Online and offline convenience facility site selection optimization method based on knowledge graph

Abstract

The invention discloses an online and offline convenient facility site selection optimization method based on a knowledge graph, and relates to the field of facility optimization site selection, and the method comprises the steps: generating an urban facility knowledge graph, employing a neural network to carry out the aggregation simulation of complex environment information in the urban facility knowledge graph, employing a binary logistic regression method, and carrying out the optimization of the site selection of the urban facility. According to consumption behaviors and feedback evaluation of urban residents in an online and offline retail platform, correlation between online and offline convenience facility high performance and urban environment factors is evaluated, positive and negative urban environment factors influencing high-performance convenience facility site selection are obtained, and a layout site selection rule of the facility is obtained through induction. And the layout recommendation system performs layout recommendation on the online and offline convenience facilities in a city domain range to obtain a layout recommendation result, and realizes layout optimization of the online and offline convenience facilities in a city. According to the method, the site selection rule of the high-performance facility can be scientifically extracted, and powerful support is provided for site selection of a new facility.

Description

Technical Field

[0001] The present invention relates to the field of facility optimization site selection, and in particular to an online and offline convenience facility site selection optimization method based on a knowledge graph. Background Art

[0002] At present, in the field of urban facility site selection, existing technologies mainly include traditional spatial analysis methods, multi-criteria decision-making methods (MCDM), and graph algorithms.

[0003] Traditional spatial analysis methods select sites based on factors such as geospatial data, population density, and traffic accessibility. Common models include the location-allocation model (LA model). These methods are usually based on the matching relationship between physical facility supply and social demand to optimize the layout of public facilities. Traditional site selection methods, such as those based on the matching relationship between accessibility and demand, usually ignore the subjective perception and behavioral feedback of users. Although these methods have certain advantages in spatial layout and resource allocation, they cannot take into account the actual needs and usage preferences of users for facilities. Especially in the O2O model, users' online evaluation, consumption behavior, and preference data are crucial to the actual operation of facilities.

[0004] Multi-criteria decision-making methods (MCDM) optimize facility location by setting weights and using decision matrices to consider multiple influencing factors (such as transportation, population, service scope, etc.). However, the decision-making process of these methods may be too subjective and difficult to apply in complex and changing urban environments. Traditional methods based on spatial analysis and multi-criteria decision-making often use unified standards to judge facility location, ignoring the personalized needs of different regions and different types of facilities. For example, retail facilities may pay more attention to the commercial potential of the surrounding environment, while public facilities focus on the fairness and accessibility of services. Existing technologies cannot effectively distinguish the location logic of different types of facilities, nor can they make accurate recommendations based on changes in demand in different regions.

[0005] In recent years, with the development of artificial intelligence technology, more and more studies have applied knowledge graphs and graph algorithms to the site selection of urban facilities. These methods can integrate multi-source heterogeneous data of cities (such as geographic information, socioeconomic data, facility service data, etc.) and use the structured representation of graphs to capture the correlation between facilities. Common methods include path-based recommendation systems and link prediction methods.

[0006] Many recommendation systems based on graph algorithms, such as link prediction methods, can provide site selection recommendations but lack sufficient interpretability. These methods mainly focus on the topological relationship of the graph but cannot provide detailed reasons for the recommendation, resulting in low trust in the recommendation results by decision makers. In a complex urban environment, lack of interpretable recommendations may affect decision makers' willingness to adopt.

[0007] The data sources involved in the site selection of urban facilities are numerous, including geospatial data, socioeconomic data, user evaluation data, etc. Existing site selection technologies are difficult to effectively integrate these heterogeneous data, which may result in the omission of important information or data conflicts in the decision-making process.

[0008] Therefore, there is an urgent need for an online and offline convenience facility location optimization method based on knowledge graph to solve the above problems. Summary of the invention

[0009] To solve the above problems, this application proposes a method for optimizing the location of online and offline convenience facilities based on knowledge graph, which includes the following steps: S1. Obtain multi-source heterogeneous urban data and integrate the multi-source heterogeneous urban data through data integration tools to obtain sample data, and input the sample data into the knowledge graph construction software to generate an urban facility knowledge graph, wherein the urban facility knowledge graph includes urban physical space, social activity space and residents' perception feedback; S2. Use neural networks to aggregate and simulate the complex environmental information in the urban facility knowledge graph to obtain the urban environmental factors where online and offline convenience facilities are located; S3. Using binary logistic regression method, based on the consumption behavior and feedback evaluation of urban residents in online and offline retail platforms, the high performance of online and offline convenience facilities and the urban environment in which they are located are evaluated, and the positive and negative urban environmental factors that affect the location of high-performance convenience facilities are obtained.

