Online car-hailing demand prediction region division method based on road network characteristics and POI characteristics

Abstract

The invention discloses an online car-hailing demand prediction region division method based on road network characteristics and POI characteristics, and the method comprises the following steps: S1, dividing a target geographic region into a plurality of basic space units, and extracting road network characteristics and POI characteristics; s2, calculating similar feature vectors of all adjacent space units; s3, determining the weight of each feature by adopting a CRITIC weight method, and obtaining the comprehensive similarity between adjacent space units; s4, constructing an undirected weighted graph based on the comprehensive similarity, and performing clustering merging by using a community discovery algorithm; and S5, obtaining a plurality of candidate partitions, and selecting an optimal partition from the plurality of candidate partitions based on the modularity, the contour coefficient and the partition stability index to serve as a final online car-hailing demand prediction region division scheme. According to the method, more scientific and reasonable demand area unit division can be realized, and the prediction precision and robustness are improved while the calculation overhead is reduced.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and data mining technology. Specifically, it relates to a method for dividing ride-hailing demand prediction areas based on road network characteristics and POI characteristics. Background Technology

[0002] With the continuous growth of urban transportation demand, ride-hailing services are playing an increasingly important role in the urban transportation structure. Ride-hailing demand forecasting is a core component of intelligent transportation systems; accurate forecasting helps optimize vehicle scheduling, reduce empty runs, and improve user experience. Existing technologies typically use regular grids (such as square grids) to divide urban space, and then perform demand forecasting within each spatial unit. This method has the advantages of being simple to implement and computationally convenient, but it suffers from spatial fragmentation, resulting in many demand gaps within the divided spatial units, leading to inefficiency and insufficient forecast accuracy. Some improved methods attempt to utilize historical trajectory data for segmentation, but these methods rely excessively on historical order distributions, making it difficult to identify potential demand areas and unable to effectively adapt to the dynamic changes in urban functional areas.

[0003] Therefore, there is an urgent need to propose a new method that can adaptively divide urban spatial features to improve prediction efficiency and accuracy. Summary of the Invention

[0004] The primary objective of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a method for dividing ride-hailing demand prediction areas based on road network characteristics and POI characteristics. This method can achieve a more scientific and reasonable division of demand area units, improving prediction accuracy and robustness while reducing computational overhead.

[0005] The objective of this invention is achieved through the following technical solution: a method for regional division of ride-hailing demand prediction based on road network characteristics and POI characteristics, comprising the following steps:

[0006] S1. Divide the target geographic area into several basic spatial units, and extract road network features and POI features for each spatial unit;

[0007] S2. Based on road network features and POI features, calculate the similarity feature vectors of all adjacent spatial units. The similarity feature vectors include regional road network density, regional activity similarity, and land use similarity.

[0008] S3. The weights of the regional road network density, regional activity similarity, and land use similarity are determined by using the CRITIC weighting method, and then integrated to obtain the comprehensive similarity between adjacent spatial units.

[0009] S4. Treat spatial units as nodes and use the comprehensive similarity as the weight of the connecting edges to construct an undirected weighted graph; use a community detection algorithm to cluster the undirected weighted graph and divide it into multiple communities, with each community corresponding to a merged ride-hailing demand prediction area.

[0010] S5. Obtain multiple candidate partitions, and select the optimal partition from the multiple candidate partitions based on modularity, profile coefficient and partition stability index, as the final ride-hailing demand prediction area division scheme.

[0011] Preferably, in step S1, the basic spatial unit is a regular hexagonal grid, and the road network features include the number of connected roads, road network density, and intersection density; the POI features include POI density, POI type proportion vector, and land mixed use degree.

[0012] Preferably, in step S2, the land use similarity is obtained by calculating the Jensen-Shannon divergence (JSD) of the proportion vector of POI types of adjacent spatial units, and the calculation formula is:

[0013] POI(i,j)=1-JSD(I||J),

[0014] Where i and j are a pair of adjacent spatial units, I and J are the POI type proportion distribution vectors of spatial units i and j respectively, and JSD(I||J) represents the Jensen-Shannon divergence between distributions I and J.

