Bi-View Embedding Fusion: A Hybrid Learning Approach for Knowledge Graph's Nodes Classification Addressing Problems with Limited Data

AI summary

Overview

Research area: Graph Machine Learning (GML) and Knowledge Graph (KG) representation learning, specifically node classification on graphs with weak or sparse node features.

Technical level: Intermediate. The paper assumes familiarity with graph embeddings, random-walk methods, message-passing neural networks, and standard centrality metrics, but it explains its core ideas in explicit mathematical definitions.

Scope: The paper proposes and empirically tests a hybrid "Bi-View" embedding pipeline that fuses unsupervised Node2Vec structural embeddings with supervised GraphSAGE embeddings (themselves fed enriched features) to improve node classification on a small, highly imbalanced healthcare knowledge graph.

What This Paper Is About

Traditional machine learning needs large amounts of labelled data, and graph-based models on Knowledge Graphs often struggle because the useful information is implicit in the graph's semantics rather than written into node features. The authors' goal is to improve node classification on such "poor-feature" KGs without generating additional synthetic data, by squeezing more signal out of the features that already exist in the graph. Their answer is to combine two complementary embedding techniques — one that sees global topology, one that sees local neighbourhoods — into a single adaptive representation.

Key Contributions

1. A hybrid Bi-View embedding architecture. The pipeline chains five components: a Node2Vec encoder, a centrality vector module, a feature aggregator, a GraphSAGE aggregator, and a learnable embedding fusion module that produces a "Bi-View" representation.

2. Feature enrichment without synthetic data. Node2Vec embeddings are concatenated with classical centrality metrics (degree, PageRank, Betweenness) and used as the input feature vector for GraphSAGE, increasing feature dimensionality using only information already present in the graph.

3. An adaptive, node-specific fusion mechanism. Rather than a fixed concatenation, a learnable coefficient α_i in (0,1) is computed per node via a sigmoid over the concatenated Node2Vec and GraphSAGE vectors, producing a weighted mixture z_fused = α_i · z_n2v + (1 − α_i) · z_sage. A separate supervised fusion model (FusionNet) concatenates both views and projects them through an MLP for classification.

4. Formal validation lemmas. Three lemmas argue that structural enrichment increases neighbourhood discriminability (Lemma 1, via the Pythagorean theorem on the concatenated feature distances), amplifies the neighbourhood gradient during message passing (Lemma 2), and lets the fused embedding space capture richer interactions than either view alone (Lemma 3).

Main Findings

• Bi-View outperforms both single-view baselines. Across Accuracy, Precision, Recall and F1-Score, Bi-View outperformed Node2Vec and GraphSAGE, reaching accuracy close to 94%.

• Node2Vec lands in the middle. Node2Vec's scores were consistently higher than GraphSAGE's but "markedly below" Bi-View.

• GraphSAGE was the weakest baseline. It reported the lowest values on all metrics, with a particularly reduced F1-score of around 64%.

• Latent space structure differs qualitatively. In the PCA visualisations, Node2Vec showed extensive class overlap and poor separability (including for the AgeGroup and Case classes); GraphSAGE arranged embeddings along a "smooth crescent-shaped curve" without compact clusters; Bi-View formed well-separated, compact clusters, particularly for the Case class and moderately for AgeGroup and Outcome.

• Bi-View had the fewest cross-category errors. Per the confusion matrices, Bi-View handled closely related classes such as Drug and Therapy best; Node2Vec did well on Case and Reaction but misclassified drugs and reactions heavily; GraphSAGE was more balanced but still confused Therapy and Reaction.

• Evaluation was deliberately restricted. Because the dataset is highly imbalanced — class samples differ by "up to three orders of magnitude" — the evaluation was limited to four classes with the highest number of instances (stated as Case, Drug, Reaction and Therapy), and metrics were computed on classes with at least 200 samples.

• Excluded classes are not merely rare — they are structurally central. Plots of average PageRank and average Betweenness showed that underrepresented classes have exceptionally high centrality scores, acting as hubs shared across classes; the authors argue including them would inject noise rather than discriminative power.

• A stated inconsistency worth noting: the experimental setup says the goal was to classify the most frequent node classes as "Patient, Drug, Therapy and Manufacturer," while the evaluation section names the four highest-instance classes as "Case, Drug, Reaction and Therapy."

Methodology in Plain English

The authors start from a real Knowledge Graph built from the FDA Adverse Event Reporting System (FAERS), a post-marketing drug safety surveillance database, loaded into Neo4j 5.24.0 with the Graph Data Science (GDS) library version 2.12.0. The graph is small — roughly 11,000 elements — and its classes (Case, Drug, Reaction, Outcome, ReportSource, Therapy, Age Group) are heavily imbalanced.