[0010] S4. According to the positive and negative urban environmental factors that affect high performance, the rules between facility layout and location selection and urban business district, human flow activities, and land use classification are summarized and recorded as layout and location selection rules; S5. Collaborative filtering recommendation system for layout site selection rules. The layout recommendation system recommends the layout of online and offline convenience facilities within the city area to obtain layout recommendation results. S6. Based on the layout recommendation results, optimize the layout of online and offline convenience facilities in the city.

[0011] Preferably, the neural network is a Weight Relation GCN, i.e., a WR-GCN neural network, which aggregates neighbor node information through convolution operations, aggregates node and edge information on a graph structure, incorporates the type of relationship into the message passing process, and synchronously propagates weight information between relationships.

[0012] Preferably, the mathematical expression of the WR-GCN neural network is: ; In the formula, Is a node The features after aggregating neighbor node information, R is the total number of relations, Is a node In relationship r The neighbor node set under It's a relationship r The corresponding weight matrix has the same size as the input dimension Output dimension, Neighbor node j Features, It is a slave node j To Node The edge weights of is an optional bias term, σ is the activation function, and the ReLU function is selected.

[0013] Preferably, the WR-GCN neural network convolution operation includes three layers, specifically: The first layer is to aggregate the relationships of all nodes at the physical space level; The second layer aggregates the neighbor nodes of each block onto the block, so that each block carries the information of the urban environment on which it is located; The third layer aggregates the neighboring block information of each block.

[0014] Preferably, the layout recommendation system of the layout location rule collaborative filtering recommendation system recommends the layout of online and offline convenience facilities within the city area to obtain the specific contents of the layout recommendation results including: The filter recommendation system makes site recommendations for specific facilities by comparing the similarity between potential locations and the urban locations of facilities in the layout site selection rules; When the filtering recommendation system makes “facility-location” recommendations, it uses the excellent location nodes in the learning set as a benchmark, compares them with the potential location nodes in the test set, and verifies the recommendation accuracy by comparing them with the excellent location nodes in the test set; The filtering recommendation system uses Softmax to learn and adjust the weights of each metric method, and uses BCE WithLogits Loss as the loss function to evaluate the model performance. It comprehensively gives the similarity scores between potential blocks and excellent site selection blocks, and finally selects the facility with the highest similarity as the recommendation result based on the similarity calculation results.

[0015] Preferably, the loss function expression of the filtering recommendation system is: ; in, It represents the prediction score (logits) of the recommendation model for “facility-location”, reflecting the prediction performance of the facility at each candidate location; It indicates the actual performance of the facility at a certain location, which is a binary value (0 or 1), indicating whether it is a “high-performance location”; is the total sample size, max( ) is used for calculation stability to avoid possible numerical overflow in exponential operations; is a penalty term that measures the degree of match between the predicted value and the true value, ensuring that the predicted score of the facility is higher in the true high-performance location and lower in the non-high-performance location; The predicted score is converted into a probability value of [0,1] by smoothing, and the recommended facility location results are output.

[0016] In summary, the online and offline convenience facility location optimization method based on knowledge graph of the present invention has the following beneficial effects compared with the traditional technology: 1. By constructing a knowledge graph to efficiently integrate multi-source heterogeneous data such as urban physical space, social activity space, and residents' perception feedback, the data processing capability of the decision support system is improved, making the site selection results more scientific and reasonable; 2. Using graph neural network technology, we can accurately simulate complex environmental information in the city, including residents’ perception and feedback on the use of facilities, which enhances the understanding and grasp of urban environmental factors and provides a more accurate decision-making basis for facility site selection; 3. Through neural networks, the correlation between online and offline convenience facilities and the urban environment is comprehensively evaluated, and the positive and negative urban environmental factors that affect high performance are identified, providing quantitative evaluation results for facility site selection, which helps to identify and optimize key influencing factors; 4. Combined with the collaborative filtering recommendation system, the present invention can intelligently recommend potential site selection locations for various online and offline convenience facilities, thereby improving the accuracy and efficiency of site selection recommendations and reducing human subjective errors; 5. By optimizing site selection, the present invention helps to improve the economic benefits of online and offline convenience facilities, such as increasing customer flow and sales, and provides facility operators with a site selection plan that maximizes economic benefits; 6. Taking residents’ subjective feedback into consideration, the layout of facilities is more in line with their actual needs and expectations, which improves residents’ satisfaction and loyalty to online and offline convenience facilities.

