FGAT: Fuzzy Graph Attention for Wireless Imputation

Updated 8 July 2026

FGAT is a fuzzy graph attention network that dynamically constructs spatial graphs using fuzzy rough set approximations.
It computes connectivity scores between sensor nodes by combining two directional fuzzy lower approximations and retains only the top-K neighbors.
Integrated into the FGATT framework alongside a Transformer encoder, FGAT enhances robustness under high missing data rates in wireless settings.

FGAT, in the context of wireless data imputation, denotes a Fuzzy Graph Attention Network: a graph-attention–based neural layer whose graph structure and edge strengths are derived from fuzzy rough sets rather than from fixed, predefined spatial information such as sensor coordinates or physical topology. Within the hybrid FGATT framework, FGAT serves as the spatial modeling component, while a Transformer encoder models temporal dependencies in time series. The design targets two recurring difficulties in wireless and IoT data: uncertain or missing spatial relationships among sensors, and the need for robustness under incomplete or noisy observations (Xing et al., 2024).

1. Definition, scope, and acronym usage

FGAT was introduced in FGATT to replace the assumption, common in standard GAT and many ST-GNN variants, that a reliable adjacency graph is already known. In this formulation, connectivity is learned through fuzzy rough set–based similarity measures, after which a GAT layer operates on the resulting fuzzy graph. A central clarification is that FGAT is not a different attention formula per se; it is a GAT layer operating on a fuzzy rough set–based, dynamically constructed graph, with normalization, dropout, and residual connections for robustness (Xing et al., 2024).

The acronym is also used for several unrelated constructs in recent arXiv literature. This makes terminological precision necessary when comparing papers.

Usage	Meaning	Representative paper
FGAT	Fuzzy Graph Attention Network for wireless imputation	(Xing et al., 2024)
FGAT	Fuzzy Graph Attention Networks for link prediction	(Xing et al., 2024)
MFGAT	Multi-view Fuzzy Graph Attention Network	(Xing et al., 2024)
FGAT	Feature-Guided Attention in occupancy flow prediction	(Chen et al., 2024)
FGAT	Fourier Graph Attention in dynamic graphs	(Xu et al., 23 Feb 2025)
FGAT	Fashion Graph Attention Network	(Saed et al., 14 Aug 2025)
FGAT / FGaT	Fe $_3$ GaTe $_2$ , a van der Waals ferromagnet	(Chang et al., 10 May 2026)

Within FGATT, the term specifically refers to fuzzy-rough-set–driven spatial graph construction for sensor-node representations, not to these other usages. This distinction matters because the mathematical role of FGAT in FGATT is graph construction plus attention-based aggregation, whereas the other works use the same acronym for deformable attention, spectral attention, recommendation models, or even a magnetic material abbreviation.

2. Fuzzy rough set formulation and dynamic connectivity

FGAT is built on fuzzy rough set notions of lower and upper approximation. Let $U$ be the universe of samples, $d_i$ a fuzzy decision membership function over $U$ , and $B$ an attribute set. The fuzzy lower and upper approximations are defined as

$\underline{R_B}d_i(x) = \inf_{y\in U}\max(1 - R(x,y), d_i(y)),\qquad \overline{R_B}d_i(x) = \sup_{y\in U}\min(R(x,y), d_i(y)),$

where $R(x,y)$ is typically a kernel-based fuzzy similarity between samples $x$ and $y$ (Xing et al., 2024).

In this construction, $_2$ 0 quantifies how strongly $_2$ 1 certainly belongs to $_2$ 2, while $_2$ 3 quantifies possible membership. FGATT uses these fuzzy approximations to define connectivity scores between nodes at each time step rather than relying on Euclidean distance or fixed topology.

For nodes $_2$ 4 with embeddings $_2$ 5 at time $_2$ 6, the per-time connectivity score is

$_2$ 7

where $_2$ 8 balances the two directional fuzzy lower approximations. Over a temporal window of length $_2$ 9, the final aggregated connectivity score is obtained by mean pooling: $U$ 0

This yields a time-aggregated fuzzy similarity between nodes. The resulting graph is then sparsified by retaining only the top $U$ 1 neighbors per node, and self-loops are removed. The paper states: “To optimize graph construction, only the top $U$ 2 edges based on connectivity scores are retained… Additionally, self-loops are removed, as they may lead to error accumulation…” (Xing et al., 2024).

The resulting graph $U$ 3 is therefore self-adaptive: it is recomputed from data embeddings and fuzzy approximations rather than inherited from external geometry. A useful way to view this mechanism is as a learned notion of spatial proximity under uncertainty. If two node embeddings mutually reinforce one another through fuzzy neighborhood membership, the connectivity score becomes high even when precise coordinates or reliable physical topology are unavailable.

3. Attention, spatial encoding, and implementation structure

Once the fuzzy graph has been constructed, FGAT applies a standard GAT aggregation rule on the selected neighbors: $U$ 4 where $U$ 5 is the current embedding of neighbor node $U$ 6, $U$ 7 is a learnable weight matrix, and $U$ 8 is the fuzzy-graph neighborhood defined by the top- $U$ 9 edges (Xing et al., 2024).

The raw attention coefficient formula is not restated in the FGATT paper, but the standard GAT form is given contextually as

$d_i$ 0

In FGAT, the key modification is not the coefficient formula itself but the fact that $d_i$ 1 and the edge set are derived from fuzzy rough sets rather than from a static adjacency (Xing et al., 2024).

