2.1. Street-network structure and commercial land-use transformation

Residential-to-commercial transformation is closely associated with the spatial organization of street networks. Classical urban economic theories explain commercial concentration in terms of accessibility, land value, and locational advantage [1,2]. More recent configurational approaches extend this perspective by demonstrating that the position of streets within the network strongly influences pedestrian movement, visibility, and commercial attractiveness [3–7]. Consequently, commercial corridors are increasingly understood as products of both economic forces and street-network configuration rather than simply responses to land demand.

Numerous empirical studies have confirmed significant relationships between street-network centrality and commercial development. Degree, betweenness, closeness, PageRank, and straightness centrality have been associated with retail concentration, commercial facilities, employment distribution, and land-use intensity in a variety of urban contexts [8,10–14]. Collectively, these studies demonstrate that streets occupying strategically important positions within the urban network are more likely to experience commercial growth because they provide greater accessibility, movement opportunities, and spatial exposure.

Although these findings establish the importance of street-network structure, most existing studies examine commercial development at the city or district scale. Consequently, they explain general spatial trends rather than the selective transformation of individual corridors. Much less attention has been devoted to understanding why one corridor experiences commercial conversion while neighboring streets with similar physical characteristics remain predominantly residential. This corridor-scale perspective represents an important gap in the current literature.

2.2. Logistic regression

Within the broader land-use change literature, logistic regression remains one of the most widely used methods for modeling binary urban transformation [15,16]. Its appeal lies in its interpretability, enabling researchers to estimate how explanatory variables influence the probability of land conversion and to compare the relative importance of different predictors. For this reason, it has been widely adopted as a benchmark model in studies of urban growth and land-use change [17–19].

However, the strengths of logistic regression are accompanied by important limitations. Although it is effective for estimating variable effects and interpreting how predictors influence the probability of land-use conversion, conventional regression models generally treat observations as independent units and do not explicitly model spatial dependence or autocorrelation [20–22].This limitation is especially important in corridor-specific urban transformation, where change is shaped not only by local node-level attributes but also by adjacency, connectivity, and relational dependence across the street network. Therefore, logistic regression is best treated here as a strong interpretable benchmark rather than a complete model of highly relational urban processes. Because of these limitations, recent urban research has increasingly adopted graph-based approaches, including Graph Neural Networks, which are better suited to learning from connected spatial structures.

2.3. Graph Neural Networks (GNN)

GNNs are designed for graph-structured data and can learn simultaneously from node attributes and network topology, this makes them particularly suitable for urban systems, where the significance of a location depends not only on its own characteristics but also on its position within a wider network [23,24]. This shift is especially relevant for residential-to-commercial corridor transformation. The issue is not only whether an individual node changes function, but also whether the predicted pattern reproduces the observed continuity of a corridor. Graph-based models make it possible to combine conventional predictors, such as centrality measures, with the adjacency structure of the street network itself. As a result, they are better suited to representing inter-node dependence and spatial continuity than conventional regression models, which may still produce fragmented predictions even when their predictors are statistically significant [24,25].

In summary, the literature indicates that commercial activity tends to concentrate on more accessible and better-connected streets and that the underlying processes of change are spatially dependent rather than isolated. It also suggests that, although logistic regression remains a valuable benchmark, graph-based models are more consistent with the relational nature of street-network processes [23,25,26]. What remains underdeveloped, however, is a corridor-scale, historically grounded explanation of why one corridor transformed commercially while nearby streets did not.

2.4. Research gap

The Previous studies have established that street-network accessibility and centrality influence pedestrian movement, commercial concentration, retail location, and land-use intensity [8,10–15]. However, most of this research has examined citywide or district-scale spatial relationships using contemporary or cross-sectional data, providing limited explanation of why commercial transformation develops along one specific corridor while nearby streets with comparable physical characteristics remain predominantly residential. Conventional land-use models, particularly Logistic Regression, offer interpretable estimates of how individual variables affect the probability of urban transformation [18–20], but they do not explicitly learn from the adjacency, connectivity, and relational dependence embedded within street networks [21–23]. Graph-based models can address this limitation by combining node-level attributes with network topology; nevertheless, their application to historically grounded, corridor-specific residential-to-commercial transformation remains insufficiently investigated. Accordingly, the research gap is threefold: limited attention has been given to the historical network conditions preceding corridor transformation; few studies have directly compared conventional statistical and graph-based models using identical street-network predictors; and previous evaluations have largely emphasized node-level classification accuracy while overlooking whether predicted changes reproduce the connected spatial structure of the observed corridor. To address these gaps, this study reconstructs Erbil’s 2004 street network and compares Logistic Regression with a Graph Convolutional Network to explain the residential-to-commercial transformation observed between 2004 and 2024, using both conventional classification metrics and spatial-coherence indicators.