[0017] The technical method of the present invention is further described in detail below through the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A step diagram of an online and offline convenience facility location optimization method based on a knowledge graph according to the present invention; Figure 2 The factors that positively and negatively affect the excellent location of the chain stores of the present invention; Figure 3 Cosine similarity of location selection for each chain store; Figure 4 The contribution weights corresponding to the dimensions before and after the graph convolution; Figure 5 The score distribution of recommended neighborhoods is given for each chain store. DETAILED DESCRIPTION

[0019] The technical method of the present invention is further described below by means of the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of the components and steps, numerical expressions and numerical values ​​described in these embodiments do not limit the scope of the present application.

[0020] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present application, its application, or uses.

[0021] Technologies, systems, and devices known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, systems, and devices should be considered part of the specification.

[0022] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0023] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0024] Embodiment 1 like Figure 1 As shown, the present invention provides an online and offline convenience facility location optimization method based on a knowledge graph, comprising the following steps: S1. Obtain multi-source heterogeneous urban data and integrate the multi-source heterogeneous urban data through data integration tools to obtain sample data, input the sample data into the knowledge graph construction software to generate the urban facility knowledge graph, which includes urban physical space, social activity space and residents' perception feedback. It requires data collection capabilities (such as API interface, database access), data integration tools (such as ETL tools), and knowledge graph construction software (such as Apache Jena, Neo4j); Compared with traditional site selection methods, multi-source data fusion integrates a wider range of data types through knowledge graph technology, including physical space data, socioeconomic data, and user perception data, providing a more comprehensive site selection perspective.

[0025] The networked structure of the knowledge graph allows for comprehensive analysis of multi-source urban data and structuring of the relationships between various elements of the city.

[0026] S2. Use a neural network to aggregate and simulate the complex environmental information in the urban facility knowledge graph to obtain residents' perceived feedback on the use of facilities. Use a neural network to aggregate and simulate the complex environmental information in the urban facility knowledge graph to obtain the urban environmental factors of online and offline convenience facilities. This requires a graph neural network framework (such as Py Torch Geometric, DGL) and high-performance computing resources (such as GPU). The neural network will aggregate the information in the knowledge graph through convolution and output it as vector indicators of several dimensions. These indicators each represent different aspects of urban environmental information, which are referred to as urban environmental factors here.

[0027] Innovative application of graph neural networks: This invention uses graph neural networks to not only capture the spatial relationship between facilities and the environment, but also simulate the perceptual feedback of residents, which is difficult to achieve with traditional methods.

[0028] Feature aggregation capabilities of graph neural networks: Capturing complex semantic relationships involved in facility location selection and generating explanatory paths.

[0029] Preferably, the neural network is a Weight Relation GCN, i.e., a WR-GCN neural network, which aggregates neighbor node information through convolution operations, aggregates node and edge information on a graph structure, incorporates the type of relationship into the message passing process, and synchronously propagates weight information between relationships.

[0030] Preferably, the mathematical expression of the WR-GCN neural network is: ; In the formula, Is a node The features after aggregating neighbor node information, R is the total number of relations, Is a node In relationship r The neighbor node set under It's a relationship r The corresponding weight matrix has the same size as the input dimension Output dimension, Neighbor node j Features, It is a slave node j To Node The edge weights of b is an optional bias term, σ is the activation function, and the ReLU function is selected.

[0031] Preferably, the WR-GCN neural network convolution operation includes three layers, specifically: The first layer is to aggregate the relationships of all nodes at the physical space level; The second layer aggregates the neighbor nodes of each block onto the block, so that each block carries the information of the urban environment on which it is located; The third layer aggregates the neighboring block information of each block.

