Intelligent action planning recommendation method and system based on neighborhood enhancement

Abstract

The invention relates to the technical field of electric digital data processing, in particular to an action planning intelligent recommendation method and system based on neighborhood enhancement, and the method comprises the following steps: constructing a corresponding local action sequence diagram; constructing global user cross action site map features and global transfer map features; converting each local action sequence diagram into a local action sequence diagram vector containing a local action transfer feature; obtaining a local action sequence diagram with enhanced virtual neighborhood feature vector semantic representation capability; and sorting and selecting an optimal place as the action planning recommendation of the intelligent navigation and travel service according to the action planning prediction scores in the planning recommendation list. According to the method and the system provided by the invention, the user behavior pattern description is more objective, comprehensive and accurate, and the recommendation effect of action planning is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical digital data processing, and in particular, to an intelligent recommendation method and system for action planning based on neighborhood enhancement. Background Art

[0002] Applications based on location services enable users to share their current locations in real time, which provides an important data source for the construction of personalized services and recommendation systems. By collecting the action data of a large number of users, the system can analyze and mine the behavior patterns of users, so as to provide accurate personalized suggestions in the action planning recommendation task. This data-driven action planning not only improves the user experience, but also brings new opportunities to the business field.

[0003] However, although these location data have significant value, they do not fully represent the action characteristics of users. Since the action data of users are uploaded by users themselves and are limited to the locations they choose to share, the incompleteness of the data has become the main obstacle to accurately capturing the behavior patterns of users. This incompleteness of the data may cause the recommendation system to fail to fully understand the needs of users, thus affecting the effect of action planning. Therefore, how to effectively process these incomplete data and fill in the missing action information has become a major challenge to improve the accuracy of action planning recommendations.

[0004] Through the statistics of user action data, it can be found that increasing data integrity has a significant effect on improving the effect of action planning recommendations. First of all, the complete user action sequence provides richer behavior information, enabling the recommendation system to more comprehensively capture the action characteristics and interest preferences of users, thus reducing the deviation of recommendation results and making the recommendation more personalized and closer to the actual needs of users. Secondly, improving data integrity is crucial for learning the collaborative relationship between users. When the system has more user action data, it can more effectively mine the similarities between different users, thus improving the effect of collaborative filtering recommendations and providing more accurate recommendations. In addition, data integrity can also reduce the negative impact of the sparsity problem on the model performance. In many recommendation tasks, sparse data will cause the model to fail to fully learn the preference patterns of users, thus affecting the accuracy of recommendations. By increasing the data integrity, the recommendation model can better fill in the data gaps and improve the learning effect and generalization ability of the recommendation model.

[0005] Many existing recommendation methods enhance the feature representation in user-generated behaviors by constructing neighborhood information, thereby improving the recommendation effect of the model. Neighborhood information is usually generated by mining other user behaviors that are similar to the target user's behavior. These similarities can be achieved through collaborative filtering methods. The current user's behavior and preferences are predicted by the behavior of similar users. However, the use of neighborhood information only stays at the level of feature enhancement of the original user-generated data, and does not fundamentally solve the problem of incomplete data. Therefore, how to use neighborhood information to fill in the missing information in user action data and comprehensively improve data integrity is a key issue that needs to be solved urgently. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide an action planning intelligent recommendation method and system based on neighborhood enhancement, which introduces a virtual neighborhood feature representation combined with the action features generated by the user to enhance the integrity of the user action sequence, and forms a more objective, comprehensive and accurate description of the user behavior pattern by learning the local and global feature representation of the user behavior sequence containing the virtual neighborhood features, thereby improving the action planning recommendation effect.

[0007] The present invention is achieved through the following technical solutions: An action planning intelligent recommendation method based on neighborhood enhancement comprises the following steps: S1: Add the corresponding virtual neighborhood action set to the action sequence composed of each single-point action in the user shared action dataset, and construct the corresponding local action sequence graph; S2: Construct global user cross-action location map features and global transition map features; S3: Introduce the constructed global user cross-action location graph features and global transition graph features into each local action sequence graph, and transform each local action sequence graph into a local action sequence graph vector containing local action transition features; S4: Input each local action sequence graph vector containing local action transfer features into the contrastive learning framework for contrastive learning, and obtain a local action sequence graph with enhanced semantic representation ability of the virtual neighborhood feature vector; S5: A multi-task learning method is used to train the feature representation of the local action sequence graph with enhanced semantic representation capability of the virtual neighborhood feature vector, and its action sequence graph vector is converted into an action plan recommendation list. The action plans in the plan recommendation list are sorted according to their predicted scores, and the optimal location is selected as the action plan recommendation for intelligent navigation and travel services.

[0008] Furthermore, each single-point action in step S1 is expressed as formula (1), and the action sequence composed of each single-point action is expressed as formula (2): (1); (2); Wherein: represents a single-point action, represents a user, represents the action location, represents the action time, represents the action location category, represents the geographical information corresponding to the action location, represents a tuple, represents an action sequence composed of each single-point action, represents the th single-point action, represents the total number of single-point actions in the action sequence composed of each single-point action, represents a set.

[0009] Further, in step S1, adding a corresponding virtual neighborhood action set to the action sequence composed of each single-point action of the user's shared action data and constructing a corresponding local action sequence graph includes the following steps: S11: Using the k-means clustering method to cluster the two-dimensional coordinates in the geographical information corresponding to the action location in each single-point action; S12: Combining each action location in each single-point action with its corresponding cluster to represent the geographical information of each action location; S13: Using each single-point action in the action sequence composed of each single-point action as a node of the local action sequence graph generated by the user to generate a set of nodes generated by the user; S14: Generating paired virtual neighborhood nodes for each node in the set of nodes generated by the user; S15: Introducing an edge between each user-generated node and its corresponding virtual neighborhood node, introducing an edge between each time-adjacent single-point action among all nodes, and introducing an edge between two nodes with a set distance, to form a corresponding local action sequence graph.

