing approaches face two fundamental limitations:

Euclidean Methods: Standard Graph Neural Networks (GNNs) operate in Euclidean space, which struggles to capture the inherent hierarchical structures and power-law distributions found in user-item interaction networks.
Hyperbolic Methods: While hyperbolic geometry naturally models hierarchical structures, current hyperbolic recommender systems lack semantic understanding of user and item profiles (often relying solely on collaborative signals) and fail to provide a principled, user-adjustable mechanism for balancing exploration and exploitation. Furthermore, integrating Large Language Model (LLM) derived semantic features into hyperbolic space remains challenging due to the mismatch between Euclidean text encoders and hyperbolic collaborative signals.

Methodology: The HERec Framework

The authors propose HERec, a hyperbolic framework designed to unify collaborative filtering, semantic understanding, and hierarchical structure learning. The framework consists of three primary components:

1. Hyperbolic Graph Collaborative Filtering

HERec utilizes the Lorentz model of hyperbolic space to encode user-item interactions.

Initialization: Embeddings are initialized in Euclidean space via Gaussian sampling and projected onto the tangent space of the hyperbolic origin, then mapped to hyperbolic space using the exponential map.
Message Passing: A hyperbolic neighborhood aggregation mechanism updates embeddings by summing neighbor representations in the tangent space before projecting them back to hyperbolic space.
Optimization: The model employs a Hyperbolic Margin Ranking Loss (an extension of Bayesian Personalized Ranking) combined with Hyperbolic-aware Margin Learning (HAML) and Hyperbolic Informative Negative Sampling (HINS). These mechanisms dynamically adjust margins and prioritize negative samples near the origin or positive items to better model head and tail items.

2. Semantic-Enhanced Hierarchical Alignment

To address the lack of semantic understanding, HERec integrates textual profiles generated by LLMs (e.g., item descriptions, user reviews) into the hyperbolic space.

Profile Generation: LLMs generate rich sentence-level descriptions for users and items.
Alignment: A Multi-Layer Perceptron (MLP) aligns the dimensionality of semantic embeddings with collaborative embeddings. These are then projected into hyperbolic space.
Hyperbolic Alignment Loss: A loss function minimizes the hyperbolic distance between the collaborative embedding ( $h_i$ ) and the projected semantic embedding ( $s_i$ ).
Theoretical Advantage: Through gradient analysis (Proposition 1), the authors demonstrate that hyperbolic alignment offers adaptive gradient magnitudes. Nodes with large norms (fine-grained preferences) receive smaller updates to preserve local structure, while nodes with small norms (abstract preferences) receive larger updates to refine global relationships. This contrasts with Euclidean space, which applies constant gradient magnitudes, potentially disrupting hierarchical structures.

3. Automatic Hierarchical Clustering for Exploration-Exploitation

HERec constructs a hierarchy tree from the learned hyperbolic embeddings without predefined hyperparameters.

Clustering Mechanism: The framework iteratively groups embeddings into clusters, using cluster centroids as pseudo-nodes for the next layer. This process is theoretically optimized by minimizing Dasgupta's cost, encouraging edge cuts to be pushed as far down the tree as possible. The algorithm enforces a binary tree structure ( $k=2$ ).
User-Adjustable Trade-off: The resulting hierarchy allows users to control the exploration-exploitation balance via two parameters:
- Temperature ( $\tau$ ): Controls the proportion of original recommendations to replace.
- Hierarchy Layer ( $l$ ): Determines the depth from which alternative items are sampled.
- Mechanism: A portion of top-k recommendations is replaced with items sampled from ancestor clusters at a specified layer. Lower layers (closer to leaves) offer specific, exploitative recommendations, while higher layers (closer to the root) offer broader, exploratory content.

Key Contributions

Semantic-Collaborative Alignment in Hyperbolic Space: The paper introduces a mechanism to align rich textual descriptions with collaborative signals directly in hyperbolic space, supported by theoretical gradient analysis showing superior hierarchical preservation compared to Euclidean alignment.
Parameter-Free Hierarchical Clustering: The authors propose a novel clustering mechanism based on Dasgupta's cost that automatically discovers hierarchical structures without requiring manual hyperparameter tuning, enabling dynamic user control over recommendation diversity.
Unified Performance: HERec is presented as a single model that simultaneously achieves state-of-the-art performance in both utility (accuracy) and diversity metrics, addressing the common trade-off where improving one often degrades the other.

Experimental Results

Extensive experiments were conducted on three datasets: Amazon-Books, Yelp, and Google-Reviews, including large-scale variants.

Utility and Diversity: HERec consistently outperformed both Euclidean (e.g., LightGCN, LightGCL) and Hyperbolic (e.g., HGCF, HICF) baselines.
- Achieved up to a 5.49% improvement in utility metrics (Recall, NDCG).
- Achieved up to an 11.39% increase in diversity metrics (Distance Diversity, Shannon Entropy, Expected Popularity Complement).
Head vs. Tail Items: The model showed significant gains in recommending tail (cold-start) items. Semantic alignment particularly boosted performance for tail items, with improvements reaching 9.71% to 14.42% over hyperbolic baselines without semantic alignment.
Robustness: The framework maintained superiority even when Euclidean baselines were enhanced with feature information, validating the theoretical advantage of hyperbolic alignment.
Ablation Studies: Removing either the margin ranking loss or the semantic alignment loss resulted in decreased utility, confirming the necessity of both components.

Significance and Claims

The paper claims that HERec effectively mitigates information cocoons by providing a principled, user-adjustable mechanism for balancing exploration and exploitation. Unlike previous methods that rely on static trade-offs or manual hierarchy definitions, HERec leverages the intrinsic properties of hyperbolic geometry to:

Capture latent hierarchical structures in user preferences automatically.
Integrate semantic insights to improve cold-start item recommendations.
Allow users to dynamically shift between exploiting known preferences and exploring new categories via the hierarchical tree structure.

The authors position this work as a significant step toward more personalized and diverse recommendation systems, demonstrating that a single model can excel in both accuracy and diversity without compromising either.