GeoDFNet: a point-of-interest classification algorithm with dual fusion of geospatial local neighborhood features

The third law of geography

Geospatial similarity, manifested through the widely recognized Third Law of Geography (Zhu et al., 2018), establishes that ‘geographic entities in analogous spatial environments exhibit congruent characteristics’, with universality and generalizability empirically confirmed (Zhu et al., 2020). This implies consistent attribute manifestation among entities within homogeneous geospatial neighborhoods, regardless of location. We operationalize this principle through categorical neighborhood congruence: By quantifying adjacent POI similarity as a neighborhood proxy, cross-location analogy emerges when proximate entities show categorical alignment. This defines analogous neighborhoods implying intrinsic anchor POI similarity, transforming geographical principles into computable features for neighborhood topology integration in POI classification.

Graph neural networks

The inherent graph structure of geospatial data—where POIs constitute nodes and spatial relationships (e.g., proximity, adjacency) form edges—establishes graph neural networks (GNNs) as a natural paradigm for modeling POI dependencies. Foundational architectures including graph convolutional networks (GCN) (Kipf & Welling, 2016), GraphSAGE (Hamilton, Ying & Leskovec, 2017), and GAT enable direct operation on irregular spatial topologies, adaptively capturing neighborhood dependencies essential for resolving POI ambiguities (e.g., classifying stationery stores adjacent to schools vs. commercial zones, or distinguishing medical facilities proximate to pharmacies from retail clusters). This capability originates from core GNN principles: The message-passing mechanism models neighborhood influence dynamics wherein POI semantics are contextually refined by local environments, while permutation invariance guarantees robustness to irregular geospatial point patterns—fulfilling fundamental requirements for geographic knowledge discovery.

Multimodal fusion

Multimodal fusion entails joint modeling of complementary attributes across heterogeneous data streams. While current research predominantly integrates vision-language modalities (e.g., images, text) (Zhou et al., 2023; Xu et al., 2023), this study pioneers geospatial-textual fusion to bridge neighborhood topology and POI name semantics. Inspired by cross-modal reinforcement paradigms—demonstrated in Liu et al.’s (2024b) fusion of remote sensing and trajectory data for road-aware POI identification—we develop a dual-stream framework: Geospatial embeddings encode Neighborhood attributes through spatial aggregation while textual embeddings capture lexical patterns. By dynamically harmonizing spatial neighborhood and textual semantics, the model capitalizes on their discriminative synergies, enabling spatial clusters to resolve textual ambiguities (e.g., ‘Dragon’ in restaurant vs. apparel contexts) and lexical cues to interpret spatial anomalies, significantly enhancing classification robustness across urban morphologies.

Overall architecture

The architecture of the proposed GeoDFNet model is illustrated in Fig. 1, which consists of four principal components: the Geospatial Neighborhood (GeosN) module, Data Matching module, Textual-Geographical Feature Fusion (T-G Feature Fusion) module, and Enhanced Geospatial Local Neighborhood module. This hierarchical design integrates geospatial neighborhood understanding with cross-modal feature interaction, followed by localized neighborhood refinement to enhance geographical representation learning.

GeosN module: This module harnesses the message propagation and neighborhood aggregation paradigm of graph neural networks. Specifically, it incorporates a first-layer of GAT with disabled self-loop connections to mitigate feature contamination from the target POI. By excluding the target node’s ego features during aggregation, the module ensures that geospatial representations are exclusively neighbor by spatially adjacent nodes. A multi-head attention mechanism is employed to adaptively recalibrate attention weights across neighbors, thereby capturing the comprehensive local geospatial neighborhood features. Finally, a linear transformation layer projects these features into a latent space, producing refined local geospatial neighborhood embeddings for downstream classification.

Data Matching Module: The GeosN module generates embeddings for all nodes; however, computational constraints necessitate batch-wise processing of textual features during training iterations, as handling the full dataset exceeds hardware capacity. This batch partitioning creates a mismatch between the quantity of textual features and geospatial local neighborhood features within individual batches. To address this discrepancy, the Data Matching module selectively aligns the geospatial local neighborhood features of corresponding POI nodes with their textual counterparts in the current batch. Specifically, this module matches both the complete and compressed geospatial local neighborhood features of geographic entities present in the batch, ensuring cross-modal consistency. The implementation employs a dual-indexing mechanism and feature similarity thresholds to dynamically establish correspondences, thereby mitigating computational overhead while preserving critical spatial-textual relationships for downstream fusion processes.

