A Lightweight LLM-Based Semantic–Spatial Inference Framework for Fine-Grained Urban POI Analysis

Highlights

What are the main findings?

• A fine-tuned open large language model (Qwen3-8B + LoRA) achieves superior fine-grained classification of unstructured POI texts (macro-F1 = 0.843), outperforming BERT and commercial LLM baselines while running on consumer-grade hardware.
• The framework expands usable fine-category POI labels by ≈14–15× in Guangzhou and Shenzhen, enabling robust 500 m grid residual–hotspot analyses that reveal structural cultural preferences and urban diversity patterns.

What is the implication of the main finding?

• The pipeline transforms unstructured urban text into interpretable spatial evidence, providing a reproducible and resource-efficient method for service-equity assessment and cultural-space analysis in planning.
• Its domain-agnostic, lightweight design can be extended to other urban functions (e.g., healthcare, education), offering a scalable tool for fine-resolution spatial governance in smart-city contexts.

Abstract

Unstructured POI name texts are widely used in fine-grained urban analysis, yet missing labels and semantic ambiguity often limit their value for spatial inference. This study proposes a large language model-based semantic–spatial inference framework (LLM-SSIF), a lightweight semantic–spatial pipeline that translates POI texts into interpretable, fine-grained spatial evidence through an end-to-end workflow that couples scalable label expansion with scale-controlled spatial diagnostics at a 500 m resolution. A key advantage of LLM-SSIF is its deployability: LoRA-based parameter-efficient fine-tuning of an open LLM enables lightweight adaptation under limited compute while scaling fine-label coverage. Trained on a nationwide cuisine-labeled dataset (~220,000 records), the model achieves strong multi-class short-text recognition (macro-F1 = 0.843) and, in the Guangzhou–Shenzhen demonstration, expands usable fine-category labels by ~14–15× to support grid-level inference under long-tail sparsity. The spatial module then isolates cuisine-specific over/under-representation beyond overall restaurant intensity, revealing contrasting cultural configurations between Guangzhou and Shenzhen. Overall, LLM-SSIF provides a reproducible and transferable way to translate unstructured POI texts into spatial–statistical evidence for comparative urban analysis.

1. Introduction

With the rapid development of big data and artificial intelligence, urban research increasingly relies on multi-source heterogeneous data to analyze complex spatial structures and social phenomena [1,2]. Among these sources, points of interest (POIs) provide high-coverage, fine-location records of urban functions such as dining, retail, healthcare, and education [3]. These records embed rich patterns related to activity rhythms, functional zoning, and cultural representation [4,5], patterns that are core to foundational theories of urban spatial structure and cultural embeddedness. POI names are not only clues for facility identification but also reflect layers of local culture, consumer preferences, and population mobility [6,7]. In this sense, extracting information from unstructured POI texts is a prerequisite for turning “urban semantics” into interpretable spatial evidence for planning and geography research [8].

A substantial body of research has sought to enrich POI semantics for fine-grained urban analysis, and existing approaches can be grouped into three streams. The first stream develops POI classification and semantic enrichment using keyword heuristics, topic models, embeddings, and supervised text classifiers [9,10,11], but it often degrades on real-world POI names due to ambiguity, naming conventions, and class imbalance [12,13,14], especially for long-tail categories [15]. The second stream uses enriched POI categories for urban function mapping and interpretation of consumption and cultural landscapes [9,15], yet grid-or community-scale inference is frequently limited by missing labels and inconsistent taxonomies across providers [16,17,18,19], undermining high-resolution and cross-city comparability [20,21]. The third stream applies LLMs to urban text understanding and short-text disambiguation [22,23,24], but practical deployment remains constrained by reliance on costly, opaque commercial APIs and the computational burden of full-parameter tuning for open models [25,26,27]. Consequently, a lightweight, open, and scalable pathway remains lacking for expanding fine-grained POI labels and converting semantic outputs into interpretable spatial evidence under realistic resource constraints [28,29].

To address this gap, this study proposes an LLM-based semantic–spatial inference framework (LLM-SSIF) for unstructured POI text. A PEFT (LoRA)-based semantic expansion module enables scalable fine-grained label expansion on an open LLM under practical compute budgets [30,31]. A spatial inference module combines scale-controllable residualization with local clustering detection to translate expanded labels into spatial evidence and distinguish structural cultural signals from density-driven scale effects [32,33]. Conceptually, the framework follows an inferential chain from POI text semantics to fine-grained labels and then to scale-adjusted spatial deviations that reveal patterns of urban cultural differentiation. More broadly, by transforming short POI names into structured signals suitable for spatial–statistical inference [34], the framework enables “urban semantics” to enter mainstream spatial analytic workflows and supports more comparable evidence on cultural embeddedness and polycentric differentiation [7,35].

Accordingly, this study addresses the following three interrelated questions: (1) How can lightweight PEFT enable accurate, fine-grained labeling of unstructured POI texts under limited compute [30]? (2) Can the framework scale labels under incomplete platform categorization and limited annotation while remaining robust for long-tail cuisines and transferable across cities? (3) How can LLM-derived labels be translated into interpretable spatial evidence of cultural differentiation [36]?

To address these questions, the Tongyi Qianwen 3-8B (Qwen3-8B) model is fine-tuned with LoRA on consumer-grade hardware and evaluated on a nationwide restaurant POI dataset (≈220,000 labeled records), followed by cross-city case studies in Guangzhou and Shenzhen. Beyond classification performance, a downstream spatial inference module is implemented to generate 500 m grid-level spatial evidence through GLM residualization (with an offset for total counts) and Getis–Ord Gi* hotspot detection.

The methodological contributions are threefold. First, the study explicitly links LLM-based label expansion with scale-controlled spatial inference, bridging the gap between semantic enrichment and theory-relevant urban interpretation. Second, a lightweight and reproducible PEFT (LoRA) strategy is demonstrated as a practical alternative to closed commercial APIs and computationally intensive full-parameter tuning, enabling fine-grained POI text recognition under typical research resources. Third, the framework operationalizes a scale-controlled “residual–hotspot” workflow that converts fine-grained labels into comparable spatial evidence for assessing cross-city cultural differentiation.