[0032] S3. Using the binary logistic regression method, the correlation between the high performance of online and offline convenience facilities and urban environmental factors was evaluated based on the consumption behavior and feedback evaluation of urban residents in online and offline retail platforms, and the positive and negative urban environmental factors that affect the location of high-performance convenience facilities were obtained.

[0033] Statistical analysis software (such as R and Python's Pandas library) and machine learning libraries (such as scikit-learn) are required. Feedback evaluation Here, the monthly sales volume, number of followers, comprehensive score and comment content of each convenience facility in the platform represent consumers' consumption behavior and feedback evaluation in the store.

[0034] Positive means that the factor plays a positive role in the site selection process, such as agglomeration and attraction; negative means that the factor plays a negative role in the site selection process, such as exclusion and alienation. The urban environmental factors here are all derived from the results of the aggregated output of the S2 neural network. In the S3 step, they are divided into positive or negative effects on the site selection of convenience facilities through regression analysis.

[0035] Based on the frequency of use, evaluation, and attention of residents to the facilities, some facilities are selected as high-performance facilities to evaluate the correlation between their high performance and the urban environment. This is a relationship between multivariate independent variables (several urban environmental influencing factors output by the graph neural network) and binary dependent variables (whether it is high performance), so a binary logistic regression (BLR) model is used. The mathematical expression is: ; in, , ... It is the urban environment factor where online and offline convenience facilities are located. , ... is the regression coefficient corresponding to the regression coefficient of the urban environmental factor, Indicates whether it is a high-performance site selection. The logical conversion value can be expressed as follows: ; in, The probability of high-performance site selection, Functions are used to express and conversion relationship.

[0036] Based on the regression calculation, the ROC curve and AUC value are used to evaluate whether the location selection of high-performance convenience facilities of different brands has a specific relationship and pattern with the urban environment in which they are located. Specifically, when the AUC value is higher than 0.7, it can be considered that the location selection logic of high-performance convenience facilities of the brand is consistent.

[0037] S4. Based on the positive and negative urban environmental factors that affect high performance, the rules between facility layout and location selection and urban business districts, human flow activities, and land use classification are summarized. This requires data mining tools (such as Python's MLlib) and visualization tools (such as Tableau and Python's Matplotlib library); According to the weight matrix in the convolution process of the graph neural network The corresponding relationship between the output urban environmental factors and specific urban environmental elements can be reversed and restored. At the same time, based on the regression analysis of the BLR model, the correlation between the high performance of convenient facilities and the urban environment can be obtained, and the layout and site selection rules of high-performance convenient facilities can be obtained.

[0038] In S2, a neural network is used to obtain objective and quantitative urban environmental factors; In S3, the relationship between urban environmental factors and high performance of online and offline convenience facilities is evaluated through regression analysis, and the urban environmental factors that have a positive or negative impact on the performance of each type of facility are obtained; In s4, based on the specific meanings of various urban environmental factors, the relationship between high performance of facilities and various urban environments such as urban business districts and human flow activities is summarized. The positive and negative urban environmental factors that affect the location of high-performance amenities can reflect the correlation between high performance and the urban environment and then be summarized.

[0039] S5. Collaborative filtering recommendation system for layout site selection rules The layout recommendation system recommends the layout of online and offline convenience facilities within the city to obtain the layout recommendation results, which requires a recommendation system algorithm library (such as Apache Mahout, Python's Surprise library) and a database management system; Compared with the traditional multi-criteria decision-making method, the collaborative filtering recommendation system of the present invention can provide more personalized and accurate site selection recommendations.

[0040] Preferably, the layout recommendation system of the layout location rule collaborative filtering recommendation system recommends the layout of online and offline convenience facilities within the city area to obtain the specific contents of the layout recommendation results including: The filter recommendation system makes site recommendations for specific facilities by comparing the similarity between potential locations and the urban locations of facilities in the layout site selection rules; When the filtering recommendation system makes “facility-location” recommendations, it uses the excellent location nodes in the learning set as a benchmark, compares them with the potential location nodes in the test set, and verifies the recommendation accuracy by comparing them with the excellent location nodes in the test set; The filtering recommendation system uses Softmax to learn and adjust the weights of each metric method, and uses BCE WithLogits Loss as the loss function to evaluate the model performance. It comprehensively gives the similarity scores between potential blocks and excellent site selection blocks, and finally selects the facility with the highest similarity as the recommendation result based on the similarity calculation results.