3.1. Study design and temporal framework

The analytical logic is temporal. A case-study plan representing the study area in 2004 was reconstructed using historical spatial references. Google Earth Pro was consulted solely as a historical visual reference for identifying the street configuration existing in 2004; no Google Earth, Google Maps, or Google Street View imagery is reproduced in any figure in this manuscript.

3.2. Data preparation and variable definition

The working geospatial dataset was prepared from the 2004 case study plan, which reconstructed the historical street structure of the study area at a time when the area was still predominantly residential. The original street network was first represented as street edges, and these edges were then converted into nodes using QGIS 3.40 Bratislava. Each node in the resulting point-based layer represented a location derived from the 2004 street network and served as the basic observation unit for subsequent analysis.

All node coordinates were projected to EPSG:32638 (WGS 84/ UTM zone 38N), a projected coordinate reference system with units in meters based on the Universal Transverse Mercator method. This ensured that all distance-based operations were performed consistently in metric units. The dependent variable was defined as a binary land-use outcome based on the observed condition in 2024. Nodes representing locations along the corridor that had transformed into commercial use were assigned a value of 1, whereas nodes that remained residential were assigned a value of 0. S1 Fig, presents the case-study area.

The 2004 explanatory network was reconstructed by digitizing the historical street configuration identified from the case-study plan and historical reference imagery. Google Earth Pro was used only as a visual reference during this reconstruction process and no Google imagery is reproduced in the manuscript or Supporting Information. The resulting street centerlines were subsequently cleaned and processed in QGIS 3.40 Bratislava. The cleaned edges were converted into nodes, and only valid point geometries were retained for analysis. The attribute fields used in modeling were then checked so that Label and Weight were numeric, while rows with invalid labels were excluded. The binary outcome was assigned from the observed 2024 land-use condition: nodes located on the corridor that had converted to commercial use were coded 1, whereas nodes on adjoining streets that remained predominantly residential were coded 0. After filtering, the final analytical dataset contained 279 valid nodes, which formed the common observation base for both the Logistic Regression and GCN models.

In addition to the binary target variable, each node retained a weight attribute derived from the street width in the prepared spatial dataset in order to preserve local spatial importance. In this study, street width was classified into three ordinal weight levels: streets with a width of 10 m were assigned a weight of 1, streets with a width of 15 m were assigned a weight of 2, and streets with a width of 30 m were assigned a weight of 3. This weighting scheme assumes that wider streets have greater movement capacity, frontage visibility, and commercial exposure within the corridor structure. Ordinal weights were preferred to raw width values because they preserve the local street hierarchy while avoiding an unnecessarily strong scale effect from absolute width magnitudes in a relatively small case-study network. After the network centrality measures were computed, each centrality value was multiplied by the corresponding node weight so that the final predictors reflected both the structural position of the node in the network and its local spatial significance.

The analysis was implemented in Python using libraries aligned with each stage of the workflow. GeoPandas supported spatial data handling and coordinate transformation, while Pandas and NumPy were used for data preparation, weighting, feature construction, and z-score standardization. NetworkX was used to construct the graph and compute degree, betweenness, closeness, PageRank, straightness, and eigenvector centrality. Scikit-learn supported k-nearest-neighbor graph construction, the stratified 70:30 train–test split, logistic regression estimation, and performance evaluation using accuracy, precision, recall, F1-score, Jaccard similarity (IoU), ROC–AUC, the confusion matrix, and the classification report. The GCN model was implemented in PyTorch for tensor construction, model definition, optimization, loss computation, and probability estimation. Matplotlib was used for figures, and Pandas and GeoPandas were used to export tabular and spatial outputs. Because eigenvector centrality showed negligible explanatory contribution during preliminary feature screening, the final LR and GCN models were estimated using five predictors: degree, betweenness, closeness, PageRank, and straightness. The research code and minimal dataset supporting this study are publicly available in Zenodo at doi.org/10.5281/zenodo.21343914.

