A method for selecting locations based on the conversion of pedestrian traffic in a street block, based on machine learning algorithms.

By using machine learning algorithms to analyze street traffic flow maps and reverse customer acquisition paths, the problem of difficulty in assessing the customer flow conversion potential of candidate locations under complex street traffic flow patterns has been solved, enabling more accurate site selection recommendations.

CN123118167APending Publication Date: 2026-11-17ZHEJIANG KESHU STORE TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611184562.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-06
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to accurately assess the effective customer flow conversion potential of candidate locations in scenarios such as open blocks, multi-entrance complex circulation routes, dense corner nodes, and frequent changes in consumer travel intentions. Traditional statistical units are also inadequate for expressing path directionality, upstream customer flow contribution, and dynamic transition relationships.

Method used

Based on machine learning algorithms, a street operation flow map is constructed. Candidate locations are used as reverse absorption and convergence nodes to search for reverse customer acquisition paths, calculate the customer acquisition contribution value of the path, generate location flow characteristics, and generate conversion potential energy fields through customer flow capture values ​​and intent tags, and finally generate a site selection recommendation list.

Benefits of technology

It enables precise quantification of the direction of passenger flow and customer acquisition contribution of candidate sites, improving the reliability and accuracy of site selection recommendations and reflecting the real potential for passenger flow dissemination and conversion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118167A_ABST
    Figure CN123118167A_ABST
Patent Text Reader

Abstract

This invention discloses a method for selecting street-level locations based on machine learning algorithms, belonging to the field of smart location selection technology. The method includes: calculating passenger flow capture values ​​based on location movement characteristics, mapping these values ​​to a location spatial grid to generate a location conversion potential energy field, generating passenger flow intention labels based on continuous trajectory changes, assembling these labels in chronological order to form an intention transition chain, and dynamically correcting the spatial potential energy value based on forward and reverse transitions to generate effective passenger flow characteristics; generating brand location characteristics based on the effective passenger flow characteristics, fusing brand location characteristics, location movement characteristics, and basic location characteristics to generate location prediction characteristics, and inputting these prediction characteristics into a location selection learning model for conversion rate regression and risk classification to generate passenger flow conversion values. This invention improves the accuracy of expressing the conversion potential of candidate locations and the reliability of location recommendations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent site selection technology, and in particular to a street-level passenger flow conversion site selection method based on machine learning algorithms. Background Technology

[0002] With the development of digital operation of commercial districts, mobile trajectory perception and machine learning modeling technologies, brand site selection is gradually shifting from experience-based judgment to data-driven analysis. Existing related technologies usually build evaluation index systems based on data such as business district popularity, store attributes, customer flow statistics, competitor distribution and consumer profiles, and predict the operating potential of candidate stores through regression models, classification models or comprehensive scoring models, thereby providing quantitative basis for brand store expansion, street business format adjustment and investment promotion decisions.

[0003] In the aforementioned applications of technology, existing methods mostly use static passenger flow scale or regional average popularity as the basis for site selection evaluation, and divide the street space into several statistical units before conducting passenger flow aggregation analysis. In scenarios such as open streets, multi-entrance complex circulation routes, dense corner nodes, and frequent changes in consumer travel intentions, this approach is prone to the problem of difficulty in distinguishing the contribution of the same passenger flow in different paths, directions, and conversion stages. This is because traditional statistical units usually cannot simultaneously express the directionality of street paths, the reverse customer acquisition contribution of upstream passenger flow to candidate locations, and the dynamic transition relationship of intention from weak to strong or from strong to weak in continuous trajectories. Consequently, the assessment of the conversion potential of candidate locations does not adequately depict the actual operating circulation and effective passenger flow capture capabilities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a site selection method for street passenger flow conversion based on machine learning algorithms to solve the problem that it is difficult to accurately evaluate the effective passenger flow conversion potential of candidate locations under complex street traffic patterns.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a method for site selection based on machine learning algorithms for converting street traffic flow. The method includes: cleaning and encoding the brand data and target street data to be selected, extracting candidate locations, and generating initial site selection data; extracting street objects and street paths from the initial site selection data, mapping street objects to movement diagram nodes, and mapping street paths to directed movement diagram edges to generate a street operation movement diagram; using candidate locations as reverse absorption and convergence nodes, searching for reverse customer acquisition paths along the reverse direction of traffic flow propagation; calculating the path customer acquisition contribution value based on the node conversion probability; and then, based on the path customer acquisition contribution value, reinjecting the contribution of movement diagram nodes in the reverse customer acquisition path to generate location movement characteristics; and based on the location movement characteristics... The system calculates passenger flow capture values ​​and maps them to a location spatial grid, generating a location conversion potential energy field. Based on continuous trajectory changes, it generates passenger flow intent labels, which are then assembled in chronological order to form an intent transition chain. The spatial potential energy value is dynamically adjusted based on forward and reverse transitions to generate effective passenger flow characteristics. Brand location characteristics are generated based on these effective passenger flow characteristics. These characteristics, along with location flow features and basic location features, are fused to generate location prediction features. These prediction features are then input into a site selection learning model for conversion rate regression and risk classification, generating passenger flow conversion values. Finally, a location selection score is calculated based on the passenger flow conversion value. Candidate locations are ranked according to their scores, generating a recommended site selection list.

[0008] As a preferred embodiment of the street-level customer flow conversion site selection method based on machine learning algorithms described in this invention, the specific steps for cleaning and encoding the brand data and target street data to be selected, extracting candidate locations, and generating initial site selection data are as follows:

[0009] Obtain the brand data and target street data for the proposed site selection. Remove invalid records and invalid shop records from the brand data and target street data respectively. Convert the cleaned brand data and target street data into brand constraint codes and shop codes.

[0010] The store codes are matched and filtered according to the brand constraint code. The store records that meet the brand constraint code are identified as candidate locations. The candidate locations are then numbered and spatially associated to generate initial site selection data.

[0011] As a preferred embodiment of the street-level customer flow conversion site selection method based on machine learning algorithms described in this invention, the specific steps for extracting street objects and street paths from initial site selection data, mapping street objects to movement map nodes, and mapping street paths to directed movement map edges to generate a street-level business movement map are as follows:

[0012] Street objects are extracted from the initial site selection data, and the street objects are segmented into nodes according to the location of the passenger flow, the location of the passenger flow turning point, and the location of the business entrance. The segmented street objects are mapped to nodes of the flow diagram. Street paths are extracted from the initial site selection data, and the street paths are processed into edges according to the direction of passenger flow. The edge-directed street paths are mapped to edges of the directed flow diagram.

[0013] Hampel local median filtering is applied to the passenger flow records corresponding to the directed flow line graph edges to remove abnormal flow records that deviate from the local median by more than the median deviation threshold. The flow intensity of the directed flow line graph edges is then corrected based on the filtered passenger flow records to generate stable flow line edges.

[0014] Node indexes are created based on the edges of the directed traffic flow graph, edge connection indexes are created based on the stable traffic flow edges, and the node conversion probability of the corresponding directed traffic flow graph edge is calculated based on the outgoing edge traffic intensity of each traffic flow graph node to generate the street operation traffic flow graph.

[0015] As a preferred embodiment of the street-level customer flow conversion site selection method based on machine learning algorithms described in this invention, the specific steps of searching for reverse customer acquisition paths along the reverse direction of customer flow propagation, using candidate locations as reverse absorption and convergence nodes, are as follows:

[0016] The candidate locations are mapped to the corresponding traffic flow nodes in the street operation traffic flow map with the probability of traffic transfer, and the corresponding traffic flow nodes are set as reverse absorption and convergence nodes to generate a reverse search map for candidate locations.

[0017] Starting from the reverse absorption and convergence node, the upstream flow graph nodes are read layer by layer in the opposite direction of the directed flow graph edge in the reverse search graph of the candidate point, and the first path to be expanded is extended by using the contribution upper bound priority queue to generate the reverse customer acquisition path.

