A method for evaluating urban land use distribution based on a knowledge embedding model

Abstract

The present invention discloses a method for evaluating urban land use distribution based on a knowledge embedding model, including: obtaining urban geographical data; dividing urban geographical units in the form based on road networks; structuring heterogeneous information with triples, and constructing a semantic network integrating multi-modal geographical information according to the obtained data types; generating an embedding representation of each geographical unit by learning entity-entity relationships and entity-attribute relationships in the semantic network through a knowledge embedding model; using a label distribution learning model to map the embedding representation of each urban geographical unit to the land use distribution to evaluate the area proportion of each land use type within each geographical unit. By introducing a semantic network as a carrier for carrying POI and other external information, the present invention uses a knowledge embedding model to learn deep and complex spatial features and semantic associations inside the semantic network, so as to achieve the purpose of enhancing the input of the downstream evaluation model and improving the evaluation accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of urban land use type identification, and in particular to an urban land use distribution evaluation method based on a knowledge embedding model. Background Art

[0002] In the information age, urban economic activities and human activities influence each other, constantly reshaping the spatiotemporal relationship and functional types in the city. As one of the application problems of urban computing, the study of urban functional distribution can help understand the "people-land" relationship, assist in urban resource allocation, and evaluate existing planning schemes. It has important research value in human mobility research and urban structure theory. However, the functional types in a region are usually non-unique and mixed. Research not only needs to identify the main functional types in the region, but more importantly, it needs to evaluate the distribution ratio of each functional type in the region.

[0003] Traditional methods are limited by the quality and quantity of samples and it is difficult to accurately characterize the relationship between regions and functional types. The emergence of urban big data provides a new perspective for interpreting the distribution of urban functions. Among them, Point of Interest (POI) data has become the main data type for solving this problem because of its easy access and the characteristics of taking into account both spatial and semantic information. At present, most of the studies on exploring urban functions based on POI are based on the division of cities into geographical units. The spatial and semantic information contained in the POI in the geographical unit is represented by features and then input into the downstream machine learning model to obtain the distribution of urban functions. Therefore, how to obtain a reasonable and complete feature representation is one of the key links in solving this problem.

[0004] Early studies often used manual feature engineering to construct features. This method is relatively simple, but it relies on manually selected quantitative indicators (such as frequency, density, and importance evaluation indicators), has poor versatility, and is difficult to take into account the intrinsic associations and potential semantic information of POIs. In order to solve the above problems, researchers introduced the Representation Learning (RL) model to automatically learn the input data to obtain dense low-dimensional feature vectors. For example, some scholars used the idea of the word vector model to use Word2vec to represent the POI type, and then obtained the feature vector of the region by weighted average method, taking into account the semantic information of POI; in addition, some scholars used random walks to capture the spatial co-occurrence relationship between POIs, and used manifold learning to capture the hierarchical semantics between POI categories, while considering the spatial and semantic information of POIs. Although the above methods can retain the spatial and semantic information of POIs, the functional distribution of cities is not only related to the number and category of POIs, but also to factors such as the proximity relationship between geographical units, administrative divisions, and road connectivity. Therefore, there is still a lack of feature representation methods that can combine multi-source heterogeneous information. Summary of the Invention

[0005] In view of the deficiencies of single POI data in feature representation and the lack of depth in feature extraction by conventional representation learning models, the present invention proposes a method for evaluating urban land use distribution based on a knowledge embedding model. Considering the superiority of knowledge graphs in heterogeneous information expression in recent years, the present invention introduces a semantic network as a carrier for carrying POI and other external information, and uses a knowledge embedding model to learn the deep and complex spatial features and semantic associations within the semantic network, so as to enhance the input of the downstream evaluation model and improve the evaluation accuracy.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for evaluating urban land use distribution based on a knowledge embedding model, comprising:

[0008] Step 1: Obtain urban geographic data, where the urban geographic data includes POI data, road network data, and administrative division data, and preprocess the data;