T-G Feature Fusion module: This module synthesizes spatial-semantic representations by integrating textual semantic embeddings and geospatial neighborhood embeddings. The Transformer encoder processes textual information from POI names to generate semantic embeddings, while the Data Matching Module provides task-aligned geospatial neighborhood features. By orchestrating cross-modal fusion between these modalities, the module leverages their discriminative synergies to construct a joint latent representation. This fusion mechanism enables joint modeling of semantic and spatial dependencies, significantly improving classification accuracy through enriched neighborhood.

Enhanced Geospatial Local Neighborhood module: In the text-geospatial local neighborhood features generated by the T-G Feature Fusion module, the contribution of text features significantly outweighs that of geospatial local neighborhood features. To address this imbalance, the Geospatial Local Neighborhood Enhancement module utilizes the compressed geospatial local neighborhood features from the GeosN module, which are filtered and aligned by the Data Matching module. These external features are then integrated through a decision fusion approach, effectively enhancing the representation of geospatial neighborhood. This process ensures a more balanced and robust feature set, strengthening the model’s ability to leverage spatial information for improved classification performance.

Finally, the fusion matrix, enriched with enhanced features, serves as the input to the Softmax layer. The subsequent section provides a detailed description of the model architecture and its components.

GeosN module

The geospatial local neighborhood module, based on graph neural networks, constructs a graph representation where real geographic entities serve as nodes, and the adjacency relationships between these entities form the edges of the graph. This module maps the geospatial neighborhood of the target POI into a high-dimensional vector space. By leveraging the message-passing and aggregation mechanism of graph neural networks, it effectively captures the geographic environment in accordance with the principle of geographical similarity. This process results in the formation of comprehensive local neighborhood features that encapsulate the spatial neighborhood of the target POI.

In this module, the GAT framework is employed to derive the geospatial local neighborhood feature vector representation. The aggregated features within its spatial domain are illustrated in Fig. 2.

In Fig. 2, the nodes V0, V1, V2, V3, V4 and V5 represent specific geographic entities, including a school, snack bars, a bookstore, and a bus station, respectively. Meanwhile, the nodes V6, V7, V8, V9 and V10 correspond to other geographic elements within the local spatial neighborhood.

The input to the GAT can be formally expressed as Eq. (1):

(1) $$h=\left\{{\overrightarrow{h}}_1,{\overrightarrow{h}}_2,\cdots ,{\overrightarrow{h}}_n\right\},{\overrightarrow{h}}_i\in {\mathbb{R}}^F.$$In Eq. (1), $n$ denotes the number of nodes in the graph, which corresponds to the number of POIs, while $F$ represents the initial feature dimension of each node.

A multi-head graph attention layer implemented without self-loops (as visualized in Fig. 3) aggregates contextual features from each node’s neighborhood. This process transforms the initial input features $h$ into higher-order knowledge representations $h^{\mathrm{\prime }}$, designated as the complete geospatial local neighborhood features. This design intentionally omits self-loop operations, distinguishing it from conventional GAT implementations.

(2) $$h^{\mathrm{\prime }}=\mathrm{G}\mathrm{A}{\mathrm{T}}_{no\mathrm{\_}self}\left(h\right).$$

(3) $$h^{\mathrm{\prime }}=\left\{{\overrightarrow{h^{\mathrm{\prime }}}}_1,{\overrightarrow{h^{\mathrm{\prime }}}}_2,\cdots ,{\overrightarrow{h^{\mathrm{\prime }}}}_n\right\}.$$In Eqs. (2) and (3), ${\mathrm{G}\mathrm{A}\mathrm{T}}_{no\mathrm{\_}self}$ denotes the graph attention operation excluding self-loops, ${\overrightarrow{h^{\mathrm{\prime }}}}_{\mathrm{i}}$ represents the complete geospatial local neighborhood feature for node $i$.

Finally, the aggregated local neighborhood features of all geospatial regions are compressed to the same dimension as the number of classifications through linear layers, and the probability distribution $h^g$ of geospatial local neighborhood features is obtained, that is, to compress the local neighborhood features of the geospatial space.