FGAT further includes linear layers after GAT to transform embeddings, layer normalization, dropout, and residual connections to facilitate gradient flow and stabilize training. The paper describes these as the components that make FGAT a robust spatial encoder over a fuzzy, dynamically learned graph (Xing et al., 2024).

This architecture induces what the paper explicitly characterizes as a two-stage mechanism. First, fuzzy rough sets determine which neighbors are plausible and how strong the underlying similarity is through edge selection. Second, neural attention learns task-driven weights $d_i$ 2 over those selected neighbors. The first stage encodes uncertainty and soft similarity; the second stage performs adaptive feature weighting.

In practical terms, the principal hyperparameters are $d_i$ 3, the number of neighbors retained per node; $d_i$ 4, the balance between the two directional fuzzy lower approximations; and the temporal window length $d_i$ 5, which is set to $d_i$ 6 in the experiments. Graph construction can involve potentially $d_i$ 7 similarity computation, but top- $d_i$ 8 sparsification, self-loop removal, and subsequent sparse GAT operations mitigate computational cost and noise (Xing et al., 2024).

4. Position within the FGATT architecture

FGAT is the spatial half of a hybrid architecture in which a Transformer encoder provides temporal modeling. The full FGATT pipeline for a wireless dataset proceeds in five stages: input data and preprocessing, dynamic graph construction, FGAT spatial encoding, Transformer temporal encoding, and imputation output (Xing et al., 2024).

The input consists of time-series sensor readings with missing values. The preprocessing includes min-max normalization, and the context length is set to 16 timesteps. Dynamic graph construction then computes fuzzy lower approximations and connectivity scores $d_i$ 9, aggregates them by mean pooling to $U$ 0, and constructs a fuzzy graph with top- $U$ 1 edges and no self-loops.

FGAT is then applied to this graph to produce spatially enriched node features. These features serve as the input to the Transformer encoder, which models temporal dependencies through self-attention: $U$ 2 The Transformer uses multi-head attention, feedforward layers, layer normalization, and residual connections to model complex temporal dependencies (Xing et al., 2024).

The final output of the Transformer is used to predict missing values for each sensor and time step. In this decomposition, FGAT is responsible for capturing relations among sensors or nodes, while the Transformer encodes the evolution of those spatially enriched features across time. The architecture is therefore explicitly spatial-temporal rather than purely temporal or purely graph-based.

5. Wireless imputation behavior and reported empirical properties

FGAT is motivated by wireless sensor networks and IoT settings in which exact spatial information may be unavailable or unreliable, sensors may move or be replaced, and missing data can distort simple distance-based graphs. In such settings, the ability to learn connectivity directly from observed data through fuzzy rough sets is the central design choice (Xing et al., 2024).

The paper reports experiments on SWaT.A7.22 and SWaT.A7.29. FGATT is trained at 50% missing and tested up to 80% missing. The reported results state that FGATT shows the best average performance across MSE, MAE, and RMSE for missing rates 20–80% on both datasets. The model also maintains relatively stable errors around the training missing rate of $U$ 3, which the paper interprets as robust generalization (Xing et al., 2024).

The comparison set includes FFN, BGRU, vanilla Transformer, and TGCN. According to the paper, these competing methods have higher error and steeper degradation, with TGCN affected particularly strongly because it relies on static spatial structures. The authors attribute FGATT’s robustness specifically to “its fuzzy rough sets-based dynamic graph and FGAT framework” (Xing et al., 2024).

The significance of these observations is methodological as much as empirical. FGAT introduces a spatial inductive bias without requiring sensor coordinates or a reliable predefined topology. Temporal-only models such as FFN, BGRU, or vanilla Transformer cannot exploit such relational patterns directly, while static-graph models can become brittle when the graph prior is unavailable or mismatched to the data regime. FGAT therefore occupies a middle position: it preserves graph-based relational reasoning while learning the graph itself under uncertainty.

The wireless-imputation FGAT is part of a broader line of fuzzy graph attention research. In link prediction, Fuzzy Graph Attention Networks combine a GAT-based encoder with Fuzzy Negative Sampling (FNS), which uses fuzzy rough sets to select high-quality negative edges based on fuzzy similarities rather than uniform random sampling. In that formulation, the link probability is computed as

$U$ 4

and the reported experiments on two research collaboration networks show higher link prediction accuracy than MLP, GCN, GraphSAGE, and GAT baselines (Xing et al., 2024).

A later extension, Multi-view Fuzzy Graph Attention Network (MFGAT), argues that vanilla FGAT and FGATT do not consider multi-view dependencies. MFGAT therefore adds a Transformation Block that constructs multiple feature views,

$U$ 5

aggregates them by a weighted sum mechanism, and then feeds the unified representation into FGAT layers, followed by a learnable global pooling mechanism. On graph classification tasks, MFGAT is reported to outperform GCN, GAT, GraphSAGE, and FGAT on PROTEINS, NCI1, and Mutagenicity, with 3 views giving the best performance among the tested view counts $U$ 6 (Xing et al., 2024).

These extensions clarify the conceptual core of FGAT across tasks. The persistent elements are fuzzy rough set–based treatment of uncertainty and graph attention–based adaptive aggregation. What changes across applications is the locus of fuzzy intervention: in wireless imputation it defines a self-adaptive spatial graph; in link prediction it also governs negative sampling; in graph classification it becomes a backbone for multi-view graph learning. A plausible implication is that FGAT is best understood not as a single closed-form layer family but as a design pattern: attention over graphs whose edge semantics are softened, filtered, or constructed through fuzzy rough set theory.