Fusing Complementary Views in Graph Representation Learning

21 concepts19 papers1 proposal

Graphs are inherently multi-faceted data structures. To map a graph's nodes into a low-dimensional vector space, an embedding algorithm must capture two fundamentally different types of information: topology (how nodes connect to one another) and attributes (the intrinsic features of the nodes themselves).

Relying on a single encoder or a single hyperparameter configuration often introduces severe inductive biases. For instance, a model optimized to detect tight-knit communities might completely miss individual structural roles, while a model that over-emphasizes local connectivity might fail when encountering "heterophilic" networks where connected nodes are intentionally dissimilar.

To overcome these limitations, a prominent research direction focuses on multi-view embedding fusion. Instead of forcing a single model to capture every nuance of a network, these techniques generate multiple distinct "views" of the graph and systematically fuse them into a unified, robust representation.

Balancing Topology and Attributes

As highlighted in a comprehensive survey by Ke Sun et al. in Network Representation Learning: From Traditional Feature Learning to Deep Learning, modern graph learning requires bridging irregular topological structures with high-dimensional auxiliary attributes. The simplest way to achieve this is through joint encoding. For example, the adversarially regularized frameworks (ARGA and ARVGA) developed by Shirui Pan et al. in Adversarially Regularized Graph Autoencoder for Graph Embedding and Learning Graph Embedding With Adversarial Training Methods utilize graph convolutional layers to process the adjacency matrix and node features simultaneously, forcing the latent vectors to capture both spaces.

However, topology and attributes represent fundamentally different data modalities. Compressing them too early can cause one signal to overwhelm the other. To prevent this, researchers have designed architectures that preserve the independence of each view before fusion:

Ensembling Algorithmic and Structural Perspectives

Multi-view fusion extends beyond the division between topology and attributes; it can also combine different structural scales or algorithmic biases. Different graph embedding algorithms capture orthogonal properties of a network.

To exploit this, Palash Goyal et al. introduced a greedy ensemble framework in Graph Representation Ensemble Learning that combines multiple distinct embedding techniques. By merging views that focus on community membership with views that focus on localized structural roles, the ensemble provides a more comprehensive representation than any single method can manage alone. This strategy is mirrored in biological domains by Chengshuai Zhao et al. in Graph embedding ensemble methods based on the heterogeneous network for lncRNA-miRNA interaction prediction, where combining local and global structural information from multiple embedding spaces significantly improves interaction predictions.

+-----------------------------------------------------------------+
|                       Graph Data Source                         |
+-----------------------------------------------------------------+
                                 |
       +-------------------------+-------------------------+
       |                         |                         |
       v                         v                         v
+--------------+          +--------------+          +--------------+
|   View 1:    |          |   View 2:    |          |   View 3:    |
| Local Nbrs   |          | Struct Peers |          | Attr Clusters|
+--------------+          +--------------+          +--------------+
       |                         |                         |
       v                         v                         v
+--------------+          +--------------+          +--------------+
|  Encoder A   |          |  Encoder B   |          |  Encoder C   |
+--------------+          +--------------+          +--------------+
       |                         |                         |
       +-------------------------+-------------------------+
                                 |
                                 v
              +-------------------------------------+
              |           Fusion Layer              |
              | (Concatenation / Tensor Decomp.)    |
              +-------------------------------------+
                                 |
                                 v
              +-------------------------------------+
              |      Unified Node Embedding         |
              +-------------------------------------+

Even when using a single algorithm, varying the hyperparameters (such as random walk lengths or window sizes) can yield wildly different embeddings. Jia Chen et al. addressed this model-selection dilemma in Ensemble Node Embeddings using Tensor Decomposition: A Case-Study on DeepWalk. Their framework, TENSEMBLE2VEC, uses PARAFAC2 tensor decomposition to fuse diverse node representations generated across different hyperparameter settings, effectively smoothing out the noise and biases of individual runs.

Instead of varying algorithms, one can vary the structural "lens" applied to each node. In Contrastive Representation Learning Based on Multiple Node-centered Subgraphs, Dong Li et al. represent a single node through multiple distinct subgraph perspectives—including its immediate local neighborhood, its structural peers, and its attribute-based clusters. Maximizing the mutual information across these diverse contexts allows the model to learn a highly holistic embedding.

Disentangling and Fusing Specialized Signals

Why does this multi-view approach work so well? The core intuition is that complex graph phenomena are often conflated within a single graph view. By explicitly disentangling these phenomena into separate channels, processing them with specialized filters, and then fusing them, models avoid the information loss inherent in compressed, single-vector architectures.

Overcoming Heterophily

In a standard graph convolutional network, neighborhood aggregation acts as a low-pass filter, smoothing features across connected nodes. This works beautifully for homophilic graphs (where similar nodes connect) but fails catastrophically on heterophilic graphs (where connected nodes are different, such as fraudsters connecting to victims).

To solve this, Zengyi Wo et al. developed GCL-IVG in Graph Contrastive Learning via Interventional View Generation. The framework uses causal intervention to separate homophilic and heterophilic structures into independent views. It processes the homophilic view with a low-pass filter (to capture smoothness) and the heterophilic view with a high-pass filter (to capture sharp differences), concatenating the final embeddings to preserve both signals.

Multi-Modal and Multi-Network Alignment

Multi-view fusion is also essential when graphs are constructed from rich, multi-modal data or span multiple networks:

  • Hyperspectral and Texture Views: In Multi-level Graph Subspace Contrastive Learning for Hyperspectral Image Clustering, Jingxin Wang et al. construct separate graph representations from distinct feature sets—such as spectral-spatial and texture features—fusing them to capture diverse image characteristics simultaneously.
  • Vision-Language Integration: Junlong Ke et al. designed IVSGMV in Integrating Vision-Language Semantic Graphs in Multi-View Clustering, which uses a shared deep auto-encoder to integrate primary node features with relational edge features derived from vision-language semantic graphs (like CLIP), bridging raw visual similarity with deep semantic meaning.
  • Granular Cross-Network Matching: When aligning identical entities across different heterogeneous networks, compressing a node's information into a single vector can blur vital distinctions. In Type-Aware Anchor Link Prediction across Heterogeneous Networks Based on Graph Attention Network, Xiaoxue Li et al. maintain composite node structures consisting of both a global fusion vector and a set of type-specific local vectors. This allows for granular, multi-view similarity comparisons that prevent information loss during cross-network alignment.

Key Foundational Work

Adversarially Regularized Graph Autoencoder for Graph Embedding (2018) — contributed 1 concept elements to this direction.
Network-Specific Variational Auto-Encoder for Embedding in Attribute Networks (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multi_view_embedding_fusion_a50c15f7,
  title = {multi-view embedding fusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-view-embedding-fusion/rdr_a50c15f7b6434a85b943e5cd3518c716},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_multi_view_embedding_fusion_a50c15f7_v1,
  title = {multi-view embedding fusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-view-embedding-fusion/rdr_a50c15f7b6434a85b943e5cd3518c716/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake