Store location recommendation method in instant delivery scenario

Abstract

The present invention provides a store location recommendation method in an instant delivery scenario, comprising the following steps: Step 1: Data preprocessing module: comprising obtaining rider data, order data and context data to be processed, and obtaining processed data using feature extraction and graph construction technology; Step 2: Rider capacity modeling: comprising obtaining a preliminary processed rider mobility graph, a rider capacity model embedded with fine-grained rider capacity, and using this as input of a heterogeneous multi-graph of regional types; Step 3: Recommendation module based on heterogeneous multi-graphs: based on the heterogeneous multi-graph of regional types and edge embedding in the rider capacity model, to capture user preferences affected by multiple factors and interaction information between store areas and store types; adopting the method, a reasonable multi-graph attention network algorithm is provided to achieve better store location recommendation results on data sets with different distributions.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method for recommending store locations in an instant delivery scenario. Background Art

[0002] First, we briefly introduce the online-to-offline (O2O) model, existing store location recommendation techniques, and the application of graph neural networks in recommendation techniques.

[0003] 1. Online to offline (O2O) model;

[0004] The online-to-offline (O2O) model refers to the combination of offline commerce and the Internet, making the Internet a platform for offline transactions. This is an increasingly important business model in recent years. Generally speaking, the O2O model first has an O2O platform with a large number of users, and users can use the platform to shop online. After the user completes the purchase, the subsequent processes are completed by the platform, such as payment, logistics and delivery.

[0005] A typical example of the O2O model is on-demand delivery, where users buy online through the O2O platform, and then the courier picks up the order from the physical store and delivers it to the user. Compared with traditional stores that only accept orders from users who come in person, stores under the O2O model cooperate with delivery platforms, providing users with great shopping convenience.

[0006] 2. Store location recommendation technology;

[0007] Small-scale survey-based methods and large-scale data-driven methods are two major existing store location recommendation techniques. Survey-based methods design questionnaires and interviews to obtain feedback from small-scale volunteers. Data-driven methods build feature-based machine learning models.

[0008] However, the survey method is based on the feedback of participants and inevitably has personal bias. Existing research on data-driven methods mainly focuses on traditional physical stores, that is, stores, warehouses, and factories. This research is not suitable for solving the store location recommendation problem under the O2O model.

[0009] 3. Application of neural networks in recommendation technology;

[0010] Graph neural network work is usually divided into two categories: spectral-based GNNs and spatial-based GNNs. However, traditional graph neural networks cannot be directly applied to heterogeneous graphs. Existing studies either use meta-paths to process heterogeneous graphs, converting heterogeneous graphs into homogeneous graphs, or directly process heterogeneous graphs without meta-paths.

[0011] Recommendation technology is an important application of graph neural networks. The core information is the interaction between users and items, which can build a user-item bipartite graph. In addition, the recommendation system also adds other auxiliary information (such as social networks and knowledge graphs) to solve data sparsity and cold start problems.

[0012] However, due to the different data characteristics, graph-based recommendation methods cannot be directly applied to store location recommendation scenarios under the O2O model; research on traditional physical stores cannot solve new problems under the O2O model, including dynamic supply caused by delivery capabilities and delivery strategies, and different user needs caused by delivery distances and user preferences. Summary of the invention

[0013] To solve the above problems, the present invention discloses a store location recommendation method in an instant delivery scenario, which adopts a rider capacity model based on a multi-semantic relationship graph attention network and a recommendation model based on heterogeneous multi-graphs, which integrates rider capacity, user preferences and contextual features. This method outperforms the most advanced baselines in various indicators.

[0014] The store location recommendation method in the instant delivery scenario includes the following steps:

[0015] Step 1: Data preprocessing module: including obtaining the rider data, order data and context data to be processed, and using feature extraction and graph construction techniques to obtain the processed data;

[0016] Step 2: Rider capacity modeling: This includes obtaining a preliminarily processed rider mobility graph and a rider capacity model embedded with fine-grained rider capacity as input for heterogeneous multi-graphs of regional types;

[0017] Step 3: Recommendation module based on heterogeneous multi-graph: Based on the heterogeneous multi-graph of regional types and edge embedding in the rider capacity model, a recommendation model based on heterogeneous multi-graph with node-level and semantic-level aggregation is developed to capture user preferences affected by multiple factors and the interaction information between store areas and store types; using the method, a reasonable multi-graph attention network algorithm is provided to achieve better store location recommendation results on data sets with different distributions.

