Method for identifying urban functional areas using fusion area representation

Abstract

The application provides a city functional area identification method based on a fusion area representation, based on mobile phone positioning data and city POI data, based on Node2vec, a region multi-period space interaction feature is extracted, a GloVe model is used to extract POI features and a TF-IDF index is introduced to extract the semantic features of the region; a multi-head attention mechanism is used for the fusion of the region features to identify the functional categories. While capturing the feature relationship, the information sharing between the regions is realized, the feature redundancy is maximally reduced, and therefore the accuracy of the city functional area identification is effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of spatial information technology, and particularly relates to a city functional area identification method fusing regional representation. BACKGROUND

[0002] A city is an open and complex huge system composed of various forms of functional areas. The city functional area not only carries various functions of social and economic development, but also provides specific functions for human activities as a basic unit of city population activities, and undertakes specific travel and activity demands. Precise identification of the city functional area can help decision makers better understand the city functional structure, so as to promote the coupling of the human-earth system and the sustainable development of the city. The daily activities of human beings in the city area and the moving behaviors among the areas can potentially reflect the functional structure of the city. Therefore, from the perspective of the crowd movement, the interaction characteristics among the city areas and the semantic characteristics in the area can be explored, and through feature representation learning, a new path can be provided for the city functional area identification.

[0003] In the identification of urban functional areas, traditional methods mainly use spatial morphological features such as spectrum, texture, and radar scattering coefficient of remote sensing images, or use social and economic features reflected by POIs to identify urban functional areas. Currently, most studies begin to use social perception big data such as social media, mobile positioning, and taxi trajectories to reveal urban functions from the perspective of crowd movement patterns. For example, "Ma Shu-hong, Zhang Jun-jie, Chen Xi-fang, et al. Identification of urban functional areas using taxi time-series data [J]. Journal of Jilin University (Engineering Version): 1-10." constructs a time-series vector for the region based on trajectory data, and then uses an improved DTW algorithm and clustering algorithm to divide the urban functional attributes. "Xiao Rui, Guo Yu-xiang, Li Xing-hua. Dynamic extraction of activity semantics of urban blocks based on topic model [J]. Remote Sensing Technology and Application, 2023, 38(3): 649-661." converts taxi trajectory data into travel mode text and uses a topic model to determine the semantics of urban functional areas. Some scholars also use POI data and social perception data to identify urban functional areas. For example, "Yang Zhen-shan, Su Jin-hua, Yang Hang, et al. Fine urban functional area research based on multi-source data: A case study of Beijing [J]. Geographical Research, 2021, 40(2): 477-494." determines the urban functional area categories by weighting and fusing the quantized POI category vector and population heat. "Jing C, Hu Y, Zhang H, et al. Context-Aware Matrix Factorization for the Identification of Urban Functional Regions with POI and Taxi OD Data [J]. ISPRS International Journal of Geo-Information, 2022, 11(6): 351. DOI: 10.3390 / ijgi11060351" uses a method of constructing a decomposition feature matrix of POI data and taxi OD trajectory data and deriving a fusion matrix to fuse the data and then divide the urban functional areas through clustering.

[0004] The shortcomings of the prior art mainly include: (1) Most studies only simply extract the time-varying features of human activities in trajectory data or use traffic features to express the spatial interaction of regions, so the human activity features reflected in trajectory data are not sufficiently mined; (2) When using POI data to obtain regional semantic features, the global features and the public awareness of POI points are not considered, which may lead to misjudgment of functional area categories; (3) Some existing studies often ignore the similarity between different data when fusing multi-source data, leading to feature redundancy and difficulty in capturing the relationship between data, which needs to be further improved. SUMMARY

[0005] To this end, in view of the defects and deficiencies existing in the prior art, the present application aims to provide a city functional area recognition method fusing regional representation. The method is based on mobile positioning data and POI data, and through the Node2vec algorithm, the spatial interaction features of the region in multiple periods are deeply mined, the GloVe model is used to extract the POI features, and the TF-IDF index representing the public awareness of the POI is introduced to extract the semantic features of the region. Finally, the multi-head attention mechanism is used to realize the effective fusion of the regional features, the information sharing between regions is realized while the feature relationship is captured, and the feature redundancy is minimized, thereby effectively improving the accuracy of city functional area recognition.

[0006] The technical scheme specifically adopted by the present application to solve its technical problems is:

[0007] A city functional area recognition method fusing regional representation: based on mobile positioning data and city POI data, based on Node2vec, the spatial interaction features of the region in multiple periods are extracted, the GloVe model is used to extract the POI features, and the TF-IDF index is introduced to extract the semantic features of the region; the multi-head attention mechanism is used for the fusion of the regional features to perform functional category recognition.