[0041] Preferably, the loss function expression of the filtering recommendation system is: ; in, It represents the prediction score (logits) of the recommendation model for “facility-location”, reflecting the prediction performance of the facility at each candidate location; It indicates the actual performance of the facility at a certain location, which is a binary value (0 or 1), indicating whether it is a “high-performance location”; is the total sample size, max( ) is used for calculation stability to avoid possible numerical overflow in exponential operations; is a penalty term that measures the degree of match between the predicted value and the true value, ensuring that the predicted score of the facility is higher in the true high-performance location and lower in the non-high-performance location; The predicted score is converted into a probability value of [0,1] by smoothing, and the recommended facility location results are output.

[0042] S6. Based on the layout recommendation results, the layout optimization of online and offline convenience facilities in the city is realized, which requires geographic information system (GIS) software (such as ArcGIS, QGIS) and decision support system.

[0043] Embodiment 2 This paper takes Nanjing as the research object, divides the research area into 9454 urban blocks according to the urban road network, and constructs the knowledge graph of urban facilities in Nanjing under the guidance of the knowledge graph ontology. The number of entities and relationship instances is shown in Tables 1 and 2. Table 1 shows the entity types and numbers (individuals) in the urban facility knowledge graph, and Table 2 shows the relationship types and number of instances (individuals) in the urban facility knowledge graph. This paper effectively organizes and integrates massive multi-source heterogeneous urban data through knowledge graphs, and represents the physical space elements and conceptual space elements related to facilities in the city in the graph.

[0044] Table 1 Entity types and numbers in the urban facilities knowledge graph ;

[0045] Table 2 Relationship types and number of instances in the urban facilities knowledge graph ;

[0046] The relationship between location and facility performance is an empirical relationship between a multivariate independent variable (several dimensions of the convolution output) and a binary dependent variable (whether it is an excellent location). Here, the Binary Logistic Regression (BLR) model is used to determine the positive and negative impact relationship between the excellent location of each brand and different factors. When calculating, first assume that the probability of an excellent location is P, and the probability of a non-excellent location is 1-P. The commonly used BLR calculation formula is as follows: ; in, , ... It is the urban environment factor where online and offline convenience facilities are located. , ... is the regression coefficient corresponding to the regression coefficient of the urban environmental factor, Indicates whether it is a high-performance site selection. The logical conversion value can be expressed as follows: ; in, The probability of high-performance site selection, Functions are used to express and conversion relationship.

[0047] ; In the formula, Y represents the logical conversion value of whether the location is advantageous; P represents the probability of the occurrence of advantageous location, and Z represents the binary response of the dependent variable (0 or 1 value), that is, Z=1.

[0048] The ROC (Receiver Operating Characteristic) curve is a graphical tool for evaluating binary classification problems. It shows the performance of the model in distinguishing between True Positive Rate (TPR) and False Positive Rate (FPR) at different classification thresholds. The AUC (Area Under the Curve) value represents the area under the ROC curve, ranging from 0 to 1. The closer the AUC value is to 1, the stronger the model's ability to distinguish. Now use these two tools to analyze the results of the BLR model, and statistically analyze the consistency of the excellent store locations of each chain store as shown in Table 3, which shows the discrimination of excellent locations of each brand chain store.

[0049] Table 3 Consistency of location selection of excellent stores in each chain ;

[0050] The AUC value reflects the complexity of different brand chain stores in site selection and the significance of influencing factors. Specifically, when the AUC value is low, it means that the existing variables have limited discrimination between the advantages and disadvantages of site selection, and cannot distinguish between potential site selection blocks and excellent performance blocks. According to the AUC value statistics, seven chain store brands such as Apple authorized stores and Watsons are above 0.8. It can be considered that the site selection location characteristics of excellent performance stores of such chain stores are relatively obvious and easy to distinguish through the urban environment selected in this study. At the same time, it shows that the performance of its stores is highly correlated with the physical and social environment of the city. The AUC values ​​of three brands such as Neptune Star are lower than 0.7 (a common model performance standard in business analysis), especially the two brands of Neptune Star and Convenience Bee, which are close to random distribution under the existing urban environmental factors. This shows that the location characteristics of excellent store site selection of such brands are not obvious enough on the knowledge graph given by the present invention, and there may be layout logic outside the location environment. In general, the comprehensive discrimination of the model for a given chain store reaches 0.77, which performs well in the field of marketing and has certain reliability and effectiveness.