[0010] Further, in step S2, constructing global user cross-action location graph features and global transfer graph features through the following method: S21: Using the propagation and aggregation functions in the lightweight graph convolutional layer, updating the user embedding vector according to Equation (3), updating all action location embedding vectors according to Equation (4), and performing average aggregation on the user embedding vector and all action location embedding vectors respectively according to Equation (5) to obtain the final global user embedding vector and the embedding vectors of the action locations and their related attributes in the global user cross-action location graph: (3); (4); (5); Wherein: represents the index layer serial number, represents the number of index layers, represents the user embedding vector of the index layer, represents the user embedding vector of the index layer, represents the set of neighbors of the user in the adjacency matrix ; represents the set of neighbors of the action location in the adjacency matrix ; represents the action location 's embedding vector, represents the action location embedding vector of the index layer, represents taking the absolute value, represents the action location embedding vector of the index layer, represents the global user embedding vector, represents the embedding vector of the action location and its related attributes in the global user cross-action location graph; S22: Construct a global transfer graph on the embedding vector of the action location and its related attributes in the global user cross-action location graph, and generate a global transfer graph node pairwise connection matrix, where the global transfer graph node pairwise connection matrix includes a global transfer graph in-degree matrix and a global transfer graph out-degree matrix; S23: Assign values to the nodes of each action sequence in the action sequence graph, and normalize the global transfer graph in-degree matrix between the nodes of the action sequence graph and the global transfer graph out-degree matrix between the nodes of the action sequence graph row by row; S24: Update according to Equation (6) through a gated graph neural network based on long short-term memory units to construct global user cross-action location graph features and global transfer graph features: (6); Wherein: represents the input feature of the node in the global transfer graph at time, represents the in-degree parameter matrix for the linear transformation of the global transfer graph, represents the embedding vector of the action location and its related attributes in the global user cross-action location graph at time, represents the global transfer graph in-degree matrix, represents the in-degree bias matrix for the linear transformation of the global transfer graph, represents the out-degree parameter matrix for the linear transformation of the global transfer graph, represents the out-degree matrix of the global transfer graph, represents the out-degree bias matrix for the linear transformation of the global transfer graph, represents the value of the input gate of the long short-term memory network of the global transfer graph at time represents the sigmoid function, represents the weight matrix of the input gate of the long short-term memory network of the global transfer graph, represents the weight matrix of the input gate of the long short-term memory network of the global transfer graph for the node state at the previous time step, represents the bias vector of the input gate of the long short-term memory network of the global transfer graph, represents the global transfer graph the value of the forget gate of the long short-term memory network at time represents the weight matrix of the forget gate of the long short-term memory network of the global transfer graph, represents the weight matrix of the forget gate of the long short-term memory network of the global transfer graph for the node state at the previous time step, represents the bias vector of the forget gate of the long short-term memory network of the global transfer graph, represents the global transfer graph the memory cell state of the long short-term memory network at time represents the global transfer graph the memory cell state of the long short-term memory network at time represents the global transfer graph the candidate memory cell state of the long short-term memory network at time represents the weight matrix of the candidate memory cell state of the long short-term memory network of the global transfer graph, represents the weight matrix of the candidate memory cell of the long short-term memory network of the global transfer graph for the node state at the previous time step, represents the bias vector of the candidate memory cell of the long short-term memory network of the global transfer graph, represents element-wise multiplication, represents the global transfer graph the value of the output gate of the long short-term memory network at time represents the weight matrix of the output gate of the long short-term memory network of the global transfer graph, represents the weight matrix of the output gate of the long short-term memory network of the global transfer graph for the node state at the previous time step, represents the bias vector of the output gate of the long short-term memory network of the global transfer graph, represents the embedding vector of the action location and its related attributes in the global user cross-action location graph at time

[0011] Further, in step S3, the following method is used to convert each local action sequence diagram into a local action sequence diagram vector containing local action transfer features: S31: Use the embedding layer to convert the elements in each user-generated node and virtual neighborhood node of the given local action sequence diagram into embedding vectors respectively, and splice the embedding vectors to form the latent representation of the nodes; S32: Add the latent representation of the nodes to the representation of the nodes in the global user cross-action location diagram with the same action location, and fuse the time information as the initial state of the nodes in the local action sequence diagram; S33: Construct a local action sequence diagram pairwise connection matrix, which includes the in-degree matrix of the local action sequence diagram and the out-degree matrix of the local action sequence diagram. Then, assign values to two adjacent nodes, two non-adjacent nodes, and virtual neighborhood nodes in the action sequence of the local action sequence diagram in the in-degree matrix of the local action sequence diagram and the out-degree matrix of the local action sequence diagram, and perform row normalization on the assigned in-degree matrix of the local action sequence diagram and the out-degree matrix of the local action sequence diagram; S34: Update the node states in the local action sequence diagram through a gated graph neural network based on long short-term memory units, and convert each local action sequence diagram into a local action sequence diagram vector containing local action transfer features.

[0012] Further, in step S34, the node states in the local action sequence diagram are updated according to Equation (7): (7); Where: represents the input features of the nodes in the local action sequence diagram at time represents the weight matrix for the linear transformation of the in-degree matrix of the local action sequence diagram, represents the in-degree matrix of the local action sequence diagram, represents the bias vector for the linear transformation of the in-degree matrix of the local action sequence diagram, represents the weight matrix for the linear transformation of the out-degree matrix of the local action sequence diagram, represents the out-degree matrix of the local action sequence diagram, represents the bias vector for the linear transformation of the out-degree matrix of the local action sequence diagram, represents the value of the input gate of the long short-term memory network of the local action sequence diagram at time represents the weight matrix for the current input in the input gate, represents the weight matrix for the state at the previous time step in the input gate, represents the bias vector of the input gate in the process of updating the node states of the local action sequence diagram, Indicates the local action sequence diagram The value of the forget gate of the long short-term memory network at a certain moment, Indicates the weight matrix of the forget gate of the long short-term memory network of the local action sequence diagram, Indicates the weight matrix for the node state of the previous time step in the forget gate of the long short-term memory network of the local action sequence diagram, Indicates the bias vector of the forget gate of the long short-term memory network of the local action sequence diagram, Indicates the local action sequence diagram The candidate memory cell state of the long short-term memory network at a certain moment, Indicates the weight matrix of the candidate memory cell state of the long short-term memory network of the local action sequence diagram, Indicates the weight matrix for the node state of the previous time step in the candidate memory cell of the long short-term memory network of the local action sequence diagram, Indicates the bias vector of the candidate memory cell of the long short-term memory network of the local action sequence diagram, Indicates the local action sequence diagram The memory cell state of the long short-term memory network at a certain moment, Indicates the local action sequence diagram The memory cell state of the long short-term memory network at a certain moment, Indicates the node state after the global transfer relationship is exchanged, Indicates the local action sequence diagram The value of the output gate of the long short-term memory network at a certain moment, Indicates the weight matrix of the output gate of the long short-term memory network of the local action sequence diagram, Indicates the weight matrix for the node state of the previous time step in the output gate of the long short-term memory network of the local action sequence diagram, Indicates the bias vector of the output gate of the long short-term memory network of the local action sequence diagram, Indicates The state of the node of the local action sequence diagram at a certain moment, Indicates The state of the node of the local action sequence diagram at a certain moment, Indicates the sigmoid function.