[0008] Further, the Node2vec-based extraction of the spatial interaction features of the region in multiple periods comprises:

[0009] Divide the research area: use OSM data to divide the research area into road network regions;

[0010] Construct a user travel weighted directed graph: divide the user travel into several specific periods, then match the user travel OD to the road network region, and use the travel volume and travel distance to construct the user travel weighted directed graph in each specific period on weekdays and weekends;

[0011] Use the Node2vec model to obtain the representation vector of each road network region, thereby obtaining the multi-period spatial interaction features of the city region.

[0012] Further, the Node2vec-based extraction of the spatial interaction features of the region in multiple periods is specifically:

[0013] Map the city crowd activities into the travel directed graph in the mobile network space, take the city road network region as the node, and construct the crowd travel network with the flow of the city crowd between regions as the edge, and use the Node2vec algorithm to obtain the spatial interaction features of different regions in different periods:

[0014] User travel is divided into multiple distinct time periods; then, user travel origin-destination (OD) is matched to road network areas, constructing a multi-time period directed graph set {G1, G2, G3, G4, G5, G6}; the directed graph of user travel in time period t is defined as: G t =(V t E t ), where V t Let E represent the set of nodes in time period t. t E represents the set of directed edges in the graph at time t. t Defined as: Among them, R i ,R j ∈V t , representing the starting and ending nodes connected by the edge, respectively. Indicates the edge weight;

[0015] Each urban road network area is a node in a directed travel graph. The greater the travel volume between two areas, the stronger the connection, which is measured by the weight of the edges.

[0016] In time period t, from node R i To node R j Boundary rights Determined by formula (1):

[0017]

[0018] in, This indicates that during time period t, from node R i Reaching node R j Travel volume, This indicates that during this time period, from node R... i Reaching node R j The average distance of all trips, while This indicates the distance from all nodes in the graph to node R within that time period. j The total number of trips;

[0019] Based on the constructed weighted directed graph of user travel, Node2vec is introduced to perform a strategic random walk on the graph to simulate the movement trajectory of people in different areas of the city: assuming the current user was at node s in the previous step and is currently at node i, then the transition probability P(v) to node j in the next step is... j |v i The calculation formula is as follows:

[0020]

[0021] In the formula, Z is the normalization operator, and π i,j The non-normalized transition probability is calculated using the following formula:

[0022] π i,j = α pq (s,j) * w i,j (3)

[0023] w i,j in formula (1) is the edge weight between nodes, α pq (s,j) is calculated by formula (4):

[0024]

[0025] where d s,j is the shortest path between node s and node j, parameters p and q control whether the walk is in the direction of the already visited area or the area not visited; when node s coincides with node j, the shortest path is 0, and the return parameter p controls the direction inward; when node s and node j are directly connected, the shortest path is 1; when node s and node j need two hops to be connected, the shortest path is 2, and the in-out parameter q controls the walk direction outward.

[0026] Further, the POI features are extracted by using the GloVe model and the semantic features of the region are extracted by introducing the TF-IDF index, including: constructing a POI buffer zone document of a shortest path sequence, and constructing a city function corpus on this basis; training the city function corpus by using the GloVe model to obtain a POI category feature vector; introducing a TF-IDF weight calculation method to evaluate the importance of each POI category in the region, and calculating a representation vector of the region, i.e. the semantic features of the region.

[0027] Further, the POI features are extracted by using the GloVe model and the semantic features of the region are extracted by introducing the TF-IDF index, including:

[0028] First, a city function corpus is constructed by using a POI buffer zone document based on a shortest path sequence; the specific steps are as follows: a buffer zone is constructed with each POI as the center, the distances between all POI points in the buffer zone are calculated, the two POIs with the farthest distance are selected as the starting point S p and the ending point D p , and a set sequence {S p , D p} is constructed; then the remaining POI points in the buffer zone are placed in a to-be-inserted set W={P1, P2, P3, …, P n}, the POIs in W are inserted into the set sequence one by one by using a greedy algorithm, and the path distance formed after the insertion is ensured to be the shortest; the process is repeated until the to-be-inserted set is empty;

[0029] Then the GloVe model is used to train the city corpus, in which the POI buffer is regarded as a document, the category name of the third-level classification of the POI in the buffer is regarded as a word in the document, and the sentence in the document is composed of these words in the order of the shortest path sequence; the finally trained representation vector of each POI category can effectively reflect the semantic information of the region;

[0030] The POI is reclassified into several POI types, and the TF-IDF weight is introduced to calculate the importance of different POI types; finally, the calculation formula of the region representation vector is as follows:

[0031]

[0032] wherein, represents the feature vector of the ith region, type vec (P i,j ) represents the representation vector of the POI category j in the ith region, TF-IDF i,j represents the TF-IDF weight value corresponding to the POI type of category j in the ith region; N represents the number of POIs in the region.