[0051] The BLR model can also be used to determine the positive and negative impact relationship between the excellent location of each chain store and different factors. Figure 2 The importance of each factor to the location selection of different brands is shown. It can be seen that the impact of each factor on different brands is significantly different. For example, dimension *21 has a negative impact on most brands, while dimension *52 has a positive impact. This difference shows the guiding or restrictive role of factors in brand location selection. At the same time, different brands have different sensitivities to different factors, and the level of this influence may be a key factor in store location strategy.

[0052] Further statistics on the cosine similarity of the influencing factors of each chain store brand are as follows: Figure 3As shown in the figure, the similarities and differences between the brands in terms of location selection strategies are observed. First, it is observed that three highly similar combinations are formed among the twelve brands, such as Apple authorized stores, Xiaomi Home and Watsons, which have extremely high similarities among each other, indicating that the urban environments where the locations of the three excellent stores are located are extremely similar. Similarly, such combinations include Yipin Fresh Supermarket, Neptune Star and Yifeng combination, Xiansheng Zaikang, Baiguoyuan and Suguo Convenience combination. In addition, there are some transitional brands outside the high-similarity combination, such as the location of "Huarun Suguo and RT-Mart" has a moderate similarity relationship with the combination of "Apple authorized stores, Xiaomi Home and Watsons", and also has a moderate similarity relationship with "Yipin Fresh Supermarket, Neptune Star and Yifeng", "Xiansheng Zaikang, Baiguoyuan and Suguo Convenience", which may have adopted a certain comprehensive location selection strategy. In addition, the two combinations of "Yipin Fresh Supermarket, Neptune Star and Yifeng" and "Xiansheng Zaikang, Baiguoyuan and Suguo Convenience" also have a moderate similarity relationship with each other. This suggests that the two have some overlapping considerations when choosing locations. Finally, Convenience Bee is only moderately similar to the combination of "Innovent, Baiguoyuan and Suguo Convenience", which shows that although the remaining brands have some commonalities in site selection, the overall strategy may be more unique and specific, and it may pay special attention to certain urban environmental factors that this invention does not pay attention to.

[0053] Finally, the present invention takes into account the factor contribution weights of the convolution process, such as Figure 4As shown. That is, reverse restore the real environment information represented by each factor after convolution, and analyze the real influencing factors of each brand store location. Here, the present invention selects several positive and negative factors that have the greatest impact on each brand and reversely analyzes them. In general, high-end electronic products and daily boutique combinations such as "Apple Authorized Stores, Xiaomi Home and Watsons" are mainly concentrated in high-end consumer markets such as large shopping districts. Pay attention to regional traffic, traffic convenience and consumer purchasing power. Combinations such as "Yipin Fresh Supermarket, Neptune Star and Yifeng" that focus on community convenience and health services are mostly located in residential areas and living circles. Because their main products are daily necessities, they pay more attention to the distribution of residents and population density within the impact range. "Xiansheng Zaikang, Baiguoyuan and Suguo Convenience" are very similar to the previous combination, and also use extensive coverage of residential areas as the main strategy, but at the same time they also cover office areas and commercial areas, which can attract more types of urban residents. The transitional combination of "Huarun Suguo and RT-Mart" has the characteristics of two highly similar combinations in terms of location selection. They are often located at the junction of large commercial districts and residential areas. They focus on both the high consumption potential of commercial areas and the daily shopping needs of residential areas. Such a location selection strategy can also take into account the diversified product lines of its stores to ensure that customers with different consumption levels and needs are covered. Finally, commercial convenience stores such as "Bianlifeng" tend to open small-area, high-density stores in order to form a wide service network. This also leads to the fact that its location selection strategy does not tend to be inclined to a certain type of urban environment, or has location selection considerations beyond the physical and social environment of the city.