[0013] Furthermore, in step S4, the following method is used to obtain the local action sequence diagram with enhanced semantic representation ability of the virtual neighborhood feature vector: S41: Take each local action sequence diagram vector containing local action transfer features as the input of the action plan recommendation and input it into the contrast learning framework for contrast learning; S42: For each local action sequence graph, compare the action sequence representation with the target virtual neighborhood nodes of the last single-point action and the source virtual neighborhood nodes of the next single-point action, and calculate the total contrast loss according to Equation (8): (8); Where: represents the loss of the target virtual neighborhood nodes, represents the vector of the source virtual neighborhood nodes of the next single-point action after linear transformation, represents the positive sample local action sequence graph vector, represents the negative sample local action sequence graph vector, represents the loss of the source virtual neighborhood nodes, represents the vector of the target virtual neighborhood nodes of the last single-point action after linear transformation, represents the total contrast loss, represents the similarity function for pairwise sample comparison, represents the sigmoid function; S43: Update the learning parameters through backpropagation algorithm according to the total contrast loss, and obtain the updated matrices for linearly transforming the vector of the source virtual neighborhood nodes for the next single-point action and the matrix for linearly transforming the vector of the target virtual neighborhood nodes for the last single-point action; S44: Update the vector of the source virtual neighborhood nodes of the next single-point action after linear transformation using the updated matrix for linearly transforming the vector of the source virtual neighborhood nodes for the next single-point action, and update the vector of the target virtual neighborhood nodes of the last single-point action after linear transformation using the updated matrix for linearly transforming the vector of the target virtual neighborhood nodes for the last single-point action, so as to obtain the local action sequence graph with enhanced semantic representation ability of the virtual neighborhood feature vectors.

[0014] Further, the method for training the feature representation of the local action sequence graph with enhanced semantic representation ability of the virtual neighborhood feature vectors in step S5 is as follows using the multi-task learning method: S51: Calculate the embedding vector of the candidate action location generated by the user according to Equation (9): (9); Where: represents the embedding vector of the candidate action location generated by the user, represents the embedding vector of the candidate action location generated by the user before fusing relevant attributes, represents the location category embedding vector of the candidate action location generated by the user, represents the geographical cluster embedding vector of the candidate action location generated by the user, Represents the user-generated candidate action locations and their associated attribute embeddings, Represents the candidate vector of the user-generated nodes in the global user cross-action location graph, Represents the candidate vector of the user-generated nodes in the global transfer graph; S52: Calculate the predicted score of the user-generated subsequent action location according to Equation (10) based on the embedding vector of the user-generated candidate action locations: (10); Where: Represents the predicted score of the user-generated subsequent action location, Represents the normalized exponential function, Represents the th feature vector of the user-generated node, Represents the linear transformation matrix after aggregating the node vectors, Represents the set of indices related to the user-generated nodes in the corresponding local action sequence graph; S53: Calculate the predicted loss of the user-generated subsequent action location according to Equation (11) based on the predicted score of the user-generated subsequent action location: (11); Where: Represents the predicted loss of the user-generated subsequent action location, Represents the one-hot encoded vector specific to the subsequent user-generated true action location; S54: Calculate the embedding vector of the candidate action location of the source virtual neighborhood node according to Equation (12): (12); Where: Represents the embedding vector of the candidate action location corresponding to the source virtual neighborhood node, Represents the embedding vector of the candidate action location corresponding to the source virtual neighborhood node before fusing the relevant attributes, Represents the location category embedding vector of the candidate action location corresponding to the source virtual neighborhood node, Represents the geographical cluster embedding vector of the candidate action location corresponding to the source virtual neighborhood node, Represents the embedding of the candidate action location corresponding to the source virtual neighborhood node and its associated attributes, Represents the candidate vector of the source virtual neighborhood node in the global user-action location graph, Represents the candidate vector of the source virtual neighborhood node in the global transfer graph; S55: Calculate the predicted score of the action location corresponding to the source virtual neighborhood node according to Equation (13) based on the embedding vector of the candidate action location of the source virtual neighborhood node: (13); Where: Represents the predicted score of the action location corresponding to the source virtual neighborhood node, Represents the -th feature vector of the source virtual neighborhood node, Represents the linear transformation matrix after aggregation of the source virtual neighborhood node vectors, Represents the index set related to the source virtual neighborhood node in the corresponding local action sequence graph; S56: Calculate the predicted loss of the action location corresponding to the source virtual neighborhood node according to Equation (14) based on the predicted score of the action location corresponding to the source virtual neighborhood node: (14); Where: Represents the predicted loss of the action location corresponding to the source virtual neighborhood node, Represents the one - hot encoded vector specific to the true action location corresponding to the source virtual neighborhood node; S57: Calculate the embedding vector of the candidate action location of the target virtual neighborhood node according to Equation (15): (15); Where: Represents the embedding vector of the candidate action location of the target virtual neighborhood node, Represents the embedding vector of the candidate action location corresponding to the target virtual neighborhood node before fusing relevant attributes, Represents the location category embedding vector of the candidate action location corresponding to the target virtual neighborhood node, Represents the geographical cluster embedding vector of the candidate action location corresponding to the target virtual neighborhood node, Represents the embedding of the candidate action location corresponding to the target virtual neighborhood node and its relevant attributes, Represents the candidate vector of the target virtual neighborhood node in the global user cross - action location graph, Represents the candidate vector of the target virtual neighborhood node in the global transfer graph; S58: Calculate the predicted score of the action location corresponding to the target virtual neighborhood node according to Equation (16) based on the embedding vector of the candidate action location of the target virtual neighborhood node: (16); Where: Represents the predicted score of the action location corresponding to the target virtual neighborhood node, Represents the -th feature vector of the target virtual neighborhood node, Represents the index set related to the target virtual neighborhood node in the corresponding local action sequence graph, Represents the linear transformation matrix after the aggregation of the target virtual neighborhood node vectors; S59: Calculate the prediction loss of the action location corresponding to the target virtual neighborhood node according to Equation (17) based on the prediction score of the action location corresponding to the target virtual neighborhood node: (17); Where: Represents the prediction loss of the action location corresponding to the target virtual neighborhood node, Represents the one-hot encoded vector specific to the true action location corresponding to the target virtual neighborhood node; S510: Calculate the comprehensive loss function according to Equation (18): (18); Where: Represents the comprehensive loss function, Represents the influence coefficient used to control the impact of the total contrast loss on the comprehensive loss; S511: With the goal of minimizing the comprehensive loss function, repeatedly iterate the prediction process and the backpropagation process during training until the prediction result converges, and complete the training of the feature representation method of the local action sequence graph.