[0018] As a preferred embodiment of the street-level customer flow conversion site selection method based on machine learning algorithms described in this invention, the specific steps of calculating the path customer acquisition contribution value based on the node conversion probability, and then using the path customer acquisition contribution value to perform contribution feedback on the movement map nodes in the reverse customer acquisition path to generate point movement features are as follows:

[0019] Read the directed flow graph edges that each reverse customer acquisition path passes through in the forward order of customer flow propagation, and calculate the path customer acquisition contribution value based on the node conversion probability and path length attenuation coefficient corresponding to the directed flow graph edge.

[0020] Based on the contribution value of the path customer acquisition, the movement map nodes in the reverse customer acquisition path are fed back with contributions. The feed-back weight is determined according to the path distance from the movement map node to the reverse absorption and convergence node. The path customer acquisition contribution value is allocated to the corresponding movement map node according to the feed-back weight. Based on the allocation results, the reverse customer acquisition intensity, path continuity, upstream node concentration and key node dependence of the candidate points are extracted to generate point movement characteristics.

[0021] As a preferred embodiment of the street passenger flow conversion site selection method based on machine learning algorithm described in this invention, the specific steps of calculating passenger flow capture values ​​based on location movement characteristics and mapping these values ​​to a location spatial grid to generate a location conversion potential energy field are as follows:

[0022] The location flow characteristics are normalized, and the normalized location flow characteristics are weighted and fused to generate passenger flow capture values. A location spatial grid is established based on the boundary coordinates of the target block. The entrance coordinates of the candidate locations are projected onto the corresponding central grid, and the passenger flow capture values ​​are written into the central grid to generate the location capture grid.

[0023] Using the point capture grid as the diffusion starting point, the passenger flow capture value is diffused along the stable movement line edge connected to the candidate point. The adjacent grids with connected movement lines are smoothed on both the spatial and movement lines to generate the point conversion potential energy field.

[0024] As a preferred embodiment of the street passenger flow conversion site selection method based on machine learning algorithms described in this invention, the steps of generating passenger flow intention tags based on continuous trajectory changes, splicing them in chronological order to form an intention transition chain, and dynamically correcting the spatial potential energy value based on forward and reverse transitions to generate effective passenger flow characteristics are as follows:

[0025] Read continuous trajectory points of the same passenger flow object, segment the continuous trajectory points into trajectory segments according to the timestamp order, generate trajectory segment features based on the continuous changes of adjacent trajectory points, and generate passenger flow intention labels through time-series classification processing;

[0026] Passenger flow intent tags that are continuous in time and reachable by grid lines are spliced ​​together to generate intent transition chains. The direction of intent change between adjacent passenger flow intent tags is determined according to the preset intent level. When adjacent passenger flow intent tags change from low to high intent level and the spatial potential energy value of the grid where the corresponding trajectory point is located increases in time sequence, it is marked as a positive transition.

[0027] When the adjacent passenger flow intent label changes from a high conversion intent level to a low conversion intent level, and the spatial potential energy value of the corresponding trajectory point's grid decreases sequentially over time, it is marked as a reverse transition.

[0028] The corresponding spatial potential energy value in the potential energy field of the point transformation is enhanced and corrected based on the positive transition, and the corresponding spatial potential energy value in the potential energy field of the point transformation is attenuated and corrected based on the reverse transition. Effective passenger flow characteristics are generated based on the corrected spatial potential energy values.

[0029] As a preferred embodiment of the street-level passenger flow conversion site selection method based on machine learning algorithms described in this invention, the specific steps for generating passenger flow conversion values ​​are as follows:

[0030] Based on the brand constraint codes of the brands to be selected, the effective customer flow characteristics are matched and calibrated to generate brand location characteristics. The brand location characteristics are then stabilized by quantile truncation and standardized to generate stable brand location characteristics.

[0031] Using the candidate location number as an index, the stable brand location features, location flow features, and location basic features are sequentially written into the feature record of the corresponding candidate location. The features are then vectorized and concatenated according to a fixed field order, and correlation redundancy removal is performed to generate the location prediction features.

[0032] The predicted features of the locations are input into the site selection learning model for conversion rate regression and risk classification, generating predicted conversion rates and risk classification results. Based on the risk classification results, the risk level of the corresponding candidate locations is determined, and the predicted conversion rate is adjusted by decreasing the risk adjustment coefficient corresponding to the risk level to generate passenger flow conversion values.

[0033] As a preferred embodiment of the street passenger flow conversion site selection method based on machine learning algorithm described in this invention, the site selection learning model is composed of an input layer set up according to a fixed field order, using location prediction features as the model input object, a feature encoding layer set after the input layer, a shared learning layer based on gradient boosting tree set after the feature encoding layer, a conversion rate regression branch and a risk classification branch set in parallel after the shared learning layer, and a passenger flow conversion value output layer.

[0034] As a preferred embodiment of the street-level passenger flow conversion site selection method based on machine learning algorithms described in this invention, the specific steps of calculating a site selection score based on passenger flow conversion values, sorting candidate sites according to the site selection scores, and generating a site selection recommendation list are as follows:

[0035] The customer flow conversion value is normalized to generate a standardized conversion value. The candidate locations are constrained and verified according to the brand constraint code of the brand to be selected, generating a valid score set. The location carrying capacity value is calculated based on the basic characteristics of the location. The standardized conversion value and the location carrying capacity value are weighted and calculated to generate a location selection score.

[0036] Candidate sites are sorted from highest to lowest according to their site selection scores. If sites have the same site selection score, the order is adjusted according to the customer flow conversion value from highest to lowest to generate a site selection recommendation list.

[0037] The beneficial effects of this invention are as follows: by constructing street objects and street paths into a directed business flow map with conversion probability, and using candidate locations as reverse absorption and convergence nodes to search for reverse customer acquisition paths, and then feeding back the contribution of the flow map nodes in the path according to the customer acquisition contribution value of the path, the structured quantification of the source of incoming traffic, path continuity and key node dependence is realized. This makes the candidate locations no longer just represented by local traffic heat, but can form location flow characteristics that reflect the real direction of traffic spread and the strength of customer acquisition contribution. These features are used for subsequent traffic capture values, conversion potential fields and site selection learning model inputs, thereby improving the accuracy of candidate location conversion potential expression and the reliability of site selection recommendation. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of a street-level customer flow conversion site selection method based on machine learning algorithms.

[0040] Figure 2 A flowchart for constructing initial site selection data and street operation flow diagrams.

[0041] Figure 3 This is a flowchart for reverse customer acquisition path search and point-of-sale flow feature generation.

[0042] Figure 4 A flowchart for generating the potential energy field and effective passenger flow characteristics of a location.

[0043] Figure 5 A standardized trend comparison chart of candidate point movement characteristics and actual conversion rates.

[0044] Figure 6 This is a magnified comparison chart of actual conversion rate rankings and site selection recommendation rankings. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Reference Figures 1-6 This is one embodiment of the present invention, which provides a method for selecting street locations based on machine learning algorithms for converting pedestrian traffic, including the following steps:

[0049] S1. Clean and encode the brand data and target street data to be selected, extract candidate locations, and generate initial site selection data.

[0050] S1.1 Obtain the brand data and target street data for the site selection, and remove invalid records and invalid shop records from the brand data and target street data respectively. Convert the cleaned brand data and target street data into brand constraint codes and shop codes.

[0051] Furthermore, data on potential brands and target neighborhoods is obtained. The data on potential brands includes information on their operational needs, site selection restrictions, business type attributes, space requirements, rent affordability, floor preferences, target customer groups, and requirements of neighboring businesses. The data on target neighborhoods includes information on the status of shops within the target neighborhood, shop location, shop operating conditions, neighborhood boundaries, neighborhood paths, business entrance location, and customer traffic.

[0052] Read the data of brands to be selected one by one, and mark the data records that express the business needs and site selection restrictions of the brands to be selected as invalid records. After removing the invalid records from the data of brands to be selected, encode and transform the remaining data records of brands to be selected so that the business needs and site selection restrictions of the brands to be selected can be formed into brand constraint codes that can be used for store matching and screening.

