KG-Enhanced Zero-Shot Semantic Communication

• The KGZS-SC network is a semantic communication system that fuses structured knowledge graphs with zero-shot learning to enable reasoning for unseen data classes.
• It uses a two-stage training process to align raw features with graph-based semantic embeddings, ensuring retrain-free and noise-resilient classification.
• Empirical results demonstrate superior generalization and efficiency, making the system ideal for dynamic, resource-constrained applications like IoT and AR/VR.

A Knowledge Graph-Enhanced Zero-Shot Semantic Communication (KGZS-SC) Network is an advanced semantic communication system that integrates structured knowledge graphs with zero-shot learning principles to enable efficient, robust, and explainable semantic transmission, especially in scenarios involving unseen data categories. By leveraging a shared knowledge graph-based semantic knowledge base (KG-SKB), the KGZS-SC network aligns transmitted semantic features in a unified category semantics embedding space and provides direct reasoning for previously unobserved classes. This design supports highly generalized semantic representations at the transmitter and efficient, retrain-free zero-shot classification at the receiver, significantly enhancing both communication performance and adaptability under dynamic or resource-constrained conditions (2507.02291).

1. Knowledge Graph-Based Semantic Knowledge Base (KG-SKB)

The KG-SKB forms the core semantic infrastructure of a KGZS-SC network. It is constructed by extracting category subgraphs from a large commonsense knowledge graph for each seen category $y \in 𝒴ₛ$. For each category, both its one-hop ($\mathcal{N}^1(y)$) and two-hop ($\mathcal{N}^2(y)$) neighbors are collected to form a subgraph $\mathcal{G}_s = (\mathcal{V}_s, \mathcal{E}_s)$. The global KG-SKB is then composed by unifying all per-category subgraphs:

$$\mathcal{V}_S = \bigcup_{y \in 𝒴ₛ} \mathcal{V}_s, \quad \mathcal{E}_S = \bigcup_{y \in 𝒴ₛ} \mathcal{E}_s$$

Each node $v \in \mathcal{V}_S$ is initialized with a pretrained GloVe embedding $d_v^{(0)} = \text{GloVe}(v) \in \mathbb{R}^D$.

The KG-SKB embeds both category-level semantics and inter-category relational knowledge into a shared high-dimensional space. Semantic feature alignment is achieved as follows: after initialization, the edge weight matrix is Laplace smoothed and row-normalized (i.e., for edge $(u, v)$: $$\tilde{C}_{uv} = C_{uv} + 1, s_{uv} = \tilde{C}_{uv} / (\sum_{w \in \mathcal{V}_S} \tilde{C}_{uw})$$). A Graph Convolutional Network (GCN) is then applied, recursively combining neighbor features,

$$z_v^{(l)} = \sum_u \frac{1}{m_{vu}} W_\delta^{(l)} d_u^{(l-1)}, \quad \hat{z}_v^{(l)} = \text{LN}(z_v^{(l)} + d_v^{(l-1)}), \quad d_v^{(l)} = \varphi(\hat{z}_v^{(l)}),$$

where $m_{vu} = \sqrt{\deg(v)\,\deg(u)}$ and $\varphi(\cdot)$ is ReLU. The output $d_v^{(L)}$ serves as the semantic embedding $\varphi(y)$ for class $y$.

2. Zero-Shot Learning Mechanism and Semantic Alignment

Zero-shot learning (ZSL) within the KGZS-SC network leverages the KG-SKB's GCN-enhanced semantic representations for robust, unified embeddings of both seen and unseen categories. The transmitter aligns the extracted visual (or generic data) features with the KG-based semantic space via a two-stage training:

• Stage 1: The semantic encoder $S_\alpha$ and GCN parameters $\delta$ are optimized to maximize similarity between the encoded input $S_\alpha(x)$ and the graph-derived embedding $\varphi(y)$, using a loss:

$$\mathcal{L}_1(\alpha, \delta) = -\log\left( \frac{e^{\text{Sim}(S_\alpha(x), \varphi(y))}}{\sum_{y' \in 𝒴ₛ} e^{\text{Sim}(S_\alpha(x), \varphi(y'))}} \right)$$

with $\text{Sim}(a,b) = \| a - b \|_2$.