[0015] Preferably, before calculating the Jensen-Shannon divergence, the POI type proportion vector is smoothed, and the formula for calculating the probability distribution after smoothing is as follows:

[0016]

[0017] Where p(k) is the original proportion of the k-th type of POI, ε is a constant, and n is the number of POI types.

[0018] Preferably, in step S2, the region activity similarity is obtained through the following steps:

[0019] S21. For spatial units i and j, define regional active feature vectors and perform normalization processing respectively. The regional active feature vectors include road network density, intersection density, POI density and land mixed use degree.

[0020] S22. Calculate the Euclidean distance between the normalized active feature vectors of spatial units i and j.

[0021] S23. Using the Gaussian kernel function, the Euclidean distance is converted into regional active similarity.

[0022] Preferably, in step S3, the process of determining the weights using the CRITIC weighting method includes:

[0023] The information content of each of the three indicators—regional road network density, regional activity similarity, and land use similarity—is calculated, and the weight of each indicator is obtained through normalization. The formula for calculating the information content is as follows:

[0024]

[0025] Where, σ j Let |r| be the standard deviation of index j, representing the contrast intensity. jk | represents the absolute value correlation coefficient between index j and index k. The sum of the conflict between indicator j and all other indicators.

[0026] Preferably, in step S4, the community detection algorithm is the Leiden algorithm, which includes three iterative stages: local movement, community refinement, and network aggregation, and takes maximizing modularity as the objective function.

[0027] Preferably, in step S5, the multiple candidate regions are obtained by adjusting the resolution parameters of the community detection algorithm and performing multiple random initializations.

[0028] Preferably, the partition stability is evaluated by calculating the normalized mutual information (NMI) or information variability (VI) of the partition results under multiple random initializations.

[0029] Preferably, the formula for calculating the modularity is:

[0030]

[0031] Among them, A ij Let k be the weight between spatial unit i and spatial unit j, i.e., the comprehensive similarity. The weighted degree of spatial unit i is k. i =∑ j A ij m = 1 / 2∑ i k i γ is the resolution parameter, δ(c i c j ) indicates whether spatial unit i and spatial unit j are in the same community.

[0032] The present invention has the following advantages and effects compared with the prior art:

[0033] (1) This invention provides a method for dividing ride-hailing demand prediction areas based on road network characteristics and POI characteristics. This method deeply integrates multi-source geospatial data (road network characteristics and POI characteristics). The areas divided by this invention are highly consistent in terms of physical connection and functional attributes, which fundamentally ensures the "homogeneity" of ride-hailing demand within the area. This lays a solid foundation for high-precision prediction and overcomes the limitations of traditional methods that rely solely on historical order data or simple spatial division. This makes the division results more consistent with the actual distribution of urban functional areas. At the same time, an advanced community detection algorithm is used for adaptive merging, which transforms the urban spatial division problem into a community detection problem in a complex network. A graph structure is constructed with comprehensive similarity as the edge weight. By quantifying the functional similarity between spatial units and using optimization algorithms for intelligent merging, a more scientific and reasonable division of demand prediction units is finally achieved. This effectively avoids the subjectivity of manual division and the large number of demand gaps generated by regular grids in dividing urban space, and significantly reduces the computational complexity of subsequent prediction models.

[0034] (2) The present invention uses the CRITIC weighting method to determine the weight of each similarity index, comprehensively considers the comparative strength and conflict of the index, avoids subjective weighting bias, and improves the scientificity and reliability of the comprehensive similarity calculation.

[0035] (3) The regions divided by the method of the present invention can be used as inputs and are applicable to various time series prediction models such as RandomForest, XGBoost, LSTM, and Transformer, with wide applicability and good compatibility. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the method for regional division of ride-hailing demand prediction based on road network characteristics and POI characteristics according to the present invention.

[0037] Figure 2 This is a block diagram illustrating the logical implementation of the ride-hailing demand prediction area division method based on road network characteristics and POI characteristics of the present invention. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0039] Example 1

[0040] like Figure 1 The diagram shows a flowchart of a method for dividing ride-hailing demand forecasting regions based on road network characteristics and POI characteristics, including the following steps:

[0041] S1. Divide the target geographic area into several basic spatial units, and extract road network features R for each spatial unit.i and POI features P i ;

[0042] Specifically, in this embodiment, a regular hexagon with a certain side length is used as the basic spatial unit to divide the study area into full coverage areas.

[0043] The road network characteristics (road network properties) include:

[0044] Number of connected roads: p ij = The number of road segments that cross the boundary between spatial unit i and spatial unit j;

[0045] Road network density: p i = Total road length within spatial unit i (m) / Area of ​​spatial unit (m²) 2 );

[0046] Intersection density; τ i = Number of intersections within spatial unit i (number) / Area of ​​spatial unit (m²) 2 ).

[0047] The POI features (point of interest features / POI characteristics) include:

[0048] POI density: poi i = Number of POIs in spatial cell i (number of POIs) / Grid area (m²) 2 );

[0049] POI type proportion vector: poi it =[p i (1),p i (2),…,p i (k),…,p i (n)];

[0050] Land Use Mixing Ratio: LUI i =-∑p i (k)·ln[p i (k)];

[0051] Where: p i (k) represents the proportion of POI type k in spatial unit i to all POIs in that spatial unit. These characteristics together constitute a quantitative description of the region's traffic carrying capacity and functional attributes.

[0052] S2. Based on road network features and POI features, calculate the similarity feature vector of all adjacent spatial units. The similarity feature vector F(i,j) = [P(i,j),D(i,j),POI(i,j)]. The specific calculation process is as follows:

[0053] (1) Regional road network density P(i,j): Calculate the number of road segments p(i,j) that cross the boundary of the spatial unit. According to the characteristics of the data distribution, select a normalization method that is appropriate to it. The normalization method includes, but is not limited to, Max normalization, Max-min normalization and logarithmic normalization.

[0054] (2) Region activity similarity D(i,j): Calculate the Gaussian similarity based on the difference of feature vectors of spatial units according to Euclidean distance. The specific calculation steps include:

[0055] S21. Define the regional activity feature vector F of a single spatial unit. i =[ρ i , τ i poi i LUI i ], ρ i , τ i poi i LUI i These represent road network density, intersection density, POI density, and land use mix, respectively. Since these four features may have different dimensions and orders of magnitude, direct distance calculation would be dominated by large numerical features. Therefore, each feature is first normalized across all spatial units. Similarly, based on the data distribution characteristics, a normalization method appropriate to it is adopted to obtain the normalized value ρ. i ’ , τ i ',poi i ',LUI i ';

[0056] S22. Calculate the Euclidean distance between the normalized active feature vectors of spatial unit i:

[0057]

[0058] 23. Calculate region activity similarity based on Gaussian kernel function:

[0059]

[0060] Here, σ is a scale parameter, which can be selected as the mean of the Euclidean distances of all adjacent spatial unit pairs.

[0061] (3) Land use similarity POI(i,j):

[0062] This indicator can accurately measure the similarity of regional functions and, compared to traditional methods such as Euclidean distance, can more effectively capture the shape differences in probability distributions. It is obtained by calculating the Jensen-Shannon divergence (JSD) of the proportion vector of POI types among adjacent spatial units, with the formula: POI(i,j)=1-JSD(I||J).

[0063] Where i and j are a pair of adjacent spatial units, I and J are the POI type proportion distribution vectors of spatial units i and j respectively, and JSD(I||J) represents the Jensen-Shannon divergence between distributions I and J.

[0064] JSD divergence does not essentially measure "similarity," but rather "difference" or "distance." It measures the amount of information lost when approximating one probability distribution with another. The derivation and calculation process of the above formula is as follows:

[0065] First, calculate the intermediate distribution M:

[0066]

[0067] Next, calculate the JSD divergence:

[0068]

[0069] Among them, D KL Let KL divergence be the KL divergence.

[0070]

[0071] Substituting the KL divergence into the JSD(I||J) calculation formula, we get:

[0072]

[0073] When using base-2 logarithms, the JSD divergence ranges from [0,1]. Since JSD measures dissimilarity, the larger the value, the greater the dissimilarity. Therefore, the region utilization similarity can be obtained through a simple transformation: POI(i,j)=1-JSD(I||J).