[0009] Step 2: Divide urban geographic units in the form based on the road network;

[0010] Step 3: Structurally organize heterogeneous information in triples, and select four factors, namely POI semantic relationship, road connectivity relationship, administrative division relationship, and spatial proximity relationship, to construct a semantic network integrating multi-modal geographic information according to the types of the obtained urban geographic data;

[0011] Step 4: Generate the embedding representation of each geographic unit by learning the entity-entity relationship and entity-attribute relationship in the semantic network through a knowledge embedding model;

[0012] Step 5: Use a label distribution learning model to map the embedding representation of each urban geographic unit to the land use distribution to evaluate the area proportion of each land use type within each geographic unit.

[0013] Further, the POI data is obtained from Amap, and the POI data includes the longitude, latitude, name, and classification information of POI points; the road network data and administrative division data are both obtained from OSM.

[0014] Further, the preprocessing of the data includes:

[0015] Step 1-1: If there are blank attributes and records in the data, delete the corresponding data;

[0016] Step 1-2: If there are multiple records with exactly the same attribute items in the data, only keep one record and delete the rest.

[0017] Further, step 2 includes:

[0018] Using a division method based on a three - level road network, divide the urban space R into n urban geographical units according to the three - level main roads.

[0019] Further, step 3 includes:

[0020] Step 3 - 1: Define the triple f P constituted by the POI semantic relationship as:

[0021]

[0022] In the formula, the head entity e region in the triple represents a geographical unit; r category represents the category of the POI; the tail entity represents the quantity of the POIs of this category;

[0023] Step 3 - 2: Define the triple f S constituted by the spatial proximity relationship as:

[0024] f S =<e region ,r nearby ,e region >

[0025] In the formula, the head entity e region and the tail entity e region represent two adjacent geographical units; r nearby represents the adjacent relationship;

[0026] Step 3 - 3: Define the triple f R constituted by the road connectivity relationship as:

[0027] f R =<e region ,r contain ,e road >

[0028] In the formula, the head entity e region in the triple represents a geographical unit; the tail entity e road represents the road name; r contain represents the inclusion relationship between the geographical unit and the road;

[0029] Step 3 - 4: Define the triple f A constituted by the administrative division relationship as:

[0030] f A =<e region ,r belongsto ,e district >

[0031] In the formula, the head entity e region is the divided geographical unit, and the tail entity e district is the administrative region of the research area; r belongsto represents the belonging relationship between the geographical unit and the administrative region.

[0032] Furthermore, the step 4 includes:

[0033] Divide the triples in the semantic network into two types, entity-entity triples E and entity-attribute triples A;

[0034] Define the objective function in the knowledge embedding model as the maximum joint probability P(E,A|X) of the entity-entity triples E and the entity-attribute triples A, given the embedding representation X:

[0035]

[0036] In the formula, e h represents the set of geographical units R = {R1, R2, …, R s}, e t represents the geographical entity object, a t represents the set of values of the number corresponding to the POI category P = {P1, P2, …, P m}; P(<e h , r e , e t >|X) represents the conditional probability of the entity-entity triple; P(<e h , r a , a t >|X) represents the conditional probability of the entity-attribute triple.

[0037] Furthermore, generate the conditional probability of the entity-entity triple based on TransE:

[0038]

[0039]

[0040] In the formula, g(<e h , r e , e t >) is the scoring function for calculating the correlation between the relationship r e and the entity pair <e h , e t >; b1 represents the deviation, and L1, L2 represent the L1, L2 norms.

[0041] Furthermore, generate the conditional probability of the entity-attribute triple through the classification model:

[0042]

[0043]

[0044] Wherein, h() is the scoring function of each attribute value of the entity, f() is the activation function, is the embedding vector of the attribute value, represents a linear transformation, b2 is the bias.