(4) $$h^g=\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}\left(h^{\mathrm{\prime }}\right)=\left\{{\overrightarrow{h^g}}_1,{\overrightarrow{h^g}}_2,\cdots ,{\overrightarrow{h^g}}_n\right\}.$$

Data matching module

During the construction of textual and graph datasets for POIs, each POI entity is assigned a unique identifier to synchronize its textual descriptors and graph node representations. For a training batch with textual inputs $T=\left[T_1,T_2,\dots ,T_{bs}\right],$ where $bs$ denotes the batch size, the module first extracts the unique POI identifiers $idx$ within the current batch. These identifiers are then cross-referenced with the node indices in the geospatial embeddings $h^{\mathrm{\prime }}$ and $h^g$ generated by the GeosN module. The aligned features are aggregated using a tensor stacking operation, yielding the batch-specific complete geospatial local neighborhood features $G^{\mathrm{\prime }}$ and compressed counterparts $G^g$, These aligned features serve as inputs to the T-G Fusion module and Enhanced Geospatial Local Neighborhood module, enabling joint optimization of cross-modal interactions as defined in Eqs. (5)–(7).

(5) $$idx=\mathrm{I}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}\left(T\right)$$

(6) $$G^{\mathrm{\prime }}=\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{c}\mathrm{k}\left(\mathrm{M}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}\left(idx,h^{\mathrm{\prime }}\right)\right)=\left\{{\overrightarrow{h^{\mathrm{\prime }}}}_1,{\overrightarrow{h^{\mathrm{\prime }}}}_2,\cdots ,{\overrightarrow{h^{\mathrm{\prime }}}}_{bs}\right\}$$

(7) $$G^g=\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{c}\mathrm{k}\left(\mathrm{M}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}\left(idx,h^g\right)\right)=\left\{{\overrightarrow{h^g}}_1,{\overrightarrow{h^g}}_2,\cdots ,{\overrightarrow{h^g}}_{bs}\right\}.$$Here, the $\mathrm{I}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{x}$ operation selects the identifier of the target sample $T$ within the current mini-batch, while the $T$ operation retrieves the corresponding encoded tensors from the heterogeneous feature spaces $h^{\mathrm{\prime }}$and $h^g$, ensuring cross-modal alignment.

T-G feature fusion module

In this research, the Transformer Encoder architecture is employed to process text embeddings and positional encoding through stacked self-attention layers and feedforward neural networks. A multi-head attention mechanism is utilized to integrate the extracted feature vectors for text classification tasks. However, this approach fails to account for spatial dependencies in POI classification, resulting in incomplete feature representation. To mitigate this limitation, we propose a hybrid framework that synergistically combines textual features with local geospatial neighborhood features. This integration significantly enhances the semantic characterization of POIs, with the detailed architectural design illustrated in Fig. 4.

For the T-G Feature Fusion module, the textual input batch processed in each training iteration is mathematically formulated as:

(8) $$\begin{array}{rl}T& =\left[T_1,T_2,\dots ,T_{bs}\right]\\ & =\left[\left[t_{11},t_{12},\dots ,t_{1c}\right],\cdots ,\left[t_{bs1},t_{bs2},\dots ,t_{bsc}\right]\right]\\ & =\left[X_1,X_2,\dots ,X_{bs}\right]\\ & =\left[\left[e_{11},e_{12},\dots ,e_{1c}\right],\cdots ,\left[e_{bs1},e_{bs2},\dots ,e_{bsc}\right]\right].\end{array}$$The local neighborhood features, aggregated through graph convolutional operations to encapsulate integrated geospatial neighborhood information, are mathematically formulated as:

(9) $$G^{\mathrm{\prime }}=\left\{{\overrightarrow{h^{\mathrm{\prime }}}}_1,{\overrightarrow{h^{\mathrm{\prime }}}}_2,\cdots ,{\overrightarrow{h^{\mathrm{\prime }}}}_{bs}\right\}.$$The resultant fused feature embeddings, generated through the Transformer Encoder’s multi-head attention mechanisms and subsequent multi-modal feature fusion layers, are formally expressed as:

(10) $$T^{tg}=\left[X_1^{tg},X_2^{tg},\dots ,X_{bs}^{tg}\right].$$For a textual sample $T_i=\left[t_{i1},t_{i2},\dots ,t_{ic}\right]$ with sequence length $c$, each token $t_{ij}$ is encoded into a vector $e_{ij}\in {\mathbb{R}}^d$ via word embedding and positional encoding, forming the tokenized matrix $X_i=\left[e_{i1},e_{i2},\dots ,e_{ic}\right]\in {\mathbb{R}}^{c\times d}$. Generate $X_i^t$ by passing $X_i$ through the Transformer Encoder.