[0053] Read the target street data one by one, focusing on the information that can express the status, location, and operating conditions of the shops. Mark the target street data records that cannot participate in site selection, cannot complete spatial association, or cannot meet the operating needs of the selected brand as invalid shop records. After removing the invalid shop records from the target street data, encode and transform the remaining target street data records so that the shop status, location, and operating conditions can form shop codes that can match the brand constraint codes.

[0054] S1.2 Match and filter the shop codes according to the brand constraint code, determine the shop records that meet the brand constraint code as candidate locations, and number and spatially associate the candidate locations to generate initial site selection data.

[0055] Furthermore, using brand constraint codes as a screening criterion, the shop codes are matched item by item. First, shop codes that do not meet the hard site selection restrictions in the brand constraint codes are excluded, and then shop codes that meet the operational suitability requirements in the brand constraint codes are retained. The shop records corresponding to the retained shop codes are determined as candidate locations.

[0056] Subsequently, candidate location numbers are generated according to the source records of candidate locations in the target street data. The business entrance location of the candidate location is associated with the street spatial location in the target street data, and the shop spatial location of the candidate location is associated with the street path in the target street data, generating initial site selection data containing candidate location numbers, brand constraint codes, shop codes, and spatial relationships.

[0057] S2. Extract street objects and street paths from the initial site selection data, map street objects to movement map nodes, map street paths to directed movement map edges, generate street operation movement map, use candidate locations as reverse absorption and convergence nodes, search for reverse customer acquisition paths along the reverse direction of customer flow propagation, calculate the path customer acquisition contribution value based on the node conversion probability, and then perform contribution feedback on the movement map nodes in the reverse customer acquisition path based on the path customer acquisition contribution value to generate location movement characteristics.

[0058] S2.1 Extract street objects from the initial site selection data, and divide the street objects into nodes according to the location of passenger flow dwelling, the location of passenger flow turning and the location of business entrance. Map the divided street objects as nodes of the flow diagram. Extract street paths from the initial site selection data, perform edge-direction processing on the street paths according to the direction of passenger flow, and map the edge-directed street paths as edges of the directed flow diagram.

[0059] Furthermore, street objects are extracted from the initial site selection data. Each street object is then segmented according to the location where customers stop, forming independent spatial objects that can represent customer behavior. The street objects are then further segmented according to the location where customers turn, forming independent spatial objects that can represent changes in customer direction. Finally, the street objects are further segmented according to the location of the business entrance, forming independent spatial objects that can represent the relationship between shops.

[0060] After completing the node segmentation, each independent spatial object is mapped to a traffic flow node, and the spatial relationship between the independent spatial object and the candidate point number is preserved. Then, street paths are extracted from the initial site selection data, and the starting and ending positions of the street paths are determined according to the direction of passenger flow. Street paths with a clear direction of passenger flow are processed into directional path segments.

[0061] After completing the edge-direction process, each directional path segment is mapped to a directed path graph edge, and the starting path graph node and the destination path graph node connected by the directed path graph edge are recorded.

[0062] It should also be noted that the location of passenger flow is obtained by identifying the passenger flow in the target street data. The identification of the location is based on the passage location of the same passenger flow object with little change in location within a continuous time and the length of stay meets the criteria for determining the location.

[0063] The turning point of passenger flow is obtained by identifying the direction change of passenger flow in the target block data. The direction change identification is extracted according to the change of passage direction between adjacent passage positions of the same passenger flow object.

[0064] The location of the business entrance is obtained by extracting the spatial relationships in the initial site selection data. The location corresponding to the entry relationship of the candidate point in the spatial relationships is determined as the business entrance location.

[0065] The direction of passenger flow is obtained by connecting adjacent passage positions in the target block data according to the time sequence of passenger flow information. The adjacent passage positions point from the position corresponding to the earlier time to the position corresponding to the later time.

[0066] S2.2 Perform Hampel local median filtering on the passenger flow records corresponding to the directed flow line graph edges, remove abnormal flow records that deviate from the local median by more than the median deviation threshold, and correct the flow intensity of the directed flow line graph edges based on the filtered passenger flow records to generate stable flow line edges.

[0067] Furthermore, the passenger flow records corresponding to each directed flow line edge are arranged in chronological order, and a local time range is selected centered on the current time position of the passenger flow record. Within this local time range, the local median of the passenger flow records is calculated, and then the median deviation is calculated based on the degree of deviation between the passenger flow records and the local median. The expression is as follows: ;

[0068] in: Directed movement graph edges At the current time location The local median of the corresponding passenger flow records; Directed movement graph edges Within a local time range Internal Time Position Corresponding passenger flow records; Local time range Any time location within; Based on the current time and location A local time range selected centered on; ;

[0069] in: Directed movement graph edges At the current time location The corresponding median deviation is used to represent the local time range. Internal passenger flow records relative to the local median Typical degree of deviation;

[0070] The median deviation is scaled up and used as the median deviation threshold. When the deviation between the current passenger flow record and the local median exceeds the median deviation threshold, the current passenger flow record is marked as an abnormal passage record and removed. After Hampel local median filtering is completed, the passage intensity of the directed flow graph edges is recalculated based on the filtered passenger flow records, so that the passage intensity of the directed flow graph edges is represented by passenger flow records with less abnormal fluctuations, and the directed flow graph edges with corrected passage intensity are generated as stable flow graph edges.

[0071] It should also be noted that the passenger flow records corresponding to the directed movement line graph edges are obtained by matching the passenger flow data in the target block data to the corresponding directed movement line graph edges;

[0072] The median deviation threshold serves as the boundary for identifying abnormal passage records. It is used to determine whether the deviation of the current passenger flow record from the local median exceeds the local normal fluctuation range. Based on the median deviation corresponding to the directed flow graph edge and the local fluctuation of the passenger flow record, the range of values ​​for the median deviation threshold is determined as an example value. to ,in Represents the edges of a directed motion graph. In time location The corresponding median deviation.

[0073] S2.3. Establish node indexes based on the directed movement line graph edges, establish edge connection indexes based on the stable movement line edges, and calculate the node conversion probability of the corresponding directed movement line graph edge based on the outgoing edge passage intensity of each movement line graph node to generate the street operation movement line graph.

[0074] Furthermore, based on the directed movement graph edges and stable movement graph edges generated in the previous step, a node index is first established according to the starting movement graph nodes and arriving movement graph nodes connected by the directed movement graph edges, so that each movement graph node can be queried for the connected directed movement graph edges; then, an edge connection index is established according to the starting movement graph nodes, arriving movement graph nodes, and traffic intensity connected by the stable movement graph edges, so that each stable movement graph edge can represent the passenger flow relationship between two movement graph nodes.

[0075] Taking each node in the flow graph as the calculation object, the flow intensity of all outgoing edges corresponding to each node is summarized. The flow intensity of each directed flow graph edge is then proportionally calculated to the total flow intensity of all outgoing edges of the node containing it, yielding the node conversion probability for that directed flow graph edge. The expression is as follows: ;

[0076] in: From the nodes of the motion diagram Point to the node of the movement diagram The probability of node transition corresponding to the edge of the directed motion graph; From the nodes of the motion diagram Point to the node of the movement diagram The traffic intensity corresponding to the edge of the directed movement graph; From the nodes of the motion diagram Point to the node of the movement diagram The traffic intensity corresponding to the edge of the directed movement graph; It is a node in the movement diagram. Any corresponding node on the arrival path diagram; Based on the nodes of the movement diagram The set of all reachable nodes in the initial path diagram node; It is the starting node of the directed path graph corresponding to the edge of the path graph;

[0077] Write the node index, edge connection index, and node flow probability into the connection relationship between the nodes in the flow graph and the edges in the directed flow graph to generate the street operation flow graph.

[0078] S2.4 Map the candidate locations to the corresponding traffic flow nodes in the street operation traffic flow map with the probability of traffic transfer, and set the corresponding traffic flow nodes as reverse absorption and convergence nodes to generate a reverse search map for candidate locations.

