Dual-Axis Transformer-GNN Framework for Touchless Finger Location Sensing by Using Wi-Fi Channel State Information

Abstract

Camera, lidar, and wearable-based gesture recognition technologies face practical limitations such as lighting sensitivity, occlusion, hardware cost, and user inconvenience. Wi-Fi channel state information (CSI) can be used as a contactless alternative to capture subtle signal variations caused by human motion. However, existing CSI-based methods are highly sensitive to domain shifts and often suffer notable performance degradation when applied to environments different from the training conditions. To address this issue, we propose a domain-robust touchless finger location sensing framework that operates reliably even in a single-link environment composed of commercial Wi-Fi devices. The proposed system applies preprocessing procedures to reduce noise and variability introduced by environmental factors and introduces a multi-domain segment combination strategy to increase the domain diversity during training. In addition, the dual-axis transformer learns temporal and spatial features independently, and the GNN-based integration module incorporates relationships among segments originating from different domains to produce more generalized representations. The proposed model is evaluated using CSI data collected from various users and days; experimental results show that the proposed method achieves an in-domain accuracy of 99.31% and outperforms the best baseline by approximately 4% and 3% in cross-user and cross-day evaluation settings, respectively, even in a single-link setting. Our work demonstrates a viable path for robust, calibration-free finger-level interaction using ubiquitous single-link Wi-Fi in real-world and constrained environments, providing a foundation for more reliable contactless interaction systems.

1. Introduction

Gesture and human movement recognition technologies have become increasingly important in applications such as smart homes, human–computer interaction (HCI), and health care. Existing sensing modalities capture motion information through different mechanisms, including camera-based, radar/lidar-based, and wearable approaches [1,2,3]. However, despite their effectiveness in controlled settings, these modalities commonly suffer from practical limitations such as sensitivity to environmental conditions, deployment complexity, and user burden, which hinder scalable deployment in real-world environments. Wi-Fi-based contactless sensing has emerged as a promising alternative that satisfies these requirements by leveraging variations in the signal, whose amplitude and phase are modulated by environmental factors and human movement [4]. While CSI provides both amplitude and phase information, the phase component is highly unstable for sensing, leading many works to focus on amplitude information. Owing to its ability to capture signal variations caused by human motion, CSI enables gesture sensing without requiring users to wear or hold any devices. However, the received CSI is influenced not only by gestures, but also by multiple external factors. Radio waves are reflected by walls, furniture, and human bodies. These effects vary with the physical structure and material of the environment. People also differ in body size, motion speed, and posture. Moreover, changes in sensor placement or orientation can significantly alter the received signal patterns. As a result, identical gestures can produce different CSI responses, depending on environmental, human, and structural variations. In this paper, we collectively define these variations as domains. A domain consists of factors such as hardware configuration, user physical characteristics, measurement position, orientation, and surrounding reflectors. Models trained in one domain often suffer from performance degradation when deployed in another domain, a problem known as domain shift.

Existing approaches to address this issue can be categorized into three main categories. First, some studies attempt to improve model generality by constructing large-scale datasets that span diverse users and environments [5,6]. Although such datasets can enhance robustness, this strategy is impractical because it cannot cover all possible conditions and incurs substantial data collection costs. Second, adversarial and transfer learning methods [7,8] have been proposed to reduce interdomain discrepancies. However, these approaches remain inherently dependent on the domains seen during training and often struggle to generalize to unseen ones. Third, multi-link systems employing multiple transmitters and receivers have been explored to obtain spatial diversity [9]. Nevertheless, multi-link Wi-Fi sensing systems require complex hardware configurations and inter-device calibration. Specifically, their deployment is further constrained by strict placement requirements, which significantly limits their practical scalability [10]. Moreover, large-scale Wi-Fi sensing systems are typically composed of numerous low-cost microcontroller-based nodes, where adopting multi-antenna or multi-link configurations can substantially increase overall system installation cost and deployment burden [11]. Motivated by this perspective, we focus on improving the reliability of gesture recognition in single-link Wi-Fi environment. To address the variations inherent to CSI signals, we construct a preprocessing pipeline that reduces fluctuations induced by hardware as well as the environment and develop a feature-modeling strategy that preserves gesture-consistent patterns across different measurement conditions. The proposed framework is evaluated in both in-domain and cross-domain settings to assess its robustness in real-world scenarios.

The major contributions of this study are summarized as follows:

1.

Single-link verification: We validate the proposed system using a practical single-link Wi-Fi setup with commercial hardware.

2.

Multi-domain combination strategy: We construct mini-batches by combining segments from multiple domains within each gesture class, explicitly forcing the model to observe diverse domain manifestations during training.

3.

Transformer–GNN hybrid learning: We design a Dual-Axis Transformer and a GNN-based integration module to learn domain-invariant representations.

2. Related Works

Recent Wi-Fi CSI–based gesture recognition has increasingly adopted deep learning approaches that learn gesture-relevant features directly from raw CSI data. Unlike traditional methods that rely on handcrafted statistical features [12,13], deep learning models automatically extract discriminative temporal and spectral patterns using neural network architectures, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers. These models have demonstrated strong capabilities in modeling complex CSI dynamics and capturing fine-grained gesture-induced variations without relying on manual feature engineering.

Deep learning–based approaches can be divided into two categories depending on whether training and test data share the same measurement conditions: in-domain and cross-domain settings. The in-domain setting assumes that the training and testing data are collected under the same users, environment, and device configuration. This allows the model to learn gesture patterns under stable conditions. In contrast, the cross-domain setting evaluates the robustness of the model when the CSI measurement conditions differ, such as changes in users, locations, orientations, or measurement days.

Following this distinction, this section reviews representative deep learning–based works that target performance improvement in both in-domain and cross-domain settings.

2.1. In-Domain Measurement Setting

Early work, such as [14], introduces a deep learning architecture designed to model temporal correlations in the CSI time series for activity recognition. The raw amplitude sequences are normalized and fed into an LSTM network, enabling the model to learn gesture-induced temporal dynamics. In [15], spatial and temporal information are jointly modeled using a CNN–GRU–Attention (CGA) architecture. Amplitude and phase-difference inputs undergo phase calibration and adaptive truncation, are converted into multichannel images, and are processed by CNN and GRU layers to capture complementary spatial–temporal features. To improve computational efficiency, the authors in [16] incorporate depthwise separable convolution (DS-Conv), feature attention (FA) modules, and residual blocks (RB), enabling lightweight yet accurate gesture recognition. Similarly, in [17], the CSI is compressed using a variational autoencoder (VAE), and low-dimensional latent vectors are used for multitask learning, simultaneously optimizing reconstruction and classification to maintain discriminative information with reduced transmission overhead. Fine-grained micro-motion recognition is also investigated. In [18], CSI amplitude is transformed into time–frequency spectrograms, and an inception-based architecture is employed to model cross-frequency interactions. Additionally, the authors in [19] introduce a Kolmogorov–Arnold Network (KAN)-based classifier for keypad-input recognition, where spline-based nonlinear functions are leveraged to learn complex decision boundaries and reduce intra-class variance. Recently, a GNN-based dual-attention fusion network, known as Wi-DualAtt [20], is introduced for CSI-based activity recognition. This approach adopts a dual-stream architecture that separately captures hierarchical subcarrier correlations and temporal dynamics through graph attention and temporal attention mechanisms, respectively, and integrates the two streams via a prediction fusion module.