[0045] Furthermore, the said step 5 includes:

[0046] Step 5-1: Input the embedding representation X of the geographical unit into a two-layer neural network to obtain the initial prediction distribution Y ～ ;

[0047] Step 5-2: Calculate the correlation matrix C between land use types through the inverse Euclidean distance, and then multiply C by the initial prediction distribution Y ～ to optimize the initial prediction distribution;

[0048] Step 5-3: Select the KL divergence as the loss function, and obtain the prediction result with the sum of 1 through the activation function softmax for the upper-layer result.

[0049] Furthermore, the calculation formula for the correlation between land use types is as follows:

[0050]

[0051] Wherein, c(l i ,l j ) is the correlation between land use types l i and l j ; n is the data volume of the training set; is the proportion of land use type l i in the kth record.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] By introducing a semantic network as the carrier for carrying POI and other external information, the present invention uses a knowledge embedding model to learn the deep and complex spatial features and semantic associations inside the semantic network (fusing multi-modal spatio-temporal semantic information), thereby enhancing the input of the downstream evaluation model and achieving the purpose of improving the accuracy of land use distribution evaluation. The present invention can predict the area proportion of each land use type. The experimental results on real urban geographical datasets illustrate the effectiveness of the method of the present invention. Description of the Drawings

[0054] Figure 1It is one of the flowcharts of a method for evaluating urban land use distribution based on a knowledge embedding model according to an embodiment of the present invention;

[0055] Figure 2 It is the second flowchart of a method for evaluating urban land use distribution based on a knowledge embedding model according to an embodiment of the present invention;

[0056] Figure 3 It is a schematic diagram for evaluating the land use distribution of geographical units based on a label distribution learning model according to an embodiment of the present invention. Detailed implementation manners

[0057] The following further explains the present invention in conjunction with the accompanying drawings and specific embodiments:

[0058] As Figure 1 , Figure 2 shown, a method for evaluating urban land use distribution based on a knowledge embedding model includes:

[0059] Step 1: Obtain urban geographical data. As an implementable manner, the data types of the present invention mainly include POI data, road network data, and administrative division data. As an implementable manner, the POI data is obtained from Amap and contains the longitude, latitude, name, and classification information of POI points; both the road network data and the administrative division data are obtained from OSM (Open street map), and the vector data of the first, second, and third-level main roads in the city and the vector data of the administrative divisions at the district (county) level are selected.

[0060] Step 2: Divide urban geographical units. The evaluation of urban land use distribution is to evaluate each geographical unit in the city. Therefore, it is first necessary to divide the geographical units. In the present invention, the geographical units adopt a division method based on the three-level road network, and the urban space R is divided into n by the third-level main roads, R = {r1, r2,..., r n};

[0061] Step 3: Construct a semantic network G integrating multi-modal geographical information. According to the types of urban geographical data obtained in the present invention, four factors affecting the distribution of land use types are considered, namely POI semantic relationship, spatial proximity relationship, road connectivity relationship, and administrative division relationship. The semantic network G integrating multi-modal geographical information is formally expressed as:

[0062] G = (f P , f S , f R , f A )

[0063] In the formula, f P represents a triple formed by the POI semantic relationship; f SRepresents a triple formed by spatial proximity relationships; f R Represents a triple formed by road connectivity relationships; f A Represents a triple formed by administrative division relationships. The specific definitions of these four influencing factors are in the following forms:

[0064] Step 3-1: Define the triple f formed by POI semantic relationships P as:

[0065]

[0066] In the formula, the head entity e in the triple region represents a geographical unit; r category represents the category of the POI; the tail entity represents the number of POIs of this category.

[0067] Step 3-2: Define the triple f formed by spatial proximity relationships S as:

[0068] f S =<e region ,r nearby ,e region >

[0069] In the formula, the head entity e region and the tail entity e region represent two adjacent geographical units; r nearby represents the adjacent relationship.

[0070] Step 3-3: Define the triple f formed by road connectivity relationships R as:

[0071] f R =<e region ,r contain ,e road >

[0072] In the formula, the head entity e in the triple region represents a geographical unit; the tail entity e road represents the road name; r contain represents the inclusion relationship between the geographical unit and the road.