[0079] Furthermore, based on the spatial relationships in the street operation flow map and the initial site selection data, the operation entrance location of the candidate location is extracted using the candidate location number as an index. Flow map nodes with spatial locations consistent with the operation entrance location are retrieved in the street operation flow map. If a flow map node with a consistent spatial location exists, it is determined as the flow map node corresponding to the candidate location. If no flow map node with a consistent spatial location exists, the spatial distance between the operation entrance location and each flow map node in the street operation flow map is calculated. Flow map nodes that cannot be connected to the operation entrance location through directed flow map edges are eliminated, and the flow map node with the smallest spatial distance that can be connected through directed flow map edges is determined as the flow map node corresponding to the candidate location.

[0080] Write the corresponding flow graph nodes of the candidate points into the spatial association relationship of the candidate point numbers, and set the corresponding flow graph nodes of the candidate points as reverse absorption and convergence nodes, which are used as the reverse search starting point in the reverse search graph of the candidate points, and represent the convergence position of the candidate points receiving the contribution of the upstream tourist flow in the street operation flow graph; write the candidate point number and the corresponding flow graph node identifier of the candidate points into the reverse absorption and convergence node, retain the directed flow graph edges, stable flow edges, edge connection index and node flow probability in the street operation flow graph, and generate the candidate point reverse search graph;

[0081] It should be noted that the candidate site reverse search graph is used to set the corresponding flow graph nodes of the candidate site as reverse absorption and convergence nodes, so that the candidate site can trace the upstream flow graph nodes and directed flow graph edges in the reverse direction of passenger flow propagation. Instead of judging the site selection value based solely on the passenger flow around the candidate site, it tracks the source of upstream passenger flow and the relationship of passenger acquisition contribution of the candidate site through the reverse path with the probability of flow transfer.

[0082] S2.5. Starting from the reverse absorption and convergence node, read the upstream movement graph nodes layer by layer in the opposite direction of the directed movement graph edge in the reverse search graph of the candidate point, and use the contribution upper bound priority queue to expand the first path to be expanded to generate the reverse customer acquisition path.

[0083] Furthermore, using the reverse absorption convergence node as the search starting point, upstream flow graph nodes that can reach the reverse absorption convergence node are searched in the opposite direction of the directed flow graph edge. The reverse absorption convergence node and the upstream flow graph nodes form the path to be expanded. Based on the flow conversion probability of the nodes traversed by the path to be expanded, the flow conversion probability of the nodes is continuously recursively extrapolated in the forward order of passenger flow propagation. Combined with the flow conversion probability of the node with the highest value among the upstream flow graph nodes that can be connected to the path to be expanded, the upper limit of the maximum customer acquisition contribution that the path to be expanded may reach is estimated, and the upper limit of the contribution of the path to be expanded is obtained. Then, the paths to be expanded are arranged from high to low according to the upper limit of contribution, and the upper limit priority queue is obtained. This queue is used to arrange the paths to be expanded from high to low according to the upper limit of contribution, and the first path to be expanded is determined from the paths to be expanded, so that the path to be expanded with the highest potential for customer acquisition contribution in the reverse search graph of the candidate point is expanded first.

[0084] The upper contribution bound of each path to be expanded in the upper contribution bound priority queue is compared numerically, and the path to be expanded with the largest upper contribution bound value is determined as the first path to be expanded. In the candidate point reverse search graph, the upstream path graph node is searched in the opposite direction of the directed path graph edge corresponding to the end path graph node of the first path to be expanded. The upstream path graph node found and the first path to be expanded form a new path to be expanded.

[0085] The contribution upper bound priority queue is updated based on the contribution upper bound of the new path to be expanded. When there is no upstream path node that can be reached from the first path to be expanded, or the contribution upper bound of the first path to be expanded does not meet the conditions for continued expansion, the expansion of the first path to be expanded is stopped, and the path from the upstream path node to the reverse absorption and convergence node is generated as the reverse customer acquisition path. This means that the upstream path node reaches the reverse absorption and convergence node through the directed path edge, so that the candidate point can be mapped to the specific upstream passenger flow source, direction of travel and path connection relationship, providing a path basis for subsequent calculation of the path customer acquisition contribution value.

[0086] It should also be noted that the reverse customer acquisition path takes the candidate location as the reverse absorption and convergence node, and traces the upstream flow path nodes that can send customer flow to the candidate location in the reverse direction of the passenger flow propagation, rather than only counting the passenger flow scale around the candidate location, so as to express the absorption range of the candidate location to the upstream passenger flow and the dependence of the customer acquisition path.

[0087] The conditions for further expansion are determined by the upper bound of the contribution of the path to be expanded and the lowest effective path customer acquisition contribution value among the currently generated reverse customer acquisition paths.

[0088] S2.6 Read the directed flow graph edges that each reverse customer acquisition path passes through in the forward order of customer flow propagation, and calculate the path customer acquisition contribution value based on the node conversion probability and path length attenuation coefficient corresponding to the directed flow graph edge.

[0089] Furthermore, the node order of each reverse customer acquisition path is restored to the forward order of customer flow propagation, so that the reverse customer acquisition path starts from the upstream flow diagram node and ends at the reverse absorption and convergence node; then, according to the connection relationship between adjacent flow diagram nodes, the directed flow diagram edges traversed by the reverse customer acquisition path are determined in sequence, and the node conversion probability corresponding to each directed flow diagram edge is obtained; then, the number of directed flow diagram edges traversed by the reverse customer acquisition path or the travel distance corresponding to the reverse customer acquisition path is subjected to exponential decay processing to obtain the path length decay coefficient, so that the shorter and continuously connected reverse customer acquisition paths retain high contribution intensity, and the longer and more connected reverse customer acquisition paths form low contribution intensity.

[0090] The conversion probabilities of nodes corresponding to each directed path edge in the reverse customer acquisition path are processed using continuous probability, and the content after continuous probability processing is attenuated and corrected using a path length attenuation coefficient to obtain the path customer acquisition contribution value, expressed as: ;

[0091] in: It is a reverse customer acquisition path The corresponding path customer acquisition contribution value is used to represent the upstream traffic flow node along the reverse customer acquisition path. The intensity of customer acquisition contribution to reaching the reverse absorption convergence node; It is a reverse customer acquisition path The corresponding path length decay coefficient is used to represent the reverse customer acquisition path. The degree to which the path length diminishes the contribution of customer acquisition. It is a reverse customer acquisition path The set of edges of the directed flow graph traversed in the forward order of passenger flow propagation;

[0092] The path acquisition contribution value represents the intensity of customer acquisition contribution from upstream nodes along the reverse acquisition path to the reverse absorption and convergence node.

[0093] S2.7. Based on the path customer acquisition contribution value, the movement map nodes in the reverse customer acquisition path are fed back with contributions. The feed-back weight is determined according to the path distance from the movement map node to the reverse absorption and convergence node. The path customer acquisition contribution value is allocated to the corresponding movement map node according to the feed-back weight. Based on the allocation result, the reverse customer acquisition intensity, path continuity, upstream node concentration and key node dependence of the candidate point are extracted to generate point movement characteristics.

[0094] Furthermore, taking the nodes in the flow chart of each reverse customer acquisition path as the contributing feedback objects, first determine the path distance from each flow chart node to the reverse absorption and aggregation node. Flow chart nodes with shorter path distances are assigned higher feedback weights, and flow chart nodes with longer path distances are assigned lower feedback weights.