[0018] The present invention further optimizes that O2-SiteRec is a store location recommendation technology based on multiple graphs of attention networks under the O2O model. On the supply side, it uses an attention network rider capacity model based on multiple semantic relationship graphs to obtain the rider capacity of the O2O platform and uses delivery time to quantify the express delivery capacity. At the same time, a recommendation model based on heterogeneous multiple graphs is adopted, which integrates rider capacity, user preferences and contextual features.

[0019] In step 1:

[0020] Order data includes spatial information, rider’s temporary information and environmental information; spatial information includes the source location (i.e. store location) and destination location (i.e. user location) of order delivery;

[0021] The temporary information reported by the riders includes the order node time: order creation time, order pickup time and delivery time; the environmental information includes the distance between the user and the store and the store type;

[0022] The rider’s trajectory data refers to the continuous location information obtained from the rider’s smartphone while he is working. With the rider’s consent, this location information will be uploaded to the platform every 20 seconds. The information includes the rider’s ID, GPS location and timestamp.

[0023] Contextual data includes points of interest (POI) and road network information; POI refers to meaningful points on the map with non-geographical significance, and POI data based on open API can be obtained from online map service providers; POI and road network information can describe the characteristics and significance of different areas in the city, and play a very important role in store location recommendations.

[0024] Furthermore, in step 1:

[0025] Order data includes spatial information, rider’s temporary information and environmental information; spatial information includes the source location (i.e. store location) and destination location (i.e. user location) of order delivery;

[0026] The temporary information reported by the riders includes the order node time: order creation time, order pickup time and delivery time; the environmental information includes the distance between the user and the store and the store type;

[0027] The rider’s trajectory data refers to the continuous location information obtained from the rider’s smartphone while he is working. With the rider’s consent, this location information will be uploaded to the platform every 20 seconds. The information includes the rider’s ID, GPS location and timestamp.

[0028] Contextual data includes points of interest (POI) and road network information; POI refers to meaningful points on the map with non-geographical significance, and POI data based on open API can be obtained from online map service providers; POI and road network information can describe the characteristics and significance of different areas in the city, and play a very important role in store location recommendations.

[0029] The present invention further preferably, step 2: rider capacity modeling: for each subgraph of period t First, two kinds of neighborhoods of each node are found from the regional geographic graph and the rider mobility graph respectively: geographic semantic neighborhood and mobility semantic neighborhood; node embeddings are obtained through geographic semantic aggregation and mobility semantic aggregation, and they are connected as edge embeddings to reconstruct the properties of the mobility edge of the rider mobility graph; then the edge embeddings are used as the input of the region-type heterogeneous multigraph; the training of this module is used as an auxiliary training for the main task.

[0030] In the present invention, the geographic semantic aggregation is further preferably:

[0031] Each region i has an initial embedding where d1 is the embedding size; for the geographic neighborhood of region i We use the distance between regions to calculate the weight, which is defined as: where dis(i, j) represents the distance between region i and region j. After the lth geosemantic aggregation, the embedding of region i can be obtained: Defined as:

[0032]

[0033] in, represents the embedding of region j after (l-1) geo-semantic aggregation, and σ(·) is an activation function.

[0034] In the present invention, the mobile semantic aggregation refers to: for the semantic neighborhood of region i Calculate the weights α between different neighborhoods according to GAT mob (i,j), which is defined as Then we get the embedding b of region i after move semantic aggregation. s,i , is defined as:

[0035]

[0036] where ψ is the parameterized attention vector; and are the initial embeddings of region j and region i, respectively.