[0033] Further, the fusion of the region features by using the multi-head attention mechanism for function category recognition comprises: inputting the spatial interaction and semantic features of each period into the model, calculating the weight through the multi-head attention mechanism, and generating the final feature representation; recognizing the urban functional area: the model outputs the function type probability of each region, and selects the highest probability as the recognition result.

[0034] Further, the fusion of the region features by using the multi-head attention mechanism for function category recognition specifically comprises:

[0035] The spatial interaction features of the region of each period and the semantic features of the region are input into the model, and the model is trained by using the labeled data of the functional area sample; through the multi-head attention mechanism, these features are integrated into a comprehensive region representation vector, so as to obtain the recognition result of the urban functional area:

[0036] Input various types of spatial interaction features and semantic features of the region;

[0037] The input vector is subjected to a series of linear transformations through the encoding layer, so as to obtain the Query value, Key value and Value value of the multi-head attention mechanism;

[0038] The weight of each input feature vector is independently calculated through each head, and multiplied by the corresponding Value value, so as to obtain the output of multiple attention heads; after splicing and linear transformation, the final feature representation is formed;

[0039] The probability of each region belonging to each functional area category is calculated by a full connection layer FCN and a Softmax function; when the probability of a certain category is the highest and higher than 50%, the region function is attributed to the category; if the probability of all categories is lower than 50%, the region is marked as a mixed functional area.

[0040] Further, it also includes preprocessing of data, specifically:

[0041] Collect and preprocess mobile positioning data and urban POI data:

[0042] Mobile positioning data preprocessing: eliminate missing data, duplicate data, invalid data, redundant data and data outside the study area; eliminate data with ping-pong switching phenomenon and drift point data; eliminate data with less than 10 records recorded by a user in a day;

[0043] Urban POI data preprocessing: eliminate data records outside the study area and duplicates; convert the POI coordinate system to the WGS84 coordinate system;

[0044] Identify the stay point and extract the travel OD data:

[0045] Stay point identification: use the ST-DBSCAN algorithm to mark the stay cluster data and eliminate noise data, take the earliest time and the latest time in the same cluster of the same user as the start time and end time of the stay point respectively, calculate the geometric center of the points in the cluster to get the latitude and longitude of the stay point; filter records with stay duration greater than 10 minutes as valid records, eliminate records with 1 stay point and more than 10 stay points;

[0046] Travel OD extraction: a segment of trajectory between two consecutive stay points of the same user is regarded as a travel OD; eliminate OD pairs with a travel distance less than 2km and a travel duration less than 15min, and finally obtain valid travel OD data.

[0047] And a city functional area identification system fusing regional representation, comprising:

[0048] A multi-period spatial interaction feature extraction module for extracting regional multi-period spatial interaction features based on Node2vec;

[0049] A semantic feature extraction module for extracting POI features using the GloVe model and introducing the TF-IDF index to extract the semantic features of the region;

[0050] And an identification module for fusing regional features using a multi-head attention mechanism to identify functional categories.

[0051] An electronic device comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the urban functional area identification method of the fused regional representation when executing the program.

[0052] Compared with the prior art, the present application and its preferred schemes have at least the following outstanding advantages: (1) The travel volume and travel distance are introduced to construct six user travel weighted directed graphs in different time periods, and based on this, the Node2vec algorithm is trained, which can effectively capture the dynamic connection between regions in different time periods, and obtain a distinctive regional representation vector, providing rich feature information for subsequent functional area identification. (2) After obtaining the POI category feature vector by using the GloVe model, the TF-IDF weight calculation method is introduced to evaluate the importance of each POI category in the region, which can extract more representative regional semantic feature vectors, thereby more accurately reflecting the actual influence of different POIs in the region and improving the accuracy of functional area identification. (3) The fusion method proposed can fully utilize the complementarity of different types of data, has higher classification accuracy, and has higher improvement in the identification accuracy of residential areas and mixed areas. BRIEF DESCRIPTION OF DRAWINGS

[0053] The present application will be further described in detail below in combination with the drawings and specific embodiments:

[0054] Figure 1 The technical framework of the urban functional area identification of the fusion regional representation of the embodiment of the present application is shown in the figure;

[0055] Figure 2 The travel directed graph example of the embodiment of the present application is shown in the figure;

[0056] Figure 3 The random walk strategy diagram of the embodiment of the present application is shown in the figure;

[0057] Figure 4 The regional feature fusion and urban functional area identification flowchart based on the multi-head attention mechanism of the embodiment of the present application is shown in the figure;

[0058] Figure 5 The identification result of the urban functional area of the embodiment of the present application is shown in the figure;

[0059] Figure 6 The confusion matrix of the four comparative experiments of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0060] In order to make the features and advantages of the present patent more obvious and easy to understand, the following embodiments are specifically described as follows:

[0061] It should be noted that the following detailed description is illustrative in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0062] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, they refer to the presence of a feature, step, operation, device, component, and / or combinations thereof.