2.2. Cross-Domain Measurement Setting

Although in-domain methods achieve high accuracy under controlled settings, real-world deployments inevitably involve variations in users, locations, orientations, and measurement periods. These domain shifts cause significant performance degradation, leading to extensive research on cross-domain generalization. In [21], a transformer-based few-shot transfer learning framework is introduced, where a pretrained feature extractor is adapted to a new domain using only a small number of samples. In [22], meta-learning is applied to one-shot gesture recognition by constructing diverse domain combinations during training, enabling generalization with minimal target data. However, such a few-shot and meta-learning-based approaches inherently depend on the availability and quality of target-domain samples, which may limit their applicability in practical scenarios where collecting reliable target data is difficult or costly. Several methods aim to learn domain-invariant representations without using target-domain samples. In [23], a transformer architecture with global self-attention minimization is utilized to reduce inter-domain variability. Similarly, the authors in [24] employ amplitude–phase joint modeling with self- and relation-attention modules to suppress domain-sensitive components. These studies are primarily evaluated under multi-link or multi-receiver settings, whereas our work focuses on the more challenging single-link CSI sensing scenario, where domain shifts are harder to mitigate due to limited spatial diversity. The approach in [25] explicitly models domain factors by combining adversarial domain discrimination with feature disentanglement. Other studies enhance generalization by increasing data diversity. According to [7], statistically consistent noise injection and multipath-based signal transformations are applied to synthesize diverse variations. In [26], virtual data are generated using a VAE-based method, and episodic training with domain-specific subnetworks is adopted to learn domain-independent representations. While these methods aim to suppress domain-specific characteristics through data augmentation, our approach instead employs a GNN to explicitly model the relationships among gesture segments originating from different domains, enabling structured cross-domain feature integration rather than implicit invariance learning. The method in [27] performs subdomain-guided feature augmentation by perturbing features along statistically significant directions of each subdomain and employs a cooperative learning structure in which gesture and domain classifiers reinforce each other. A graph-based approach has also been explored to improve cross-domain generalization by leveraging spatial topology information. Specifically, WiGNN [28] explicitly models the spatial relationships among multiple receivers by representing them as nodes in a graph and aggregates Doppler-based features across receivers using a graph neural network. Although WiGNN similarly adopts a GNN-based formulation, its graph nodes correspond to physical receivers, whereas our work fundamentally differs by modeling domains as graph nodes, allowing explicit reasoning over cross-domain relationships even in single-link CSI sensing environments.

Table 1. Comparison of deep learning–based CSI gesture recognition methods (# Links: total number of transmitter–receiver antenna pairs, defined as the product of the number of transmit antennas and receive antennas).

Domain	Ref.	Feature	Preprocessing	Methodology	# Links
In-domain	[14]	Amplitude	PCA	LSTM	1
	[15]	Amplitude Phase Diff	Butterworth Filter, Phase Unwrapping	CNN + GRU + Attention	3
	[16]	Amplitude	Butterworth Filter, Gaussian Smoothing	Depthwise Separable Convolution, Feature Attention, Residual Block	1
	[17]	Amplitude	-	Variational Autoencoder	9
	[18]	Amplitude	Weighted MA, STFT	CNN-based Inception Block	8
	[19]	Amplitude	Hampel Filter	CNN + KAN-based Feature Regularization	1
	[20]	Amplitude	Sliding Window	Dual-stream Attention + GNN	6
Cross-domain	[21]	Phase	PCA, DFS	Transformer + Few-shot Learning	12
	[22]	Amplitude Phase	Spectrogram	Autoencoder + Few-shot Learning	3
	[23]	Amplitude Phase	PCA, BVP	Transformer-based Multi-branch Attention	3
	[24]	Amplitude Phase	CSI Ratio	ResNet-18-based Self-Attention Relation Attention	18
	[25]	Phase	PCA, DFS	CNN + GRU Domain Network, Adversarial Attention Fine-tuning	6
	[7]	Amplitude	-	Autoencoder, Feature Augmentation, Adversarial Regularization	3
	[26]	Amplitude, Phase, Spectrogram	Median Filter, Phase Unwrapping, DFS	VAE-based Data Augmentation, Episodic Training	9
	[27]	Amplitude	Butterworth Filter, Phase Unwrapping	1D-CNN + Residual Block Subdomain-guided Perturbation	1, 3
	[28]	Amplitude	DFS, Cyclic Shift Augmentation	CNN + GNN	6
	Ours	Amplitude	Hampel Filter, Decibel Conversion, Segment Combination	Transformer + GNN	1

summarizes these deep learning–based approaches in terms of CSI information used, preprocessing methods, deep learning methods, and measurement configurations. While existing deep learning–based Wi-Fi sensing methods have achieved strong performance, several limitations remain in terms of their design assumptions. From the experimental setting perspective, many prior studies assume multi-link or multi-antenna measurement configurations to exploit spatial diversity, which may limit their applicability in practical deployments. From the model architecture perspective, Transformer-based approaches have been primarily used to learn temporal dependencies in CSI, and although some methods jointly model spatiotemporal information, they often rely on single-domain training or few-shot fine-tuning to adapt to new domains, rather than explicitly learning common patterns across multiple domains. Meanwhile, graph-based models have mainly been employed to capture spatial relationships by treating subcarriers as nodes or to model link-level relationships in multi-link environments; while effective for representing physical measurement structures, such designs are limited in modeling shared gesture patterns observed across different domains. In contrast, our work assumes a realistic single-link setting and introduces a dual-axis Transformer to independently learn temporal and spatial features from CSI, followed by a GNN that treats different domains as nodes to explicitly model common gesture patterns shared across domains. This design enables robust cross-domain gesture recognition while maintaining practicality in real-world measurement environments.

3. Methods

3.1. System Overview

We propose a gesture recognition module that identifies finger location gestures using Wi-Fi CSI data. The proposed system first converts the raw CSI amplitude into fixed-length segment units using a preprocessing pipeline based on Hampel filtering, decibel conversion, and sliding-window segmentation, which suppresses impulsive noise, stabilizes signal dynamics, and enables consistent representation of fine-grained finger motions. Each segment is represented as a two-dimensional time–subcarrier matrix and is processed by a transformer-based network that jointly models temporal dynamics and subcarrier-level spatial patterns, allowing the model to capture both motion-induced temporal variations and frequency-dependent propagation characteristics inherent in CSI. Subsequently, the GNN integrates inter-segment relationships to capture domain-induced variations, thereby enhancing the consistency of feature representations across segments. Finally, the classifier improves class separability by combining the cross-entropy loss with the center loss, encouraging intra-class compactness while maintaining inter-class discrimination under cross-domain conditions. The overall architecture is shown in Figure 1, and the following sections describe each component of the proposed system in detail.