[0037] The present invention further preferably obtains edge embedding by reconstructing the properties of the moving edges in the rider movement graph: the final embedding b of region i i It is calculated by combining the geographic embedding and the mobile embedding, defined as:

[0038]

[0039] Where W b weighted for trainability; we merge the two region embeddings into an edge embedding It is then fed into the MLP for delivery time prediction, which is defined as

[0040]

[0041] W1 is a trainable weight; the loss function is defined as:

[0042]

[0043] where |E| is the number of edges observed in the rider movement graph; DT i,j is the ground truth delivery time in the rider mobility graph; edge embeddings containing rider capabilities are used as input to the regional heterogeneous multigraph.

[0044] The present invention is further preferred that in the region type heterogeneous graph in step 3: each region has two views: a store view and a user view; in the store view, the store region is defined as a region under the store view, indicating the types of stores available in the region; in the user view, the user region is defined as a region under the user view, indicating which type of store is preferred by users in this region; both regions are subsets of the total region;

[0045] Using a heterogeneous graph to model the relationship between regions in different views and store store types in cycle t; where V h Contains store area node s∈S, customer area node u∈U, store type node a∈E; is the node feature, edge Include E S-A ,

[0046] (1) Indicates that user area u is within the delivery range of store area s during period t. Based on the historical delivery range and distance, the probability is calculated to determine which nodes have edges, while considering the impact of courier capacity on store delivery range. su,t It is a multidimensional vector representing attributes, consisting of various factors that affect the interaction between customers and stores.

[0047] (2)E S-A (s,a) means that there is a store of store type a in store area s. Edge attribute φ sa is a multidimensional vector containing business features and historical order numbers;

[0048] (3) Indicates that users in user region u prefer stores of store type a in period t. Edge attribute φ ua,t is the number of historical transactions;

[0049] In addition, user order preferences and the interactions between user regions and store regions change over time, which means that these edges are different in different periods. Because of the variability of edges, we define a region-type mixed multigraph, i.e., a multigraph in Represents edge sets at different periods;

[0050] Therefore, the store location recommendation problem under the O2O model can be expressed as follows:

[0051] The input of this formula is a heterogeneous multi-graph G of region type h , Riders move multiple images G c and regional geographic map G ge , in order to learn a model F θ This model can predict the number of orders of target type a in region s Through this model, for a given store type a, it can be predicted that in all candidate areas S = (s1, s2, ..., s j ) and select the candidate area with the largest number of orders, that is, the top ranking, as the recommendation result.

[0052] Beneficial effects of the present invention:

[0053] 1. This is the first study on store location recommendation technology under the O2O model, and it analyzes the unique characteristics of stores under the O2O model from two aspects: supply (e.g., additional delivery capacity) and demand (e.g., changing user consumption patterns), both of which play an important role in store location recommendation and make the research work not limited to traditional physical store location recommendation.

[0054] 2. Application of graph neural networks in recommendation systems; heterogeneous multi-graphs are used to model the relationships between store areas, user areas, and store types, taking into account the impact of multi-graph structures and edge attributes. On the supply side, in order to obtain rider capacity, a rider movement multi-graph is constructed, and delivery time is used to quantify express delivery capabilities. Based on this graph, a delivery capability model based on a multi-semantic relationship graph attention network is designed to obtain fine-grained express delivery capability features. On the demand side, a heterogeneous multi-graph of regional types is constructed, which simulates the complex semantic relationships between multiple views of regions and storage types in different periods.

[0055] 3. As the first store location recommendation model in the O2O mode, O2-SiteRec has better technical advantages than commonly used methods. The model is compared with three types of baselines, including store site recommendation methods, general graph-based recommendation methods, and heterogeneous graph methods. Experimental results show that the O2-SiteRec method outperforms other state-of-the-art methods, improving the performance by 12.18% over the NDCG@3 method and 9.01% over the Precision@3 method.

[0056] 4. A heterogeneous multi-graph based recommendation model is designed to capture user preferences and other interaction information affected by multiple heterogeneous factors for site recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the system structure of the method provided by the example of this application;

[0058] Figure 2 It is a processing flow chart of the rider capacity model provided by this application;

[0059] Figure 3 This is a processing flow chart of the recommendation model based on heterogeneous multi-graphs provided in this application. DETAILED DESCRIPTION

[0060] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. It should be noted that the words "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to directions in the accompanying drawings, and the words "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0061] This embodiment provides a store location recommendation method in an instant delivery scenario, such as Figure 1 As shown: The following steps are included:

[0062] Step 1: Data preprocessing module: Step 2: Rider capacity modeling; Step 3: Recommendation module based on heterogeneous multi-graphs.