[0063] The embodiment of the application is based on mobile positioning data, urban POI data and OSM road network data, uses Node2vec algorithm and GloVe model to extract multi-period spatial interaction features of the region and semantic features of the region respectively, and uses a multi-head attention mechanism to fully integrate these features for functional category identification. The main implementation process includes:

[0064] (1) Collect and pre-process mobile positioning data and urban POI data;

[0065] (2) Identify the stay point and extract the trip OD data;

[0066] (3) Extract multi-period spatial interaction features of the region based on Node2vec;

[0067] (4) Extract semantic features of the region based on GloVe;

[0068] (5) Identify the urban functional area by fusing the multi-head attention mechanism;

[0069] The technical solutions of the embodiments of the application are described in detail below.

[0070] The purpose of the application is to propose a method for identifying urban functional areas by fusing regional representation. The basic framework of the model is shown in Figure 1 , which includes the following steps:

[0071] First, collect and pre-process mobile positioning data and urban POI data;

[0072] Mobile positioning data preprocessing: eliminate missing data, duplicate data, invalid data, redundant data and data outside the study area; eliminate data with ping-pong switching phenomenon and drift point data; eliminate data with too few records in a day (less than 10). A single track data is shown in Table 1.(2) Urban POI data preprocessing: eliminate data records outside the study area and duplicates; convert the POI coordinate system to the WGS84 coordinate system.

[0073] Table 1 Single trajectory data example

[0074]

[0075]

[0076] Second step, identify the stay point and extract the trip OD data;

[0077] (1) Stay point identification: ST-DBSCAN algorithm is used to mark out stay cluster data and eliminate noise data. The earliest time and the latest time in the same cluster of the same user are taken as the start time and end time of the stay point respectively. The geometric center of the points in the cluster is calculated to obtain the latitude and longitude of the stay point. The records with stay duration longer than 10 min are selected as valid records. The records with stay point number equal to 1 and too many stay points (more than 10) are eliminated. (2) Trip OD extraction: the trajectory between two consecutive stay points of the same user is regarded as a trip OD. The OD pairs with too short distance (less than 2 km) and too short duration (less than 15 min) are eliminated. Finally, the valid trip OD data is obtained.

[0078] Third step, extract regional multi-period spatial interaction features based on Node2vec;

[0079] (1) Divide the study area: use OSM data to divide the study area into road network regions, which are used as the urban unit regions for subsequent experimental analysis; (2) Construct the user trip weighted directed graph: divide the user trip into six specific periods, and then match the user trip OD to the road network region. Introduce the trip volume and trip distance to construct the user trip weighted directed graph in six specific periods of weekdays and weekends; (3) Use Node2vec model to obtain the representation vector of each road network region, so as to obtain the multi-period spatial interaction features of the urban region.

[0080] Fourth step, extract regional semantic features based on GloVe;

[0081] (1) Construct the POI buffer zone document of the shortest path sequence, and construct the urban function corpus on this basis; (2) Use GloVe model to train the urban function corpus to obtain the POI category feature vector; (3) Introduce the TF-IDF weight calculation method to evaluate the importance of each POI category in the region, and calculate the representation vector of the region, i.e. the semantic features of the region.

[0082] Fifth step, city functional area identification based on multi-head attention mechanism;

[0083] (1) Fusion regional features: input the spatial interaction and semantic features of each period into the model, calculate the weight through the multi-head attention mechanism, and generate the final feature representation; (2) Identify urban functional areas: the model outputs the probability of the functional type of each area, and selects the highest probability as the identification result.

[0084] In the above steps, the extraction of regional multi-period spatial interaction features based on Node2vec, the extraction of regional semantic features based on GloVe, and the identification of urban functional areas by fusing multi-head attention mechanism are the focus of the present application, which will be discussed in detail in the following subsections.(1) Extraction of regional multi-period spatial interaction features based on Node2vec.

[0085] The activities of urban population in real physical space can be mapped into the travel directed graph in mobile network space. Based on this, the present application constructs a population travel network with urban road network areas as nodes and the flow of urban population between areas as edges, and uses the Node2vec algorithm to explore the spatial interaction features of different areas at different times.