[0074] Since some regions may not have a certain type of POI, therefore p i When (k) = 0, substituting it into the formula will result in the form log = 0, which has no mathematical meaning. Therefore, the following rule is adopted:

[0075] Before calculating the Jensen-Shannon divergence, the POI type proportion vector is smoothed. Smoothing adds a very small constant (pseudo-count) to each probability value to avoid zero values, and simultaneously re-normalizes to ensure that the smoothed vector still represents a probability distribution (summing to 1). The specific formula is:

[0076]

[0077] Where, p i (k) represents the original proportion of the k-th type of POI in spatial unit i, ε is a small constant, which can be set to 1e-10, and n is the number of POI types.

[0078] S3. The weights of the regional road network density, regional activity similarity, and land use similarity are determined by using the CRITIC weighting method, and then integrated to obtain the comprehensive similarity between adjacent spatial units.

[0079] Specifically, the CRITIC weighting method is a more comprehensive approach than the entropy weighting method. It considers both the comparative strength and conflict of indicators simultaneously. This method overcomes the biases of subjective weighting methods and the shortcomings of entropy weighting methods in ignoring indicator correlation, ensuring the scientific nature of weight allocation. Comparative strength is measured by standard deviation; the larger the standard deviation, the greater the data fluctuation, the more information the indicator provides, and therefore it should be given a higher weight. Conflict is measured by the correlation between indicators (absolute value correlation coefficient). If an indicator is highly positively correlated with other indicators, it indicates a high degree of overlap in the information they reflect, and its weight should be reduced.

[0080] The CRITIC method combines the above-mentioned contrast strength and conflict to calculate an information-rich value for each indicator:

[0081]

[0082] Where, σ j Let |r| be the standard deviation of index j, representing the contrast intensity. jk | represents the absolute value correlation coefficient between index j and index k. The sum of the conflict between indicator j and all other indicators is represented by the absolute value correlation coefficient |r|. jk This is because both positive and negative correlations indicate overlapping information. Ultimately, the weights are determined by C. j The relative size determines C j The larger the value, the greater the amount of comprehensive information contained in indicator j, and the higher its weight.

[0083] The specific calculation process includes:

[0084] S31. Construct a feature matrix consisting of the feature vectors of all adjacent spatial unit pairs;

[0085] Assuming there are s pairs of adjacent spatial grids, then construct the feature matrix X based on the feature vectors F(i,j)=[P(i,j),D(i,j),POI(i,j)] between adjacent spatial cells.

[0086]

[0087] Since P(i,j), D(i,j), and POI(i,j) have all been calculated using normalization, there is no need to normalize them again.

[0088] In this embodiment, the similarity feature vectors F(i,j) of all adjacent spatial unit pairs are combined to construct an s-row, 3-column feature matrix X, which serves as the input to the CRITIC weighting method.

[0089] S32. Calculate the standard deviation of each indicator in the feature matrix as the contrast strength;

[0090] Calculate the standard deviation of each column (i.e., each indicator) in the feature matrix X:

[0091]

[0092] in, Let j be the mean of index j, and finally we get a vector [σ1,σ2,σ3] containing 3 standard deviations.

[0093] S33. Calculate the correlation coefficient matrix between indicators in the feature matrix, and calculate the conflict between each indicator and other indicators based on the correlation coefficient.

[0094] The correlation coefficient matrix R of the feature matrix X is usually calculated using the Pearson correlation coefficient r. jk :

[0095]

[0096] R is a 3×3 pairwise matrix, where the elements on the diagonal are r ii =1.

[0097] S34. Multiply the contrast strength and conflict of each indicator to obtain the information content of that indicator;

[0098] For indicator j, its information content C j for:

[0099] S35. Normalize the information content to obtain the final weight of each indicator.

[0100] The information content of each indicator is normalized to obtain the final weight: The final weight vector [w1, w2, w3] is obtained and used to calculate the overall similarity:

[0101] f(i,j)=w1×P(i,j)+w2×D(i,j)+w3×POI(i,j),

[0102] S4. Treat spatial units as nodes and use the comprehensive similarity as the weight of the connecting edges to construct an undirected weighted graph; use a community detection algorithm to cluster the undirected weighted graph and divide it into multiple communities, with each community corresponding to a merged ride-hailing demand prediction area.