The structure of the paper is as follows: Section 2 introduces the research methods and data; Section 3 presents classification results and performance evaluation; Section 4 conducts spatial analysis based on case studies of Guangzhou and Shenzhen, revealing structural distributions of cuisines and their cultural implications; and Section 5 concludes with a discussion and future directions.

2. Methodology

2.1. Research Framework

To bridge unstructured POI text with interpretable urban spatial evidence, this study proposes an LLM-based semantic–spatial inference framework (LLM-SSIF) and operationalizes the cross-disciplinary goal of “unstructured text → urban spatial evidence” as the following four-stage pipeline (see Figure 1): (i) data acquisition, taxonomy harmonization, and cleaning; (ii) supervised PEFT (LoRA) of an open LLM; (iii) model evaluation via macro-averaged metrics and error profiling; and (iv) aggregation of labels to 500 m grids and coupling GLM residuals with Gi* hotspots to yield interpretable spatial evidence. The design prioritizes deployability and interpretability, stably expanding fine-label coverage under limited compute.

First, this study compiles approximately 220,000 cuisine-labeled dining POIs from across China. After data cleaning and multi-class taxonomy harmonization, the dataset is split into training, validation, and test sets. The nationwide coverage increases category and regional diversity, strengthening generalization and providing a robust basis for subsequent city-level spatial analyses and case studies.

Second, to achieve accurate fine-grained classification within limited computational resources, this study adopts the open LLM Qwen3-8B and fine-tunes it with PEFT (LoRA). This enables lightweight training and efficient inference on consumer-grade GPUs, achieving a practical performance–compute–deployability trade-off.

Third, the model’s effectiveness is evaluated through a focus on robustness and coverage. This study prioritizes macro-averaged precision, recall, and F1-score to ensure balanced assessment across imbalanced cuisine classes, supplemented by the Valid Coverage Rate (VCR) to quantify the expansion of analytically usable labels. A detailed examination of the confusion matrix offers further insight into model behavior, particularly for long-tail and semantically ambiguous categories.

Fourth, to transform the expanded labels into evidence of cultural spatial differentiation, this study connects the classification output to spatial econometric analysis. For our case studies in Guangzhou and Shenzhen, classification outputs are aggregated to 500 m × 500 m grid cells and analyzed using descriptive mapping, offset-GLM residualization (with total count as offset), and Getis–Ord Gi* hotspot detection. This reveals structural preferences across cuisines and the spatial manifestations of urban cultural differences.

In sum, LLM-SSIF couples PEFT-tuned LLM labeling with offset-GLM residualization and Gi* hotspot detection on 500 m grids, converting unstructured POI names into analysis-ready inputs and interpretable city-scale spatial evidence. The cross-city case studies demonstrate transfer with minimal adaptation while maintaining robust macro-level performance.

2.2. Data and Preprocessing

This study uses a nationwide dataset of restaurant points of interest (POIs) in China for 2023 and selects Guangzhou and Shenzhen as case cities for the spatial analysis. Restaurant POIs were collected from the Amap (Gaode Map) Open Platform via Python 3.7 scripts using its Web Service API (POI Search, v3). After acquisition, the raw POI records were systematically cleaned and organized. Duplicate entries were identified and removed by jointly applying fuzzy matching on POI names and a spatial proximity constraint (within 50 m), retaining only the most information-complete record for each duplicate group. This procedure provided a relatively stable data foundation for constructing the classification taxonomy.

Although Amap’s official POI categories offer relatively fine-grained delineation of food-service venues, cuisine-specific labels are frequently missing in practice. In our dataset of approximately 5.10 million restaurant POIs, only about 220,000 records contain cuisine annotations—roughly 4.3% of the total—rendering the labeled subset insufficient for fine-grained spatial analyses of culinary–cultural heterogeneity.

To support cuisine identification from restaurant POIs, we develop an LLM-based classification label system. To ensure analytical reliability and interpretability, we focus on 15 cuisine categories with both strong identification performance and adequate sample sizes (see

Table 1. Cuisine classification with abbreviations and description.

ID	Cuisine Name	Abbreviation	Description
1	Northeastern Chinese Cuisine	NE Chinese	Hearty dishes, wheat-based staples
2	Yunnan–Guizhou Cuisine	Yunnan–Guizhou	Spicy, sour, ethnic minority flavors
3	Sichuan Cuisine	Sichuan	Famous for chili, numbing spices
4	Huizhou Cuisine	Huizhou	Anhui region, stews, umami flavors
5	Japanese Cuisine	Japanese	Sushi, ramen, seafood
6	Thai–Vietnamese Cuisine	Thai–Viet	Spicy, sour, fresh herbs
7	Hunan Cuisine	Hunan	Hot, sour, bold flavors
8	Cantonese Cuisine	Cantonese	Dim sum, mild seasoning, seafood
9	Jiangsu Cuisine	Jiangsu	Freshwater fish, light sweet taste
10	Northwestern Chinese Cuisine	NW Chinese	Lamb, wheaten foods, hearty flavors
11	Western Cuisine	Western	European and American styles, grilled meats, dairy
12	Hubei Cuisine	Hubei	Freshwater fish, noodles
13	Fujian Cuisine	Fujian	Seafood, soups, mild sweetness
14	Korean Cuisine	Korean	Kimchi, barbecue, rice dishes
15	Shandong Cuisine	Shandong	Seafood, wheat, light salty flavors

).