3.3. Street-network centrality measures

Following the logic of multiple centrality assessment, the combined set of measures captures complementary dimensions of the 2004 street system, including local connectivity, intermediary control, global accessibility, route efficiency, and recursive importance [27]. Six centrality measures were initially calculated because no single measure can fully represent the structural characteristics of a street network. Together, degree, betweenness, closeness, PageRank, straightness, and eigenvector centrality capture complementary dimensions of local connectivity, intermediary control, global accessibility, route directness, and recursive network importance [11,28]. Following preliminary feature screening, eigenvector centrality was excluded from the final models because it consistently showed the weakest explanatory and predictive contribution compared with the other measures. The remaining five predictors were retained to reduce unnecessary model complexity, preserve interpretability, and ensure that Logistic Regression and the GCN were evaluated using the same predictor set.

1. Degree centrality was used to represent immediate local connectivity, that is, how many direct links each node has to adjacent nodes. In street networks, it reflects local permeability and the number of immediate movement options around a node [27,28].

(1)

where: = degree of node , = element of the adjacency matrix and = set of nodes

2. Betweenness centrality was used to capture intermediary or corridor control, since it identifies nodes that lie on many shortest paths between other nodes. This is especially relevant for commercial-corridor formation because highly traversed nodes tend to concentrate through-movement and exposure [27,29–31].

(2)

where is the betweenness centrality of node , and are source and destination nodes, denotes the number of shortest paths from to , and denotes the number of those shortest paths that pass through node .

3. Closeness centrality was used to measure global accessibility, meaning how efficiently a node can reach all other nodes in the network. In urban analysis, this helps identify locations that are more centrally positioned within the overall street system [27,32,33].

(3)

where: = closeness centrality of node , = total number of nodes in the network and = shortest-path distance between node and node

4. PageRank centrality was included to represent recursive structural importance, where a node is important not only because it has connections, but because it is connected to other important nodes. This adds a hierarchical dimension of network influence beyond simple connectivity [34–37].

(4)

where is the PageRank of node i, alpha is the damping factor, N is the total number of nodes, N(i) is the set of nodes connected to node i, and is the weight of the edge between nodes j and is the sum of the weights of all edges connected from node j to its neighboring nodes and is the PageRank value of node .

5. Straightness centrality was used to evaluate route directness, that is, how closely movement through the network approximates straight-line travel. This is particularly suitable for street networks because commercial activity often favors routes that are not only connected, but also spatially efficient and legible [27,38].

(5)

Where: represents the straightness centrality of node . is the total number of nodes in the network. refers to the Euclidean distance between nodes and , represents the graph distance between nodes and .

After the centrality measures were computed on the reconstructed network, each centrality score was multiplied by the corresponding node weight so that the final predictors reflected both network position and local node significance [39–41]. The resulting weighted centrality variables were then standardized using z-score normalization before model training.

(6)

where is the raw centrality score and is the node-level street-width weight.

The weighted variables were subsequently standardized using z-score normalization based on the mean and standard deviation of the training subset, and the same scaling parameters were applied to the full feature set used in model training and evaluation.

(7)

where is the standardized weighted value of variable j for node i, is the weighted feature value, is the mean of variable j in the training set, and is the standard deviation of variable j in the training set. [42].

3.4. Feature scaling, main train–test split, and repeated validation design

To minimize information leakage, the dataset was divided into training and test subsets before final feature standardization. Preprocessing parameters were therefore estimated from the training data only and then transferred to the held-out test data. A stratified 70:30 train–test split was first adopted using a fixed random seed of 42. This main split was used for the primary model comparison and for producing the spatial prediction maps, ensuring direct comparability between Logistic Regression and the GCN.

For each split, the weighted centrality predictors were standardized using the mean and standard deviation of the training subset only. The same scaling parameters were then applied to the corresponding test subset. This procedure ensured that test nodes did not influence either scaling or class-weight estimation during model fitting, thereby reducing the risk of information leakage and overstated model performance.

In parallel, the graph supplied to the GCN was constructed as an undirected Euclidean k-nearest-neighbor network, with k capped at 8 and reduced automatically when the number of observations was smaller. Edge weights were defined by Euclidean distance; inverse distance was used where recursive centrality computation required it; and the GCN adjacency matrix was represented as a binary connectivity matrix with self-loops and symmetric normalization. The graph structure remained available in the transductive GCN setting, while the supervised loss was evaluated only on the training nodes.