[0103] Specifically, in this embodiment, the community detection algorithm is the Leiden algorithm. The Leiden algorithm includes three iterative stages: local movement, community refinement, and network aggregation. This ensures community connectivity and effectively avoids inferior local optima, thus maximizing the modularity Q. The formula for calculating the modularity is:

[0104]

[0105] Among them, A ij Let k be the weight between spatial unit i and spatial unit j, i.e., the comprehensive similarity. The weighted degree of spatial unit i is k. i =∑ j A ij m = 1 / 2∑ i k i γ is the resolution parameter, δ(c i c j The value indicates whether spatial unit i and spatial unit j are in the same community. It is 1 if they are in the same community and 0 otherwise.

[0106] Specifically, by adaptively merging the initial fine-grained grid through the community detection algorithm, the number of spatial units is significantly reduced, thereby lowering the computational complexity and training time of the subsequent time series prediction model and making large-scale city-level ride-hailing demand prediction more efficient.

[0107] S5. Obtain multiple candidate partitions, and select the optimal partition from the multiple candidate partitions based on modularity, profile coefficient and partition stability index, as the final ride-hailing demand prediction area division scheme.

[0108] Specifically, in step S5, the multiple candidate regions are obtained by adjusting the resolution parameters of the community detection algorithm and performing multiple random initializations: by setting different resolution parameters γ (e.g., 0.5, 0.8, 1.0, 1.2, 1.5) and random seeds, a large number of candidate partitions are generated. For each candidate partition, its modularity Q, silhouette coefficient, normalized mutual information NMI, and information difference metric VI are calculated to evaluate its stability.

[0109] Modularity is a metric for the density of nodes within a community, comparing a partitioned network to a random network. Specifically, modularity measures whether the number of edges within a community in the partitioned network significantly exceeds the expected number of edges for that community in a random network. Higher modularity indicates better clustering, higher inter-cluster discrimination, and better intra-cluster density. Lower modularity indicates poorer clustering and lower cluster quality.

[0110] The silhouette coefficient is a metric used to evaluate the quality of clustering results. It measures the similarity of each node to its own community and its nearest neighbor community, thus evaluating the density and separation of the clusters. The silhouette coefficient ranges from [-1, 1], with higher values ​​indicating better clustering results.

[0111] Stability measures the reliability of a partitioning result, referring to whether the community partitioning remains consistent under multiple random initializations (different random seeds). Higher stability indicates that the partitioning result can obtain similar community partitioning under different initializations, demonstrating better stability and robustness.

[0112] Normalized Mutual Information (NMI) is a metric used to assess the similarity between two partitions. Its value ranges from [0, 1], where 1 represents identical partitions and 0 represents completely different partitions. NMI evaluates the consistency of two partitions by comparing the information sharing between them.

[0113] The Variation of Information (VI) is another metric for measuring the difference between two communities. It measures the “distance” between the two partitions based on the information difference in information theory. The larger the value, the greater the difference.

[0114] In this embodiment, partitions with high modularity (Q) and silhouette score (Silhouette Score), high normalized mutual information (NMI) values ​​under different initializations, and low information difference metric (VI) are selected as the final partitioning results. This invention evaluates and optimizes candidate partitions using multi-dimensional indicators such as modularity, silhouette score, and partition stability (e.g., NMI, VI), ensuring that the final partitioning results possess high cohesion, strong separability, and good generalization ability, making them suitable for large-scale city-level ride-hailing demand forecasting.

[0115] The core of this invention's method includes: by deeply integrating road network characteristics and POI characteristics, the regions delineated by this invention exhibit high consistency in physical connectivity and functional attributes, fundamentally ensuring the "homogeneity" of ride-hailing demand within these regions. This lays a solid foundation for high-precision prediction and overcomes the limitations of traditional methods that rely solely on historical order data or simple spatial division, making the delineation results more consistent with the actual distribution of urban functional areas. The method transforms the urban spatial delineation problem into a community discovery problem within a complex network, constructing a graph structure using comprehensive similarity as edge weights. By quantifying the functional similarity between spatial units and employing optimization algorithms for intelligent merging, a reasonable prediction region is formed, effectively avoiding the subjectivity of manual delineation and significantly reducing the computational complexity of subsequent prediction models. The regions delineated by this invention can be used as input and are applicable to various time series or machine learning prediction models such as RandomForest, XGBoost, LSTM, and Transformer, demonstrating broad applicability and good compatibility.