As shown in Figure 2a, cuisine-labeled POIs cover almost all prefecture-level cities, ensuring nationwide representativeness. Figure 2b shows that most names are 3–8 characters, reflecting typical naming but also creating ambiguity in short texts. At the categorical level (Figure 2c), Sichuan, Hunan, and Cantonese dominate, yet even the smallest categories contain over a thousand samples, providing sufficient data to support model training across 15 cuisines. Meanwhile, to further verify the effectiveness of the LLM-SSIF, Guangzhou and Shenzhen are selected as case studies for in-depth exploration.

The final dataset is split into training, validation, and test sets in a 64%:16%:20% ratio—first dividing the data 80:20 into training and test sets, and then randomly sampling 20% from the training set for validation. Class balance is maintained across all subsets to ensure consistency during model training and evaluation.

2.3. Model and Lightweight Fine-Tuning

This study adopts Parameter-Efficient Fine-Tuning (PEFT) to enable efficient model adaptation under constrained hardware resources (see Figure 3). By updating only a small subset of parameters in a pre-trained model, PEFT substantially reduces computational and GPU memory costs relative to full-parameter fine-tuning [31]. Among available PEFT strategies, we employ Low-Rank Adaptation (LoRA) [30,37], which injects trainable low-rank matrices into Transformer attention layers. This design achieves effective task adaptation with a modest parameter increase and, in practice, can attain performance close to full fine-tuning.

Hyperparameter choices are guided by LoRA’s low-rank adaptation mechanism and training stability. We set the rank to 8, providing sufficient adaptation capacity while keeping the parameter increment manageable. The scaling factor (α) is set to 16 to balance the numerical scale between the frozen pre-trained weights and the low-rank updates, thereby improving training stability [30]. For regularization, we set LoRA dropout to 0, motivated by a combined strategy of “structural constraint + training-time monitoring.” Specifically, LoRA introduces only a limited number of trainable parameters while the backbone weights remain frozen, which already restricts model degrees of freedom; in addition, we mitigate overfitting through validation-based monitoring and early stopping. For our short-text multi-class classification task, additional dropout may weaken the effective update signal in the low-rank branch and is unlikely to yield substantial generalization gains.

We optimize the model using AdamW (learning rate: 5 × 10⁻⁵) with a cosine annealing schedule to balance convergence speed and stability, and apply gradient clipping (max norm = 1.0) to reduce instability risks under mixed-precision training. The per-device batch size is 16, combined with gradient accumulation to increase the effective batch size, alleviating memory constraints and reducing gradient noise. Training is conducted in bfloat16 precision with a maximum sequence length of 512 tokens, which covers the primary semantic content in POI names and prompt templates. All configurations are summarized in

Table 2. Fine-tuning hyperparameters.

Hyperparameter	Value
Model	Qwen3-8B
Fine-tuning method	LoRA
LoRA rank (r)	8
LoRA α	16
LoRA dropout	0
LoRA target	All attention modules
Trainable parameters	21,823,488 (≈0.27% of 8.21B backbone)
Optimizer	AdamW
Learning rate	5 × 10−5
Scheduler	Cosine decay
Precision	bfloat16
Batch size	16
Max sequence length	512 tokens
Epochs	max 10 (early stopping at ~4.66)
Training duration (wall-clock)	~13.25 h
Dataset split	Train:Val:Test = 64%:16%:20%

.

The maximum number of epochs is set to 10, with early stopping enabled [38]. The training loss decreases rapidly in the early stage and then stabilizes (Figure 4). The validation loss reaches its minimum at approximately 1300 steps (≈4.7 epochs; loss ≈ 0.0724) and subsequently increases (Figure 5); accordingly, we select the best checkpoint between epochs 4 and 5 as the final model. All lightweight fine-tuning experiments are conducted on consumer-grade hardware (NVIDIA RTX 4090 GPU, 24 GB VRAM, designed by NVIDIA Corporation, Santa Clara, USA), indicating that effective training and deployment for fine-grained text classification can be achieved in this setting without access to high-performance computing infrastructure.

2.4. Evaluation Metrics

To comprehensively evaluate the effectiveness and deployability of the proposed framework for classifying unstructured urban POI text, this study constructs a multi-dimensional evaluation system. The main metrics include overall accuracy, macro-averaged precision, macro-averaged recall, macro-averaged F1-score, and per-class F1 scores [39].

Due to the significant imbalance in POI category distribution (e.g., Cantonese and Hunan cuisine samples far outnumber those of niche cuisines), relying solely on overall accuracy or weighted metrics can lead to large-category dominance and obscure performance on underrepresented classes [40]. Therefore, macro-averaged metrics are used, where precision, recall, and F1-score are computed for each class and then averaged arithmetically, giving each category equal weight in the evaluation [41]. Compared to micro- or weighted averages, macro-averaging is better suited for assessing model performance across diverse classes in imbalanced short-text classification tasks [42].

Additionally, to reflect the model’s effective coverage of the full dataset, this study introduces Valid Coverage Rate (VCR) as a supplementary metric.

Overall accuracy is defined as follows:

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}\mathrm{ }=\mathrm{ }{\displaystyle \frac{\mathrm{T}\mathrm{P}\mathrm{ }+\mathrm{ }\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}\mathrm{ }+\mathrm{ }\mathrm{T}\mathrm{N}\mathrm{ }+\mathrm{ }\mathrm{F}\mathrm{P}\mathrm{ }+\mathrm{ }\mathrm{F}\mathrm{N}}}$$

(1)

where $\mathrm{T}\mathrm{P}$ (True Positives) refers to correctly predicted positive samples, $\mathrm{T}\mathrm{N}$ (True Negatives) to correctly predicted negative samples, $\mathrm{F}\mathrm{P}$ (False Positives) to samples incorrectly predicted as positive, and $\mathrm{F}\mathrm{N}$ (False Negatives) to samples incorrectly predicted as negative.