[0015] Optimized, in step S5, obtain the action plan prediction score in the planning recommendation list according to Equation (19): (19); Where: Represents the representation prediction score corresponding to the target virtual neighborhood node of the last single-point action in the self-action sequence, Represents the representation corresponding to the target virtual neighborhood node of the last single-point action in the self-action sequence, Represents the action plan prediction score in the planning recommendation list.

[0016] A neighborhood-enhanced action plan intelligent recommendation system for executing a neighborhood-enhanced action plan intelligent recommendation method as described in any one of the above, including a local action sequence graph construction module, a global user cross-action location graph construction module, a global transition graph construction module, a local action sequence graph feature representation module, a contrast learning module, a multi-task learning module, and an action plan recommendation module; The local action sequence graph construction module is used to add corresponding virtual neighborhood action sets to the action sequence composed of each single-point action set to construct a local action sequence graph; The global user cross-action location graph construction module is used to construct a global user cross-action location graph to spread the collaborative information between action locations; The global transfer graph construction module updates the nodes of the global transfer graph based on the gated graph neural network of the long short-term memory unit, thereby constructing the global transfer graph; The local action sequence graph feature representation module is used to fuse the global user cross-action location graph feature and the global transfer graph feature into the local action sequence graph, and update the node state of the local action sequence graph based on the gated graph neural network of the long short-term memory unit, obtaining a local action sequence vector containing local action transfer features; The contrast learning module is used to enhance the semantic representation ability of the virtual neighborhood feature vector of the local action sequence graph, obtaining a local action sequence graph with enhanced semantic representation ability of the virtual neighborhood feature vector; The multi-task learning module is used to train the feature representation of each local action sequence graph with enhanced semantic representation ability of the virtual neighborhood vector; The action planning recommendation module uses the local action sequence graph obtained by the multi-task learning module to predict the subsequent action location, and selects the optimal location for recommendation according to the prediction score ranking.

[0017] Advantages of the invention: An action planning intelligent recommendation method and system based on neighborhood enhancement provided by the present invention, based on the problem of incomplete user action records, introduces action virtual neighborhood features into the local action sequence graph to simulate potential action records that are not uploaded by users, thereby increasing the integrity of the action sequence.

[0018] Specifically, the action sequence neighborhood feature obtains the collaborative influence feature from the action sequences of all users, and is optimized through a contrast learning framework, enabling it to effectively capture the local and global action sequence features of users.

[0019] In terms of action sequence feature learning, by integrating local and global action sequence graph features, the integrity of the individual user action sequence is supplemented through virtual neighborhood nodes, thereby effectively learning the personalized user action sequence representation. Among them, the local action sequence feature representation realizes the node transfer information interaction in the local action sequence graph, and the collaborative feature of the global action sequence captures the cross-user collaborative relationship through the user, action location, and global transfer graph, enhancing the adaptability of the model to diverse user behaviors, improving the accuracy of action planning recommendation, and also providing an effective technical path for solving the missing user action records. Description of the drawings

[0020] Figure 1 is the schematic flowchart of the present invention.

[0021] Figure 2 is the schematic structural diagram of the present invention. Detailed implementation manners

[0022] An intelligent recommendation method for action planning based on neighborhood enhancement, which includes the following steps, and its flowchart is as Figure 1 shown below: S1: Add the corresponding virtual neighborhood action set to the action sequence composed of each single-point action in the user-shared action data set, and construct the corresponding local action sequence graph; The user-shared action data set refers to the action data shared by users in intelligent navigation and travel service software such as navigation applications, taxi-hailing applications, and travel planning applications. The goal of this stage is to add the corresponding virtual neighborhood action set to the action sequence composed of each single-point action set, and further construct the local action sequence graph. Use the user action records in the three-city data sets NYC, TKY, and CA of the location-based social network as training data. Among them, the NYC data set contains 5,099 action locations, 1,075 users, 318 location categories, 104,074 action records, and 14,160 action sequences; the TKY data set contains 7,844 action locations, 2,281 users, 291 location categories, 361,430 action records, and 44,692 action sequences; the CA data set contains 9,923 action locations, 4,318 users, 301 location categories, 250,780 action records, and 32,920 action sequences.

[0023] Furthermore, each single-point action is in formula (1), and the action sequence composed of each single-point action is in formula (2): (1); (2); Where: represents a single-point action, represents a user, represents an action location, represents an action time, represents an action location category, represents the geographical information corresponding to the action location, represents a tuple, represents the action sequence composed of each single-point action, represents the th single-point action, represents the total number of single-point actions in the action sequence composed of each single-point action, represents a set.

[0024] Assume that the user-shared action data set is , , represents the user set, , represents the action location set, , represents the set of action times, , represents the set of categories, and represents the set of action locations; in the user - shared action data set , each user has a single - point action , denoted as , each single - point action contains multiple action sequences, and the action sequences composed of each single - point action , represents the number of single - point actions in the action sequence composed of single - point actions and records the single - point actions of the current user within a specific time period.

[0025] For an action sequence composed of a single - point action , each single - point action in it is used as a node of the local action sequence graph generated by the user. The corresponding set of nodes generated by the user is denoted as . For each node in , paired virtual neighborhood nodes are generated, including the source node pointing to and the target node pointed to by , , and are used to simulate the single - point actions that have not been recorded by the user.

[0026] Specifically, adding the corresponding virtual neighborhood action set to the action sequence composed of each single - point action of the user - shared action data and constructing the corresponding local action sequence graph includes the following steps: S11: Using the k - means clustering method to cluster the two - dimensional coordinates in the geographical information corresponding to the action locations in each single - point action; S12: Representing the geographical information of each action location by connecting each action location in each single - point action with its corresponding cluster; S13: Using each single - point action in the action sequence composed of each single - point action as a node of the local action sequence graph generated by the user to generate a set of nodes generated by the user; S14: Generating paired virtual neighborhood nodes for each node in the set of nodes generated by the user; S15: Introducing an edge between each user - generated node and its corresponding virtual neighborhood node, introducing an edge between each pair of single - point actions adjacent in time among all nodes, and introducing an edge between two nodes with a set distance, thus forming the corresponding local action sequence graph.

[0027] For a pair of single-point actions with an access order , the chronological order corresponding to the local action sequence diagram nodes is as follows: .

[0028] Specifically, when introducing an edge between two nodes at a set distance, it can be added with a time jump when the distance between the nodes is 2 or 3 to further improve connectivity. The constructed local action sequence diagram will serve as the basic data structure for subsequent action planning recommendations.