[0095] The path customer acquisition contribution value is allocated to the corresponding movement map nodes in the reverse customer acquisition path according to the backflow weight, forming the contribution allocation content of each movement map node; then, the path customer acquisition contribution value of all reverse customer acquisition paths corresponding to the candidate point is summarized to extract the reverse customer acquisition intensity of the candidate point; the connection relationship between adjacent movement map nodes in each reverse customer acquisition path is verified, and the proportion of node relationships with continuous directed movement map edges is counted in all node relationships to extract the path continuity of the candidate point; the contribution allocation content of each upstream movement map node is statistically analyzed, and the upstream node concentration of the candidate point is extracted based on the concentration of contribution allocation content among upstream movement map nodes; the upstream movement map node with the highest contribution allocation content is extracted, and the key node dependency of the candidate point is extracted based on the proportion of the upstream movement map node with the highest contribution allocation content in the contribution allocation content of all upstream movement map nodes; the point movement characteristics are generated by reverse customer acquisition intensity, path continuity, upstream node concentration, and key node dependency.

[0096] like Figure 5 As shown, the reverse customer acquisition intensity, path continuity, local passenger flow, and actual conversion rate, after standardization, can be compared in trend on the same coordinate system. The curve changes reveal that the reverse customer acquisition intensity curve and the actual conversion rate curve maintain high consistency at multiple candidate locations, while the local passenger flow curve shows a significant misalignment with the actual conversion rate curve, indicating that the conversion potential of candidate locations does not solely depend on the surrounding passenger flow scale. This invention characterizes the direction of passenger flow propagation through node conversion probability, traces the source of upstream passenger flow through reverse customer acquisition paths, and utilizes contribution feedback to form location movement characteristics. This allows the reverse customer acquisition intensity to comprehensively reflect the strength of path contribution and passenger flow accessibility, thus more accurately expressing the customer acquisition capacity of candidate locations.

[0097] S3. Calculate passenger flow capture values ​​based on point-to-point movement characteristics, map the passenger flow capture values ​​to point-to-point spatial grids, generate point-to-point transformation potential energy fields, generate passenger flow intention labels based on continuous trajectory changes, splice them in time sequence to form intention transition chains, and dynamically correct spatial potential energy values ​​based on forward and reverse transitions to generate effective passenger flow characteristics.

[0098] S3.1 Normalize the location flow characteristics, and perform weighted fusion on the normalized location flow characteristics to generate passenger flow capture values. Establish a location spatial grid based on the target street boundary coordinates, project the entrance coordinates of candidate locations onto the corresponding central grid, and write the passenger flow capture values ​​into the central grid to generate a location capture grid.

[0099] Furthermore, the location flow characteristics are normalized to convert features with different value scales into a unified numerical range. Based on the correlation strength between location flow characteristics and actual conversion rates in historical candidate locations, a flow characteristic weighting coefficient is determined. The normalized location flow characteristics are then weighted and fused according to this coefficient to generate a passenger flow capture value for each candidate location, representing its comprehensive ability to absorb incoming passenger flow. Subsequently, the coverage area of ​​the location spatial grid is determined based on the target street boundary coordinates, and continuous grids are divided according to the passable space within the target street boundary coordinates. The entrance coordinates of each candidate location are matched to their corresponding grid position within the location spatial grid, and this grid position is designated as the central grid. The passenger flow capture value corresponding to each candidate location is assigned to the central grid, enabling it to carry the passenger flow capture capability of the candidate location, thus generating a location capture grid.

[0100] S3.2. Using the point capture grid as the diffusion starting point, the passenger flow capture value is diffused along the stable movement line side connected to the candidate point, and the adjacent grids with connected movement lines are smoothed on both sides of space and movement line to generate the point conversion potential energy field.

[0101] Furthermore, the central grid, which carries the passenger flow capture value, is used as the starting point for the diffusion of potential energy guided by the movement path. The passenger flow capture value of the central grid is used as the initial spatial potential energy value. The diffusion direction is determined according to the stable movement path edges connected to the candidate points, and the grids connected to the central grid are determined sequentially along the passenger flow direction of the stable movement path edges. For each grid connected by the movement path, the initial spatial potential energy value is attenuated according to the movement path distance between the grid connected by the movement path and the central grid to obtain the spatial potential energy value corresponding to the grid connected by the movement path.

[0102] Using the grid with obtained spatial potential energy values ​​as the grid to be smoothed, adjacent grids that are spatially adjacent to the grid to be smoothed and connected by stable movement lines are selected. Spatial weights are formed based on the spatial distance between the grid to be smoothed and the adjacent grids, and movement line weights are formed based on the passage intensity of the stable movement lines between the grid to be smoothed and the adjacent grids. Then, the spatial potential energy values ​​of the adjacent grids are weighted and smoothed according to the spatial weights and movement line weights to obtain the smoothed spatial potential energy values ​​corresponding to the grid to be smoothed. A correspondence is established between the grids within the target block boundary coordinate coverage area and the smoothed spatial potential energy values ​​to generate a point-to-point potential energy field.

[0103] It should be noted that the point conversion potential energy field can extend the passenger flow capture value of candidate points from the central grid to the grids connected to the passenger flow lines of the candidate points, so that the passenger acquisition capacity of the candidate points can form a continuous distribution within the target block space. It is used to represent the potential intensity of passenger flow conversion of candidate points by each grid within the boundary coordinate coverage area of ​​the target block. The higher the spatial potential energy value, the higher the potential intensity of passenger flow conversion of the corresponding grid to the candidate point; the lower the spatial potential energy value, the lower the potential intensity of passenger flow conversion of the corresponding grid to the candidate point.

[0104] The potential energy field of the point conversion uses the passenger flow direction on the stable circulation line as the basis for the diffusion of passenger flow capture value, so that the passenger flow capture value is preferentially transmitted to the grid with circulation line connection relationship with the candidate point. Through spatial-circuit bilateral smoothing, a continuous spatial potential energy value distribution is formed between grids that are spatially adjacent and have a high degree of circulation line connection, thereby more accurately expressing the conversion potential influence range of the candidate point in the target block.

[0105] S3.3 Read the continuous trajectory points of the same passenger flow object, divide the continuous trajectory points into trajectory segments according to the timestamp order, generate trajectory segment features based on the continuous changes of adjacent trajectory points, and generate passenger flow intention labels through time-series classification processing.

[0106] Furthermore, firstly, continuous trajectory points of the same passenger flow object are collected based on the passenger flow object identifier, and then the continuous trajectory points of the same passenger flow object are organized according to the time stamp order. When there is a long time interval between adjacent continuous trajectory points, the continuous trajectory points are split at the time interval position; when there is a sudden change in spatial distance between adjacent continuous trajectory points, the continuous trajectory points are split at the spatial change position; when there is an unreachable path between adjacent continuous trajectory points, the continuous trajectory points are split at the path break position, generating trajectory segments.

[0107] For each trajectory segment, firstly, position change features are generated based on the position changes between adjacent trajectory points; then, movement speed change features are generated based on the movement speed changes between adjacent trajectory points; then, movement direction change features are generated based on the movement direction changes between adjacent trajectory points; then, dwell time change features are generated based on the continuity of trajectory points in the same location; and finally, distance change features are generated based on the degree of proximity of trajectory points to the business entrance. These are used to form trajectory segment features. The trajectory segment features are then subjected to temporal classification processing, and a passenger flow intent label is output for each trajectory segment.

[0108] S3.4. Connect the passenger flow intention tags that are continuous in time and reachable by the grid movement line to generate an intention transition chain, and determine the direction of intention change between adjacent passenger flow intention tags according to the preset intention level. When the adjacent passenger flow intention tags change from low conversion intention level to high conversion intention level, and the spatial potential energy value of the grid where the corresponding trajectory point is located increases in time sequence, it is marked as a positive transition.

[0109] Furthermore, the passenger flow intent tags corresponding to the same passenger flow object are organized in chronological order; when adjacent passenger flow intent tags are continuous in time, it is further verified whether the grids where the trajectory points corresponding to the adjacent passenger flow intent tags are located can be reached through stable movement lines; when the time is continuous and the grid movement lines are reachable, the adjacent passenger flow intent tags are spliced ​​together in chronological order to generate intent transition chains.