For multi-class classification, the precision, recall, and F1-score of class $\mathrm{i}$ are defined as follows:

$$\begin{array}{c}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}_i={\displaystyle \frac{\mathrm{T}{\mathrm{P}}_i}{\mathrm{T}{\mathrm{P}}_i+\mathrm{F}{\mathrm{P}}_i}}\end{array}$$

(2)

$$\begin{array}{c}{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}_i={\displaystyle \frac{\mathrm{T}{\mathrm{P}}_i}{\mathrm{T}{\mathrm{P}}_i+\mathrm{F}{\mathrm{N}}_i}}\end{array}$$

(3)

$${\mathrm{F}1}_i={\displaystyle \frac{2\cdot {\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}_i\times {\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}_i}{{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}_i+{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}_i}}$$

(4)

where $T{P}_i$, $F{P}_i$, and $F{N}_i$ are the true positives, false positives, and false negatives for class $i$, respectively. Precision measures the model’s ability to avoid false positives, recall measures its ability to identify all relevant instances, and F1-score balances both.

Macro-averaged metrics are calculated as the arithmetic mean across all classes, as follows:

$$\begin{array}{c}\mathrm{M}\mathrm{a}\mathrm{c}\mathrm{r}\mathrm{o}-\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{1}{N}}{\sum _{i=1}^N\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}_i\end{array}$$

(5)

$$\mathrm{M}\mathrm{a}\mathrm{c}\mathrm{r}\mathrm{o}-\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{1}{N}}{{\sum }_{i=1}^N\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}_i$$

(6)

$$\begin{array}{c}\mathrm{M}\mathrm{a}\mathrm{c}\mathrm{r}\mathrm{o}-\mathrm{F}1={\displaystyle \frac{1}{N}}{{\sum }_{i=1}^N\mathrm{F}1}_i\end{array}$$

(7)

where $N$ is the total number of classes (excluding the Other category). Macro-averaging mitigates the bias introduced by class imbalance and better reflects model performance across different classes.

The Valid Coverage Rate (VCR) is defined as follows:

$$\begin{array}{c}\mathrm{V}\mathrm{C}\mathrm{R}={\displaystyle \frac{N_{\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{d}}}{N_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}}}\end{array}$$

(8)

where $N_{\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{d}}$ denotes the number of samples assigned to any fine-grained class (i.e., predicted as non-Other) and $N_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}$ is the total number of samples. Thus, VCR quantifies the proportion of the dataset that receives a usable non-Other label from the model.

Overall, macro-averaging (excluding “Other”) curbs head-class dominance and reflects performance on long-tail categories. Valid Coverage Rate (VCR) measures the proportion of samples assigned non-Other labels, i.e., the effective sample density for 500 m grid econometrics. In practice, macro-F1 captures label reliability, whereas VCR captures analytical coverage; together, they determine the robustness of fine-scale spatial evidence.

2.5. Residuals and Spatial Analysis

In order to transform the expanded labels into interpretable evidence of structural cultural preferences and to distinguish these from patterns driven solely by general commercial density, this study employs a two-step spatial econometric approach. The core analytical challenge lies in determining whether the clustering of a cuisine stems from a locally embedded cultural preference or merely mirrors the overall restaurant supply.

After obtaining cuisine distribution data classified by the LLM, this study further computes standardized residuals at the 500 m × 500 m grid level to reveal deviations of different cuisines in urban space. Specifically, let the observed count of cuisine $c$ in grid $j$ be as follows:

$$y_{jc}=observed count of cuisine c in grid j$$

and the total number of restaurants in grid $j$ be as follows:

$$n_j=total number of restaurants in grid j$$

Let the overall proportion of cuisine $c$ in the entire city be $p_c$. Under scale control, the expected count is as follows:

$$\begin{array}{c}{\mu }_{jc}=n_j\cdot p_c\end{array}$$

(9)

To control for the scale effect dominated by the total number of restaurants $n_j$, this study employs a Generalized Linear Model (GLM) to estimate the expected count ${\mu }_{jc}$ of cuisine $c$ in grid $j$. The model incorporates $\mathit{log}\left(n_j\right)$ as a fixed offset (not an estimated variable) [43]. The core purpose of using an offset is to remove the ‘expected’ component due to variations in commercial density from the observed counts, thereby isolating the relative over- or under-concentration specific to a cuisine. The model is fitted as a Poisson or Negative Binomial regression with only an intercept, as follows:

$$\begin{array}{c}\log ({\mu }_{jc})={\beta }_{0c}+\log (n_j),\end{array}\hspace{1em}$$

(10)

$${\mu }_{jc}=n_j\cdot e^{{\beta }_{0c}}$$

(11)

If the Negative Binomial model converges, its variance function is $\begin{array}{c}\mathrm{V}\mathrm{a}\mathrm{r}\left({\mu }_{jc}\right)={\mu }_{jc}+\kappa \cdot {\mu }_{jc}^2\end{array}$. If it fails to converge, it degenerates to a Poisson model: $\begin{array}{c}\mathrm{V}\mathrm{a}\mathrm{r}\left({\mu }_{jc}\right)={\mu }_{jc}\end{array}$.

Based on the expected value and variance, the standardized Pearson residual is calculated as follows:

$$\begin{array}{c}{r}_{jc}={\displaystyle \frac{y_{jc}-{\mu }_{jc}}{\sqrt{\mathrm{V}\mathrm{a}\mathrm{r}\left({\mu }_{jc}\right)}}}\end{array}$$

(12)

This residual measures the relative deviation between local observations and citywide expectations. If $r_{jc}>0$, the grid has more instances of a cuisine than expected, suggesting possible spatial clustering; if $r_{jc}<0$, the cuisine is underrepresented, indicating possible spatial insufficiency. Larger absolute values of $\mid r_{jc}\mid$ indicate greater deviation. Residuals for cuisines with extremely low citywide frequencies or insufficient variance are set to zero to avoid estimation bias.