[0029] S2: Construct the global user cross-action location graph features and the global transfer graph features; Specifically, the global user cross-action location graph features and the global transfer graph features can be constructed by the following method: S21: Use the propagation and aggregation functions in the lightweight graph convolutional layer to update the user embedding vector according to Equation (3), update all action location embedding vectors according to Equation (4), and perform average aggregation on the user embedding vector and all action location embedding vectors respectively according to Equation (5) to obtain the final global user embedding vector and the embedding vectors of the action locations and their related attributes in the global user cross-action location graph: (3); (4); (5); Where: Where represents the index layer number, represents the number of index layers, represents the index layer user embedding vector, represents the index layer user embedding vector, represents the user in the adjacency matrix 's neighbor set, represents the action location in the adjacency matrix 's neighbor set, represents the action location 's embedding vector, represents the index layer action location embedding vector, represents taking the absolute value, represents the index layer action location embedding vector, represents the global user embedding vector, represents the embedding vectors of the action locations and their related attributes in the global user cross-action location graph; S22: Construct a global transition graph on the embedding vectors of the action locations and their related attributes in the global user cross-action location graph, and generate a pairwise connection matrix of the global transition graph nodes. The pairwise connection matrix of the global transition graph nodes includes the in-degree matrix of the global transition graph and the out-degree matrix of the global transition graph; The global user cross-action location graph depicts the interaction relationship between users and action locations formed by all user action sequences. The propagation and aggregation functions in the lightweight graph convolutional layer are used to update the user embedding vectors , all action location embedding vectors and the embedding vectors of their related attribute elements , , where is the category embedding vector, is the geographical cluster embedding vector, represents the vector concatenation operation.

[0030] Through the above update, it is beneficial to mine the high-order connectivity in the global action sequence graph.

[0031] S23: Assign values to the nodes of each action sequence in the action sequence graph, and normalize the in-degree matrix of the global transition graph between the nodes of the action sequence graph and the out-degree matrix of the global transition graph between the nodes of the action sequence graph row by row; S24: Update according to Equation (6) through the gated graph neural network based on the long short-term memory unit to construct the global user cross-action location graph features and the global transition graph features: (6); where: represents the input features of the nodes in the global transition graph at time represents the in-degree parameter matrix for the linear transformation of the global transition graph, represents the embedding vectors of the action locations and their related attributes in the global user cross-action location graph at time represents the in-degree matrix of the global transition graph, represents the in-degree bias matrix for the linear transformation of the global transition graph, represents the out-degree parameter matrix for the linear transformation of the global transition graph, represents the out-degree matrix of the global transition graph, represents the out-degree bias matrix for the linear transformation of the global transition graph, represents the value of the input gate of the global transition graph long short-term memory network at time represents the sigmoid function, represents the weight matrix of the input gate of the global transition graph long short-term memory network, The weight matrix of the input gate of the global transfer graph long short-term memory network for the node state at the previous time step, The bias vector of the input gate of the global transfer graph long short-term memory network, Denotes the global transfer graph The value of the forget gate of the long short-term memory network at time The weight matrix of the forget gate of the global transfer graph long short-term memory network, The weight matrix of the forget gate of the global transfer graph long short-term memory network for the node state at the previous time step, The bias vector of the forget gate of the global transfer graph long short-term memory network, Denotes the global transfer graph The memory cell state of the long short-term memory network at time Denotes the global transfer graph The memory cell state of the long short-term memory network at time Denotes the global transfer graph The candidate memory cell state of the long short-term memory network at time The weight matrix of the candidate memory cell state of the global transfer graph long short-term memory network, The weight matrix of the candidate memory cell of the global transfer graph long short-term memory network for the node state at the previous time step, The bias vector of the candidate memory cell of the global transfer graph long short-term memory network, Denotes element-wise multiplication, Denotes the global transfer graph The value of the output gate of the long short-term memory network at time The weight matrix of the output gate of the global transfer graph long short-term memory network, The weight matrix of the output gate of the global transfer graph long short-term memory network for the node state at the previous time step, The bias vector of the output gate of the global transfer graph long short-term memory network, Denotes The state of the node in the global transfer graph after update at time

[0032] In the first update step, Is initialized as a zero vector, and after update, Is denoted as , Denotes the state of the global transfer graph node after all update steps. In the process of representing the local action sequence graph features in subsequent steps, the node state after the global transfer relationship exchange is used As the initial state of the global transfer graph node, , where Index using the ID in the action location in a single-point action, and extract from The identifier representing the time step.

[0033] S3: Introduce the constructed global user cross-action location graph feature and global transfer graph feature into each local action sequence graph, and convert each local action sequence graph into a local action sequence graph vector containing local action transfer features; Specifically, the following method can be used to convert each local action sequence graph into a local action sequence graph vector containing local action transfer features: S31: Use the embedding layer to convert the elements in each user-generated node and virtual neighborhood node of the given local action sequence graph into embedding vectors respectively, and splice the embedding vectors to form the latent representation of the node; Due to this operation, in addition to each user-generated node having its embedding vector, each virtual neighborhood node is also converted into an embedding vector, and the embedding matrix cardinality of each type of element is three times the original.

[0034] S32: Add the latent representation of the node and the node representation in the global user cross-action location graph with the same action location, and fuse the time information as the initial state of the local action sequence graph node; S33: Construct a local action sequence graph pairwise connection matrix, which includes the local action sequence graph in-degree matrix and the local action sequence graph out-degree matrix. Then, assign values to two adjacent nodes, two non-adjacent nodes, and virtual neighborhood nodes in the action sequence of the local action sequence graph in the local action sequence graph in-degree matrix and the local action sequence graph out-degree matrix, and perform row normalization on the assigned local action sequence graph in-degree matrix and local action sequence graph out-degree matrix; The local action sequence graph in-degree matrix and the local action sequence graph out-degree matrix are used to describe the transfer relationship between action locations.

[0035] For two adjacent nodes in the action sequence, set their values in the local action sequence graph in-degree matrix and the local action sequence graph out-degree matrix to 1. If there is a virtual neighborhood node between two adjacent nodes, multiply their values in the local action sequence graph in-degree matrix and the local action sequence graph out-degree matrix by 0.5 again, indicating that the transfer relationship between these two nodes is virtual.

[0036] S34: Update the node state in the local action sequence graph through a gated graph neural network based on long short-term memory units, and convert each local action sequence graph into a local action sequence graph vector containing local action transfer features.

[0037] ​In the forget gate of the long short-term memory network, set the initial state of the memory cell as the node state after the global transfer relationship is exchanged , so as to adaptively retain the node features from the global transfer relationship in the forget gate without overly affecting the dynamic features in the current local action sequence graph.