(11) $$X_i^t=\mathrm{E}\mathrm{n}\mathrm{c}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}\left(X_i\right).$$Cross-modal feature fusion:

(12) $$X_i^{t_g}=\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{a}\mathrm{r}\left(\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}\left(X_i^t,{\overrightarrow{h^{\mathrm{\prime }}}}_i\right)\right).$$Here, ${\overrightarrow{h^{\mathrm{\prime }}}}_i$ denotes the geospatial local neighborhood feature of the $i$-th POI. By concatenating the POI name text feature $X_i^t$ with its corresponding geospatial local neighborhood feature ${\overrightarrow{h^{\mathrm{\prime }}}}_i$ along the feature dimension, all modalities are projected into a shared latent space to ensure cross-modal compatibility and meaningful interaction, while preserving modality-specific characteristics and preventing critical unimodal information loss. A linear layer then compresses the concatenated features to match the dimensionality of the classification task, generating the fused text- neighborhood joint representation $X_i^{t_g}$.

Enhanced geospatial local neighborhood module

To reinforce the geospatial local neighborhood, the outputs of the T-G Fusion module ( $X_i^{t_g}$) and the GeosN module ( ${\overrightarrow{h^g}}_{\mathrm{i}}$) are integrated via a decision fusion strategy. The fused features are then fed into a Softmax layer for final classification.

For a batch of size bs, the textual-geospatial features are aggregated:

(13) $$T^{tg}=\left[X_1^{tg},X_2^{tg},\dots ,X_{bs}^{tg}\right].$$And the complete geospatial local neighborhood features are represented:

(14) $$G^g=\left\{{\overrightarrow{h^g}}_1,{\overrightarrow{h^g}}_2,\cdots ,{\overrightarrow{h^g}}_{bs}\right\}.$$The element-wise summation in the decision fusion strategy is adopted to fuse multimodal features, generating the input $f$ for the classification layer. This approach preserves the raw modality-specific information, $T^{tg}$ and $G^g$, while enabling effective cross-modal relational learning. The linear combination ensures compatibility between heterogeneous representations and retains critical unimodal characteristics, thereby enhancing the discriminative power of the fused feature $f$.

(15) $$f=T^{tg}+G^g.$$Here, the summation is performed element-wise to amplify geospatial neighborhood salience. The fused feature $f$ is normalized via a Softmax layer to generate the final POI classification probabilities:

(16) $$P(y|f)=Softmax\left(W_{c}f+b_c\right)$$where $W_c$ and $b_c$ are the classification head’s weight matrix and bias term, respectively.

Experiment

Dataset

To comprehensively validate the feasibility and superiority of the proposed method, we constructed three heterogeneous POI datasets derived from distinct data sources and geographical regions:

Shanghai POI dataset: The dataset is derived from the Shanghai subset of the POI data for key Chinese cities, available on the Geographic Data Sharing Infrastructure, global resources data cloud (http://www.gis5g.com/). It encompasses selected administrative districts under Shanghai’s jurisdiction, covering a total of 14 distinct categories. The dataset comprises 218,022 POIs, with detailed category distributions provided in Table 4.

AutoNavi Beijing POI dataset: The dataset originates from the Peking University Open Research Data Platform (opendata.pku.edu.cn), specifically comprising POI data for Beijing. It includes a total of 22 categories and 12,397 POIs, providing a comprehensive representation of geographic entities within the region.

OSM Guangdong POI dataset: The dataset is sourced from the OpenStreetMap website (www.openstreetmap.org), specifically focusing on Guangdong Province. After data selection and filtering, it comprises 93 categories and 29,265 POIs, offering a diverse and extensive representation of geographic entities within the region.

For the three datasets, the POI latitude and longitude information, category information, and name information are utilized for data processing.

Graph dataset construction: we uniformly use the WGS 84 (EPSG:4326) geographic coordinate system. Using the latitude and longitude information, each node is connected to its five nearest neighbors (k = 5) to form the graph dataset. In this representation, nodes correspond to POI points, node features represent POI categories, and edges denote the adjacency relationships between POIs.

Text dataset construction: POI names constitute the textual classification dataset, with each entry assigned a category label. Unique identifiers maintain cross-modal consistency between graph data and text records. Text preprocessing involves character-level tokenization that preserves original casing and punctuation, followed by vocabulary control limiting tokens to 10,000 maximum frequencies (infrequent characters mapped to <UNK>). Sequences are normalized to fixed lengths through <PAD> padding and truncation, with randomly initialized embeddings fine-tuned during training. This pipeline ensures identifier alignment while transforming raw names into standardized character sequences for joint text-graph modeling.

