Graph Neural Networks (GNNs) have become the bedrock of modern recommendation systems. By framing user-item interactions as a bipartite graph, GNNs use iterative message-passing to capture high-order collaborative signals—discovering, for instance, that if you like an item, you might also like items purchased by users with similar tastes.
However, an architectural paradox limits these models. While deep neural networks typically perform better as layers are added, graph-based recommenders often suffer a severe drop in performance when stacked beyond a few layers. This degradation is driven by over-smoothing (or embedding convergence), a phenomenon where node representations lose their distinctiveness and collapse toward a uniform representation.
The Mechanism of Embedding Collapse
To understand why over-smoothing occurs, we must look at what a graph convolution layer actually does. At its core, message passing is a localized averaging operation: a node updates its representation by aggregating the features of its immediate neighbors.
As shown by Lei Chen et al. in Revisiting Graph Based Collaborative Filtering: A Linear Residual Graph Convolutional Network Approach, repeated neighborhood aggregation acts mathematically as a form of Laplacian smoothing. Stacking multiple layers expands the receptive field of each node exponentially. Because user-item graphs are highly interconnected "small-world" networks, the neighborhoods of completely different users begin to overlap almost entirely after only a few hops.
Consequently, iterative message passing forces node embeddings to converge toward a common mean, as noted by Kyung-Min Kim et al. in Tripartite Heterogeneous Graph Propagation for Large-scale Social Recommendation. With every additional layer, the unique variance of individual node representations shrinks, rendering them mathematically indistinguishable.
Why Personalization Suffers
In traditional graph tasks like node classification, a moderate amount of smoothing can be beneficial because it enforces homophily (the idea that connected nodes share labels). In collaborative filtering, however, over-smoothing is catastrophic.
Personalization relies entirely on variance and discriminative power. As Penghang Yu et al. explain in Mind Individual Information! Principal Graph Learning for Multimedia Recommendation, when user profiles become too similar, the system loses the ability to capture unique individual preferences.
This loss of distinctiveness erases the fine-grained preference patterns required to separate items a user genuinely desires from those they do not. Sirui Chen et al. highlight in Rankformer: A Graph Transformer for Recommendation based on Ranking Objective that this convergence directly degrades top-K ranking performance, forcing the recommender to output generic, popularity-driven results instead of personalized discoveries.
The Hub Effect: Popularity Bias and Disassortative Mixing
The structural properties of recommendation graphs accelerate this embedding collapse. Real-world interaction data is heavily skewed, characterized by a long tail of niche items and a small set of highly popular items.
In Search Behavior Prediction: A Hypergraph Perspective, Yan Han et al. identify disassortative mixing as a primary driver of over-smoothing. Popular items act as massive network hubs, connecting to a vast, highly diverse array of otherwise unrelated users. When a GNN propagates messages through these hubs, the distinct features of diverse users are mixed together.
This process is further illuminated by Jiajia Chen et al. in How Graph Convolutions Amplify Popularity Bias for Recommendation?, who demonstrate that iterative message passing inadvertently pulls user embeddings toward popular items in the representation space. Instead of preserving individual niche interests, the deep layers cause user profiles to reflect aggregate global trends, severely amplifying popularity bias and drowning out the long tail.
Architectural Strategies to Prevent Convergence
To exploit deep relational structures without triggering representation collapse, researchers have developed several distinct strategies:
1. Linear Simplification and Residuals
Traditional GNNs inherit non-linear activations and weight matrices from standard deep learning architectures. However, Ronghui Guo et al. demonstrate in A Simple Graph Convolutional Network With Abundant Interaction for Collaborative Filtering that stripping away these redundant non-linearities and incorporating initial residual connections preserves the original node identities. By mapping the initial embeddings directly to deeper layers, the model prevents the historical user context from being washed away during aggregation.
2. Interest Isolation and Subgraph Propagation
Rather than allowing messages to flow globally across the entire graph, some frameworks restrict propagation to localized structures. In Interest-aware Message-Passing GCN for Recommendation, Zhiyong Cheng et al. group users into interest-based subgraphs. By confining high-order message passing to these clusters, the model filters out the noise introduced by dissimilar users and popular hubs, allowing the network to safely scale in depth.
3. Bypassing Convolution via Contrastive Learning
Perhaps the most radical shift involves replacing explicit, multi-layer graph convolutions entirely. In Unifying Graph Convolution and Contrastive Learning in Collaborative Filtering, Yihong Wu et al. prove theoretically that contrastive learning loss functions naturally model high-order connectivity. By maximizing the agreement between different views of the same node, contrastive learning implicitly captures relational structures without the need for physical message-passing layers.
This insight is shared by Xuheng Cai et al. in LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation and further validated by Yi Zhang 0103 in Unveiling Contrastive Learning's Capability of Neighborhood Aggregation for Collaborative Filtering, demonstrating that self-supervised alignment can extract global collaborative signals while completely avoiding the over-smoothing risks inherent to deep geometric architectures.
Key Foundational Work
Revisiting Graph Based Collaborative Filtering: A Linear Residual Graph Convolutional Network Approach (2020) — contributed 3 concept elements to this direction.
Learning to Transfer Graph Embeddings for Inductive Graph based Recommendation (2020) — contributed 2 concept elements to this direction.
Tripartite Heterogeneous Graph Propagation for Large-scale Social Recommendation (2019) — contributed 1 concept elements to this direction.