[0038] Furthermore, in step S34, update the node state in the local action sequence graph according to formula (7): (7); wherein: represents the input feature of the node of the local action sequence graph at time represents the weight matrix for the linear transformation of the in-degree matrix of the local action sequence graph, represents the in-degree matrix of the local action sequence graph, represents the bias vector for the linear transformation of the in-degree matrix of the local action sequence graph, represents the weight matrix for the linear transformation of the out-degree matrix of the local action sequence graph, represents the out-degree matrix of the local action sequence graph, represents the bias vector for the linear transformation of the out-degree matrix of the local action sequence graph, represents the value of the input gate of the long short-term memory network of the local action sequence graph at time represents the weight matrix for the current input in the input gate, represents the weight matrix for the state at the previous time step in the input gate, represents the bias vector of the input gate in the process of updating the node state of the local action sequence graph, represents the local action sequence graph the value of the forget gate of the long short-term memory network at time represents the weight matrix of the forget gate of the long short-term memory network of the local action sequence graph, represents the weight matrix for the node state at the previous time step in the forget gate of the long short-term memory network of the local action sequence graph, represents the bias vector of the forget gate of the long short-term memory network of the local action sequence graph, represents the local action sequence graph the candidate memory cell state of the long short-term memory network at time represents the weight matrix of the candidate memory cell state of the long short-term memory network of the local action sequence graph, represents the weight matrix for the node state at the previous time step in the candidate memory cell of the long short-term memory network of the local action sequence graph, The bias vector of the candidate memory unit representing the local action sequence diagram long short-term memory network, Represents the local action sequence diagram The memory cell state of the long short-term memory network at time Represents the local action sequence diagram The memory cell state of the long short-term memory network at time Represents the node state after the global transfer relationship is exchanged, Represents the local action sequence diagram The value of the output gate of the long short-term memory network at time Represents the weight matrix of the output gate of the local action sequence diagram long short-term memory network, Represents the weight matrix of the output gate of the local action sequence diagram long short-term memory network for the node state at the previous time step, Represents the bias vector of the output gate of the local action sequence diagram long short-term memory network, Represents The state of the local action sequence diagram node at time Represents The state of the local action sequence diagram node at time Represents the sigmoid function; S4: Input each local action sequence diagram vector containing local action transfer features into the contrast learning framework for contrast learning to obtain a local action sequence diagram with enhanced semantic representation ability of the virtual neighborhood feature vector; Specifically, the following method can be used to obtain the local action sequence diagram with enhanced semantic representation ability of the virtual neighborhood feature vector: S41: Input each local action sequence diagram vector containing local action transfer features into the contrast learning framework for contrast learning as the input of action planning recommendation; S42: For each local action sequence diagram, compare the action sequence representation with the target virtual neighborhood node of the last single-point action and the source virtual neighborhood node of the next single-point action, and calculate the total contrast loss according to Equation (8): (8); Where: Represents the loss of the target virtual neighborhood node, Represents the source virtual neighborhood node vector of the next single-point action after linear transformation, , Represents the number of nodes corresponding to the current local action sequence diagram; Represents the corresponding trajectory The source virtual neighborhood node vector of the next single-point action in Represents for The matrix for linear transformation, Represents the positive sample local action sequence graph vector, , Represents the -th feature vector of a single-point action, Represents the negative sample local action sequence graph vector, , Represents the -th feature vector of a single-point action in the local action sequence graph randomly sampled from the training batch, Represents the loss of the source virtual neighborhood node, Represents the target virtual neighborhood node vector of the last single-point action after linear transformation, , Represents the target virtual neighborhood node vector of the last single-point action, Represents the matrix used for linear transformation of Represents the total contrast loss, Represents the similarity function used for pairwise sample comparison, Represents the sigmoid function; Encapsulating the target virtual neighborhood node of the last single-point action and the source virtual neighborhood node of the next single-point action in each action sequence from the feature representations of all single-point actions in the current action sequence can provide a more meaningful representation of the virtual neighborhood nodes in the local action sequence graph, making it more practical; S43: Update the learning parameters according to the total contrast loss through the backpropagation algorithm to obtain the updated matrices for linear transformation of the source virtual neighborhood node vector for the next single-point action and the matrix for linear transformation of the target virtual neighborhood node vector for the last single-point action; S44: Update the source virtual neighborhood node vector of the next single-point action after linear transformation using the updated matrix for linear transformation of the source virtual neighborhood node vector for the next single-point action, and update the target virtual neighborhood node vector of the last single-point action after linear transformation using the updated matrix for linear transformation of the target virtual neighborhood node vector for the last single-point action to obtain a local action sequence graph with enhanced semantic representation ability of virtual neighborhood features.

[0039] The local action sequence feature representation realizes the interaction of node transfer information in the local action sequence graph. The collaborative features of the global action sequence capture the cross-user collaborative relationship through the user, action location, and global transfer graph, enhancing the model's adaptability to diverse user behaviors, improving the accuracy of action planning recommendations, and also providing an effective technical path for solving the missing user action records.

[0040] S5: For the local action sequence graph with enhanced semantic representation ability of virtual neighborhood feature vectors, use the multi-task learning method to train the feature representation of the local action sequence graph, convert its action sequence graph vector into an action planning recommendation list, sort according to the prediction scores in the planning recommendation list, and select the optimal location as the action planning recommendation for intelligent navigation and travel services.

[0041] The tasks here include predicting the node vector of the target virtual neighborhood and the action location corresponding to the node of the target virtual neighborhood related to the subsequent action location, predicting the source virtual neighborhood node corresponding to the source virtual neighborhood node related to the subsequent action location, and predicting the user-generated subsequent action location for the user-generated node vector.

[0042] Specifically, the method for training the feature representation of the local action sequence graph with enhanced semantic representation ability of virtual neighborhood feature vectors using the multi-task learning method is as follows: S51: Calculate the embedding vector of the user-generated candidate action location according to Equation (9): (9); Where: represents the embedding vector of the user-generated candidate action location, represents the embedding vector of the user-generated candidate action location before fusing relevant attributes, represents the location category embedding vector of the user-generated candidate action location, represents the geographical cluster embedding vector of the user-generated candidate action location, represents the embedding of the user-generated candidate action location and its relevant attributes, represents the user-generated node candidate vector of the global user cross-action location graph, represents the user-generated node candidate vector of the global transfer graph; S52: Calculate the prediction score of the user-generated subsequent action location according to Equation (10) based on the embedding vector of the user-generated candidate action location: (10); Where: represents the prediction score of the user-generated subsequent action location, represents the normalization exponential function, represents the th feature vector of the user-generated node, represents the linear transformation matrix after node vector aggregation, represents the index set related to the user-generated node in the corresponding local action sequence graph; S53: Calculate the prediction loss of the user-generated subsequent action location according to Equation (11) based on the prediction score of the user-generated subsequent action location: (11); Wherein: represents the predicted loss of the subsequent action location generated by the user, represents a one-hot encoded vector specific to the subsequent true action location generated by the user; Through steps S51 to S51, the subsequent action location generated by the user can be predicted.