The purpose of residualization is to control for grid-level scale/exposure effects, thereby preventing mechanical influences of overall restaurant volume on cuisine counts and shifting the analytical focus from “spatial agglomeration of total restaurants” to “structural deviations relative to a citywide baseline.” Accordingly, Pearson residuals capture the relative over- or under-representation of a given cuisine within local grids compared with its mean citywide share, rather than differences in absolute counts. Importantly, residualization is not intended to eliminate spatial structure. On the contrary, if spatial autocorrelation persists after removing scale effects, this remaining dependence constitutes the key signal of interest in this study, pointing to structural spatial organization potentially shaped by embedded cultural preferences. This rationale is consistent with the core design logic of residual analysis in spatial point processes proposed by Baddeley et al. (2005) [32]: residuals are meant to “remove model-explained effects and reveal systematic departures not captured by the model,” rather than to eradicate all spatial association.

To assess whether residualization effectively attenuates scale-driven spatial dependence, we compute Moran’s I for both the raw counts and the residuals (

Table 3. Global Moran’s I diagnostics for observed cuisine counts and GLM residuals at the 500 m grid level.

City	Cuisine	Moran’s I (Raw)	Moran’s I (Residual)	ΔI (Residual—Raw)
Guangzhou	Cantonese	0.2471	0.0615	−0.1856
	Hunan	0.1255	0.0848	−0.0407
	Western	0.2405	0.0187	−0.2218
Shenzhen	Cantonese	0.1174	0.0862	−0.0312
	Hunan	0.1020	0.0943	−0.0077
	Western	0.1885	0.0177	−0.1708

Note: Moran’s I values for both the raw counts and residuals are statistically significant (all p < 0.001).

). The results show that, for representative cuisines in Guangzhou and Shenzhen, Moran’s I for the residuals is substantially lower than that for the raw counts (all p < 0.001), indicating that the GLM with an offset term effectively reduces spatial autocorrelation induced by overall density/scale effects. At the same time, the residuals remain significantly positively autocorrelated, suggesting that identifiable structural spatial patterns persist even after controlling for scale.

Residualization controls for grid-level exposure effects so that cuisine patterns are evaluated as structural deviations from a citywide baseline, rather than as a mechanical by-product of total restaurant volume. Pearson residuals, therefore, quantify the relative over- or under-representation of each cuisine in a grid compared with its expected citywide share. Crucially, residualization is designed to remove model-explained scale effects while retaining any remaining spatial dependence as the key signal of interest, consistent with residual diagnostics in spatial point processes [32].

This study applies the local Getis–Ord Gi* statistic to the residual surface to localize the structural signal. Gi* detects statistically significant clustering of over- or under-representation relative to the citywide baseline; negative Gi* values indicate systematically lower-than-expected shares after controlling for exposure/scale, rather than “disadvantage.”

$$\begin{array}{c}{G}_i^{\ast }={\displaystyle \frac{\sum _j{w}_{ij}{x}_j-\overline{X}\sum _j{w}_{ij}}{S\sqrt{\frac{n\sum _j{w}_{ij}^2-{\left(\sum _j{w}_{ij}\right)}^2}{n-1}}}}\end{array}$$

(13)

where $G_i^{\ast }$ denotes the local Getis–Ord Gi* statistic for grid cell $i$. Here, $x_j$ represents the standardized residual $r_{jc}$ of cuisine $i$ in grid $j$;${ w}_{ij}$ is the spatial weight between grids $i$ and $j$; $\stackrel{\overline{¯}}{X}$ and $S$ are, respectively, the mean and standard deviation of the residuals across all grids in the city; and $n$ is the total number of grid cells. ArcGIS 10.7 was employed to automatically determines the optimal spatial scale and performs Monte Carlo testing, outputting hot and coldspots at 90%, 95%, and 99% confidence levels.

Through this residual–hotspot hybrid approach, the study identifies local “excess” or “deficiency” clustering patterns of various cuisines at the 500 m × 500 m grid level, while controlling for total restaurant volume. This allows for a more accurate interpretation of the structural differences in urban culinary cultures.

2.6. Key Methodological Assumptions

To directly address our three scientific questions, the proposed LLM-based semantic–spatial inference framework (LLM-SSIF) relies on three methodological assumptions. First, short and unstructured POI name texts contain sufficient semantic cues for fine-grained cuisine recognition, such that supervised PEFT (LoRA) can learn stable category signals under limited computational resources. Second, nationwide training coverage (≈220,000 cuisine-labeled POIs) provides sufficient diversity in naming conventions and regional expressions to support practical cross-city generalization under incomplete platform categorization and limited manual annotation. Third, in spatial inference, GLM residualization with total restaurant counts as an offset can adequately control scale effects, so standardized Pearson residuals reflect relative over-/under-representation beyond citywide expectations and can be meaningfully coupled with Getis–Ord Gi* hotspot detection on 500 m grids to yield interpretable evidence of cultural spatial differentiation.

3. Results

3.1. Overall Classification Performance

Table 4. Overall model performance comparison.

Model	Qwen3-8B PEFT (LoRA)	GPT-4o	Doubao-1.5-Pro	DeepSeek-R1	BERT
Accuracy	0.893	0.861	0.875	0.881	0.817
Precision	0.835	0.777	0.777	0.790	0.753
Recall	0.853	0.838	0.880	0.887	0.678
F1-score	0.843	0.801	0.813	0.827	0.704
VCR (%)	95.82%	90.80%	86.90%	90.54%	97.02%

Note: Macro-averaged metrics are computed excluding the Other class. VCR is the proportion of all samples predicted as non-Other (usable fine-grained labels).

presents a comparative evaluation of five models on the cuisine classification task for dining POIs. Among them, the PEFT (LoRA)-tuned Qwen3-8B achieves the highest scores across key metrics (see Figure 6), including macro-averaged Precision (0.835), Recall (0.853), and F1-score (0.843), as well as overall accuracy (0.893). These results indicate that domain-specific fine-tuning significantly enhances performance on multi-class short-text classification tasks, achieving a well-balanced trade-off between precision and recall while maintaining high classification coverage and minimizing misclassification rates.