[0063] The step 1 data preprocessing module: first extract the geographical features of each region, including POI set, POI diversity, transportation convenience and store diversity. These geographical features are regarded as attribute sets of store area nodes and user area nodes. The definitions of these geographical features are as follows:

[0064] 1. POI set: Define a vector where each dimension represents the number of POIs in a specific category in each region.

[0065] 2. POI diversity: Define an information entropy as a fraction of all categories of POIs that appear in a region

[0066] 3. Traffic convenience: Define a vector consisting of the number of intersections and roads in an area.

[0067] 4. Store diversity: defined as an information entropy as a fraction of all types of stores present in an area

[0068] In addition, for a specific category a in a store area s, commercial features can be extracted, which include competitiveness and complementarity. These features are used as E S-A Attributes;

[0069] 1. Competitiveness: It reflects the competition among similar stores in the vicinity and is defined as the number of similar stores in the area divided by the total number of nearby stores.

[0070] 2. Complementarity: It reflects the cooperative relationship between different types of stores in the vicinity and is defined as in Indicates that type a is in region s * The number of stores, is of type a in region s * The average number of stores, N set (a,a * ) is the number of occurrences of pairs of types a and , N A is the number of all store types.

[0071] The extracted features can also affect the interaction between user areas and store areas, including distance and historical transaction records, which are used as Properties:

[0072] 1. Distance: This is defined as the distance between the user area and the store area

[0073] 2. Historical transaction records: defined as the number of historical transactions between the user area and the store area.

[0074] Step 2: Rider capacity model: Due to the rider's movement behavior, it is not easy to obtain fine-grained rider capacity. Intuitively, geographically adjacent areas have similar rider capacity, and areas with mobility relationships have certain correlations. We designed a rider capacity model based on a multi-semantic relationship graph attention network to obtain rider capacity.

[0075] For each subgraph of period t Execute as Figure 2The steps are shown. First, two kinds of neighborhoods of each node are found from the regional geographic graph and the rider mobility graph respectively: geographic semantic neighborhood and mobility semantic neighborhood. Node embeddings are obtained through geographic semantic aggregation and mobility semantic aggregation, and they are connected as edge embeddings to reconstruct the properties of the mobility edge of the rider mobility graph. Then the edge embeddings are used as the input of the region-type heterogeneous multigraph. In addition, the training of this module is used as an auxiliary training for the main task.

[0076] 1) Geographic semantic aggregation: Each region i has an initial embedding where d1 is the embedding dimension. For the geographic neighborhood of region i We use the distance between regions to calculate the weight, which is defined as:

[0077]

[0078] where dis(i, j) represents the distance between region i and region j. After the lth geosemantic aggregation, the embedding of region i can be obtained: Defined as:

[0079]

[0080] in, represents the embedding of region j after (l-1) geo-semantic aggregation, and σ(·) is an activation function.

[0081] 2) Mobile semantic aggregation: For the semantic neighborhood of region i We calculate the weights α between different neighborhoods according to GAT mob (i,j), which is defined as

[0082] Then we get the embedding b of region i after move semantic aggregation. s,i , is defined as:

[0083]

[0084] where ψ is the parameterized attention vector. and are the initial embeddings of region j and region i, respectively.

[0085] 3) Obtain edge embedding by reconstructing the properties of the moving edges in the rider movement graph: the final embedding b of region i i It is calculated by combining the geographic embedding and the mobile embedding, defined as:

[0086]

[0087] Where W bWe merge the two region embeddings into an edge embedding It is then fed into the MLP for delivery time prediction, which is defined as W1 is the trainable weight. The loss function is defined as:

[0088]

[0089] Where |E| is the number of edges observed in the rider movement graph. DT i,j is the ground truth delivery time in the rider mobility graph. The edge embeddings containing rider capabilities are used as the input of the regional heterogeneous multigraph.