[0116] The above embodiments are preferred embodiments of the present invention and are not intended to limit the present invention. Any changes or other equivalent substitutions made without departing from the technical solution of the present invention are included within the protection scope of the present invention.

Claims

1.A method for dividing a region for demand prediction of a ride-hailing service based on road network characteristics and POI characteristics, the method comprising: The method comprises the following steps: S1, dividing a target geographical area into a plurality of basic spatial units, and extracting road network features and POI features for each spatial unit; S2, calculating similarity feature vectors of all adjacent spatial units based on the road network features and the POI features, the similarity feature vectors comprising regional road network compactness, regional activity similarity and land use similarity; S3, determining weights of the regional road network compactness, the regional activity similarity and the land use similarity by using a CRITIC weight method, and fusing to obtain a comprehensive similarity between adjacent spatial units; S4, regarding the spatial units as nodes and regarding the comprehensive similarity as a weight of a connecting edge, constructing an undirected weighted graph, and clustering and dividing the undirected weighted graph into a plurality of communities by using a community discovery algorithm, each community corresponding to a combined online car-hailing demand prediction area; S5, obtaining a plurality of candidate partitions, and selecting an optimal partition from the plurality of candidate partitions as a final online car-hailing demand prediction area division scheme based on a modularity, a silhouette coefficient and a partition stability index. 2.The method of claim 1, wherein, In step S1, the basic spatial units are regular hexagonal grids, the road network features comprise the number of connected roads, road network density and intersection density, and the POI features comprise POI density, POI type proportion vector and land mixed use degree. 3.The method of claim 2, wherein, In step S2, the land use similarity is obtained by calculating a Jensen-Shannon divergence JSD of the POI type proportion vector of adjacent spatial units, and a calculation formula is as follows: POI(i,j)=1-JSD(I||J), wherein i and j are a pair of adjacent spatial units, I and J are POI type proportion distribution vectors of the spatial units i and j respectively, and JSD(I||J) represents the Jensen-Shannon divergence between the distributions I and J. 4.The method of claim 3, wherein, Before the Jensen-Shannon divergence is calculated, the POI type proportion vector is smoothed, and a calculation formula of the probability distribution after smoothing is as follows: wherein p(k) is an original proportion of the kth POI type, ε is a constant, and n is the number of POI types. 5.The method of claim 2, wherein, In step S2, the regional activity similarity is obtained by the following steps: S21, defining regional activity feature vectors for the spatial units i and j respectively and performing normalization processing, the regional activity feature vectors comprising road network density, intersection density, POI density and land mixed use degree; S22, calculating the Euclidean distance between the normalized regional activity feature vectors of the spatial units i and j; S23, converting the Euclidean distance into the regional activity similarity by using a Gaussian kernel function. 6.The method of claim 1, wherein, In step S3, the process of determining the weights by using the CRITIC weight method comprises: calculating information amounts of each of the three indexes of the regional road network compactness, the regional activity similarity and the land use similarity, and obtaining the weights of the indexes by normalization, and a calculation formula of the information amount is as follows: where σ j is the standard deviation of index j, represents the contrast intensity, |r jk | indicates the absolute value correlation coefficient between index j and index k, represents the sum of the conflict of index j with all other indexes. 7.The method of claim 1, wherein, In step S4, the community discovery algorithm is a Leiden algorithm, the Leiden algorithm comprises three iterative stages of local movement, community refinement and network aggregation, and a maximized modularity is taken as an objective function. 8.The method of claim 1, wherein, In step S5, the plurality of candidate regions are obtained by adjusting a resolution parameter of a community discovery algorithm and performing multiple random initializations. 9.The method of claim 1, wherein, The partition stability is evaluated by calculating a normalized mutual information NMI or a variation of information VI of the partition results under multiple random initializations. 10.The method of claim 7, wherein, The calculation formula of the modularity is: wherein A ij is the weight between spatial unit i and spatial unit j, i.e. the integrated similarity, the weighting degree of spatial unit i is k i =∑ j A ij , m = 1 / 2∑ i k i , γ is a resolution parameter, δ(c i , c j ) represents whether spatial unit i and spatial unit j are in the same community.

Images