[0086] (a) Constructing a user travel weighted directed graph

[0087] The user travel is divided into six specific time periods, which are: morning peak (7:00-9:00), morning working hours (9:00-11:00), lunch time (11:00-13:00), afternoon working hours (13:00-17:00), evening peak (17:00-19:00), and evening entertainment time (19:00-23:00). Then the user travel OD is matched to the road network area (node) to construct a multi-period user travel directed graph set {G1, G2, G3, G4, G5, G6}. The user travel directed graph at time t is defined as: G t = (V t , E t ), where V t is the set of nodes at time t, and E t is the set of directed edges of the graph at this period. E t is defined as: where R i , R j ∈ V t are the starting and ending nodes connected by the edge, respectively, represents the edge weight.

[0088] Each urban road network area is a node of the travel directed graph, and the direction and quantity of user travel between nodes are represented by a single arrow and a number. The thicker the arrow line, the larger the travel quantity between the two areas, and the stronger the connection. This connection strength is generally measured by the weight of the edge. An example of the travel directed graph is shown in Figure 2as shown.

[0089] Considering that most people have fixed travel routes and prefer to choose shorter paths, the edge weight from node R i to node R j in the tth time period can be determined by formula (1):

[0090]

[0091] wherein, represents the travel volume from node R i to node R j in the tth time period, represents the average distance of all travels from node R i to node R 的 in the tth time period, and represents the total travel volume from all nodes in the graph to node R j in the tth time period.

[0092] (b) Node2vec-based regional vector representation

[0093] Based on the constructed user travel weighted directed graph, Node2vec is introduced to perform a strategic random walk on the graph to simulate the movement trajectory of people in various regions of the city, and the random walk strategy is as shown in Figure 3 . Assuming that the current user is at node s in the last step and is at node i at present, the transition probability P(v j |v i ) to j in the next step is calculated as follows:

[0094]

[0095] In the formula, Z is a normalization operator, and π i,j is a non-normalized transition probability, which is calculated as follows:

[0096] π i,j = α pq (s,j) * w i,j (3)

[0097] In the formula, w i,j is the edge weight between nodes calculated by formula (1), and α pq (s,j) is obtained by formula (4):

[0098]

[0099] wherein d s,j ​is the shortest path between node s and node j, parameters p and q control the direction of the walk, whether it is to the area that has been gone or to the area that has not been gone. When node s coincides with node j, the shortest path is 0, and the return parameter p controls the direction inward; when node s and node j are directly connected, the shortest path is 1; when node s and node j need two hops to communicate, the shortest path is 2, and the in-out parameter q controls the direction of the walk.

[0100] (2) Extract regional semantic features based on GloVe.

[0101] Human activities are usually centered around POIs, and the distribution of POIs in a city contains a large amount of semantic information. Based on this, the present application uses the GloVe model to explore the semantic features of different regions in the city.

[0102] (a) Constructing a city function corpus

[0103] In order to reflect the distribution characteristics of POIs in geographical space and maintain the positional relationship between adjacent POIs, the present application uses a POI buffer zone document based on the shortest path sequence to construct a city function corpus. The specific steps are as follows: a buffer zone is constructed around each POI, the distances between all POI points in the buffer zone are calculated, the two POIs with the farthest distance are selected as the starting point S p and the ending point D p , and a set sequence {S p , D p} is constructed. Then the remaining POI points in the buffer zone are placed in the to-be-inserted set W = {P1, P2, P3, …, P n}, and the POIs in W are inserted into the set sequence one by one using a greedy algorithm to ensure that the path distance formed after insertion is the shortest. Repeat this process until the to-be-inserted set is empty.

[0104] (b) POI feature vector representation based on GloVe

[0105] The GloVe model is used to train the city corpus, and in this process, the POI buffer zone is regarded as a document, the categories of POIs in the buffer zone (the category name of the third-level classification of POIs) are regarded as words in the document, and the sentences in the document are composed of these words according to the shortest path sequence. The final training of each POI category representation vector can effectively reflect the semantic information of the region, providing support for subsequent calculation of regional semantic features.

[0106] (c) Calculate regional representation vector

[0107] To reduce the impact of the difference in the number of different categories of POIs, the present application reclassifies POIs into cognitive and density types, as shown in Table 2, a total of 16 POI types, and introduces a TF-IDF weight to calculate the importance of different POI types. The final calculation formula of the regional representation vector is as follows:

[0108]

[0109] wherein, represents the feature vector of the i-th region, type vec (P i,j ) represents the representation vector of the POI category j in the i-th region, TF-IDF i,j represents the TF-IDF weight value corresponding to the POI type of category j in the i-th region. N represents the number of POIs in the region.

[0110] Table 2 POI cognitive density type classification system

[0111]

[0112] (3) City functional area recognition with multi-head attention mechanism

[0113] The spatial interaction features of the region in each time period and the semantic features of the region are taken as the input of the model, and the model is trained using the labeled data of the functional area samples. Through the multi-head attention mechanism, these features are integrated into a comprehensive regional representation vector, thereby obtaining the recognition result of the city functional area. The specific workflow is shown in Figure 4 .