[0073] Step 3-4: Define the triple f formed by administrative division relationships A as:

[0074] f A =<e region ,r belongsto ,e district >

[0075] In the formula, the head entity eregion is a divided geographical unit, and the tail entity e district is the administrative region of the research area; r belongsto represents the belonging relationship between the geographical unit and the administrative region.

[0076] Step 4: Learn the latent relationships in G through the knowledge embedding model to generate the embedding representation (vector) X of each urban geographical unit.

[0077] Specifically, the triples in G can be divided into two types, namely entity-entity triples E and entity-attribute triples A. The objective function in the knowledge embedding model is defined as the maximum joint probability of entity-entity triples E and entity-attribute triples A, given the embedding representation X:

[0078]

[0079]

[0080] In the formula, e h represents the set of geographical units R = {R1, R2,..., R s}, e t represents geographical entity objects, such as roads and administrative divisions, a t represents the set of values corresponding to the number of POI categories P = {P1, P2,..., P m}; P(<e h , r e , e t >|X) represents the conditional probability of the entity-entity triple, generated by the TransE model; P(<e h , r a , a t >|X) represents the conditional probability of the entity-attribute triple, generated by the classification model.

[0081] Specifically, the present invention generates the conditional probability of entity-entity triples based on TransE. TransE, as a simple and effective knowledge representation learning model, performs well in related tasks. TransE is a distance-based knowledge embedding model that defines the distance between the relationship and the entity in the triple as a scoring function. P(<e h , r e , e t >|X) can be expressed by the following formula:

[0082]

[0083]

[0084] In the formula, g(<e h , re , e t ) is the scoring function for calculating the correlation between relation r e and the entity pair <e h , e t >; b1 represents the bias, and L1, L2 represent the L1 and L2 norms.

[0085] Furthermore, the correlation between entities and attributes can be appropriately captured through a classification method. P(<e h , r a , a t >|X) can be expressed by the following formula:

[0086]

[0087]

[0088] In the formula, h() is the scoring function for each attribute value of the entity, f() is the activation function, such as tanh, is the embedding vector of the attribute value, represents the linear transformation, b2 is the bias.

[0089] Step 5: Evaluate the land use distribution of each urban geographical unit Use the label distribution learning model to map the embedding representation X of each urban geographical unit to the land use distribution to evaluate the area proportion of each land use type within each geographical unit. The problem of predicting the proportion of each land use type in a geographical unit can be transformed into label distribution learning (LDL). Compared with multi-label learning (MLL), LDL labels instances in a more natural way and assigns a value to each possible label.

[0090] The evaluation process is as Figure 3 shown and includes:

[0091] Step 5-1: Input the embedding representation X of the geographical unit into a two-layer neural network to obtain the initial prediction distribution Y ～ ;

[0092] Step 5-2: Calculate the correlation matrix C between land use types through the inverse Euclidean distance, and then multiply C by the initial prediction distribution Y ～ to optimize the initial prediction distribution. Specifically, the formula for calculating the correlation between land use types is as follows:

[0093]

[0094] In the formula, c(li , l j ) is the correlation between land use type l i and l j ; n is the amount of data in the training set; is the proportion of land use type l in the k-th record i .

[0095] Step 5-3: Obtain the prediction results with a sum of 1 by passing the upper-layer results through the activation function softmax.

[0096] Since the Kullback-Leibler (K-L) divergence can measure the difference between probability distributions, the KL divergence is selected as the loss function, and the K-L divergence equation is as follows:

[0097]

[0098] In the formula, y i is the true land use distribution is the predicted land use distribution.

[0099] To verify the effectiveness of the present invention, the embodiments selected the K-L divergence d kl (↓), the Chebyshev distance d cheb (↓) and the cosine similarity c sim (↑) as evaluation indicators and verified them on the geographical dataset of Jinhua City, China. Two groups of experiments were set up in this embodiment: (1) Compare the accuracy of the results evaluated by the present invention with other baseline methods (Word2vec, TF-IDF, TransE). The experimental results are shown in Table 1 below; (2) Compare the accuracy of each fusion factor on the evaluation results. The experimental results are shown in Table 2 below.