Dataset splitting: A shared mask partitions both graph and text datasets into training (50%), validation (30%), and test (20%) sets. To address class imbalance, we implement dynamic resampling via quartile analysis: class stratification is performed using Q1, median, and Q3 distribution thresholds; minority classes below Q1 are oversampled to match the median size, while majority classes exceeding Q3 are undersampled to the Q3 level. Comprehensive dataset statistics are detailed in Table 5.

Evaluation metrics

To assess model performance, we employ four standard metrics: precision (P), accuracy, recall (R), and the F1-score, defined as follows:

(17) $$P={\displaystyle \frac{TP}{TP+FP}\times 100\mathrm{\%}}$$

(18) $$accuracy={\displaystyle \frac{TP+TN}{TP+FN+FP+TN}}$$

(19) $$R={\displaystyle \frac{TP}{TP+FN}\times 100\mathrm{\%}}$$

(20) $$F1={\displaystyle \frac{2\times P\times R}{P+R}}$$where TP (True Positives) denotes correctly predicted positive samples, FN (False Negatives) positive samples misclassified as negative, FP (False Positives) negative samples misclassified as positive, and TN (True Negatives) correctly predicted negative samples.

The proposed GeoDFNet model employs a weighted cross-entropy loss function during training, utilizing class-specific weights to mitigate class imbalance effects. This loss function measures the discrepancy between predicted class probabilities and ground-truth labels, with heightened penalties for misclassifying rare classes. The mathematical formulation of this weighted loss is given in Eq. (21):

(21) $$LOSS=-\sum _{c=1}^C{w}_c{y}_{c}log\left({\hat{y}}_c\right)$$where $C$ denotes the number of classes, $y_c$ represents the ground-truth indicator for class $c,$ and ${\hat{y}}_c$ is the predicted probability for class $c$.

Comparative experiments

We compare our GeoDFNet with the following baselines:

TextCNN (Kim, 2014): This model leverages a convolutional neural network (CNN) architecture, employing multiple convolutional layers with varying kernel sizes to capture key information at different granularities within the text. By extracting local features of varying lengths, TextCNN effectively encodes textual semantics for downstream tasks.

TextRNN (Liu, Qiu & Huang, 2016): This model is built on a recurrent neural network (RNN) architecture, which captures sequential information in text by incorporating recurrent connections. These connections enable the network to retain and process contextual information across sequences, making it effective for modeling dependencies in textual data.

TextDPCNN (Johnson & Zhang, 2017): This model is based on convolutional neural networks (CNNs) and enhances the extraction of long-range dependencies in text through the use of downsampling, isometric convolutions, and network deepening. These techniques enable the model to effectively capture both local and global textual patterns, improving its ability to process and understand complex textual structures.

TextRCNN (Lai et al., 2015): This model combines a recurrent convolutional neural network architecture with bidirectional RNNs and pooling layers to effectively capture contextual information and represent textual semantics with greater accuracy. By integrating recurrent and convolutional mechanisms, TextRCNN leverages both sequential and local features, enhancing its ability to model complex textual relationships.

TextRNN_ATTENTION (Zhou et al., 2016): This model employs a bidirectional LSTM network enhanced with an attention mechanism to identify and emphasize the most critical semantic information within sentences for relational classification tasks. By leveraging attention, the model dynamically focuses on the most relevant parts of the input, improving its ability to capture contextual dependencies and semantic nuances.

Transformer: This model utilizes a self-attention mechanism to process the entire input sequence simultaneously, enabling it to capture global dependencies and relationships within the data. By focusing on the interactions between all elements of the sequence, the Transformer effectively performs classification tasks while maintaining a high level of contextual understanding.

Comprehensive evaluation on the Shanghai POI dataset in Table 7 confirms GeoDFNet’s consistent superiority over six text-based baselines, with pronounced performance disparities in semantically ambiguous categories. While low-ambiguity categories like Public Utilities—where explicit lexical cues (e.g., ‘public toilet’) enable near-perfect baseline accuracy (TextCNN F1 = 99.68% ± 0.12)—high-ambiguity categories reveal critical gaps: GeoDFNet achieved 99.42% ± 0.53 $F1$ in Catering Services (vs. best baseline 92.86% ± 0.21), 91.10% ± 8.59 in Scenic Spots (vs. 61.49% ± 0.92), and 97.20% ± 2.02 in Retail Services (vs. 63.78% ± 2.12). This divergence stems from GeoDFNet’s operationalization of geospatial similarity principles, where lexically unclassifiable names (e.g., ‘Centennial Dragon Robe’) are accurately categorized through spatial neighborhood integration.