By comparison, commercial API models such as GPT-4o and Doubao-1.5-pro report similar or slightly higher Recall values (0.838 and 0.880, respectively), but their lower Precision scores (~0.777) result in reduced F1-scores (0.801 and 0.813). This suggests that while they achieve broader coverage, they are more susceptible to misclassification in categories with fuzzy boundaries or high inter-class similarity. DeepSeek-R1 achieves the highest Recall (0.887) among all models and outperforms both commercial APIs in Precision (0.790), but still lags behind Qwen3-8B in overall balance, with a final F1-score of 0.827. In contrast, the traditional BERT model records the lowest scores across all metrics, including macro Precision (0.753), Recall (0.678), F1-score (0.704), and accuracy (0.817), underscoring its limitations in complex, imbalanced, and semantically rich classification tasks.

In terms of Valid Coverage Rate (VCR), Qwen3-8B achieves 95.82%, slightly below BERT’s 97.02%. However, the model delivers a substantially higher classification accuracy, indicating its capacity to maintain high coverage while improving label precision. These results collectively validate Qwen3-8B as the best-performing model across both predictive and coverage metrics and reinforce the effectiveness and scalability of the proposed LLM-based label-expansion approach for large-scale POI applications.

3.2. Per-Class Performance

Table 5. Per-class F1 by cuisine.

Cuisine Name	Qwen3-8B PEFT (LoRA)	GPT-4o	Doubao-1.5-Pro	DeepSeek-R1	BERT
NE Chinese	0.884	0.786	0.866	0.834	0.798
Yunnan–Guizhou	0.830	0.745	0.830	0.800	0.672
Sichuan	0.920	0.902	0.915	0.913	0.856
Huizhou	0.908	0.838	0.846	0.900	0.625
Japanese	0.931	0.935	0.941	0.939	0.918
Thai–Viet	0.824	0.825	0.801	0.848	0.825
Hunan	0.899	0.885	0.899	0.904	0.815
Cantonese	0.880	0.840	0.855	0.885	0.790
Jiangsu	0.659	0.618	0.570	0.644	0.299
NW Chinese	0.800	0.736	0.722	0.690	0.778
Western	0.885	0.871	0.891	0.888	0.867
Hubei	0.751	0.747	0.714	0.774	0.362
Fujian	0.732	0.640	0.616	0.661	0.555
Korean	0.916	0.904	0.914	0.915	0.913
Shandong	0.832	0.745	0.809	0.812	0.484

presents the per-class F1-score results for major cuisine types. It can be observed that Qwen3-8B achieves the highest or near-highest F1-scores across most categories, demonstrating its robustness in multi-class short-text classification tasks. The model performs especially well in highly standardized categories such as Sichuan cuisine (0.920), Japanese cuisine (0.931), and Korean cuisine (0.916). For less represented and more heterogeneous categories like Jiangsu cuisine, the F1-score (0.659) is relatively lower but still significantly outperforms BERT (0.299), showing robustness in low-resource and semantically ambiguous scenarios. Overall, Qwen3-8B maintains a leading position in mainstream cuisines and exhibits strong adaptability in niche and complex categories.

Commercial LLMs perform comparably or slightly better in some mainstream categories (e.g., Japanese cuisine and Western cuisine), but overall lack the stability of Qwen3-8B. DeepSeek-R1 performs similarly to Qwen3-8B in categories like Hunan, Cantonese, and Western cuisine, but shows greater fluctuation in niche classes. BERT significantly lags behind in most categories, especially in cross-cultural or ambiguously named categories (e.g., Hubei and Fujian cuisine), where the gap is more pronounced. In summary, Qwen3-8B’s fine-grained classification results provide strong empirical support for the framework’s generalizability and robustness.

3.3. Confusion Analysis

To assess performance in high-complexity contexts, this study uses a row-normalized confusion matrix (i.e., per-class recall) to evaluate class-wise discrimination (see Figure 7). Qwen3-8B attains over 90% per-class recall in several categories—Sichuan (91.4%), Huizhou (92.5%), and Japanese (92.6%)—which aligns with their per-class F1-scores and indicates that the model captures nuanced cues in short POI names.

Importantly, Qwen3-8B distinguishes well between closely related classes such as Sichuan and Hunan cuisines, achieving 91.4% and 88.3% per-class recall, respectively. This discriminative capacity is particularly valuable for identifying spatial patterns of culinary diversity in urban settings.

In contrast, Jiangsu cuisine (F1 = 0.659) and Hubei cuisine (F1 = 0.751) are more prone to confusion (see Figure 7) because their restaurant names provide weaker name-level cues. Sichuan and Hunan names often contain explicit markers, such as regional abbreviations “川/蜀” (Sichuan) and “湘” (Hunan) and flavor cues like “麻辣” (mala; numbing–spicy). By contrast, Jiangsu names more frequently use cuisine-agnostic generic expressions, including common suffix patterns such as “X记/家” (X as a placeholder name) and broad terms like “龙虾” (lobster), “土菜馆” (home-style eatery), and “小厨” (small kitchen). When distinctive markers are absent, predictions become more sensitive to inter-class semantic overlap and co-occurrence in the training data, leading to more frequent shifts toward high-co-occurrence categories such as Sichuan cuisine.

Overall, the confusion matrix exhibits strong diagonal dominance, with per-class recall exceeding 80% for most categories. These results confirm the model’s stability across both high-frequency and long-tail classes and highlight its capacity for fine-grained labeling.

4. Case Study: Guangzhou and Shenzhen

4.1. Data Expansion and Classification Scaling

Although the platform provides ≈220,000 fine-grained cuisine-labeled samples nationwide (about 4.3% of all restaurant POIs), labels become rapidly fragmented in the Guangzhou and Shenzhen case studies when the unit of analysis is refined to a “single city–single cuisine–500 m grid.” Beyond a few mainstream cuisines, many long-tail categories within a city contain only dozens or even single-digit labeled records (

Table 6. Labeled vs. LLM-classified counts and lift ratios (Guangzhou, Shenzhen).