[0100] Table 1. Comparison results of the accuracy between the present invention and other methods

[0101]

[0102]

[0103] As shown in Table 1 above, the present invention achieves the best results in all three metrics. In particular, the K-L divergence of the model in the present invention is less than 0.03, and the cosine similarity is greater than 0.75, indicating that the evaluated distribution is quite close to the ground truth. In addition, the average Chebyshev distance is 0.3585, indicating that the maximum error between the evaluated distribution and the ground truth is less than 0.36. Compared with the baseline methods, since Word2vec and TF-IDF only consider the influence of POIs and ignore other factors, their accuracy is lower than that of the present invention. However, the feature vectors simply constructed by TF-IDF perform better than Word2vec, indicating that considering the significance of POI types can achieve better prediction results than considering the semantic information of POIs. At the same time, although TransE considers other factors, it has difficulties in dealing with the "many-to-many relationships" in this task if different types of relationships are not modeled separately, which further proves the necessity of the present invention.

[0104] Table 2. Influence of integrating different influencing factors on the accuracy of land use distribution evaluation results

[0105]

[0106] As shown in Table 2 above, the comparison of the land use distribution evaluation results when integrating different influencing factors is presented. The present invention considers four factors (POI semantic relationship, spatial proximity relationship, road connectivity relationship, and administrative division relationship). In order to find out whether these factors contribute to the task, the embodiments design different fusion schemes to construct semantic networks and compare their performances. The results show that: First, the best performance can be obtained by integrating the four factors, indicating that the knowledge fusion mode is better than the single mode; Second, comparing f P and f P +f S it is found that the accuracy of the latter is significantly improved, indicating that spatially adjacent regions are more likely to have similar land use distributions; Finally, comparing f P and f P +f R it is found that for the fusion of f RIt does not improve the accuracy but rather reduces it. In many tasks, such as traffic flow prediction and human mobility prediction, the model will perform better after considering road connectivity. However, in the work of the present invention, road connectivity is not positively correlated with land use distribution, but slightly negatively correlated. Road connectivity breaks through the distance limit and increases the connection between non-adjacent regions. In the past, due to inconvenient transportation, living and working places were usually close to each other. However, for two relatively distant regions (A and B) with convenient transportation, people can choose to work in region A and live in region B. It can be seen from this that strong road connectivity does not increase the similarity of land use distribution between regions, but rather makes the land use distribution more differentiated. In addition, f P The accuracy is greater than f P +f R It also shows that better results will be obtained without integrating more factors. It should be noted that before constructing the semantic network, the relevance of each selected factor to the downstream task must be verified.

[0107] In summary, the present invention introduces a semantic network as a carrier for carrying POI and other external information, and uses a knowledge embedding model to learn the deep and complex spatial features and semantic associations inside the semantic network (fusing multi-modal spatio-semantic information), so as to enhance the input of the downstream evaluation model and achieve the purpose of improving the accuracy of land use distribution evaluation. The present invention can predict the area proportion of each land use type. The experimental results on real urban geographic datasets illustrate the effectiveness of the method of the present invention.