• Stage 2: The channel encoder/decoder is fine-tuned to minimize both a recovery loss and a consistency loss:

$$\mathcal{L}_2(\beta, \hat{\beta}) = \| \hat{s} - S_\alpha(x) \|_2^2 + \lambda \| \hat{s} - \varphi(y) \|_2^2$$

ensuring that transmitted semantic vectors remain close to their KG-based category representation even under channel noise.

At the receiver, upon decoding a semantic vector $\hat{s}$, classification is conducted by directly matching $\hat{s}$ to all KG-SKB category embeddings:

$$\hat{y} = \arg\min_{y \in 𝒴} \| \varphi(y) - \hat{s} \|_2$$

This allows direct zero-shot inference for unseen classes with no retraining.

3. Generalization, Robustness, and Efficiency

Generalization is fundamentally enhanced by the structured semantic alignment between raw features and category embeddings in the KG-SKB. The GCN layers propagate both local and global semantic dependencies, addressing challenges like the hubness problem typically observed in high-dimensional ZSL setups. The two-stage procedure—strong feature encoding followed by channel adaptation—makes the system robust to both semantic variation and channel distortion.

Efficiency is achieved by transmitting only highly compressed, semantically salient vectors instead of raw data. The computational cost for online inference per instance is $O(M + M' + |\mathcal{Y}|D)$—where $M, M'$ are the extraction and channel costs, and $|\mathcal{Y}|$ is the set of candidate classes—enabling fast adaptation as category sets evolve.

4. Simulation Results and Empirical Observations

Empirical evaluation on the APY dataset (20 seen, 12 unseen classes) demonstrates the KGZS-SC network's superior generalization, particularly in low SNR conditions. At $-10$ dB SNR, for instance, the method achieves 57.43% accuracy on seen and 41.64% on unseen classes, yielding a harmonic mean of 48.28%, substantially outperforming traditional ZSL-SC baselines in both accuracy and stability across the SNR range. This robustness under increased channel noise highlights the method’s practical viability.

SNR (dB)	Seen Acc. (%)	Unseen Acc. (%)	Harmonic Mean (%)
-10	57.43	41.64	48.28
0	86.95	81.28	84.03
10	90.67	86.27	88.42
15	90.68	86.13	88.34

Efficiency is further underscored by retrain-free operation during deployment: unseen categories can be integrated by extending the KG-SKB and transmission/classification proceeds without additional computation or model updates.

5. Applications and System Implications

The KGZS-SC network is especially relevant for dynamic or resource-constrained settings such as digital twins, IoT, AR/VR, and autonomous environments, where new semantic categories can emerge unpredictably and bandwidth is limited. By reducing communication overhead—transmitting only compact, graph-aligned semantic vectors—and utilizing robust zero-shot inference, the KGZS-SC enables real-time semantic decision-making as observed entities or classes change over time.

Interpretability is also improved, since transmitted representations are aligned with explainable knowledge structures dictated by the KG-SKB. The architecture facilitates generalization and adaptability across evolving or open-set domains, mitigating the need for continual retraining and supporting resilient semantic communication in challenging, changing environments.

6. Broader Context and Technological Positioning

The KGZS-SC network synthesizes advancements from knowledge-graph-based visual understanding, graph-based zero-shot learning, and graph convolutional architectures, embedding their strengths into the semantic communication context. It addresses the key shortcomings of data-driven approaches—lack of interpretability and poor generalization to unseen cases—by grounding the semantic exchange in a structured, relational framework that directly supports reasoning and efficient, scalable communication (2507.02291). The approach unifies semantic feature extraction, graph-based knowledge propagation, and noise-resilient encoding/decoding into a single, explainable pipeline, setting a benchmark for future semantic communication systems with zero-shot generalization demands.

7. Summary Table: Key Components

Component	Functionality	Associated Methods/Formulas
KG-SKB	Encodes structured category-level semantics	GloVe embeddings, 2-hop subgraph union
GCN Layers	Propagate semantic features over KG-SKB	$z_v^{(l)}$ aggregation, residual conn.
ZSL Module	Direct classification for unseen categories	$\hat{y} = \arg\min_y \| \varphi(y) - \hat{s} \|_2$
Training	Feature alignment and channel adaptation	$\mathcal{L}_1,\, \mathcal{L}_2$
Communication	Transmits only compact semantic representations	No retraining for new categories

The Knowledge Graph-Enhanced Zero-Shot Semantic Communication network thus represents a rigorously grounded approach to flexible, interpretable, and adaptable semantic transmission in both static and highly dynamic application domains.