[0054] After analyzing the key factors of each brand's site selection, the present invention uses a site selection recommendation system to predict and recommend possible sites for each brand in the city. First, the knowledge graph of the study area is divided into two parts, north and south, as the learning set and test set respectively (see Figure 2 ), so that the urban environment in each data set is as similar as possible. The statistics of nodes in each data set are shown in Table 4. Table 4 is a statistical table of data sets (pieces).

[0055] Table 4 Dataset node statistics ;

[0056] Vector similarity is used to compare and score the excellent locations in the learning set of each chain store and the potential locations in the test set. The present invention selects the block nodes with the highest scores as the recommended locations for each chain store brand. After multiple rounds of training, the measurement method weight suitable for each chain store brand is obtained, and the corresponding block score is retained. Table 5 is the potential location recommendation index for each brand chain store when different proportions are selected as the recommended location. It can be considered that the faster the recall rate climbs, the more relevant the performance of the chain store brand store is to the urban environment.

[0057] It can be seen that the recall rates of Neptune and Yifeng climbed slowly and were always lower than those of other chain store brands. This indicates that the correlation between their store performance and the urban environment is low, showing that they may have strong market adaptability and wide flexibility in site selection. The recall rates of Watsons and Yippin Fresh climbed very quickly and could reach a relatively high value. This reflects that the marketing performance of such brands is highly dependent on the specific environment and is highly correlated with the urban environment in which they are located. The recall rate values ​​and climbing speeds of other brands are at a medium level. To a certain extent, this reflects the correlation between their store performance and the urban environment. It also shows that the operating model or service of such brands can adapt to different market environments and can find suitable locations in various site selection environments to cover a wider market.

[0058] Table 5 Potential location recommendation indicators for chain stores of various brands when different proportions are selected as recommended locations ;

[0059] The recommendation scores of each chain store brand in each block in the test set are classified according to quantiles, and the visual distribution results of the recommended locations in the city are observed as follows: Figure 5 As shown. Specifically, the recommended locations for Apple Authorized Stores, Xiaomi Home and Watsons are mainly concentrated in commercial centers and areas with high population concentration. These areas have large traffic and strong consumption power, so they are suitable for the sale of high-end products. For brands such as Convenience Bee, Suguo Convenience, Yipi Fresh, and Baiguoyuan, the recommended locations are widely distributed in residential areas and areas with convenient transportation. These brands need to cover more daily consumption needs. For large supermarkets such as RT-Mart and China Resources Suguo, the recommended locations are concentrated in large residential areas and peripheral areas of the city. These areas require a stable consumer population and a large space. A comprehensive analysis of the recommended locations of all brands shows that the commercial potential of the central urban area and major commercial areas is huge and has a strong appeal to almost all brands. Chain store brands such as convenience stores and drug stores that focus on serving the daily needs of community residents also show high recommendation scores near residential areas.

[0060] Through empirical research, the present invention finds that there are certain commonalities and differences among brands in terms of location selection strategies, brand positioning, and competitive relationships with other brands. For example, high-end electronic products and daily boutique brands (such as Apple authorized stores and Xiaomi Home) are mainly concentrated in the high-end consumer market, while community convenience and health service brands (such as Yipi Fresh Supermarket and Neptune Star) are mostly located in residential areas. These differences indicate that different brands have different focuses and strategies in site selection, and also reflect the competitive relationship and possibility between brands.

[0061] In terms of method, by combining knowledge graph and graph neural network technology, this paper successfully captures the complex relationship between urban environment and facility service performance, and uses the recommendation system to optimize the layout of 12 online and offline chain store brands in Nanjing. This method can scientifically extract the location rules of high-performance facilities and provide strong support for the location selection of new facilities.

[0062] At the same time, in terms of application, the present invention incorporates residents' evaluation feedback on facilities into the site selection and layout system, so that the site selection decision follows the logic of spatial reproduction. In addition to the traditional reliance on the physical and social environment, it also takes into account the subjective feedback of residents, thereby making the facility layout more in line with the actual needs and expectations of residents, and enhancing the rationality and humanistic care of the site selection decision.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical method of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical method of the present invention.