[0043] S54: Calculate the embedding vector of the candidate action location of the source virtual neighborhood node according to Equation (12): (12); Wherein: represents the embedding vector of the candidate action location corresponding to the source virtual neighborhood node, represents the embedding vector of the candidate action location corresponding to the source virtual neighborhood node before fusing relevant attributes, represents the location category embedding vector of the candidate action location corresponding to the source virtual neighborhood node, represents the geographical cluster embedding vector of the candidate action location corresponding to the source virtual neighborhood node, represents the embedding of the candidate action location corresponding to the source virtual neighborhood node and its relevant attributes, represents the global user-action location source virtual neighborhood node candidate vector, represents the source virtual neighborhood node candidate vector in the global transfer graph; S55: Calculate the predicted score of the action location corresponding to the source virtual neighborhood node according to Equation (13) based on the embedding vector of the candidate action location of the source virtual neighborhood node: (13); Wherein: represents the predicted score of the action location corresponding to the source virtual neighborhood node, represents the th feature vector of the source virtual neighborhood node, represents the linear transformation matrix after aggregating the source virtual neighborhood node vectors, represents the index set related to the source virtual neighborhood node in the corresponding local action sequence graph; S56: Calculate the predicted loss of the action location corresponding to the source virtual neighborhood node according to Equation (14) based on the predicted score of the action location corresponding to the source virtual neighborhood node: (14); Wherein: represents the predicted loss of the action location corresponding to the source virtual neighborhood node, A one-hot encoded vector representing the actual action location corresponding to the source virtual neighborhood node specific to the source; Through steps S54 to S56, the action location corresponding to the source virtual neighborhood node related to the subsequent action location can be predicted.

[0044] S57: Calculate the embedding vector of the candidate action location of the target virtual neighborhood node according to Equation (15): (15); Where: represents the embedding vector of the candidate action location of the target virtual neighborhood node, represents the embedding vector of the candidate action location corresponding to the target virtual neighborhood node before fusing relevant attributes, represents the location category embedding vector of the candidate action location corresponding to the target virtual neighborhood node, represents the geographical cluster embedding vector of the candidate action location corresponding to the target virtual neighborhood node, represents the embedding of the candidate action location corresponding to the target virtual neighborhood node and its relevant attributes, represents the candidate vector of the target virtual neighborhood node in the global user cross-action location graph, represents the candidate vector of the target virtual neighborhood node in the global transfer graph; S58: Calculate the prediction score of the action location corresponding to the target virtual neighborhood node according to Equation (16) based on the embedding vector of the candidate action location of the target virtual neighborhood node: (16); Where: represents the prediction score of the action location corresponding to the target virtual neighborhood node, represents the th feature vector of the target virtual neighborhood node, represents the index set related to the target virtual neighborhood node in the corresponding local action sequence graph, represents the linear transformation matrix after aggregating the target virtual neighborhood node vectors; S59: Calculate the prediction loss of the action location corresponding to the target virtual neighborhood node according to Equation (17) based on the prediction score of the action location corresponding to the target virtual neighborhood node: (17); Where: represents the prediction loss of the action location corresponding to the target virtual neighborhood node, represents the one-hot encoded vector specific to the actual action location corresponding to the target virtual neighborhood node; Through steps S57 to S59, the action location corresponding to the target virtual neighborhood node related to the subsequent action location can be predicted.

[0045] S510: Calculate the comprehensive loss function according to Equation (18): (18); Where: represents the comprehensive loss function, represents the influence coefficient used to control the impact of the total contrast loss on the comprehensive loss; S511: With the goal of minimizing the comprehensive loss function, repeatedly iterate the prediction process and the backpropagation process during training until the prediction result converges, and complete the training of the feature representation method of the local action sequence graph.

[0046] Optimized, in step S5, obtain the action plan prediction score in the planning recommendation list according to Equation (19): (19); Where: represents the representation prediction score corresponding to the target virtual neighborhood node of the last single-point action in the self-action sequence, represents the representation corresponding to the target virtual neighborhood node of the last single-point action in the self-action sequence, represents the action plan prediction score in the planning recommendation list.

[0047] Specifically, by minimizing , in the backpropagation process, , and etc. will be updated, so as to optimize the feature representation of the user-generated node; by minimizing , in the backpropagation process, , and etc. will be updated, so as to optimize the feature representation of the source virtual neighborhood target node; by minimizing , in the backpropagation process, , and etc. will be updated, so as to optimize the feature representation of the target virtual neighborhood target node. By minimizing , the semantic representation ability of the virtual neighborhood node vector in the local action sequence graph is enhanced.

[0048] A neighborhood-enhanced action plan intelligent recommendation system for executing a neighborhood-enhanced action plan intelligent recommendation method as described in any one of the above, and the system structure schematic diagram is as Figure 2As shown in the figure, it includes a local action sequence graph construction module, a global user cross-action location graph construction module, a global transition graph construction module, a local action sequence graph feature representation module, a contrast learning module, a multi-task learning module, and an action planning recommendation module; The local action sequence graph construction module is used to add corresponding virtual neighborhood action sets to the action sequence composed of each single-point action set to construct a local action sequence graph; The global user cross-action location graph construction module is used to construct a global user cross-action location graph to spread the collaborative information between action locations; The global transition graph construction module updates the nodes of the global transition graph based on the gated graph neural network of the long short-term memory unit to construct a global transition graph; The local action sequence graph feature representation module is used to fuse the global user cross-action location graph features and the global transition graph features into the local action sequence graph, and update the node states of the local action sequence graph based on the gated graph neural network of the long short-term memory unit to obtain a local action sequence vector containing local action transfer features; The contrast learning module is used to enhance the semantic representation ability of the virtual neighborhood feature vector of the local action sequence graph vector to obtain a local action sequence graph with enhanced semantic representation ability of the virtual neighborhood feature vector; The multi-task learning module is used to train the feature representation of each local action sequence graph with enhanced semantic representation ability of the virtual neighborhood vector; The action planning recommendation module uses the local action sequence graph obtained by the multi-task learning module to predict the subsequent action location, and selects the optimal location for recommendation according to the prediction score ranking.