[0108] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An urban land use distribution evaluation method based on a knowledge embedding model, characterized in that Including: Step 1: Obtain urban geographic data, which includes POI data, road network data, and administrative division data, and preprocess the data; Step 2: Divide urban geographic units in the form based on the road network; Step 3: Structurally organize heterogeneous information in triples, and select four factors, namely POI semantic relationship, road connectivity relationship, administrative division relationship, and spatial proximity relationship, to construct a semantic network integrating multi-modal geographic information according to the types of the obtained urban geographic data; Step 4: Generate the embedding representation of each geographic unit by learning the entity-entity relationship and entity-attribute relationship in the semantic network through a knowledge embedding model; Step 5: Use the label distribution learning model to map the embedding representation of each urban geographic unit to the land use distribution to evaluate the area proportion of each land use type in each geographic unit; The said Step 4 includes: Dividing the triples in the semantic network into two types, entity-entity triples E and entity-attribute triples A; Defining the objective function in the knowledge embedding model as the maximum joint probability P(E,A|X) of entity-entity triples E and entity-attribute triples A, given the embedding representation X: where, e h represents the set of geographical units \(R = \{R_1, R_2, \ldots, R s \}\), e t represents a geographical entity object, a t represents the set of values of the quantity corresponding to the POI category \(P = \{P_1, P_2, \ldots, P m \}\); \(P(<e h , r e , e t >|X)\) represents the conditional probability of the entity-entity triple; \(P(<e h , r a , a t >|X)\) represents the conditional probability of the entity-attribute triple; The conditional probability of generating entity-entity triples based on TransE: where g(<e h ,r e ,e t >) is a scoring function for calculating the correlation between the relationship r e and the entity pair <e h ,e t >; b1 represents the bias, and L1, L2 represent the L1 and L2 norms; The conditional probability of generating entity-attribute triples through a classification model: where h() is the scoring function for each attribute value of the entity, f() is the activation function, is the embedding vector of the attribute value, represents a linear transformation, and b2 is the bias.

2. The urban land use distribution evaluation method based on a knowledge embedding model according to claim 1, characterized in that The said POI data is obtained from Amap, and the POI data contains the longitude, latitude, name, and classification information of POI points; both the road network data and the administrative division data are obtained from OSM.

3. The urban land use distribution evaluation method based on the knowledge embedding model according to claim 1, characterized in that The said preprocessing of the data includes: Step 1-1: If there are blank attributes and records in the data, delete this piece of data; Step 1-2: If there are multiple records with exactly the same attribute items in the data, only keep one record and delete the rest.

4. The urban land use distribution evaluation method based on a knowledge embedding model according to claim 1, characterized in that, The said Step 2 includes: Adopting a division method based on the three-level road network, dividing the urban space R into n urban geographic units according to the three-level main roads.

5. A method for evaluating urban land use distribution based on a knowledge embedding model according to claim 1, characterized in that, The said Step 3 includes: Step 3-1: Define the triple f formed by the POI semantic relationship P as: f P = <e region ,r category ,e POIfre > In the formula, the head entity e in the triple region represents a geographical unit; r category represents the category of POI; the tail entity represents the number of POIs of this category; Step 3-2: Define the triple f formed by the spatial proximity relationship S as: f S = <e region ,r nearby ,e region > In the formula, the head entity e region and the tail entity e region represent two adjacent geographical units; r nearby represents the adjacent relationship; Step 3-3: Define the triple f R constituted by the road connectivity relationship as: f R = <e region ,r contain ,e road > In the formula, the tail entity e road represents the road name; r contain represents the inclusion relationship between the geographical unit and the road; Step 3-4: Define the triple f composed of administrative division relationships A as: f A = <e region ,r belongsto ,e district > In the formula, the tail entity e district is the administrative region of the research area; r belongsto represents the belonging relationship between the geographical unit and the administrative region.

6. The urban land use distribution evaluation method based on a knowledge embedding model according to claim 1, characterized in that The said Step 5 includes: Step 5-1: Input the embedding representation X of the geographical unit into a two-layer neural network to obtain the initial prediction distribution Y ～ ; Step 5-2: Calculate the correlation matrix C between land use types through inverse Euclidean distance, and then multiply C by the initial prediction distribution Y ～ Optimize the initial prediction distribution; Step 5-3: Select the KL divergence as the loss function, and obtain the prediction result with a sum of 1 after passing the upper-layer result through the activation function softmax.

7. A method for evaluating urban land use distribution based on a knowledge embedding model according to claim 6, characterized in that The calculation formula for the correlation between land use types is as follows: where c(l i ,l j ) is the correlation between land use types l i and l j ; n is the data volume of the training set; is the proportion of land use type l in the k-th record i of.

Images