[0114] (a) Input 12 spatial interaction features (6 for weekdays and 6 for weekends) and the semantic features of the region.

[0115] (b) Perform a series of linear transformations on the input vector through the encoding layer to obtain the Query value, Key value and Value value of the multi-head attention mechanism.

[0116] (c) Independently calculate the weight of each input feature vector and multiply it with the corresponding Value value to obtain the output of multiple attention heads. After splicing and linear transformation, the final feature representation is formed.

[0117] (d) Calculate the probability of each region belonging to each functional area category through the fully connected layer (FCN) and the Softmax function. When the probability of a category is the highest and higher than 50%, the function of the region belongs to the category; if the probability of all categories is lower than 50%, the region is marked as a mixed functional area.

[0118] The efficacy of the solution provided by the present application is further demonstrated and explained below through specific test examples:

[0119] The present application is based on mobile positioning data and POI data of a certain large city in China, and adopts a regional representation fusion method to identify the city functional area. The identification result of the city functional area is shown in Figure 5 The final six types of functional areas are obtained, including residential, industrial, commercial, public management and public service, leisure and sightseeing, and mixed. For the classification result shown in Figure 5 30 regions of each type of functional area are randomly selected for testing, and the classification result is compared with the real functional area category obtained from Gaode map to verify the accuracy of the classification result. The confusion matrix result is shown in Table 3, the overall accuracy OA is 81%, and the value of Kappa coefficient is 0.77, indicating that the identification result has strong consistency with the real area function.

[0120] Table 3 Classification confusion matrix

[0121]

[0122]

[0123] To further verify the effectiveness of the method proposed by the present application, two commonly used methods in the field are selected for comparative analysis, and the accuracy results corresponding to each method are shown in Table 4. The results show that the kappa coefficient and OA value of the method of the present application are higher than those of the DTW_KNN-based method and the Word2Vec-based method. The DTW_KNN method mainly depends on the interaction between the time series vector to express the regions, but ignores the global influence of the region, and the functional area division result obtained by clustering only with the flow characteristics is difficult to distinguish the regions with complex functional structure. Although the Word2Vec method considers the number of POIs and the spatial distribution, and represents the semantic characteristics of different regions through high-dimensional vectors, it obtains better classification results than the DTW_KNN method, but it does not consider the public awareness of POI and the global characteristics, so the classification result is still not accurate enough. The method proposed by the present application improves the above shortcomings, and considers both the global influence and the local characteristics of the region when discussing the spatial interaction; in the semantic feature analysis, both the global distribution structure of POI and the public awareness are considered, so a more accurate classification result is obtained.

[0124] Table 4 Accuracy comparison of different methods

[0125] Method Kappa OA (%) DTW_KNN 0.41 0.51 Word2Vec 0.51 0.60 The method of the invention 0.77 0.81

[0126] Since the method of the present application is composed of multiple parts, ablation experiments are performed to verify the influence of different data on functional area identification and the effect of feature fusion method respectively. The first type of experiment uses single source data for functional area identification, respectively using the method based on Node2vec algorithm and mobile positioning data (method 1), and the method based on GloVe model and POI data (method 2). The second type of experiment uses a traditional weighted fusion method instead of a multi-head attention mechanism for urban functional area identification (method 3). The confusion matrix of the comparative experiment of the four methods is shown in Table 1. Figure 6 As shown in Table 1, the results show that the recognition accuracy of the method of the present application is the highest, which is 81.11%, followed by method 2 (75%), method 3 (67.78%), and method 1 (63.33%) with the lowest accuracy. Among them, method 2 performs well in identifying industrial areas and leisure and sightseeing areas, and only performs poorly in identifying residential areas and mixed areas, but is still better than methods 1 and 3, and this method also performs better than Word2Vec which also uses word embedding technology; method 1 performs well in identifying except commercial areas and leisure and sightseeing areas, and the accuracy of the remaining categories is relatively low, especially in identifying residential areas and mixed areas, but the recognition accuracy is still higher than that of the DTW_KNN algorithm; method 3 uses two kinds of data, which has a significant improvement in identifying industrial areas and mixed areas compared with method 1, a small degree of improvement in identifying residential areas, but the accuracy of identifying commercial areas is reduced, and the recognition accuracy is not as good as method 2 which uses one kind of data. Compared with the above three comparison methods, the recognition accuracy of the method of the present application in identifying residential areas and mixed areas is greatly improved, and the effect in identifying the remaining categories is also good.

[0127] In summary, the above experiments show that each link of feature extraction in the method of the present application is necessary, and the effect of fully fusing the spatial interaction features of multi-scale regions and the semantic features of the regions for urban functional area identification is better than that of single data source and simple fusion method. By fully fusing two types of data, the shortcomings of single data can be made up, and higher recognition accuracy can be achieved.