3.2. CSI Data Preprocessing

In this section, we describe the preprocessing pipeline applied to the Wi-Fi CSI measurements. For clarity, the notations used throughout the preprocessing stage are summarized in

Table 2. List of Notations Used in CSI Data Preprocessing.

Notation	Description
H	Raw complex CSI matrix.
T	Number of CSI time samples collected in a single measurement.
${\tilde{N}}_s$	Number of raw subcarriers.
$N_s$	Number of data subcarriers retained.
$\tilde{A}$	Amplitude matrix of all the raw subcarriers.
A	Amplitude matrix of data subcarriers only.
s	Subcarrier index.
$a_s\left(t\right)$	Amplitude value of subcarrier s at time index t.
w	Window length used for Hampel filter.
$N_t$	Temporal length of each segment.
R	Number of non-overlapping segments obtained from a single measurement.
$X^{\left(c\right)}$	Multi-domain input segment set for gesture class c.
D	Set of available domains.
M	Number of segments sampled to construct $X^{\left(c\right)}$.

. In an OFDM-based Wi-Fi system, the transmitter divides the frequency band into ${\tilde{N}}_s$ orthogonal subcarriers, and the receiver estimates the channel response of each subcarrier in a complex form. The channel response of the i-th subcarrier at time index $t\in \{1,2,\dots ,T\}$ is expressed as

$$H_i\left(t\right)=\left|H_i\left(t\right)\right|e^{j\angle H_i\left(t\right)},i=1,2,\dots ,{\tilde{N}}_s,$$

(1)

where $|H_i\left(t\right)|$ denotes the amplitude, $\angle H_i\left(t\right)$ denotes the phase of $H_i\left(t\right)$, and $j=\sqrt{-1}$ is an imaginary unit. Accordingly, the CSI sequence for a gesture class collected over time can be expressed as

$$H=\left[\begin{array}{cccc}{H}_1\left(1\right)& H_1\left(2\right)& \dots & H_1\left(T\right)\\ H_2\left(1\right)& H_2\left(2\right)& \dots & H_2\left(T\right)\\ ⋮& ⋮& \ddots & ⋮\\ H_{{\tilde{N}}_s}\left(1\right)& H_{{\tilde{N}}_s}\left(2\right)& \dots & H_{{\tilde{N}}_s}\left(T\right)\end{array}\right]\in {\mathbb{C}}^{{\tilde{N}}_s\times T}.$$

(2)

Each row corresponds to the temporal response of a subcarrier, and each column corresponds to the received CSI frame.

The phase component of Wi-Fi CSI is highly sensitive to carrier frequency offset (CFO), sampling frequency offset (SFO), phase wrapping, and hardware-induced phase inconsistencies [29]. These fluctuations introduce instability across the domains and degrade the robustness of the classifier. In contrast, the amplitude component $|H_i\left(t\right)|$ is more stable and directly reflects gesture-induced variations. Following prior studies [7,14,16,17,19,27], we extract only the amplitude

$$\tilde{A}=\left|H\right|\in {\mathbb{R}}^{{\tilde{N}}_s\times T}.$$

(3)

Wi-Fi CSI contains multiple subcarriers, among which only a subset is used for data transmission, while the rest serve control or guardband functions. The subcarriers are categorized into data, pilot, and null subcarriers. Because the pilot and null subcarriers do not contain gesture-related information, we remove them and retain only the data subcarriers:

$$A=\left|H\right|\in {\mathbb{R}}^{N_s\times T},$$

(4)

where $N_s$ denotes the number of data subcarriers.

The CSI amplitude may include temporary outliers caused by environmental interference or hardware noise. To mitigate this issue, we applied the Hampel filter [30]. Let the amplitude sequence of subcarrier $s\in \{1,2,\dots ,N_s\}$ be $a_s=[a_s\left(1\right),a_s\left(2\right),\dots ,a_s\left(T\right)]$. For a window of length w centered at t, the median and MAD (Median Absolute Deviation) statistics are

$$m_s\left(t\right)=\mathrm{median}\left(a_s\left(\tau \right)\mid \tau \in [t-\lfloor w/2\rfloor ,t+\lfloor w/2\rfloor ]\right),$$

(5)

$${\mathrm{MAD}}_s\left(t\right)=\mathrm{median}\left(|a_s\left(\tau \right)-m_s\left(t\right)|\mid \tau \in [t-\lfloor w/2\rfloor ,t+\lfloor w/2\rfloor ]\right).$$

(6)

An outlier is replaced according to the following equation:

$${\tilde{a}}_s\left(t\right)=\left\{\begin{array}{cc}{m}_s\left(t\right),\hfill & \mathrm{if}|a_s\left(t\right)-m_s\left(t\right)|>\gamma ·{\mathrm{MAD}}_s\left(t\right),\hfill \\ a_s\left(t\right),\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(7)

Here, ${\tilde{a}}_s\left(t\right)$ denotes the amplitude after outlier correction for subcarrier s at time t. This preserves the temporal structure, while correcting local anomalies.

Although the Hampel filter corrects localized anomalies, the amplitude scale still varies across subcarriers owing to the propagation distance, reflection strength, and antenna sensitivity. These inter-subcarrier scale differences may bias the feature distribution, causing certain frequency bands to dominate the learning process of a wi-fi sensing model. To mitigate this imbalance, the amplitude is converted to a decibel (dB) scale.

$${\widehat{a}}_s\left(t\right)=20{log}_{10}\left({\tilde{a}}_s\left(t\right)\right).$$

(8)

The resulting value ${\widehat{a}}_s\left(t\right)$ represents the dB-scale amplitude of subcarrier s at time t. This transformation stabilizes the amplitude range and emphasizes the relative variations that are useful for gesture discrimination.

After outlier correction and dB conversion, the amplitude sequence A still has a large temporal length T, making it computationally inefficient to feed the full sequence into the model and potentially diluting gesture-related temporal patterns. To address this issue, we divided the time axis into fixed-length segments of size $N_t$, where each segment captures a localized temporal window of the gesture. The r-th segment is defined as

$$A_r=A[:,(r-1)N_t:r{N}_t]\in {\mathbb{R}}^{N_s\times N_t},r=1,2,\dots ,R,$$

(9)

where $R=\lfloor T/N_t\rfloor$ denotes the total number of non-overlapping segments obtained from CSI measurements for T. These segments serve as basic temporal units that preserve localized gesture patterns while reducing the computational burden of processing a full-length CSI sequence.