To reduce dependence on a single train–test partition, an additional repeated-split robustness validation was conducted. The same stratified 70:30 split procedure was repeated across 30 random seeds. For each repetition, the training and test nodes were reselected while preserving the class distribution, training-only standardization was reapplied, and both the Logistic Regression and GCN models were re-estimated using the same five-predictor specification. The resulting classification and spatial-coherence metrics were summarized using the mean and standard deviation across the 30 repetitions.

3.5. Logistic Regression model

Logistic regression was adopted as the baseline statistical model because it is a standard and interpretable classifier for binary outcomes, making it well suited for benchmarking more complex learning models. Its coefficients provide a transparent description of the association between each predictor and the log-odds of commercial conversion, while predicted probabilities are obtained through the logistic response function [43–45]. To reduce the effect of class imbalance, weighted fitting was applied through balanced class weights, and the model was estimated with a maximum of 2,000 iterations. For node i, the probability of commercial conversion was modeled as

(8)

where denotes the probability of commercial conversion for node , is the intercept term, is the coefficient of predictor , and is the standardized value of predictor for node .

Logistic regression was used as an interpretable benchmark rather than as a complete spatial model. This choice was made because the study aims to evaluate whether graph-based learning adds explanatory and predictive value beyond a conventional binary land-use change model. Although spatial econometric models are appropriate for explicitly estimating spatial lag, spatial error, or spillover effects, they require a predefined spatial weights structure and primarily model spatial dependence through econometric parameters. In contrast, the present study focuses on a street network represented as a graph, where the relational structure among nodes is central to the research question. The GCN was therefore selected as the main relational model because it directly combines node-level predictors with adjacency-based learning. Spatial econometric modeling remains a valuable extension for future comparative work, but logistic regression provides a clearer and more widely used baseline for assessing the added value of graph-convolutional learning.

3.6. Graph convolutional network (GCN) model

A two-layer Graph Convolutional Network (GCN) was employed to examine whether the same centrality-derived predictors gained additional explanatory power when neighborhood effects were learned explicitly from the graph structure. In the final specification, each node was initialized with five standardized weighted predictors (degree_w, betw_w, close_w, pagerank_w, and straight_w). The graph was treated in a semi-supervised transductive setting: node embeddings were computed over the full k-nearest-neighbor graph, but the supervised loss was evaluated only on the training nodes using a Boolean training mask. Edge distances were used for distance-sensitive centrality calculations, whereas the GCN adjacency matrix was represented as a binary connectivity matrix with self-loops and symmetric normalization. The hidden layer contained 16 units with ReLU activation, the output layer contained 2 units for binary classification, optimization used Adam with a learning rate of 0.01, and training proceeded for 200 epochs. Class-weighted cross-entropy was used so that the minority commercial class exerted greater influence during fitting [46,47].

Accordingly, the first graph-convolution layer was defined as:

(9)

The output layer was defined as:

(10)

In these expressions, Â denotes the symmetrically normalized adjacency matrix with self-loops, X is the input feature matrix, and are trainable weight matrices, and and are bias terms [46].

Node-level class probabilities were obtained by applying a row-wise SoftMax function:

(11)

The model was therefore estimated in a semi-supervised transductive setting, which is consistent with the original GCN formulation and preserves the common graph context for both training and test nodes while restricting supervision to the training subset [46,48] Optimization was based on a class-weighted cross-entropy loss evaluated over the training nodes only:

(12)

Here, T denotes the set of training nodes, is the weight assigned to the observed class of node i, and is the predicted probability of the true class. Optimization was performed using Adam [47].

During repeated-split validation, the same GCN architecture and hyperparameters were retained for each repetition, but the training and test masks were updated according to the current stratified split. The model was reinitialized and retrained for each random seed so that the reported robustness results reflected independent train–test partitions rather than repeated evaluation of a single trained model.

Feature screening was conducted before the final comparative estimation using three complementary indicators: the bivariate correlation between each weighted predictor and the commercial label, the corresponding Logistic Regression coefficient, and the GCN feature-contribution score derived from the trained model. Eigenvector centrality consistently showed the weakest contribution across these indicators, with a bivariate correlation of −0.073, a Logistic Regression coefficient of −0.153, and a GCN contribution score of 0.089. These results indicated that eigenvector centrality added limited discriminatory information beyond the other centrality measures in this corridor-scale network. Its exclusion therefore reduced redundancy and model complexity without removing a substantively important predictor. The final Logistic Regression and GCN models were consequently estimated using weighted degree, betweenness, closeness, PageRank, and straightness centrality.

