Handling Heterophily in Recommender Systems with Wavelet Hypergraph Diffusion

Darnbi Sakong, Thanh Tam Nguyen · 2025

Read Handling Heterophily in Recommender Systems with Wavelet Hypergraph Diffusion by Darnbi Sakong, Thanh Tam Nguyen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-14399.txt.

Introduction

Darnbi Sakong and Thanh Nguyen wrote a 2025 paper on a recommender network built for messy, mixed taste data. Standard recommenders assume birds of a feather; real users often link across unlike items. Their model's name stacks its tricks: fusion, wavelets, and hypergraph diffusion.

One part learns how to pass messages when neighbours differ. Another reads the link web at several scales with wavelet layers. A third blends structure with text about items. On real data sets the stack beat leading methods on accuracy, toughness, and scale. The target is the high order web between users and items that flat models flatten away.

Catalog record

CREATORDarnbi Sakong, Thanh Tam Nguyen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-14399

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Handling Heterophily in Recommender Systems with Wavelet Hypergraph Diffusion Darnbi Sakong, Thanh Tam Nguyen Abstract Recommender systems are pivotal in delivering personalised user experiences across various domains. However, capturing the heterophily patterns and the multi-dimensional nature of user-item interactions poses significant challenges. To address this, we introduce FWHDNN (Fusion-based Wavelet Hypergraph Diffusion Neural Networks), an innovative framework aimed at advancing representation learning in hypergraph-based recommendation tasks. The model incorporates three key components: (1) a cross-difference relation encoder leveraging heterophily-aware hypergraph diffusion to adapt message-passing for diverse class labels, (2) a multi-level cluster-wise encoder employing wavelet transform-based hypergraph neural network layers to capture multi-scale topological relationships, and (3) an integrated multi-modal fusion mechanism that combines structural and textual information through intermediate and late-fusion strategies. Extensive experiments on real-world datasets demonstrate that FWHDNN surpasses state-of-the-art methods in accuracy, robustness, and scalability in capturing high-order interconnections between users and items. Journal: arXiv 1 Introduction Recommender systems often rely on Collaborative Filtering (CF) to predict items that users might prefer by analyzing their past interactions and finding similarities with others who share comparable behaviors. This approach has become essential in improving user experiences across diverse applications, such as social networking platforms [ 1 ] , streaming services [ 2 ] , and online shopping sites [ 3 ] , by filtering through large volumes of content. The advent of graph neural networks (GNNs) has inspired recent advancements in representation learning by leveraging the structural properties of user-item interaction data to capture collaborative signals inherent in graph topologies. GNN-based CF models offer significant advantages for recommendation systems by effectively capturing and representing the intricate relationships between users and items through nodes and edges. This structural representation allows the models to uncover hidden patterns, similarities, and community structures, enhancing the accuracy and relevance of recommendations. For example, models like LightGCN [ 4 ] and UltraGCN [ 5 ] simplify traditional graph convolutional networks (GCNs) to better align with the needs of recommendation systems. Such GNN-based CF models are limited in their ability to capture complex, multi-way relationships, as they primarily focus on pairwise interactions. Furthermore, these models often struggle to represent and leverage higher-order connections, which can result in less nuanced and comprehensive recommendations. To further enhance embedding expressiveness with intricate user-item relationships, hypergraphs have been integrated into the CF framework due to their ability to represent complex group-level interactions between users and items. Unlike traditional binary interaction graphs, hypergraphs connect multiple users to a specific item via hyperedges, encapsulating their multi-way interactions. A notable example is SHT [ 6 ] , which employs a integration of hypergraph transformers to capture global collaborative relationships for recommendation task. While hypergraph-based methods have achieved notable advancements, they often neglect the presence of heterophilic patterns in user-item interaction data. Figure 1 illustrates a scenario where users and items exhibit diverse characteristics – often referred to as a heterophilic pattern . On the top oval, a single user u_{i} interacts with multiple items(e.g., item_{a} , item_{b} , item_{c} , item_{d} ), each belonging to different genres (e.g., Leadership, Technology, Self-help). Meanwhile, on the bottom oval, a