Although segmenting the CSI sequence improves efficiency, training the model using segments from a single domain may lead to overfitting to domain-specific characteristics. To promote robustness against domain shifts, the model should be trained using CSI data collected from multiple domains for the same gesture class. Each training input consists of M segments selected from multiple domains. For each segment index m, a domain is sampled independently from the domain pool, denoted as $d_m\sim U\left(D\right)$, where $d_m$ represents the domain chosen for the m-th segment. A segment of gesture class c is then sampled from domain $d_m$. Repeating this procedure M times yields

$$X\left(c\right)=\{A_1^{(c,d_1)},A_2^{(c,d_2)},\dots ,A_M^{(c,d_M)}\}\in {\mathbb{R}}^{M\times N_s\times N_t}.$$

(10)

$A_m^{(c,d_m)}$ denotes the m-th segment of class c drawn from domain $d_m$. Because $X\left(c\right)$ contains segments drawn from different domains while belonging to the same gesture class, it provides a diverse multi-domain sample set for training. This multi-domain segment set serves as the final input representation for model training. The goal of the data preprocessing pipeline is to stably refine raw CSI measurements and transform them into representations that enable the model to learn characteristics across diverse domains. By focusing on the CSI amplitude, the pipeline avoids phase instability while preserving signal variations induced by gestures. Hampel filtering and dB-scale conversion suppress noise and normalize inter-subcarrier scale differences, allowing meaningful patterns to emerge across frequencies. Temporal segmentation extracts localized motion patterns while reducing computational complexity. Finally, constructing each training input by assembling segments from multiple domains encourages the model to focus on gesture-related characteristics rather than domain-specific factors, thereby improving robustness to domain shifts.

3.3. Dual-Axis Transformer for Feature Extraction

The preprocessed input $X^{\left(c\right)}$ contains M segments collected from different domains but belonging to the same gesture class c. To effectively capture both spatial and temporal relationships within each segment $A_m^{(c,d_i)}$, we propose a dual-axis transformer feature extractor (DATFE). The overall process of the proposed DATFE is illustrated in Figure 2. This module incorporates two transformer branches, each specialized for capturing complementary characteristics of the CSI data. Notably, time-wise and subcarrier-wise patterns in CSI data exhibit fundamentally different characteristics. Therefore, a single transformer that focuses on only one axis is insufficient to fully represent the heterogeneous structure of CSI signals. By jointly modeling temporal variations and subcarrier-wise responses through a dual-axis design, DATFE enables more informative feature extraction.

Based on this design, the first branch, referred to as the subcarrier-axis transformer models spatial interactions across subcarriers by treating each subcarrier as an input token to a transformer. The second branch, referred to as the time-axis transformer captures temporal variations by treating each time frame as a token. The input tokens for each branch are defined as follows.

$$x_s=A_m^{(c,d_m)}[s,:]\in {\mathbb{R}}^{1\times N_t},x_t=A_m^{(c,d_m)}{[:,t]}^{\top }\in {\mathbb{R}}^{1\times N_s}.$$

(11)

$x_s$ corresponds to the temporal sequence of subcarrier s, whereas $x_t$ corresponds to the subcarrier response vector at time t. Using these token definitions, the input sequences for each branch are constructed as

$$X_{\mathrm{sc}}={\left\{x_s\right\}}_{s=1}^{N_s},X_{\mathrm{time}}={\left\{x_t\right\}}_{t=1}^{N_t}.$$

(12)

Thus, in the Subcarrier-axis Transformer, each token represents the temporal variation of subcarrier s, whereas in the Time-axis Transformer, each token corresponds to the full subcarrier response at time t. A [CLS] token is added to the beginning of both input sequences $X_{\mathrm{sc}}$ and $X_{\mathrm{time}}$ to summarize their global contextual information. Both branches share the same transformer architecture. Each token is projected into a latent space of dimension $d_{\mathrm{model}}$ via a linear embedding layer, followed by sinusoidal positional encoding to preserve the ordering information.

Each branch passes its embedded sequence through several transformer encoder blocks. The multi-head self-attention mechanism captures complex dependencies among tokens by jointly attending to them in multiple representation spaces. Each encoder block consists of multi-head self-attention, a feedforward network, residual connections, and layer normalization. After transformer encoding, the [CLS] token in each branch serves as the global summary representation for each $X_{\mathrm{sc}}$ and $X_{\mathrm{time}}$, resulting in the following vectors:

$$z_{\mathrm{sc}}^{\left(m\right)},z_{\mathrm{time}}^{\left(m\right)}\in {\mathbb{R}}^{d_{\mathrm{model}}}.$$

(13)

Here, $z_{\mathrm{sc}}^{\left(m\right)}$ represents the global embedding of the subcarrier-axis sequence $X_{\mathrm{sc}}$, while $z_{\mathrm{time}}^{\left(m\right)}$ represents the global embedding of the time-axis sequence $X_{\mathrm{time}}$. Because these two vectors capture complementary perspectives of the same segment, they are fused as follows:

$$z_{\mathrm{joint}}^{\left(m\right)}=\left[\mathrm{LN}\left(z_{\mathrm{sc}}^{\left(m\right)}\right);\mathrm{LN}\left(z_{\mathrm{time}}^{\left(m\right)}\right)\right]\in {\mathbb{R}}^{2d_{\mathrm{model}}}.$$

(14)

first, each representation is normalized using layer normalization (LN) to ensure scale consistency across the two axes, and the normalized vectors are then concatenated to form a joint embedding. A gating network is applied to dynamically control the relative importance of the two axes. The gate values are obtained by passing the joint embedding through a gating MLP followed by a sigmoid activation:

$$[g_{\mathrm{sc}}^{\left(m\right)},g_{\mathrm{time}}^{\left(m\right)}]=\sigma \left(f_{\mathrm{gate}}\left(z_{\mathrm{joint}}^{\left(m\right)}\right)\right).$$

(15)

Here, $f_{\mathrm{gate}}(·)$ denotes the gating MLP that determines the contribution of each axis, $\sigma (·)$ is the sigmoid activation function. These gate values are then used to perform element-wise weighted fusion between each representation and its corresponding gate:

$$z_{\mathrm{fused}}^{\left(m\right)}=\left[z_{\mathrm{sc}}^{\left(m\right)}\odot g_{\mathrm{sc}}^{\left(m\right)};z_{\mathrm{time}}^{\left(m\right)}\odot g_{\mathrm{time}}^{\left(m\right)}\right].$$

(16)

Finally, the fused vector is passed through a fusion MLP to produce the final segment-level representation:

$$f^{\left(m\right)}=f_{\mathrm{fusion}}\left(z_{\mathrm{fused}}^{\left(m\right)}\right)\in {\mathbb{R}}^{d_f}.$$

(17)

$f_{\mathrm{fusion}}(·)$ represents the fusion MLP composed of Layer Normalization, a linear layer, and a GELU activation.