3.7. Model evaluation

Model performance was evaluated at two complementary levels. First, conventional predictive accuracy was assessed on the held-out test set using five standard classification metrics: accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (ROC–AUC). These measures quantified the ability of the models to distinguish between residential and commercial nodes at the individual-node level. Precision was computed as TP/ (TP + FP), recall as TP/ (TP + FN), F1-score as 2 × (Precision × Recall)/ (Precision + Recall), and Intersection over Union (IoU) as TP/ (TP + FP + FN) [49–51].

While these conventional metrics evaluate classification performance at the node level, they do not indicate whether predicted commercial nodes form a spatial pattern consistent with the observed corridor structure. Therefore, a second level of evaluation was conducted to assess spatial coherence. This was done by extracting the subgraph composed of all nodes predicted as commercial and then examining its connected structure. Three aspects were considered: the number of connected components, the fragmentation of the predicted-positive pattern, and the proportion of predicted commercial nodes contained in the largest connected component.

Spatial coherence was then examined on the subgraph of predicted commercial nodes. Fragmentation [50–53] was computed as

(13)

where denotes fragmentation, is the number of predicted positive nodes in connected component , and is the total number of predicted positive nodes in the evaluated subgraph [54,55].The Largest Connected Component Ratio (LCCR) was computed as , where is the size of the largest connected component and is the total number of predicted positive nodes. Robustness was examined in two ways. First, a specification sensitivity analysis compared the original six-predictor model with the final five-predictor model. Second, repeated-split validation was used to examine whether the findings were sensitive to the particular random partition of the data. The stratified 70:30 split was repeated across 30 random seeds, and both models were retrained and re-evaluated for each repetition. Classification metrics were calculated on the held-out test nodes for each split and summarized as mean ± standard deviation. Spatial-coherence metrics were also recalculated on the held-out test nodes for each repetition to determine whether the corridor-continuity advantage of the GCN remained stable across alternative train–test partitions.

The stratified 70:30 split was repeated across 30 random seeds, and both models were retrained and re-evaluated for each repetition. Classification metrics were calculated on the held-out test nodes for each split and summarized as mean ± standard deviation. Spatial-coherence metrics were also recalculated on the held-out test nodes for each repetition to determine whether the corridor-continuity advantage of the GCN remained stable across alternative train–test partitions.

For the main split, spatial-coherence metrics were reported for both the held-out test nodes and the full dataset. The test-set results were used as validation evidence, whereas the full-dataset results were interpreted as descriptive spatial-pattern diagnostics showing how each model represented the complete corridor structure.

3.8. Analytical interpretation

The final interpretation proceeded at two complementary levels. At the explanatory level, the Logistic Regression coefficients and the comparative feature contributions of the GCN were used to identify which centrality dimensions were most strongly associated with commercial transformation. At the classification level, the two models were compared using conventional node-level metrics, including accuracy, precision, recall, F1-score, ROC–AUC, and IoU. These metrics assessed how well each model classified individual nodes as residential or commercial.

At the spatial level, the predicted commercial nodes were evaluated as a connected subgraph to determine whether each model reproduced the corridor-like structure of transformation. This distinction was important because a model may correctly classify many individual nodes while still producing a fragmented spatial pattern. Therefore, the interpretation focused not only on whether transformed nodes were identified, but also on whether the predicted pattern reproduced the spatial continuity of the observed commercial corridor.

The repeated-split validation was used to distinguish stable model behavior from results that may depend on one particular train–test partition. Accordingly, the GCN was interpreted as providing a methodological advantage only where its spatial-coherence performance remained stronger across repeated splits, rather than being treated as universally superior across all conventional classification metrics.

3.9. Methodological contribution

Methodologically, the approach treats the 2004 street network not as a static background map but as a prior structural condition capable of shaping later land-use change. By comparing a conventional statistical classifier with a graph-based relational model on the same historical predictors, the method tests whether explicit network dependence adds explanatory value to the study of corridor-specific transformation. This design makes it possible to distinguish between explanations based only on node-level accessibility and explanations that also depend on relational embedding within the wider street system. By adding repeated stratified validation and spatial-coherence diagnostics, the workflow evaluates both predictive reliability and the spatial structure of model outputs, which is especially important for corridor-based transformation.where continuity is part of the phenomenon being explained. The methodological workflow is summarized in S2 Fig.