[0090] Step 3: Recommendation model based on heterogeneous multi-graph

[0091] User preferences determine whether a region is suitable for a particular store type. We construct a heterogeneous multigraph of region types to model the complex relationships between store regions, user regions, and store types. This heterogeneous multigraph-based recommendation model obtains user preferences and interaction information between store regions and store types to make site recommendations. The model contains node-level aggregation to consider the impact of edge attributes (e.g., multiple factors) and temporal semantic-level aggregation to consider multi-graph structures. In order to further consider multiple factors that affect the interaction between user regions and store regions (i.e., delivery distance, historical interaction records, and rider capacity), we set , as a multidimensional vector corresponding to multiple factors.

[0092] Figure 3 The design of a heterogeneous multi-graph recommendation model consisting of five main steps is described.

[0093] It includes (1) node attribute fusion; using an embedding layer to represent the potential features of each node ID. Formally, we use the initial embedding layer and denotes store region s, user region u, and store type a respectively, where d2 is the embedding layer size. Then the initial embedding layer and node attributes are fused to obtain the node fused embedding. The fused embedding of store region, user region, and store type is defined as and

[0094] (2) Facing The edge attribute fusion of We incorporate the edge attribute φ us,t With edge embedding layer A new composite edge attribute is obtained from the rider capacity model, which contains multiple factors, including courier capacity, delivery distance, and historical interactions. The new edge attribute is defined as

[0095] (3) Node-level aggregation to obtain store region embeddings and store type embeddings at different periods; for the subgraph in period t Store area node embedding and store type node embedding are obtained through store area modeling and store type modeling respectively.

[0096] (i) Store area modeling: On the one hand, for the target node store area, Learn the preferences of users in the delivery area; on the other hand, through E S-A Captures high-order interaction information between store areas and store types. Embedding layer of store area node s after the lth aggregation It is defined as:

[0097]

[0098] in Weighted for trainability. and N S-A (s) represents the neighbors of store area s, which are based on and E S-A Aggre is an aggregate function, which will be defined in detail later. and is the embedding layer of the user region node u, store type node a, and store region node s after the (l-1)th aggregation. For the customer region node u, the embedding layer Can be After l aggregations, the type of store it prefers is obtained, defined as:

[0099]

[0100] in is a trainable weight, Indicates based on The neighbors of user area u are represented by

[0101] (ii) Store type modeling: For the target node store type, we use E S-A Captures the storage area and high-level information that it interacts with. The embedding layer of the storage type node A after the first aggregation Defined as

[0102]

[0103] in, is the trainable weight, N A-S (a) indicates that the S-A Neighbors of store type a.

[0104] (iii) Aggregation Function (Aggre): Considering the different types and attributes of edges in the inverted heterogeneous graph, we design an aggregation function that estimates the importance of each source node through node attributes, edge attributes and edge type.

[0105] The multi-head attention mechanism is used to calculate the importance score. In the following, we take the store area node (i.e., the target node) and the user area node (i.e., the source node) as examples. First, in order to consider the influence of different edge types, we set a trainable parameter W e , which is shared by the same edge type. Second, to exploit edge properties, we combine the embedding layer Source node (e.g., user region u) and edge attribute φ us,t As the fusion vector. For the i-th attention head, we use linear projection Project the above fusion vector into the i-th key vector. The i-th key vector K i (u) is defined as:

[0106]

[0107] Similarly, using linear projection Project the target node (e.g., storage area node s) into the i-th query vector. The i-th query vector Q i (s) is defined as Finally, the query vector, key vector, and edge type are combined to calculate the importance of the i-th attention head and normalized using the softmax function to obtain the weight of the i-th attention head, which is defined as

[0108] α i (u,s)=softmax(σ(K i (u)W e Q i (s) T ))

[0109] Then, the information of neighboring nodes is aggregated and the outputs of multi-head attention are concatenated to obtain the final representation of Aggre, which is defined as:

[0110]

[0111] where || represents the vector concatenation operator.