3.4. Graph-Based Feature Integration

The segment-level features obtained from the DATFE represent the same gesture class but are derived from different domains. Consequently, each $f^{\left(m\right)}$ reflects domain dependent statistical variations, leading to heterogeneous feature distributions, even within the same gesture class. Even though these segment-level features correspond to the same gesture class, they are not independent but share latent relationships induced by common gesture semantics across different domains. Accordingly, we formulate the segment-level features and their cross-domain relationships within a graph-based framework. In this context, the graph is constructed in the learned feature space, where each node represents a segment-level feature embedding corresponding to the same gesture class observed under different domains. These features are high-level representations obtained by the DATFE that abstract discriminative gesture patterns and encapsulate gesture semantics shared across domains. Accordingly, the relationships between nodes are naturally characterized by similarity in the feature space, reflecting the degree to which features extracted from different domains consistently represent the same gesture semantics. An overview of the GNN-based feature integration is shown in Figure 3.

Specifically, each feature vector $f^{\left(m\right)}$ is treated as a graph node, and the pairwise relationships among the nodes are encoded as edge weights. The set of graph nodes is defined as

$$F=[f^{\left(1\right)},f^{\left(2\right)},\dots ,f^{\left(M\right)}]\in {\mathbb{R}}^{M\times d_f}.$$

(18)

The edge weight between two nodes is computed as the cosine similarity between their ${\ell }_2$-normalized feature vectors, followed by linear scaling to the range $[0,1]$. The resulting adjacency matrix $S\in {[0,1]}^{M\times M}$ is defined as

$$S_{ij}={\displaystyle \frac{1}{2}}\left({\displaystyle \frac{f^{\left(i\right)\top }{f}^{\left(j\right)}}{\parallel f^{\left(i\right)}\parallel \parallel f^{\left(j\right)}\parallel }}+1\right).$$

(19)

Hence, $S_{ij}=1$ indicates perfect alignment of the feature directions between segments i and j, whereas $S_{ij}=0$ indicates that the directions are opposite to each other. By passing messages on this graph, the GNN aggregates information from related segments, which suppresses domain-specific variations while enhancing gesture-consistent feature components shared across domains.

The adjacency matrix S and feature matrix F are then fed into a single GNN layer. In this study, we apply a simple linear GNN, as each node already represents a high-level segment embedding extracted by DATFE and the number of nodes is relatively small. Additionally, mean aggregation is employed to integrate segment-level representations, since all nodes correspond to the same gesture class and mainly differ due to domain-induced variations. The updated node representation is computed by aggregating the neighboring features using a degree-normalized adjacency matrix, followed by a linear transformation and activation:

$$H=\varphi \left(D^{-1}SF{W}_g\right)\in {\mathbb{R}}^{M\times d_g},$$

(20)

where $D=\mathrm{diag}\left(S\right)$ is the diagonal degree matrix, $W_g\in {\mathbb{R}}^{d_f\times d_g}$ is a learnable weight matrix, and $\varphi (·)$ denotes the rectified linear unit activation function.

Next, all node embeddings are average-pooled to obtain a global graph-level representation as follows:

$$g={\displaystyle \frac{1}{M}}\sum _{m=1}^M{H}_m\in {\mathbb{R}}^{d_g}.$$

(21)

The resulting vector g serves as an integrated representation of the entire gesture, capturing inter-segment relationships across domains and is subsequently used as the input to the classifier.

3.5. Classifier

We implement a fully connected neural network that learns nonlinear decision boundaries between gesture classes through a series of linear layers and nonlinear activation functions. The integrated feature vector g is first projected into a lower-dimensional latent space, followed by a nonlinear transformation using an activation function $\varphi (·)$, and finally passed through a softmax layer to produce the predicted class probability vector $\widehat{y}\in {\mathbb{R}}^{N_c}$:

$$\widehat{y}=\mathrm{Softmax}\left(W_2\varphi (W_{1}g+b_1)+b_2\right).$$

(22)

Here, $W_1\in {\mathbb{R}}^{d_g\times d_h}$ and $W_2\in {\mathbb{R}}^{d_h\times N_c}$ denote the learnable weight matrices of the linear transformations, and $b_1$ and $b_2$ represent the bias terms. The activation function $\varphi (·)$ is implemented as a ReLU, and $d_h$ denotes the dimension of the hidden layer.

To train the classifier, we combine the cross-entropy loss and center loss [31]. The cross-entropy loss encourages the model to align the predicted probabilities with the ground-truth labels:

$$L_{\mathrm{CE}}=-\sum _{k=1}^{N_c}{y}_{k}log\left({\widehat{y}}_k\right),$$

(23)

where y is the one-hot encoded ground-truth label. The center loss enhances intra-class compactness by minimizing the distance between each feature vector and its corresponding class center:

$$L_{\mathrm{Center}}={\displaystyle \frac{1}{2}}\sum _{m=1}^B{∥f^{\left(m\right)}-c_{y^{\left(m\right)}}∥}_2^2,$$

(24)

where $c_{y^{\left(m\right)}}$ denotes the learnable center of class $y^{\left(m\right)}$ and B is the batch size. The class centers $\left\{c_y\right\}$ were initialized using a random uniform distribution. The final training objective is formulated as:

$$L=L_{\mathrm{CE}}+\lambda L_{\mathrm{Center}},$$

(25)

where $\lambda$ is a balancing weight that adjusts the contribution of the center loss.

4. Experimental Evaluation

4.1. Data Collection and Implementation Details

Data were collected to acquire Wi-Fi CSI measurements for recognizing finger-location gestures. Figure 4 illustrates the relative placement of the transmitter, CSI capture device(CD), and four location target boards used as the gesture interface. Figure 5 shows the floor plan of the seminar room in which the measurements were performed. All devices operate under the IEEE 802.11n standard on a 20 MHz channel in the 2.4 GHz band. An ESP8266 microcontroller equipped with a low-power Wi-Fi chipset serves as the transmitter and is configured in the station mode to connect to the access point (AP). The transmitter continuously generated UDP packets destined for a non-existent IP address within the same subnet to avoid the need for a specific destination MAC address. Each UDP packet carried a 1-byte payload, and the transmission interval was fixed at $\Delta t=10\mathrm{ms}$. The transmitted Wi-Fi frames propagated through the indoor environment and were captured by a RaspberryPi-4B serving as the CSI capture device. The Nexmon CSI tool [32] is installed on the CD to extract the CSI measurements for every received Wi-Fi frame. Each OFDM symbol consisted of ${\tilde{N}}_s=64$ subcarriers, from which the pilot and null subcarriers were removed. As a result, only the $N_s=52$ data subcarriers were retained for the gesture analysis. The four location target board was mounted at the midpoint between the transmitter and CSI capture device, and the separation between the two devices was maintained at approximately 50 cm. Six participants voluntarily participated in the data collection process. Each participant performed independent measurement sessions for ten separate days. Each daily session included four gesture classes. For each gesture class, the participant maintained the corresponding finger location pose for one minute. Based on the packet sampling rate of the system, each one-minute recording produced approximately 6000 CSI frames per gesture class, resulting in approximately 24,000 frames per daily session for each participant. Demographic information of the participants is provided in

Table 3. Participant information.