[0110] Determine the conversion intention level corresponding to adjacent passenger flow intention tags according to the intention level, and compare the direction of change of conversion intention level of adjacent passenger flow intention tags; when the previous passenger flow intention tag corresponds to a low conversion intention level and the next passenger flow intention tag corresponds to a high conversion intention level, extract the spatial potential energy value of the grid where the trajectory point corresponding to the previous passenger flow intention tag is located and the spatial potential energy value of the grid where the trajectory point corresponding to the next passenger flow intention tag is located; when the spatial potential energy value of the grid where the trajectory point corresponding to the next passenger flow intention tag is located is higher than the spatial potential energy value of the grid where the trajectory point corresponding to the previous passenger flow intention tag is located, mark the intention change between adjacent passenger flow intention tags as a positive transition.

[0111] It should also be noted that the intent level is determined by the conversion probability corresponding to the customer intent tag, the change in distance between the customer and the business entrance, the length of time the customer stays near the business entrance, and the trend of the customer moving towards the candidate point along a stable path.

[0112] S3.5 When the adjacent passenger flow intention label changes from a high conversion intention level to a low conversion intention level, and the spatial potential energy value of the corresponding trajectory point's grid decreases sequentially over time, it is marked as a reverse transition.

[0113] Furthermore, based on the intent transition chain, the conversion intent level corresponding to adjacent passenger flow intent tags is determined according to the intent level, and the direction of intent change between adjacent passenger flow intent tags is compared; when the previous passenger flow intent tag corresponds to a high conversion intent level and the next passenger flow intent tag corresponds to a low conversion intent level, the spatial potential energy value of the grid where the trajectory point corresponding to the previous passenger flow intent tag is located and the spatial potential energy value of the grid where the trajectory point corresponding to the next passenger flow intent tag is located are extracted; when the spatial potential energy value of the grid where the trajectory point corresponding to the next passenger flow intent tag is located is lower than the spatial potential energy value of the grid where the trajectory point corresponding to the previous passenger flow intent tag is located, the intent change between adjacent passenger flow intent tags is marked as a reverse transition.

[0114] S3.6. Based on the positive transition, the corresponding spatial potential energy value in the potential energy field of the point conversion is enhanced and corrected, and based on the reverse transition, the corresponding spatial potential energy value in the potential energy field of the point conversion is attenuated and corrected. Based on the corrected spatial potential energy value, effective passenger flow characteristics are generated.

[0115] Furthermore, the spatial potential energy value of the grid where the trajectory point corresponding to the positive jump is located in the potential energy field of the point conversion is determined. The number of occurrences of the positive jump is used as the frequency basis for enhancement correction, and the magnitude of the increase in the conversion intention level is used as the intensity basis for enhancement correction. The spatial potential energy value is adjusted upward to obtain the enhanced spatial potential energy value, which is used to represent the potential increase in the intensity of passenger flow conversion of the grid where the positive jump occurs for the candidate point.

[0116] Next, determine the spatial potential energy value of the grid where the trajectory point corresponding to the reverse jump is located in the potential energy field of the point conversion. Use the number of occurrences of the reverse jump as the frequency basis for attenuation correction, and use the decrease in the conversion intention level as the intensity basis for attenuation correction. Adjust the spatial potential energy value downward to obtain the attenuated spatial potential energy value, which is used to indicate the reduced potential intensity of the grid that has undergone the reverse jump to form passenger flow conversion for the candidate point.

[0117] The spatial potential energy values ​​are summarized and corrected according to the potential energy fields corresponding to the candidate points, and the effective passenger flow scale, effective passenger flow direction, effective passenger flow dwell tendency and effective passenger flow approach tendency are extracted from the corrected spatial potential energy values ​​to generate effective passenger flow characteristics.

[0118] S4. Generate brand location features based on effective customer flow characteristics, integrate brand location features, location flow features, and location basic features to generate location prediction features, and input the location prediction features into the site selection learning model for conversion rate regression and risk classification to generate customer flow conversion values.

[0119] It should be noted that, using location prediction features as the model input, the input layer is first set up according to the fixed field order of the location prediction features, ensuring that the location prediction features corresponding to the candidate location numbers enter the input layer in a consistent field position. Following the input layer, a feature encoding layer is set up. This layer converts continuous features in the location prediction features into standardized numerical representations and categorical features into encoded numerical representations, forming a unified feature representation. After the feature encoding layer, a gradient boosting tree-based shared learning layer is set up. This layer takes the unified feature representation as input and fits the feature residual information through multiple decision trees round by round, forming a shared feature representation for conversion rate regression and risk classification. The shared feature output is used to construct the location learning model. After the shared learning layer based on the gradient boosting tree, a conversion rate regression branch and a risk classification branch are set up in parallel. The conversion rate regression branch generates the predicted conversion rate based on the shared feature output, and the risk classification branch generates the risk classification result based on the shared feature output. After the conversion rate regression branch and the risk classification branch, a passenger flow conversion value output layer is set up. The passenger flow conversion value output layer determines the risk level based on the risk classification result and adjusts the predicted conversion rate by decreasing it according to the risk adjustment coefficient corresponding to the risk level to generate the passenger flow conversion value. The input layer, feature encoding layer, shared learning layer based on gradient boosting tree, conversion rate regression branch, risk classification branch and passenger flow conversion value output layer are connected in the order of transmission of location prediction features to construct the location learning model.

[0120] The input layer ensures that the predicted features of locations enter the site selection learning model in a fixed field order, avoiding field misalignment between different candidate locations. The feature encoding layer converts the predicted features of locations into a unified feature representation, enabling continuous and categorical features to participate in subsequent learning. The gradient boosting tree-based shared learning layer learns the common feature relationships affecting conversion rate and risk classification from the unified feature representation. The conversion rate regression branch generates the predicted conversion rate based on the shared features output by the gradient boosting tree-based shared learning layer. The risk classification branch generates the risk classification result based on the shared features output by the gradient boosting tree-based shared learning layer. The passenger flow conversion value output layer determines the risk level based on the risk classification result and adjusts the predicted conversion rate by decreasing it according to the risk adjustment coefficient corresponding to the risk level, generating the passenger flow conversion value.

[0121] The training process of the site selection learning model is as follows: Historical candidate site prediction features are obtained and input into the site selection learning model according to a fixed field order; the actual conversion rate of historical candidate sites is used as the training label for the conversion rate regression branch, and the actual risk level of historical candidate sites is used as the training label for the risk classification branch; the conversion rate regression branch uses a mean squared error loss function to constrain the deviation between the predicted and actual conversion rates, and the risk classification branch uses a cross-entropy loss function to constrain the deviation between the risk classification results and the actual risk level. The mean squared error loss function and the cross-entropy loss function are then weighted and fused to form a joint loss function; based on the joint loss function, a shared learning layer based on a gradient boosting tree, the conversion rate regression branch, and the risk classification branch are trained to obtain the trained site selection learning model.

[0122] It should also be noted that the actual conversion rate of historical candidate locations is obtained by the correspondence between the number of customers visiting the historical candidate locations during the actual operating period and the number of customers passing through the associated traffic flow areas of the historical candidate locations.

[0123] The actual risk level of historical candidate sites is obtained through risk records corresponding to the rent pressure, customer flow fluctuations, operational stability, and deviations from brand constraints of historical candidate sites during their actual operating cycles.

[0124] S4.1 Match and calibrate the effective customer flow characteristics according to the brand constraint code of the brand to be selected, generate brand location characteristics, perform quantile truncation stabilization on the brand location characteristics, and perform standardization transformation on the truncated brand location characteristics to generate stable brand location characteristics.

[0125] Furthermore, using the brand constraint code of the candidate brand as the matching basis, the effective customer flow characteristics of each candidate location are matched with the operational needs of the candidate brand. When the effective customer flow characteristics match the operational needs of the candidate brand, the customer flow expression content corresponding to the effective customer flow characteristics is retained. When the effective customer flow characteristics do not match the operational needs of the candidate brand, the customer flow expression content corresponding to the effective customer flow characteristics is downgraded to generate brand location characteristics.