Claims

1. A method for optimizing the location of online and offline convenience facilities based on knowledge graph, characterized in that: The following steps are involved: S1. Obtain multi-source heterogeneous urban data and integrate the multi-source heterogeneous urban data through data integration tools to obtain sample data, and input the sample data into the knowledge graph construction software to generate an urban facility knowledge graph, wherein the urban facility knowledge graph includes urban physical space and social activity space; S2. Use a neural network to aggregate and simulate the environmental information in the urban facility knowledge graph to obtain the urban environmental factors where online and offline convenience facilities are located; S3. Using binary logistic regression method, based on the consumption behavior and feedback evaluation of urban residents in online and offline retail platforms, the correlation between high performance of online and offline convenience facilities and urban environmental factors was evaluated, and the positive and negative urban environmental factors that affect the location of high performance convenience facilities were obtained; S4. According to the positive and negative urban environmental factors that affect high performance, the rules between facility layout and location selection and urban business district, human flow activities, and land use classification are summarized and recorded as layout and location selection rules; S5. Collaborative filtering recommendation system for layout site selection rules. The layout recommendation system recommends the layout of online and offline convenience facilities within the city area to obtain layout recommendation results. S6. Based on the layout recommendation results, optimize the layout of online and offline convenience facilities in the city.

2. The method for optimizing the location of online and offline convenience facilities based on knowledge graph according to claim 1, characterized in that: The neural network is WeightRelation GCN, i.e., WR-GCN neural network. The WR-GCN neural network aggregates neighbor node information through convolution operations, aggregates node and edge information on a graph structure, incorporates the type of relationship into the message passing process, and synchronously propagates weight information between relationships.

3. The method for optimizing the location of online and offline convenience facilities based on knowledge graph according to claim 2, characterized in that: The mathematical expression of the WR-GCN neural network is: ； In the formula, Is a node The features after aggregating neighbor node information, R is the total number of relations, Is a node In relationship r The neighbor node set under It's a relationship r The corresponding weight matrix has the same dimensions as the input dimension and the output dimension. Neighbor node j Features, It is a slave node j To Node The edge weights of is an optional bias term, σ is the activation function, and the ReLU function is selected.

4. The method for optimizing the location of online and offline convenience facilities based on knowledge graph according to claim 3, characterized in that: The WR-GCN neural network convolution operation includes three layers, specifically: The first layer is to aggregate the relationships of all nodes at the physical space level; The second layer aggregates the neighbor nodes of each block onto the block, so that each block carries the information of the urban environment on which it is located; The third layer aggregates the neighboring block information of each block.

5. The method for optimizing the location of online and offline convenience facilities based on knowledge graph according to claim 4, characterized in that: The layout recommendation system recommends the layout of online and offline convenience facilities within the city area. The specific contents of the layout recommendation results include: The filter recommendation system makes site recommendations for specific facilities by comparing the similarity between potential locations and the urban locations of facilities in the layout site selection rules; When the filtering recommendation system makes "facility-location" recommendations, it uses the excellent location nodes in the learning set as a benchmark, compares them with the potential location nodes in the test set, and compares them with the excellent location nodes in the test set to verify the recommendation accuracy; The filtering recommendation system uses Softmax to learn and adjust the weights of each metric method, and uses BCE With LogitsLoss as the loss function to evaluate the model performance. It comprehensively gives the similarity scores between potential blocks and excellent location blocks, and finally selects the facility with the highest similarity as the recommendation result based on the similarity calculation results.

6. The method for optimizing the location of online and offline convenience facilities based on knowledge graph according to claim 5, characterized in that: The loss function expression of the filtering recommendation system is: ； in, It represents the prediction score of the recommended model for "facility-location", reflecting the prediction performance of the facility at each candidate location; Indicates the actual performance of the facility at a certain location, which is a binary value, indicating whether it is a "high-performance location"; is the total sample size, max( ) is used for calculation stability to avoid possible numerical overflow in exponential operations; is a penalty term that measures the degree of match between the predicted value and the true value, ensuring that the predicted score of the facility is higher in the true high-performance location and lower in the non-high-performance location; The predicted score is converted into a probability value of [0,1] by smoothing, and the recommended facility location results are output.

Images