[0128] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0129] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0130] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0131] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0132] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the present application shall have the common meaning understood by one of ordinary skill in the art to which the present application pertains. The terms "first", "second", and similar terms used in the present application do not denote any order, quantity, or importance, but are used to distinguish different components. The terms "include", "contain", and similar terms mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms "connected" or "coupled" and similar terms do not limit to physical or mechanical connections or couplings, but can include electrical connections or couplings, whether direct or indirect. The terms "upper", "lower", "left", "right", and the like are used only to indicate relative positional relationships when the absolute positions of the described objects are changed, and the relative positional relationships can also be changed accordingly.

[0133] The above merely describes preferred embodiments of the present application, but is not intended to limit the present application to other forms. Any person skilled in the art can make changes or modifications to the above disclosed technical contents to obtain equivalent embodiments. However, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the present application and according to the technical essence of the present application shall still fall within the protection scope of the present application.

[0134] The present application is not limited to the above preferred embodiments, and any person can derive other various forms of a fusion area representation method for identifying urban functional areas under the inspiration of the present application. Any equivalent change and modification made within the scope of the present application shall fall within the scope of the present application.

Claims

1. A method for identifying urban functional areas from a fusion of region representations, characterized in that: Based on mobile positioning data and urban POI data, based on Node2vec, the spatial interaction features of the region in multiple periods are extracted, the POI features are extracted by using the GloVe model, and the semantic features of the region are extracted by introducing the TF-IDF index; The multi-head attention mechanism is used for the fusion of regional features to identify the functional categories; the Node2vec-based extraction of the spatial interaction features of the region in multiple periods comprises: Divide the research area: divide the research area into road network regions by using OSM data; Construct a user travel weighted directed graph: divide the user travel into several specific periods, then match the user travel OD to the road network region, and construct the user travel weighted directed graph in each specific period on weekdays and weekends by using the travel volume and travel distance; Specifically: Map the urban population activities into the travel directed graph in the mobile network space, take the urban road network region as a node, and take the flow of the urban population between regions as an edge to construct the population travel network, and use the Node2vec algorithm to obtain the spatial interaction features of different regions in different periods: The user trip is divided into a plurality of different specific time periods; then the user trip OD is matched to the road network area, and a multi-period user trip directed graph set {G1, G2, G3, G4, G5, G6} is constructed; the user trip directed graph of the t period is defined as: Gt t = (Vt t , Et t ), wherein Vt t represents the set of nodes in the t period, Et t represents the directed edge set of the graph in the t period; Et t is defined as: wherein R i , R j ∈ Vt t respectively represent the starting and ending nodes connected by the edge, represents the edge weight; Each urban road network region is a node of the travel directed graph, and the greater the travel volume between two regions, the stronger the connection, which is measured by the weight of the edge; In time period t, from node R i To node R j Boundary rights Determined by formula (1): wherein, denotes the amount of trips from node R i to node R j in the t-th time period, denotes the average distance of all trips from node R i to node R j in the t-th time period, and denotes the sum of the amount of trips from all nodes in the graph to node R j in the t-th time period. The Node2vec model is used to obtain the representation vector of each road network region, so as to obtain the multi-period spatial interaction features of the urban region; The GloVe model is used to extract the POI features and introduce the TF-IDF index to extract the semantic features of the region, comprising: First, a POI buffer zone document based on the shortest path sequence is used to construct a city function corpus; then, the city corpus is trained by using the GloVe model, and finally, the representation vector of each POI category obtained by training can effectively reflect the semantic information of the region; POIs are reclassified into several POI types, and the importance of different POI types is calculated by introducing the TF-IDF weight; finally, the calculation formula of the region representation vector is as follows: wherein, represents the feature vector of the kth region, type vec (P k,m ) represents the representation vector of the POI class m in the kth region, TF-IDF k，m represents the TF-IDF weight value corresponding to the POI type of class m in the kth region; N represents the number of POIs in the region. 2.The method of claim 1, wherein: The Node2vec-based extraction of the spatial interaction features of the region in multiple periods specifically comprises: Based on the constructed user travel weighted directed graph, Node2vec is introduced to perform a strategic random walk on the graph to simulate the movement trajectory of people in various regions of the city: assuming that the current user is at node s in the last step and is at node i at present, then the transition probability P(v j |v i ) to j in the next step is calculated as follows: where Z is a normalization operator, π i,j is the non-normalized transition probability, and is calculated as follows: π i,j = a pq (s, j) * w i,j (3) where w i,j is the edge weight between nodes calculated by formula (1), and α pq (s, j) is calculated by formula (4): where d s,j is the shortest path between node s and node j, parameters p and q control the direction of the walk, whether it is towards already visited areas or towards unvisited areas; when node s coincides with node j, the shortest path is 0, and the return parameter p controls the direction of the walk towards the inside; when node s and node j are directly connected, the shortest path is 1 ; when node s and node j need two hops to be connected, the shortest path is 2, and the in-out parameter q controls the direction of the walk towards the outside. 3.The method of claim 1, wherein: The GloVe model is used to extract the POI features and introduce the TF-IDF index to extract the semantic features of the region, comprising: constructing a POI buffer zone document based on the shortest path sequence, and constructing a city function corpus based on the same; the city function corpus is trained by using the GloVe model to obtain the POI category feature vector; the importance of each POI category in the region is evaluated by introducing the TF-IDF weight calculation method, and the representation vector of the region, i.e. the semantic features of the region, is calculated.