[0112] (4) Fusion of results from different subgraphs through temporal semantic level aggregation: According to the embedding of store areas and store types in each subgraph during period t, we first connect them to obtain sq,t The embedding layer Considering that the importance of different regions and store types in different periods is usually different, we use a multi-head attention mechanism to calculate the importance of each period. For the i-th attention head, we project the concatenated embeddings of all periods into the i-th key vector, according to the linear projection definition for:

[0113]

[0114] Then calculate the time period t j The embedding layer Then use linear projection Projected to the i-th query vector middle:

[0115]

[0116] The attention weight is calculated as the inner product of the query vector and the key vector and normalized using the softmax function, which is defined as Finally, we calculate the key vector The weights and concatenate the outputs of the multi-head attention as the final embedding H sa , defined as:

[0117]

[0118] (5) Order quantity forecast; H sa Embedding Multilayer Perceptron (MLP) for order number Prediction, which is defined as The loss function O2 of this task is defined as:

[0119]

[0120] Where |N| is the number of observations, p sa is the ground truth sequence number of store region s on storage type a.

[0121] (6) Model training process

[0122] The objective function formula combining these two tasks is as follows:

[0123] Loss=O2+βO1

[0124] where β is a trade-off parameter. To optimize the objective function, we adopt Adam as the optimizer in our implementation. There are four initial embedding layers in our model, which are randomly initialized and learned jointly during the training phase. To mitigate overfitting, we adopt a dropout strategy in our model.

[0125] Embodiment 1: for recommending store locations;

[0126] The store location recommendation technology designed based on patented technology can analyze and calculate the locations of various optional stores on the O2O platform and give recommended locations.

[0127] For example: common restaurants and physical stores that register on the O2O platform and participate in delivery services.

[0128] Example 2: For predicting the number of orders for stores at a specified location, a recommendation model based on heterogeneous multi-graphs can provide order predictions within a period

[0129] In the recommendation model based on heterogeneous multi-graph, the order quantity is predicted. Taking the O2O platform as an example, the recommendation model can predict the order quantity within the cycle time, coordinate the orders and delivery resources in the region, and make reasonable dispatch decisions.

[0130] Embodiment 3: used to obtain user preferences and analyze user behavior;

[0131] This model analyzes the correlation between users' order preferences and store distances, integrates user information through a heterogeneous graph of regional types, and reflects the relationship between user behavior and stores; the core technology can be applied to smart delivery application scenarios to achieve intelligent recommendations for users.

[0132] Embodiment 4: Provides a method for analyzing rider capacity;

[0133] This model analyzes the rider capacity of the riders and can improve the rider allocation strategy by analyzing the delivery time and delivery range as well as the daily cycle time of the orders.

[0134] Example 5: A representation method for integrating user and store information into a graph network is provided; the core technology can be applied to construct a user-store graph, integrating user preferences and store features to obtain potential shopping tendencies

[0135] The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above technical features.

Claims

1. A store location recommendation method in an instant delivery scenario, characterized by: The following steps are involved: Step 1: Data preprocessing module: including obtaining the rider data, order data and context data to be processed, and using feature extraction and graph construction techniques to obtain the processed data; Step 2: Rider capacity modeling: including obtaining a preliminary processed rider mobility graph, a rider capacity model embedded with fine-grained rider capacity, and using it as the input of the heterogeneous multi-graph of regional type; Rider capacity modeling: for each subgraph of period t First, two kinds of neighborhoods of each node are found from the regional geographic graph and the rider mobility graph respectively: geographic semantic neighborhood and mobility semantic neighborhood; node embedding is obtained through geographic semantic aggregation and mobility semantic aggregation, and connected as edge embedding to reconstruct the properties of the mobility edge of the rider mobility graph; then edge embedding is used as the input of the heterogeneous multigraph of regional type; the training of this module is used as an auxiliary training of the main task; the geographic semantic aggregation refers to: Each region i has an initial embedding where d1 is the embedding size; for the geographic neighborhood of region i We use the distance between regions to calculate the weight, which is defined as: where dis(i, j) represents the distance between region i and region j; after the lth geosemantic aggregation, the embedding of region i can be obtained: Defined as: in, represents the embedding of region j after (l-1) times of geo-semantic aggregation, and σ(·) is an activation function; The mobile semantic aggregation refers to: for the semantic neighborhood of region i Calculate the weights α between different neighborhoods according to GAT mob (i,j), which is defined as Then we get the embedding b of region i after move semantic aggregation. s,i , is defined as: where ψ is the parameterized attention vector; and are the initial embeddings of region j and region i, respectively; Step 3: Recommendation module based on heterogeneous multi-graph: Based on the heterogeneous multi-graph of regional types and edge embedding in the rider capacity model, a recommendation model based on heterogeneous multi-graph with node-level and semantic-level aggregation is proposed to capture user preferences affected by multiple factors and the interaction information between store areas and store types. User preferences affected by multiple factors and the interaction information between store areas and store types are input; using the method, a reasonable multi-graph attention network algorithm is provided to achieve better store location recommendation results on data sets with different distributions.