Cuisine Name	Guangzhou			Shenzhen
	Label Count	LLM Count	Lift Ratio	Label Count	LLM Count	Lift Ratio
NE Chinese	105	1811	17.25	146	2557	17.51
Yunnan–Guizhou	30	718	23.93	38	876	23.05
Sichuan	879	9437	10.74	1145	13,381	11.69
Huizhou	6	91	15.17	14	143	10.21
Japanese	611	3024	4.95	493	2562	5.20
Thai–Viet	46	203	4.41	30	132	4.40
Hunan	1213	5843	4.82	1784	9926	5.56
Cantonese	2203	45,267	20.55	2084	38,162	18.31
Jiangsu	3	61	20.33	8	91	11.38
NW Chinese	72	1824	25.33	123	2698	21.93
Western	476	14,926	31.36	359	12,983	36.16
Hubei	50	906	18.12	87	1372	15.77
Fujian	46	1900	41.30	69	2331	33.78
Korean	202	2341	11.59	169	2959	17.51
Shandong	7	213	30.43	13	366	28.15
Total	5949	88,565	14.89	6562	90,539	13.80

), limiting high-resolution spatial inference. To examine cross-regional and cross-cultural transferability, we further conducted an additional case analysis in Chengdu (a Sichuan-cuisine–dominant context distinct from Cantonese cuisine); due to space constraints, the full results are reported in Appendix A.

To mitigate this sparsity, the semantic expansion module of LLM-SSIF is applied. Specifically, a PEFT (LoRA)-tuned Qwen3-8B model assigns cuisine labels to previously unlabeled restaurant POIs, improving within-city coverage and category representativeness. In Guangzhou and Shenzhen, the number of labeled records increased from 5949 and 6562 to 88,565 and 90,539—about 14.89 and 13.80 times, respectively. Importantly, the gains are not confined to dominant categories; long-tail cuisines also exhibit substantial lift ratios (see Table 6 and Figure 8), indicating robust generalization across both frequent and rare labels and enabling grid-level spatial inference on a materially less sparse dataset.

4.2. City-Level Composition and Spatial Patterns

At a macro-level, Guangzhou and Shenzhen share structural similarities in their foodscapes while also revealing distinctive differences. In both cities, local cuisine (i.e., the city’s native cuisine—Cantonese in both cases) constitutes the largest segment, underscoring the foundational role of regional culinary culture (see Figure 9). The degree of dominance, however, diverges. In Guangzhou, local cuisine accounts for 51.1% of all dining establishments (45,267 restaurants), reinforcing the city’s status as the birthplace of Cantonese cuisine and a historic gastronomic hub. In Shenzhen, the local share is 42.1% (38,162 restaurants), while domestic (non-local Chinese) cuisines reach 37.3% (33,741 restaurants), substantially higher than Guangzhou’s 25.7% (22,804 restaurants).

This divergence aligns with the cities’ socio-demographic profiles: Guangzhou’s composition reflects rootedness in traditional local culture, whereas Shenzhen, often characterized as a migrant city, shows greater interregional diversity and demand heterogeneity. Foreign (non-Chinese) cuisines are also well represented in both cities, with Guangzhou at 23.1% (20,494 restaurants) and Shenzhen at 20.6% (18,636 restaurants), indicating that globalized consumption patterns are embedded across both urban contexts.

Beyond composition, the grid-level distribution of total restaurants at a 500 m × 500 m resolution reveals contrasting exposure landscapes (Figure 10). In Guangzhou, restaurant counts concentrate in the historic core and taper toward surrounding districts, approximating a largely monocentric intensity pattern. In Shenzhen, multiple high-intensity clusters emerge across core districts and extend into several peripheral districts, reflecting a more polycentric spatial structure.

Taken together, the city-level composition (Figure 9) and grid-level exposure patterns (Figure 10) derived from the LLM-SSIF-classified POI set indicate systematic cross-city differences between Guangzhou and Shenzhen, providing a coherent baseline for subsequent within-city structural analysis.

4.3. Cultural Embedding Through Hotspot and Residual Analysis

Building on the macro baselines, this study first applies conventional hotspot analysis to assess whether the LLM-expanded dataset reproduces real-world clustering. As shown in Figure 11, Getis–Ord Gi* on raw counts identifies significant clusters across core urban areas in both cities, consistent with central advantages in density, accessibility, and commercial activity. This indicates that the LLM-classified data can reliably reproduce macro-level agglomeration.

At the same time, most central districts appear as hotspots regardless of cuisine type, reflecting scale effects from overall restaurant density and making cuisine-specific spatial preferences hard to see. To address this, this study implements a residual-based hotspot approach: expected counts are estimated via a GLM with total restaurant count as an offset, and Gi* is applied to standardized residuals (Figure 12). This isolates relative over-/under-representation, revealing structural cultural preferences beyond aggregate intensity.

The residual hotspot analysis reveals distinct cuisine-specific patterns across the three representative cuisines. After scale control, Cantonese shifts from raw-count hotspots in high-intensity central areas to relative coldspots in several central commercial zones, while emerging as hotspots in historically rooted neighborhoods and selected peripheral contexts. Hunan shows limited differentiation in the raw-count analysis but displays multiple peripheral residual hotspots, suggesting non-local clustering beyond overall supply intensity. Western remains strongly core-oriented under both analyses, forming residual hotspots in central areas and systematic under-representation toward the periphery.