Critically, GeoDFNet outperformed all baselines in every run (10/10). Wilcoxon signed-rank tests on the accuracy data in Table 8 confirmed statistically significant superiority (p = 0.001953 for all comparisons), with consistently higher terminal $F1$-scores and training accuracy. These results robustly validate the method’s efficacy in advancing POI classification.

Ablation experiments

To validate the contributions of the T-GFF and EGN modules in GeoDFNet, we conducted rigorous ablation studies by systematically isolating components. The T-GFF module performs multimodal feature-level fusion integrating textual and geospatial representations, while the EGN module executes decision-level fusion incorporating spatial neighborhood topology. Through sequential removal of these modules (Table 9), we precisely quantify their individual impacts on POI classification performance, establishing causal relationships between architectural components and model efficacy.

T-G feature fusion module evaluation

To isolate the contribution of the T-GFF module, we compared the full GeoDFNet model against a variant where the T-GFF module was removed (denoted as GeoDFNet w/o T-GFF, which retains only decision-level fusion). The results demonstrate that removing this feature-level fusion causes a significant performance drop: the $F1$-score decreases to 97.04% (±1.69) and accuracy falls to 98.01% (±1.08) from the full model’s 97.65% (±0.85) $F1$ and 98.60% (±0.45) accuracy. This decline confirms the critical role of T-GFF in effectively integrating spatial-textual features at the feature level, which is essential for the model’s understanding of POI characteristics.

Enhanced geospatial local neighborhood module evaluation

To examine the contribution of the EGN module, we compared the full model against a variant where the EGN module was removed (denoted as GeoDFNet w/o EGN, which retains only feature-level fusion). Removing this decision-level fusion module also leads to a substantial performance degradation: the $F1$-score drops to 96.81% (±1.93) and accuracy reduces to 98.12% (±1.24). This result clearly demonstrates the effectiveness of the EGN module in leveraging spatial topology, which is vital for capturing the influence of spatial neighborhood features on category characteristics.

Synergistic effect

The significant performance decline observed in both ablated models—not only in overall accuracy and $F1$ but also in P and R—underscores that the T-GFF and EGN modules are both indispensable components of GeoDFNet. The full model’s superior and more stable performance (as evidenced by lower standard deviations) arises from the synergistic effect of these complementary fusion strategies: T-GFF’s feature-level integration working in concert with EGN’s decision-level refinement.

k-value sensitivity analysis

We conducted a parametric sensitivity analysis to evaluate the impact of the number of neighboring points, k, used in constructing the geographic dataset. As summarized in the Table 10, model performance is sensitive to the choice of k, with optimal results achieved at k = 5 across all metrics: precision (97.14 ± 1.10%), recall (98.30 ± 0.66%), $F1$-score (97.65 ± 0.85%), and accuracy (98.60 ± 0.45%). This configuration also exhibited the smallest standard deviations, indicating superior stability and robustness.

Performance improved as k increased from 2 to 5, suggesting that incorporating more spatial context enhances feature representation and contextual understanding. However, beyond k = 5, all metrics gradually declined, implying that excessively large neighborhoods may introduce noise or redundant information, thereby reducing model efficacy. These results highlight the critical role of selecting an appropriate spatial scale to balance contextual information and discriminative power.

Generalization experiments

To comprehensively evaluate the generalization capability of the GeoDFNet model and validate its applicability across different regions and datasets, a series of experiments were conducted to assess its adaptability and robustness in handling diverse semantic structures and regional characteristics. Additional POI public datasets, including the AutoNavi Beijing POI dataset and the OSM Guangdong POI dataset, were introduced to perform generalization performance experiments.

During model training, each dataset was partitioned into training, validation, and test sets at a ratio of 5:3:2 to ensure consistency in model training and evaluation. Given the varying sizes of the datasets, an appropriate number of training epochs was selected for each dataset, while other hyperparameters remained consistent with the experimental settings described earlier to ensure comparability of results. The model’s classification performance was evaluated using two key metrics: accuracy and F1-score.

The classification performance of each model on datasets from different regions and sources is summarized in Table 11. The results demonstrate that the proposed GeoDFNet model achieves strong performance across diverse datasets, highlighting its generalization ability and robustness. The reason why the AutoNavi Beijing POI dataset t results are higher than the OSM Guangdong POI Dataset results is that the OSM Guangdong POI dataset has relatively fewer samples and more labels.