2. The store location recommendation method in the instant delivery scenario according to claim 1 is characterized by: In step 1: Order data includes spatial information, rider’s temporary information and environmental information; spatial information includes the source location of order delivery, i.e. store location, and destination location, i.e. user location; The temporary information reported by the riders includes the order node time: order creation time, order pickup time and delivery time; the environmental information includes the distance between the user and the store and the store type; The rider’s trajectory data refers to the continuous location information obtained from the rider’s smartphone while he is working. With the rider’s consent, this location information will be uploaded to the platform every 20 seconds. The information includes the rider’s ID, GPS location and timestamp. Contextual data includes points of interest (POIs) and road network information; POI refers to a meaningful point on a map that has non-geographical significance. POI data based on an open API is obtained from online map service providers. POI and road network information can describe the characteristics and significance of different areas in a city, and play a very important role in store location recommendations.

3. The store location recommendation method in the instant delivery scenario according to claim 1 is characterized by: The edge embedding is obtained by reconstructing the properties of the moving edges in the rider's movement graph: the final embedding b of region i i It is calculated by combining the geographic embedding and the mobile embedding, defined as: Where W b weighted for trainability; we merge the two region embeddings into an edge embedding It is then fed into the MLP for delivery time prediction, which is defined as W1 is a trainable weight; the loss function is defined as: where |E| is the number of edges observed in the rider movement graph; DT i,j is the ground truth delivery time in the rider mobility graph; edge embeddings containing rider capabilities are used as input to the regional heterogeneous multigraph.

4. The store location recommendation method in the instant delivery scenario according to claim 1 is characterized by: The heterogeneous graph of region types in step 3: each region has two views: a store view and a user view; In the store view, define the store area as the area under the store view, indicating the types of stores available in the area; in the user view, define the user area as the area under the user view, indicating which type of store is preferred by users in this area; both areas are subsets of the total area; Using a heterogeneous graph to model the relationship between regions in different views and store store types in cycle t; where V h Contains store area node s∈S, customer area node u∈U, store type node a∈A; is the node feature, edge Include E S-A , (1) Indicates that user area u is within the delivery range of store area s during period t; Based on the historical delivery range and distance, the probability is calculated to determine which nodes have edges, while considering the impact of express delivery capacity on the store delivery range; φ su,t It is a multidimensional vector representing attributes, consisting of various factors that affect the interaction between customers and stores; (2)E S-A (s,a) means that there is a store of store type a in store area s; the edge attribute φ sa is a multidimensional vector containing business features and historical order numbers; (3) Indicates that users in user region u prefer stores of store type a in period t; edge attribute φ ua,t is the number of historical transactions; In addition, user order preferences and the interactions between user regions and store regions change over time, which means that these edges are different in different periods; because of the variability of the edges, we define a region-type mixed multigraph, i.e., the multigraph in Represents edge sets at different periods; Therefore, the store location recommendation problem under the O2O model can be expressed as follows: The input of this formula is a heterogeneous multi-graph G of region type h , Riders move multiple images G c and regional geographic map G ge , in order to learn a model F θ ; This model can predict the number of orders of target type a in region s Through this model, for a given store type a, it can be predicted that in all candidate areas S = (s1, s2, ..., s j ) and select the candidate area with the largest number of orders, that is, the top ranking, as the recommendation result.

Images