	P1	P2	P3	P4	P5	P6
Gender	Male	Male	Male	Male	Male	Female
Height (cm)	180–190	170–180	170–180	170–180	160–170	150–160

. In addition, the key statistics of the collected CSI dataset are summarized in

Table 4. Summary of CSI dataset statistics.

Statistic	Value
Sampling interval	10 ms
Number of participants	6
Recording days per participant	10
CSI frames per gesture class (per day)	6 k
CSI frames per participant (per day)	24 k
Total CSI frames	1.44 M

.

The collected data were processed using the proposed preprocessing pipeline and subsequently used for model training. The main experimental hyperparameters are summarized as follows. Hampel filtering was applied for outlier removal using a window size of $w=5$ and threshold factor $\gamma =3$. The segment length was set to $N_t=52$, which corresponds to the number of data subcarriers in the 20 MHz IEEE 802.11n configuration. This choice was intentional, as it enables each segment to be represented as a square $52\times 52$ matrix after temporal segmentation, thereby allowing uniform modeling across the time and subcarrier dimensions without introducing interpolation or padding artifacts. For the transformer encoder, the model dimension was set to $d_{\mathrm{model}}=128$, the number of attention heads to 4, and the number of encoder blocks to 3. The GNN-based feature integration module operated on $M=6$ segments and employed a single nonlinear GNN layer. The value of M was determined empirically to balance temporal context aggregation and computational efficiency, providing sufficient cross-segment diversity while keeping the graph size manageable for efficient training and inference. The weighting coefficient for the center loss was set to $\lambda =0.01$. The model training was performed using the Adam optimizer with an initial learning rate of $1\times {10}^{-4}$ and a weight decay of $5\times {10}^{-5}$. The class centers for the center loss were updated using a learning rate of $1\times {10}^{-2}$. The batch size was set to 256, and training was conducted for 300 epochs, with early stopping applied when the validation loss did not improve for 20 epochs. The hyperparameters used in the experiment are summarized in

Table 5. Summary of experimental hyperparameter settings.

Component	Hyperparameter	Value
Preprocessing	w	5
	$\gamma$	3
	$N_t$	52
DATFE	$d_{\mathrm{model}}$	128
	# Attention heads	4
	# Encoder blocks	3
GNN Integration	M	6
	# GNN layers	1
Loss	$\lambda$	0.01
Training	Optimizer	Adam
	Learning rate (main optimization)	$1\times {10}^{-4}$
	Learning rate (center optimization)	$1\times {10}^{-2}$
	Batch size	256
	Max epochs	300
	Early stopping patience	20

. All experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 4070 SUPER GPU (12 GB VRAM), Intel Core i5-14600KF CPU (3.50 GHz), and 32 GB of RAM. The software environment comprised Python 3.12.7, CUDA 12.7, and PyTorch 2.5.1. The dataset splits and random seeds were fixed to ensure reproducibility. In addition to the experimental setup described above, we further discuss the practical deployment and real-time feasibility of the proposed system. In the proposed system, low-power devices such as a Raspberry Pi operate as CSI collection devices (CDs), while model inference is performed on a separate server. Each CSI frame consists of in-phase and quadrature components, and its size is $8N_s$ bytes when each component is represented using 4 bytes. With $N_s=52$, the resulting CSI frame size is only 416 bytes. Under a typical wired or wireless LAN environment with a throughput of at least 100 Mbps, transmitting a CSI frame to the server requires approximately 33.28 μs. Combined with the average server-side inference time of 1.5 ms per sample, the overall end-to-end latency remains well below the 10 ms CSI sampling interval, indicating that this sampling rate does not impose a practical constraint on system operation.

4.2. Performance Evaluation

In this section, we evaluate the effectiveness of the proposed method under two performance settings: in-domain and cross-domain. These settings reflect two distinct goals in Wi-Fi CSI–based gesture recognition: the former concerns how well the model classifies gestures when the training and test data share the same domain characteristics, whereas the latter examines the model’s ability to generalize to unseen data collected under different user- or environment-related conditions. In our dataset, variations across participants naturally arise from differences in body shape, hand geometry, motion tendencies, and resulting signal reflection patterns. Likewise, variations across measurement days occur due to changes in clothing, physical condition, ambient temperature, humidity, and subtle differences in the device placement or orientation. These two sources of variation correspond to the evaluation scenarios considered in this paper, namely, cross-user and cross-day settings.

To verify the proposed method, we compare the performance of our method to those of four representative baseline approaches introduced in Section 2. These baselines each represent a distinct methodological design choice in Wi-Fi CSI based gesture recognition. Specifically, the selected baselines cover different perspectives, including alternative representation learning paradigms, CNN-based deep learning architectures, feature space data augmentation strategies, and explicit domain generalization frameworks. This selection enables a systematic evaluation of the proposed model against existing approaches across diverse methodological dimensions. All methods are trained and evaluated using identical datasets and experimental protocols to ensure fair comparison.

In the in-domain setting, we first assess how reliably the proposed model recognizes gestures when the training and test samples originate from the same participant and measurement day.

Table 6. Five-fold In-domain performance of the proposed model.

Metric	Fold 1	Fold 2	Fold 3	Fold 4	Fold 5	Average
Acc (%)	99.01	99.63	99.14	99.38	99.38	99.31
Precision (%)	99	99.64	99.14	99.39	99.39	99.31
F1-score (%)	98.99	99.64	99.14	99.38	99.38	99.31

presents the five-fold cross-validation results, showing that accuracy, f1-score, and precision all exceed 98.99% across folds, with an average accuracy of 99.31%. These results indicate that gesture patterns within the same acquisition condition are highly consistent and that the proposed model effectively captures the intrinsic characteristics of each gesture class in the data. To further contextualize the performance,

Table 7. In-domain performance comparison with representative baselines.

Model	Kan-Sense	CSI-DeepNet	WiSGP	DGSense	Ours
Acc (%)	98.41	93.62	91.91	96.52	99.31

compares the proposed method with the four baseline approaches in the in-domain setting. Although all methods achieve high accuracy owing to the relatively stable data distribution, the proposed method attains the best overall performance with an accuracy of 99.31%, demonstrating its strong representation capacity even without domain variation.