single item item_{j} is consumed by various users user_{a} , user_{b} , user_{c} ), each from different age groups (e.g., 45-54, 18-24, 35-44). By highlighting these cross-cutting relationships – one user engaging with diverse item genres, and one item appealing to diverse user demographics – the example emphasizes the multifaceted and heterogeneous nature of user-item interactions. Such patterns make it challenging to capture all relevant and irrelevant signals (e.g., shared preferences, demographic overlaps, cross-genre interests), which is a key motivation for developing models that can effectively handle complex recommendation settings. In practical applications, users typically interact with items from a wide range of categories. To effectively learn representations, items within similar or related categories should have embeddings that are closely aligned, while embeddings for items in different categories should be more distinct. Furthermore, accurately modeling multi-hop neighborhood relationships is essential for capturing complex interactions. However, many existing models rely on deeply stacked Hypergraph Convolutional Networks (HGCN) layers, which often result in over-smoothing, making the embeddings less distinguishable. Overcoming this limitation requires a more sophisticated strategy that naturally integrates local and global connections without deep stacking, while accounting for the diversity and heterogeneity in user-item interaction graphs. Figure 1: Example of heterophilic pattern in user and item hypergraphs. To address these challenges, we propose a novel framework based on the fusion of various modality through wavelet-driven hypergraph diffusion. Specifically, our approach integrates two distinct components that are based on fusion mechanism: a Cross-Difference Relation Encoder to model heterophilic interaction patterns and a Multi-level Cluster-wise Encoder to represent localized structural relationships. The key contributions of this work are summarized as follows: 1. We propose a fusion of wavelet-based hypergraph diffusion framework, called FWHDNN , to capture multi-faceted features while capturing both heterophilic patterns and localized topological structures through the coordinated learning of two specialized encoders. 2. A Cross-Difference Relation Encoder is introduced, utilizing an equivariant operator inspired by ED-HNN [ 7 ] to effectively propagate and distinguish messages to heterogeneous nodes. 3. A Multi-level Cluster-wise Encoder is presented, combining wavelet transforms [ 8 ] with hypergraph convolutional layers to flexibly diffuse information and robustly learn structural patterns in the hypergraph. 4. To integrate features from the two encoders, we implement a multi-level fusion mechanism, incorporating intermediate fusion for partial feature interactions and late fusion to consolidate fully learned representations from each channel. 5. We employ Multi-Encoder Collaborative Supervision which adopts cross-view contrastive learning to ensure consistent embeddings for the same entities across different perspectives, improving the alignment of features from the two encoders. 6. Extensive experiments were conducted on real-world recommendation datasets, demonstrating the superior performance of our model compared to state-of-the-art baselines. 2 Model and Problem Formulation In this section, we begin by presenting the structural data, which includes user-item interactions and user-item hypergaph, followed by a formal definition of heterophilic pattern and our task. Interaction Graph. The interaction graph represents the explicit relationships between users and items. Formally, it is a bipartite graph G=(U,I,E) , where U and I are the sets of users and items, respectively, and E is the set of edges indicating interactions (e.g., clicks, purchases, or ratings) between users and items. Each edge (u,i)\in E may be associated with additional attributes, such as timestamps or interaction strengths, which can enhance the representation of the graph. User-Item Hypergraph. Hypergraphs can be divided into user hypergraph and item hypergraph. The user hypergraph H_{u}=(U,\mathcal{E}_{u}) focuses on relationships among users. Here, U is the set of user nodes, and \mathcal{E}_{u} is the set of hyperedges. Each hyperedge e\in\mathcal{E}_{u} connects a group of users sharing a common interaction pattern, such as engaging with the same item, or exhibiting similar preferences. The item hypergraph H_{i}=(I,\mathcal{E}_{i}) focuses on relationships among items. Here, I represents the set of item nodes, and \mathcal{E}_{i} is the set of hyperedges. Each hyperedge e\in\mathcal{E}_{i} connects a subset of items that are co-interacted with by the same user or group of users. This structure enables the modeling of item co-occurrences, such as items frequently bought together or shared in playlists, and captures their interdependencies in a higher-order manner. By combining the user hypergraph

Read the complete text