4. The method of claim 3, wherein: In the process of extracting the POI features by using the GloVe model and introducing the TF-IDF index to extract the semantic features of the region: The specific steps of constructing the city function corpus by using the POI buffer zone document based on the shortest path sequence are as follows: a buffer zone is constructed with each POI as the center, the distances between all POI points in the buffer zone are calculated, the two POIs with the farthest distance are selected as the starting point S p and the ending point D p , and a set sequence {S p ,D p} is constructed; then the remaining POI points in the buffer zone are placed in the set to be inserted W=P1, P2, P3, …, P n}, and the POIs in W are inserted into the set sequence one by one by using the greedy algorithm, so that the path distance formed after the insertion is the shortest; the process is repeated until the set to be inserted is empty; In the process of training the city corpus by using the GloVe model: the POI buffer zone is regarded as a document, the category names of the third-level classification of POIs in the buffer zone are regarded as words in the document, and the sentences in the document are composed of these words in the shortest path sequence.

5. The method of claim 1, wherein: The fusion of regional features by using the multi-head attention mechanism for function category recognition comprises: inputting spatial interaction and semantic features of each time period into a model, calculating weights by using the multi-head attention mechanism, and generating a final feature representation; and recognizing a city function area: the model outputs a function type probability of each area, and selects the highest probability as a recognition result.

6. The method of claim 5, wherein: The fusion of regional features by using the multi-head attention mechanism for function category recognition specifically comprises: inputting spatial interaction features of each time period and semantic features of each area into a model, training the model by using labeled data of a function area sample, integrating the features into a comprehensive area representation vector by using the multi-head attention mechanism, and obtaining a recognition result of the city function area: inputting various spatial interaction features and semantic features of each area; performing a series of linear transformations on the input vector by using an encoding layer to obtain Query values, Key values and Value values of the multi-head attention mechanism; independently calculating weights of each input feature vector by using each head and multiplying the weights with corresponding Value values to obtain outputs of multiple attention heads; and forming a final feature representation by using splicing and linear transformation; calculating probabilities of each area belonging to each function area category by using a fully connected layer FCN and a Softmax function; when a probability of a certain category is the highest and is higher than 50%, the area function belongs to the category; and if probabilities of all categories are lower than 50%, the area is marked as a mixed function area.

7. The method of claim 1, wherein: Further comprising pre-processing of data, specifically comprising: collecting and pre-processing mobile phone positioning data and city POI data: mobile phone positioning data pre-processing: eliminating missing data, duplicate data, invalid data, redundant data and data outside the study area; eliminating data with ping-pong switching phenomenon and drift point data; and eliminating data with less than 10 records recorded by a user within one day; city POI data pre-processing: eliminating data records outside the study area and duplicate data records; and converting a POI coordinate system into a WGS84 coordinate system; recognizing a stay point and extracting trip OD data: stay point recognition: marking stay cluster data and eliminating noise data by using an ST-DBSCAN algorithm, taking an earliest time and a latest time in the same cluster of a same user as a start time and an end time of the stay point respectively, calculating a geometric center of points in the cluster to obtain longitude and latitude of the stay point, and selecting records with a stay duration greater than 10 minutes as valid records, and eliminating records with one stay point and records with more than 10 stay points; trip OD extraction: regarding a section of a trajectory between two consecutive stay points of a same user as a trip OD; eliminating OD pairs with a trip distance less than 2 km and a trip duration less than 15 minutes; and finally obtaining valid trip OD data.

8. A system for urban functional area identification using a fusion of region representations, for implementing the method according to any one of claims 1 to 7, characterized in that, comprise: a multi-period spatial interaction feature extraction module configured to extract regional multi-period spatial interaction features based on Node2vec; a semantic feature extraction module configured to extract POI features by using a GloVe model and introduce a TF-IDF index to extract semantic features of an area; and And a recognition module is used to fuse regional features by using a multi-head attention mechanism to perform functional category recognition.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the urban functional area recognition method of the fused regional representation when executing the program.

Images