In contrast, the cross-domain setting evaluates the model’s robustness when the training and test data come from different participants (cross-user) or from different measurement days (cross-day). All experiments are repeated three times with different random seeds, and the mean accuracy along with the standard deviation is reported. Error bars corresponding to the standard deviation are explicitly added in Figure 6 and Figure 7. As summarized in Figure 6, the average accuracy in the cross-user case is 54.43%. The performance decreases substantially compared with the in-domain results because each participant exhibits distinct physical characteristics, hand shapes, and gesture execution styles, all of which influence the CSI patterns captured by the receiver. Figure 7 reports the results for the cross-day setting, where the model is trained on data collected over nine days and tested on the remaining day. The proposed method achieves an average accuracy of 56.61%, outperforming all baseline approaches. Day-to-day differences, such as clothing, physical state, ambient conditions, and slight variations in device placement, introduce noticeable changes in the CSI amplitude patterns, resulting in degraded performance for all methods. Nevertheless, the proposed approach consistently demonstrates superior generalization capability compared to other methods.

Among the baselines, WiSGP and DGSense are originally designed for domain-generalization purposes. However, in the single-link Wi-Fi setting used in this study, these advantages diminish. The WiSGP relies on subdomain-guided perturbations to generate virtual domain samples. However, in a single-link scenario, such perturbations inject excessive noise and reduce the training stability. Similarly, DGSense depends on multi-link information to construct 2D compressed Doppler maps, which are unavailable in a single-link environment. These limitations result in a performance drop of approximately 6–13% compared with the proposed method.

The absolute accuracy in cross-domain settings is approximately 55%, which reflects the inherent difficulty of the task in extremely challenging single-link CSI sensing conditions. However, the proposed approach achieved a minimum improvement of 3.5% in cross-user performance and 2.6% in cross-day performance over the comparative methods. These results clearly demonstrate the severe domain shift inherent in cross-domain Wi-Fi sensing and underscore the tangible effectiveness of the proposed framework in improving generalization under realistic deployment conditions. To further examine the robustness of the proposed method under unseen environmental conditions, we conducted an additional evaluation in a seminar room that was not used during training. CSI data were collected from a total of three participants, including two participants, P1 and P2, who had contributed to the training data, and one additional participant who was completely unseen during training. For each participant, approximately one minute of data per gesture class was recorded in the new environment and used exclusively for testing. Under this setting, the proposed method achieved classification accuracies of 53.23% and 62.71% for P1 and P2, respectively, while the accuracy for the unseen participant dropped to 46.88%, which was the lowest among the three cases. This performance degradation highlights the compounded difficulty introduced by simultaneous domain shifts in both environment and user identity. Nevertheless, even for the unseen participant in a completely new spatial setting, the proposed method maintains non-trivial recognition performance, suggesting that the learned representations capture gesture-related characteristics that are partially invariant to environmental and user-dependent variations.

4.3. Visualization of the Preprocessing

Figure 8 provides a visual illustration of how the proposed preprocessing pipeline affects the CSI amplitude signals. The figure presents amplitude measurements from class 1 data collected from a randomly selected participant, offering a representative example of the raw and preprocessed signal characteristics, with amplitudes from subcarriers 3 and 15 used for illustration. In Figure 8a, abrupt spikes and irregular fluctuations are clearly observed, which act as noise unrelated to the underlying gesture information and hinder the model from learning stable and consistent features. After applying the preprocessing pipeline, the signal becomes noticeably more stable. As shown in Figure 8b, local impulsive deviations are substantially reduced, resulting in a smoother and more coherent temporal pattern. This demonstrates that the preprocessing effectively suppresses abrupt disturbances while preserving the essential structure of the gesture-related signal. Figure 8c further presents the amplitude distribution after the dB-scale transformation. The dynamic range across the two subcarriers becomes significantly more uniform, and the overall amplitude scale is compressed, making relative variations more distinguishable. This reduces the bias caused by inter-subcarrier magnitude differences and prevents certain frequency bands from disproportionately influencing the feature extraction process. In conclusion, the proposed preprocessing pipeline effectively mitigates signal outliers and alleviates inter-subcarrier scale discrepancies, thereby creating a more stable foundation for learning gesture-relevant patterns.

4.4. Ablation Study

To analyze the structural effectiveness of the proposed framework, we conduct an ablation study. In particular, we explicitly disentangle the contributions of preprocessing techniques, single-axis versus dual-axis Transformer architectures, and a Transformer-only configuration versus a Transformer combined with a GNN-based integration module. The experimental results are summarized in Figure 9, which visualizes the performance degradation of each ablated component across the in-domain, cross-user, and cross-day settings. To this end, we construct multiple variant models by selectively removing individual components of the proposed pipeline. First, we analyze the effect of preprocessing by removing the dB conversion and the multi-domain segment combination strategy. Next, to evaluate the structural contribution of the Dual-Axis Transformer (DATFE), we construct single-axis Transformer variants by removing either the subcarrier-axis encoder or the time-axis encoder. Finally, to assess the importance of cross-domain feature integration, we consider a Transformer-only configuration in which the GNN-based integration module is removed. For the analysis of the multi-domain segment combination strategy, each input sample is reconstructed using only segments extracted from a single domain. In addition, in the GNN ablation experiment, instead of integrating multiple segment-level features, each segment is independently processed by the DATFE and classified without any cross-domain integration. This setting represents a pure Transformer operating at the segment level. In the in-domain setting, all variant models achieve generally high classification accuracy. This reflects the fact that gesture patterns are relatively stable when the training and test data are collected from the same participant under consistent measurement conditions. Nevertheless, removing the time-axis encoder from the DATFE results in the largest performance degradation, with an accuracy reduction of approximately 5.91%. This indicates that temporal dynamics play a critical role in gesture discrimination in the in-domain setting. In contrast, removing the subcarrier-axis encoder leads to only a minor decrease in accuracy, reaching 99.01%, which can be attributed to the relatively small spatial distributional differences across subcarriers within the same domain. In contrast, in the cross-domain experiments, performance differences among the variant models become much more pronounced. Removing the dB conversion decreases cross-user accuracy by 4.95% and cross-day accuracy by 5.96%. Since CSI amplitude values inherently exhibit large scale differences across subcarriers, omitting dB normalization allows certain subcarriers to dominate the learned representation, making the model more vulnerable to domain shifts. Similarly, eliminating the multi-domain segment combination strategy results in performance drops of 3.75% in the cross-user setting and 5.78% in the cross-day setting, indicating that using only single-domain segments provides insufficient representational diversity for cross-domain evaluation. The comparison between single-axis and dual-axis Transformer architectures further highlights the effectiveness of the proposed design. Removing the subcarrier-axis encoder results in accuracy drops of 3.78% in the cross-user setting and 3.88% in the cross-day setting, while removing the time-axis encoder leads to reductions of 2.7% and 1.2%, respectively. These results suggest that, under cross-domain conditions, Wi-Fi CSI is more sensitive to the spatial distribution of signal strength across subcarriers than to fine-grained temporal dynamics. This implies that spatial patterns captured along the subcarrier axis reflect geometry-dependent characteristics determined by body shape and reflection paths, thereby providing more robust and transferable cues across unseen users and measurement days. In contrast, temporal dynamics are relatively more vulnerable to noise induced by variations in gesture execution and environmental changes. Finally, under the Transformer-only configuration, where the GNN-based integration module is removed and each segment is independently classified using the DATFE, cross-user and cross-day accuracy decreases by approximately 2.2%. This suggests that although the Transformer-only model can effectively extract segment-level features, segments collected from different domains may still exhibit heterogeneous feature distributions within the same gesture class. In contrast, the GNN-based integration module explicitly models inter-segment relationships through similarity-based graph connectivity, thereby integrating complementary domain information and facilitating the learning of domain-invariant representations.