The pipeline works in stages. First, Node2Vec runs biased random walks over the graph to capture who sits near whom topologically, producing 64-dimensional vectors (walk length 80, 10 iterations, both the return parameter p and in-out parameter q set to 1.0). Separately, the authors compute centrality metrics per node — Betweenness and PageRank — as deterministic structural features. They concatenate the original node attributes, the Node2Vec vector, and the centrality vector into one enriched starting feature vector.

That enriched vector is then fed into GraphSAGE (dimension 64), which does supervised neighbourhood aggregation to produce a second embedding that blends local context with the injected global structure. Finally, a fusion step combines the Node2Vec view and the GraphSAGE view. Two forms of fusion appear in the paper: an adaptive per-node weighted sum controlled by a learned coefficient α_i, and a supervised "FusionNet" — an MLP with hidden dimension 128 and output dimension 64, trained with cross-entropy loss over 100 epochs.

For evaluation, a decision tree classifier was used on top, and three configurations were compared: Node2Vec embeddings only, GraphSAGE embeddings only, and the fused Bi-View embeddings. PCA was used to project the embeddings into 2D for visual inspection of class separability. The authors state the code and hyperparameter configurations will be made public upon acceptance.

Why This Matters

Impact on research. The paper's central claim is that you do not need synthetic data to compensate for sparse features — you can recover latent signal by combining a topological view and a neighbourhood-aggregation view. It also targets a specific gap the authors say is unexplored: explicitly fusing Node2Vec with centrality features as auxiliary input to GraphSAGE within low-feature graph settings.

Real-world applications:

• Pharmacovigilance and drug safety. The FAERS graph under study is precisely the kind of data used to detect adverse event patterns and medication errors.
• Healthcare knowledge graphs. The authors note the dataset was formalised as a KG from a healthcare-analytics example, a common pattern in clinical data integration.
• Fraud and anomaly detection. Relational datasets where labels are scarce but structure is rich are the paper's stated motivation alongside social networks and biological systems.
• Recommendation and entity typing in enterprise KGs. Assigning categories to entities when textual attributes are thin or missing.

Industry relevance. Practitioners working with graph databases such as Neo4j's GDS stack can implement this pipeline with off-the-shelf algorithms. The paper's framing — improving results without buying or generating more labelled data — speaks directly to the cost constraints that dominate applied graph ML projects.

Future Directions

• Test on other graph types. The authors plan to evaluate the strategy on temporal and multi-relational datasets, extending beyond a single static healthcare KG.
• Meta-learning and inductive transfer. They intend to explore transferable meta-knowledge so the approach adapts to dynamic or cross-domain graph environments.
• Better handling of highly imbalanced classes. The current work sidesteps rare but highly central classes; a principled way to use them without adding noise remains open.
• Resolving the target-class ambiguity and reporting full metrics. The paper's text names different class sets in the setup and evaluation sections, and the numeric tables appear only as figures — a fuller, reproducible account would clarify both.

Target Audience

This paper suits graduate students and applied researchers in graph machine learning, graph neural networks, and knowledge graph representation learning who are dealing with small or feature-poor datasets. It is also relevant to data scientists and engineers building node classification pipelines on graph databases (particularly Neo4j with GDS), and to researchers in biomedical informatics or pharmacovigilance interested in applying graph embeddings to drug safety data. Readers should already be comfortable with embedding models, message passing, and standard evaluation metrics.

Authors’ abstract

Traditional Machine Learning (ML) methods require large amounts of data to perform well, limiting their applicability in sparse or incomplete scenarios and forcing the usage of additional synthetic data to improve the model training. To overcome this challenge, the research community is looking more and more at Graph Machine Learning (GML) as it offers a powerful alternative by using relationships within data. However, this method also faces limitations, particularly when dealing with Knowledge Graphs (KGs), which can hide huge information due to their semantic nature. This study introduces Bi-View, a novel hybrid approach that increases the informative content of node features in KGs to generate enhanced Graph Embeddings (GEs) that are used to improve GML models without relying on additional synthetic data. The proposed work combines two complementary GE techniques: Node2Vec, which captures structural patterns through unsupervised random walks, and GraphSAGE, which aggregates neighbourhood information in a supervised way. Node2Vec embeddings are first computed to represent the graph topology, and node features are then enriched with centrality-based metrics, which are used as input for the GraphSAGE model. Moreover, a fusion layer combines the original Node2Vec embeddings with the GraphSAGE-influenced representations, resulting in a dual-perspective embedding space. Such a fusion captures both topological and semantic properties of the graph, enabling the model to exploit informative features that may exist in the dataset but that are not explicitly represented. Our approach improves downstream task performance, especially in scenarios with poor initial features, giving the basis for more accurate and precise KG-enanched GML models.