These residual patterns can be interpreted as cuisine-specific “spatial anchoring” in relative share after controlling for overall restaurant intensity. From the perspective of polycentric urban structure and service distribution, cuisine-specific residual hotspots indicate localized service specialization anchored to different centers and their surrounding consumption catchments, rather than a simple reflection of overall supply intensity. For Cantonese, residual hotspots indicate a systematically higher-than-expected local share, consistent with cultural path dependence and locally anchored daily consumption, whereas the central core’s higher cuisine diversity can dilute its relative representation despite high absolute supply. For Hunan, peripheral residual hotspots plausibly reflect migrant-linked culinary clustering: non-local food practices become locally anchored and reproduced where interregional migrants and their everyday consumption networks concentrate, rather than implying a citywide shift in taste. For Western cuisine, persistent central residual hotspots point to dependence on central consumption environments associated with higher purchasing power and service-economy activity, thereby sustaining a durable core–periphery gradient even after exposure control.

Overall, the residual–hotspot evidence demonstrates how LLM-expanded labels can be translated into interpretable, scale-controlled spatial signals suitable for cross-city comparison within the LLM-SSIF pipeline.

5. Discussion

The primary contribution of this study is the proposed LLM-based semantic–spatial inference framework (LLM-SSIF), a unified pipeline that translates unstructured POI name text into interpretable, fine-grained spatial evidence. This framework distinguishes itself from existing work in two key aspects. First, it prioritizes a practical trade-off for urban researchers by employing PEFT (LoRA) on open LLMs, making high-quality adaptation feasible on consumer hardware—contrasting with reliance on black-box commercial APIs or computationally prohibitive full fine-tuning. Second, and more critically, it is end-to-end oriented to spatial inference, embedding classification outputs within a scale-controlled residual–hotspot spatial analytic workflow to reveal structural cultural preferences beyond raw density patterns.

Empirically, LLM-SSIF is instantiated via a PEFT (LoRA) label-expansion module trained on consumer-grade GPUs [30], achieving a favorable performance–compute–deployability trade-off while retaining data control [44]. In the Guangzhou–Shenzhen cuisine case, the approach enlarges usable fine-category coverage by ~14–15× and surpasses strong baselines on macro-averaged metrics, showing improved robustness for long-tail and semantically similar classes [45]. This scaling is crucial at the “single city–single cuisine–500 m grid” resolution, where platform labels otherwise become too fragmented to support stable high-resolution analysis.

Beyond classification, the pipeline links labels to scale-controlled spatial analytic inference to produce interpretable evidence. Naïve diagnostics (intensity maps and raw-count hotspot detection) capture clustering but are dominated by scale effects. By coupling labels with GLM residualization (using total count as an offset) and Getis–Ord Gi* on 500 m × 500 m grids, this study reveals cuisine-specific structural differentiation in relative share that intensity maps alone obscure: Cantonese is over-represented in historically rooted and selected peripheral contexts but less distinctive in high-intensity central commercial areas, Hunan forms multiple peripheral migrant-linked clusters, and Western remains persistently core-oriented with under-representation toward the periphery. These patterns align with each city’s historical–cultural context and migrant dynamics, illustrating how the label-expansion → residual–hotspot chain yields interpretable urban evidence.

Building on this methodological contribution, the framework also demonstrates clear practical value. The expanded labels enable fine-grained monitoring of service content and cultural differentiation at a 500 m resolution without requiring high-performance computing infrastructure, and the modular design lowers adoption barriers in planning and operations. Although the study does not pursue causal identification [46], the observed spatial tendencies are consistent with established mechanisms in urban cultural geography [47], reinforcing the interpretability of the methodological evidence [48].

Substantively, the Guangzhou–Shenzhen results provide a consumption-content lens on urban structure: Guangzhou exhibits a more historically embedded, core-weighted culinary pattern, whereas Shenzhen shows a more polycentric, migration-linked configuration. By separating cuisine-specific deviations from overall supply intensity, the residual–hotspot maps shift interpretation from where restaurants concentrate to where specific cultural contents are structurally anchored, and the two cities serve primarily as demonstrations of the LLM-SSIF inferential chain (text → fine labels → scale-controlled deviations → localized structural signals), rather than exhaustive urban narratives.

Although illustrated with restaurants, the pipeline is domain-agnostic. Any unstructured POI dataset—clinics, schools, or retail—can be transformed into fine-grained spatial evidence through the same label-expansion and residual–hotspot chain. This extensibility highlights its value for service-structure evaluation and comparative urban profiling beyond dining.

Naturally, this study has limitations. Because LLM-SSIF is designed to translate unstructured POI texts into scale-controlled spatial–statistical evidence, rather than to identify causal effects, the underlying mechanisms are not causally established: the current analysis does not yet incorporate multi-source explanatory data (e.g., rent levels or population flows) to test and disentangle potential drivers behind the observed patterns [49]. In addition, bias assessment remains limited and should be strengthened by quantifying label noise and ambiguity through targeted spot-checks and confusion-pair diagnostics.

Nevertheless, the framework is grounded on a nationwide training dataset of more than 220,000 cuisine-labeled POIs, ensuring exposure to diverse naming conventions and cultural variations. This provides an inherent basis for cross-city generalization. The Guangzhou–Shenzhen cases are, therefore, presented as representative demonstrations of how the framework translates expanded labels into spatial evidence, rather than as the only test beds. Future work could adopt explicit cross-regional evaluations to further quantify robustness under extreme heterogeneity.

6. Conclusions

This study proposes LLM-SSIF, a resource-efficient and transferable pipeline that links LLM-based label expansion with scale-controlled residual–hotspot spatial analysis at 500 m resolution, bridging unstructured POI text with interpretable urban evidence. The framework substantially enlarges fine-category coverage and outperforms baselines on macro metrics, alleviating long-tail sparsity and enabling robust spatial analytic inference via offset-GLM residualization and Getis–Ord Gi*. The Guangzhou–Shenzhen case demonstrates how expanded labels uncover structural cultural preferences beyond raw intensity patterns, supporting planning practice with fine-grained evidence. Future work should test mechanisms using multi-source data and causal or quasi-causal designs, while extending the pipeline to other urban functions and planning-oriented tools for routine monitoring and equity assessment.