4.5. Feature Alignment Analysis

In this section, we qualitatively analyze how effectively the proposed method performs feature alignment under cross-domain conditions. To this end, we first compute the class-wise centroid using the features extracted from the training data:

$${\mu }_c={\displaystyle \frac{1}{N_c^{\mathrm{train}}}}\sum _{i\in c}{f}_i^{\mathrm{train}},$$

(26)

where ${\mu }_c$ denotes the centroid of class c in the feature space, $N_c^{\mathrm{train}}$ denotes the number of training samples belonging to class c, and $f_i^{\mathrm{train}}$ represents the feature vector extracted from the i-th sample of that class. To measure how far each test feature lies from its corresponding class center, we compute the L2 distance:

$$r_{j,c}={∥f_j^{\mathrm{test}}-{\mu }_c∥}_2.$$

(27)

Here, j denotes the index of a test sample belonging to class c, and $r_{j,c}$ indicates how well the sample aligns with the class distribution learned from the training data. A smaller value implies that the sample is better aligned with the embedding space of its corresponding class. Since embedding scales vary across models, class-wise mean normalization is applied to ensure a fair comparison of distance distributions among different methods. Figure 10 illustrates the normalized distance boxplots for each gesture class across all baseline models and the proposed method, evaluated under the cross-day setting where Day 3 serves as the test domain. In most classes, the proposed method exhibits a noticeably lower median compared with competing approaches. This indicates that the data processed by the proposed method is positioned closer to its class centroids, and implies that the model achieves effective feature alignment even in unseen domains. To further examine how these distance characteristics affect classification performance, we evaluate each method after removing test samples based on a third-quantile ($Q3$) distance threshold. For each class, the $Q3$ value is computed from the distribution of distances between test samples and the corresponding class centroid obtained from the training data. This filtering focuses the analysis on the majority of samples that are well aligned with their class centroids, enabling a clearer examination of the relationship between feature alignment and classification performance. Figure 11 illustrates the confusion matrices of the proposed method before and after outlier removal under the same cross-day evaluation setting. After discarding high distance data, the accuracies for class 2, class 3, and class 4 improve by approximately 14%, 21%, and 16%, respectively. This observation indicates that most misclassifications originate from a small number of samples mapped far from their class centers. Moreover,

Table 8. Overall classification accuracy (%) of each method before and after removing samples whose distances exceed the third quantile (All: using all test samples, Filtered: after removing test samples exceeding the third quantile).

	KAN	CSI-DeepNet	WiSGP	DGSense	Ours
All	63.05	65.32	66.67	59.10	73.35
Filtered	68.29	78.41	81.26	70.35	86.27

summarizes the overall accuracy changes for all models in this Day 3 test scenario before and after removing data above $Q3$. While all models show certain improvements, the proposed method achieves the largest performance gain of approximately 13% along with the highest final accuracy among all compared methods. Overall, these results demonstrate that the proposed method effectively aligns the majority of test data close to their corresponding class centers under cross-domain conditions, thereby directly contributing to improved classification performance.

We qualitatively analyze the feature alignment behavior using t-SNE visualization under a cross-domain evaluation setting, where Day 3 is used as the test domain. Figure 12a shows the segment-level features extracted from a model trained using only DATFE, where the features exhibit dispersed distributions across different domains even within the same gesture class. In contrast, Figure 12b illustrates the feature distribution after integrating multi-domain segments through the GNN following DATFE, where the features become more compact and form relatively better separated class-wise clusters. This qualitative observation suggests that, while DATFE extracts discriminative spatiotemporal features from individual segments, segments collected from different domains may still exhibit heterogeneous distributions. By explicitly modeling inter-segment relationships across multiple domains, the GNN aggregates information from semantically similar segments while suppressing domain-specific variations. As a result, features extracted from different domains converge toward more consistent class-specific representations, leading to improved feature alignment in the embedding space.

5. Conclusions and Future Works

In this paper, we propose a gesture recognition framework tailored for single-link Wi-Fi-based CSI sensing. The system includes a preprocessing pipeline that mitigates hardware-induced perturbations and a multi-domain segment combination strategy that exposes the model to diverse gesture manifestations. The Dual-Axis Transformer independently captures spatial and temporal patterns, while the GNN-based integration module aggregates inter-segment dependencies. Experimental results demonstrate that the proposed model achieves 99.31% accuracy in the in-domain scenario and outperforms representative baselines by at least 3.5% and 2.6% under the cross-user and cross-day settings, respectively. Despite these promising results, a notable limitation remains in that the cross-domain performance is still substantially lower than the in-domain accuracy. This gap highlights the inherent difficulty of generalizing from limited single-link CSI observations, where variations across users and measurement days significantly alter the propagation paths and lead to distributional inconsistencies. Addressing this limitation is an important direction for future research.

Several deep learning-based extensions can be pursued to further improve the generalization capability. First, incorporating advanced denoising or distortion-removal networks can enhance the CSI quality by suppressing environment and hardware-related artifacts before feature extraction. Second, employing generative models, such as variational GANs, may enable the synthesis of diverse virtual CSI samples, thereby expanding domain coverage without incurring additional data collection costs. Third, adopting a federated learning paradigm would allow for multi-user collaboration while preserving privacy, potentially improving robustness to user-specific variations in real-world scenarios. From the evaluation scenario perspective, future work can extend beyond the current static finger gesture recognition setting. Specifically, the proposed framework can be evaluated under continuous finger gesture scenarios, where gesture transitions occur sequentially over time. In addition, robustness can be further examined in environments with moving bystanders. In such settings, external motion introduces interference and noise into CSI measurements, enabling a more realistic assessment of single-link CSI-based gesture recognition systems.

Despite these remaining challenges, the proposed framework demonstrates the practical potential of single-link CSI-based gesture recognition. While this study focuses on finger-location gesture recognition, the proposed framework is not limited to this specific task and can be extended to a broader range of fine-grained gesture and interaction recognition scenarios based on single-link CSI sensing. Also, by improving robustness in single-link CSI-based sensing environments, this work takes a step toward Wi-Fi sensing that operates reliably out-of-the-box, without requiring complex hardware configurations or prior environment calibration. This capability reduces deployment cost and system complexity, while enabling new forms of accessible and privacy-conscious human–computer interaction in everyday environments.