[0126] The brand location features are stabilized by quantile truncation. Feature values ​​above the upper quantile boundary are compressed to the upper quantile boundary, and feature values ​​below the lower quantile boundary are raised to the lower quantile boundary, reducing the impact of extreme feature values ​​on candidate location comparison. The truncated brand location features are then standardized to a uniform numerical scale, generating stable brand location features.

[0127] It should also be noted that the upper quantile boundary is obtained by sorting the brand point feature values ​​of historical candidate points according to their numerical values, and then taking the feature value corresponding to the quantile position closest to the high value end, such as the 95th quantile value.

[0128] The lower quantile boundary is obtained by sorting the brand location feature values ​​of historical candidate locations by numerical value and taking the feature value corresponding to the quantile position closest to the low value end, such as the 5th quantile value.

[0129] S4.2 Using the candidate point number as an index, the stable brand point features, point movement features, and point basic features are sequentially written into the feature record of the corresponding candidate point. Vectorized concatenation and correlation redundancy removal are performed according to a fixed field order to generate point prediction features.

[0130] Furthermore, using the candidate location number as an index, the stable brand location features, location flow features, and location basic features are mapped to the feature records under the same candidate location number. After mapping, the feature records of each candidate location are organized according to a fixed field order to ensure that the stable brand location features, location flow features, and location basic features maintain consistent field positions in the feature records of each candidate location.

[0131] The processed feature records are vectorized and concatenated to form a fixed-length feature vector for each candidate point. Then, the fixed-length feature vector is subjected to correlation and redundancy removal processing, retaining the feature content that is relevant to the site selection prediction and has low repetition, and removing the feature content that is too highly correlated and has obvious repetition, thus generating the point prediction features.

[0132] It should also be noted that the basic characteristics of the site are obtained by parsing the berth codes corresponding to the candidate site numbers in the initial site selection data.

[0133] S4.3 Input the predicted features of the location into the site selection learning model to perform conversion rate regression and risk classification, generate the predicted conversion rate and risk classification results, determine the risk level of the corresponding candidate location based on the risk classification results, and adjust the predicted conversion rate according to the risk adjustment coefficient corresponding to the risk level to generate the passenger flow conversion value.

[0134] Furthermore, the location prediction features corresponding to each candidate location are input into the location learning model in a fixed field order; the location prediction features first enter the feature encoding layer, which performs continuous feature standardization and category feature encoding on the location prediction features to form a unified feature representation;

[0135] The unified feature representation enters the shared learning layer based on gradient boosting trees. The shared learning layer based on gradient boosting trees extracts shared feature outputs that affect the conversion performance and risk performance of candidate locations through multiple decision trees in rounds. The shared feature outputs simultaneously enter the conversion rate regression branch and the risk classification branch. The conversion rate regression branch performs continuous value regression on the conversion probability of candidate locations based on the feature expressions related to customer flow absorption capacity, brand matching degree, and location operating conditions in the shared feature output, generating a continuously valued predicted conversion rate. The risk classification branch classifies and judges the operating risk categories of candidate locations based on the feature expressions related to customer flow fluctuations, path dependence, brand constraint deviation, and location risk in the shared feature output, generating risk classification results.

[0136] Subsequently, the risk level of the candidate locations is determined based on the risk classification results, and the corresponding risk adjustment coefficient is obtained according to the risk level. When the risk level is high, the risk adjustment coefficient causes a strong decreasing adjustment of the predicted conversion rate, and when the risk level is low, the risk adjustment coefficient causes a weak decreasing adjustment of the predicted conversion rate. Finally, the passenger flow conversion value is generated based on the predicted conversion rate and the risk adjustment coefficient, so that the passenger flow conversion value simultaneously reflects the conversion rate regression judgment and risk classification judgment of the candidate locations.

[0137] S5. Calculate the site selection score based on the passenger flow conversion value, sort the candidate sites according to the site selection score, and generate a site selection recommendation list.

[0138] S5.1 Normalize the passenger flow conversion value to generate a standardized conversion value. Perform constraint verification on the candidate locations based on the brand constraint code of the brand to be selected, generate a valid score set, calculate the location carrying capacity value based on the basic characteristics of the location, and perform weighted calculation on the standardized conversion value and the location carrying capacity value to generate a location selection score.

[0139] Furthermore, the passenger flow conversion value corresponding to each candidate location is normalized to convert the passenger flow conversion values ​​between different candidate locations to a unified numerical scale, generating standardized conversion values.

[0140] Subsequently, the candidate locations are constrained and verified according to the brand constraint code of the brand to be selected. If the candidate location meets the location restrictions in the brand constraint code, the candidate location is retained. If the candidate location does not meet the location restrictions in the brand constraint code, the candidate location is removed. The retained candidate locations are used to generate a valid score set.

[0141] Based on the basic characteristics of the locations, the operational carrying capacity of the candidate locations is extracted, and a comprehensive score is given to generate the location carrying capacity value. The standardized conversion value and location carrying capacity value corresponding to each candidate location in the effective score set are weighted and calculated to generate the location selection score, so that the location selection score reflects both the passenger flow conversion capacity and the operational carrying capacity of the candidate location.

[0142] S5.2 Sort the candidate sites according to the site selection score from high to low, and when the site selection scores are the same, adjust the order according to the passenger flow conversion value from high to low, and generate a site selection recommendation list.

[0143] Furthermore, the candidate locations in the valid score set are sorted from highest to lowest according to their location selection scores, placing the candidate locations with higher location selection scores first. When two or more candidate locations have the same location selection score, the customer flow conversion value corresponding to the candidate locations is compared, and the order of the candidate locations with the same location selection score is adjusted according to the customer flow conversion value from highest to lowest. After completing the sorting and order adjustment, the candidate location number, location selection score, customer flow conversion value, and risk level corresponding to each candidate location are retained to generate a location recommendation list.

[0144] like Figure 6 As shown in the overview diagram, the ranking curve recommended by this invention is closer to the actual conversion rate ranking curve, while the ranking curve of the local heat method deviates significantly. The enlarged view further reveals that within areas with concentrated high-conversion-potential locations, this invention's method more consistently ranks high-quality candidate locations at the top, while the local heat method is prone to misjudgments due to its focus on only local customer traffic. This figure illustrates that by constructing a street-level business flow map with conversion probability, searching for reverse customer acquisition paths, and reinforcing the path customer acquisition contribution value, this invention can generate location flow characteristics that more accurately reflect the actual direction of customer flow and the strength of customer acquisition contribution, thereby improving the reliability of the site selection recommendation list.

[0145] In summary, this invention constructs a directed business flow map with conversion probability from street objects and street paths, uses candidate locations as reverse absorption and convergence nodes to search for reverse customer acquisition paths, and then feeds back the contribution of the flow map nodes in the path based on the customer acquisition contribution value of the path. This achieves structured quantification of the source of incoming traffic, path continuity, and key node dependencies, so that candidate locations are no longer represented only by local traffic heat, but can form location flow characteristics that reflect the real direction of traffic spread and the strength of customer acquisition contribution. These characteristics are used as input for subsequent traffic capture values, conversion potential fields, and site selection learning models, thereby improving the accuracy of candidate location conversion potential expression and the reliability of site selection recommendations.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for selecting locations based on street-level customer flow conversion using machine learning algorithms, characterized in that, include: The data on the brand to be selected and the data on the target street are cleaned and coded, and candidate locations are extracted to generate initial data for site selection; Street objects and street paths are extracted from the initial site selection data. Street objects are mapped to movement map nodes, and street paths are mapped to directed movement map edges to generate street operation movement map. Candidate locations are used as reverse absorption and convergence nodes. Reverse customer acquisition paths are searched along the reverse direction of customer flow propagation. The path customer acquisition contribution value is calculated based on the node conversion probability. Then, the contribution of the movement map nodes in the reverse customer acquisition path is fed back based on the path customer acquisition contribution value to generate point movement characteristics. The passenger flow capture value is calculated based on the location movement characteristics, and the passenger flow capture value is mapped to the location spatial grid to generate the location transformation potential energy field. Passenger flow intention labels are generated according to continuous trajectory changes, and they are spliced ​​in time sequence to form an intention leap chain. The spatial potential energy value is dynamically corrected according to the forward and reverse leaps to generate effective passenger flow characteristics. Brand location features are generated based on effective customer flow characteristics. Brand location features, location flow features, and basic location features are then integrated to generate location prediction features. These location prediction features are then input into a site selection learning model for conversion rate regression and risk classification to generate customer flow conversion values. The site selection score is calculated based on the passenger flow conversion value. Candidate sites are then sorted according to the site selection score to generate a site selection recommendation list.

2. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 1, characterized in that, The process of cleaning and encoding the brand data and target street data to be selected, extracting candidate locations, and generating initial site selection data involves the following steps: Obtain the brand data and target street data for the proposed site selection. Remove invalid records and invalid shop records from the brand data and target street data respectively. Convert the cleaned brand data and target street data into brand constraint codes and shop codes. The store codes are matched and filtered according to the brand constraint code. The store records that meet the brand constraint code are identified as candidate locations. The candidate locations are then numbered and spatially associated to generate initial site selection data.

3. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 2, characterized in that, The steps for extracting street objects and street paths from the initial site selection data, mapping street objects to movement map nodes, and mapping street paths to directed movement map edges to generate a street operation movement map are as follows: Street objects are extracted from the initial site selection data, and the street objects are segmented into nodes according to the location of the passenger flow, the location of the passenger flow turning point, and the location of the business entrance. The segmented street objects are mapped to nodes of the flow diagram. Street paths are extracted from the initial site selection data, and the street paths are processed into edges according to the direction of passenger flow. The edge-directed street paths are mapped to edges of the directed flow diagram. Hampel local median filtering is applied to the passenger flow records corresponding to the directed flow line graph edges to remove abnormal flow records that deviate from the local median by more than the median deviation threshold. The flow intensity of the directed flow line graph edges is then corrected based on the filtered passenger flow records to generate stable flow line edges. Node indexes are created based on the edges of the directed traffic flow graph, edge connection indexes are created based on the stable traffic flow edges, and the node conversion probability of the corresponding directed traffic flow graph edge is calculated based on the outgoing edge traffic intensity of each traffic flow graph node to generate the street operation traffic flow graph.

4. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 1, characterized in that, The specific steps for searching reverse customer acquisition paths along the reverse direction of passenger flow propagation, using candidate locations as reverse absorption and convergence nodes, are as follows: The candidate locations are mapped to the corresponding traffic flow nodes in the street operation traffic flow map with the probability of traffic transfer, and the corresponding traffic flow nodes are set as reverse absorption and convergence nodes to generate a reverse search map for candidate locations. Starting from the reverse absorption and convergence node, the upstream flow graph nodes are read layer by layer in the opposite direction of the directed flow graph edge in the reverse search graph of the candidate point, and the first path to be expanded is extended by using the contribution upper bound priority queue to generate the reverse customer acquisition path.

5. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 4, characterized in that, The steps are as follows: First, the path customer acquisition contribution value is calculated based on the node conversion probability. Then, the contribution of the path customer acquisition contribution value is used to feed back the nodes in the flow map of the reverse customer acquisition path to generate point flow features. Read the directed flow graph edges that each reverse customer acquisition path passes through in the forward order of customer flow propagation, and calculate the path customer acquisition contribution value based on the node conversion probability and path length attenuation coefficient corresponding to the directed flow graph edge. Based on the contribution value of the path customer acquisition, the movement map nodes in the reverse customer acquisition path are fed back with contributions. The feed-back weight is determined according to the path distance from the movement map node to the reverse absorption and convergence node. The path customer acquisition contribution value is allocated to the corresponding movement map node according to the feed-back weight. Based on the allocation results, the reverse customer acquisition intensity, path continuity, upstream node concentration and key node dependence of the candidate points are extracted to generate point movement characteristics.

6. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 5, characterized in that, The specific steps for calculating passenger flow capture values ​​based on point-to-point movement characteristics and mapping these values ​​to a point-to-point spatial grid to generate a point-to-point potential energy field are as follows: The location flow characteristics are normalized, and the normalized location flow characteristics are weighted and fused to generate passenger flow capture values. A location spatial grid is established based on the boundary coordinates of the target block. The entrance coordinates of the candidate locations are projected onto the corresponding central grid, and the passenger flow capture values ​​are written into the central grid to generate the location capture grid. Using the point capture grid as the diffusion starting point, the passenger flow capture value is diffused along the stable movement line edge connected to the candidate point. The adjacent grids with connected movement lines are smoothed on both the spatial and movement lines to generate the point conversion potential energy field.

7. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 1, characterized in that, The process involves generating passenger flow intent tags based on continuous trajectory changes, assembling them in chronological order to form intent transition chains, and dynamically adjusting spatial potential energy values ​​based on forward and reverse transitions to generate effective passenger flow characteristics. The specific steps are as follows: Read continuous trajectory points of the same passenger flow object, segment the continuous trajectory points into trajectory segments according to the timestamp order, generate trajectory segment features based on the continuous changes of adjacent trajectory points, and generate passenger flow intention labels through time-series classification processing; Passenger flow intent tags that are continuous in time and reachable by grid lines are spliced ​​together to generate intent transition chains. The direction of intent change between adjacent passenger flow intent tags is determined according to the preset intent level. When adjacent passenger flow intent tags change from low to high intent level and the spatial potential energy value of the grid where the corresponding trajectory point is located increases in time sequence, it is marked as a positive transition. When the adjacent passenger flow intent label changes from a high conversion intent level to a low conversion intent level, and the spatial potential energy value of the corresponding trajectory point's grid decreases sequentially over time, it is marked as a reverse transition. The corresponding spatial potential energy value in the potential energy field of the point transformation is enhanced and corrected based on the positive transition, and the corresponding spatial potential energy value in the potential energy field of the point transformation is attenuated and corrected based on the reverse transition. Effective passenger flow characteristics are generated based on the corrected spatial potential energy values.

8. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 7, characterized in that, The specific steps for generating customer flow conversion values ​​are as follows: Based on the brand constraint codes of the brands to be selected, the effective customer flow characteristics are matched and calibrated to generate brand location characteristics. The brand location characteristics are then stabilized by quantile truncation and standardized to generate stable brand location characteristics. Using the candidate location number as an index, the stable brand location features, location flow features, and location basic features are sequentially written into the feature record of the corresponding candidate location. The features are then vectorized and concatenated according to a fixed field order, and correlation redundancy removal is performed to generate the location prediction features. The predicted features of the locations are input into the site selection learning model for conversion rate regression and risk classification, generating predicted conversion rates and risk classification results. Based on the risk classification results, the risk level of the corresponding candidate locations is determined, and the predicted conversion rate is adjusted by decreasing the risk adjustment coefficient corresponding to the risk level to generate passenger flow conversion values.

9. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 1, characterized in that, The site selection learning model is constructed by using location prediction features as the model input object, setting an input layer according to a fixed field order, setting a feature encoding layer after the input layer, setting a shared learning layer based on gradient boosting tree after the feature encoding layer, setting a conversion rate regression branch and a risk classification branch in parallel after the shared learning layer, and setting a passenger flow conversion value output layer.

10. The street-level customer flow conversion site selection method based on machine learning algorithms as described in claim 8, characterized in that, The specific steps for calculating site selection scores based on passenger flow conversion values, sorting candidate sites according to these scores, and generating a recommended site selection list are as follows: The customer flow conversion value is normalized to generate a standardized conversion value. The candidate locations are constrained and verified according to the brand constraint code of the brand to be selected, generating a valid score set. The location carrying capacity value is calculated based on the basic characteristics of the location. The standardized conversion value and the location carrying capacity value are weighted and calculated to generate a location selection score. Candidate sites are sorted from highest to lowest according to their site selection scores. If sites have the same site selection score, the order is adjusted according to the customer flow conversion value from highest to lowest to generate a site selection recommendation list.