[0049] In summary, an action planning intelligent recommendation method and system provided by the present invention combines virtual neighborhood feature representation with the action features generated by users, enhances the integrity of the user action sequence, and forms a more objective, comprehensive, and accurate description of the user behavior pattern through local and global feature representation learning of the user behavior sequence containing virtual neighborhood features, improving the recommendation effect of action planning.

[0050] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An action planning intelligent recommendation method based on neighborhood enhancement, characterized in that: The steps include: S1: Add the corresponding virtual neighborhood action set to the action sequence composed of each single-point action in the user shared action dataset, and construct the corresponding local action sequence graph; S2: Construct global user cross-action location map features and global transition map features; S3: Introduce the constructed global user cross-action location graph features and global transition graph features into each local action sequence graph, and transform each local action sequence graph into a local action sequence graph vector containing local action transition features; S4: Input each local action sequence graph vector containing local action transfer features into the contrastive learning framework for contrastive learning, and obtain a local action sequence graph with enhanced semantic representation ability of the virtual neighborhood feature vector; S5: A multi-task learning method is used to train the feature representation of the local action sequence graph with enhanced semantic representation capability of the virtual neighborhood feature vector, and its action sequence graph vector is converted into an action plan recommendation list. The action plans in the plan recommendation list are sorted according to their predicted scores, and the optimal location is selected as the action plan recommendation for intelligent navigation and travel services.

2. The method for intelligently recommending action plans based on neighborhood enhancement according to claim 1, characterized in that: Each single-point action in step S1 is expressed as formula (1), and the action sequence composed of each single-point action is expressed as formula (2): （1）； （2）； in: Indicates a single point of action. Represents the user, Indicates the location of the action. Indicates the action time. Indicates the type of action location. Indicates the geographical information corresponding to the action location, Represents a tuple, Represents the action sequence composed of each single-point action, Indicates A single point of action, Indicates the total number of single-point actions in the action sequence composed of each single-point action, Represents a collection.

3. The method for intelligently recommending action plans based on neighborhood enhancement according to claim 2, characterized in that: In step S1, the action sequence composed of each single-point action of the user's shared action data is added to the corresponding virtual neighborhood action set, and the corresponding local action sequence graph is constructed, which includes the following steps: S11: Use k-means clustering method to cluster the two-dimensional coordinates in the geographic information corresponding to the action location in each single-point action; S12: connecting each action location in each single-point action with its corresponding cluster in parallel to represent the geographic information of each action location; S13: taking each single-point action in the action sequence composed of each single-point action as a node of the user-generated local action sequence graph, and generating a user-generated node set; S14: Generate a pair of virtual neighbor nodes for each node in the user-generated node set; S15: Introduce an edge between each user-generated node and its corresponding virtual neighborhood node, introduce an edge between each temporally adjacent single-point action between all nodes, and introduce an edge between two nodes at a set distance apart, to form a corresponding local action sequence graph.

4. The method for intelligently recommending action plans based on neighborhood enhancement according to claim 1, characterized in that: In step S2, the global user cross-action location map feature and the global transition map feature are constructed by the following method: S21: Using the propagation and aggregation function in the lightweight graph convolutional layer, the user embedding vector is updated according to formula (3), the embedding vectors of all action locations are updated according to formula (4), and the user embedding vector and the embedding vectors of all action locations are averaged and aggregated according to formula (5) to obtain the final global user embedding vector and the embedding vectors of the action locations and their related attributes in the global user cross-action location graph: (3)； (4)； （5）； Among them: Indicates the index level number. Indicates the number of index levels, express Index layer user embedding vector, express Index layer user embedding vector, Represents the user in the adjacency matrix The neighbor set of Represents the action location in the adjacency matrix The neighbor set of Indicates the location of the action The embedding vector of express Index-level action location embedding vector, Indicates taking the absolute value, express Index-level action location embedding vector, represents the global user embedding vector, Embedding vectors representing action locations and their related attributes in the global user cross-action location graph; S22: constructing a global transition graph on the embedded vectors of the action locations and their related attributes in the global user cross-action location graph, generating a global transition graph node pairwise connection matrix, wherein the global transition graph node pairwise connection matrix includes a global transition graph in-degree matrix and a global transition graph out-degree matrix; S23: assigning values ​​to the nodes of each action sequence in the action sequence graph, and normalizing the in-degree matrix of the global transfer graph between the action sequence graph nodes and the out-degree matrix of the global transfer graph between the action sequence graph nodes after the assignment by row; S24: The global user cross-action location map feature and the global transition map feature are constructed by updating the gated graph neural network based on long short-term memory units according to formula (6): （6）； in: express The input features of the nodes in the global transition graph at each moment, represents the in-degree parameter matrix for the linear transformation of the global transfer graph, express The embedding vector of the action location and its related attributes in the global user cross-action location graph at each moment, represents the in-degree matrix of the global transition graph, represents the in-degree bias matrix for linear transformation of the global transfer graph, represents the out-degree parameter matrix for the linear transformation of the global transition graph, represents the out-degree matrix of the global transition graph, represents the out-degree bias matrix used for linear transformation of the global transfer graph, express The value of the input gate of the long short-term memory network of the global transfer graph at the moment, represents the sigmoid function, represents the weight matrix of the input gate of the LSTM network of the global transition graph, The weight matrix representing the input gate of the LSTM network of the global transition graph for the node state at the previous time step, represents the bias vector of the input gate of the LSTM network of the global transition graph, Represents the global transition graph The value of the forget gate of the short-term memory network at time length, represents the weight matrix of the forget gate of the LSTM network of the global transition graph, represents the weight matrix for the node state at the previous time step in the forget gate of the long short-term memory network of the global transition graph, represents the bias vector of the forget gate of the long short-term memory network of the global transition graph, Represents the global transition graph The state of the memory unit of the short-term memory network, Represents the global transition graph The state of the memory unit of the short-term memory network, Represents the global transition graph The candidate memory unit state of the short-term memory network at time length, The weight matrix representing the candidate memory cell states of the LSTM network of the global transition graph, represents the weight matrix for the node state at the previous time step in the candidate memory unit of the LSTM network of the global transition graph, The bias vector representing the candidate memory unit of the LSTM network of the global transition graph, represents element-wise multiplication, Represents the global transition graph The value of the output gate of the short-term memory network at time length, represents the weight matrix of the output gate of the LSTM network of the global transition graph, represents the weight matrix for the node state at the previous time step in the output gate of the LSTM network of the global transition graph, represents the bias vector of the output gate of the LSTM network of the global transition graph, express Embedding vectors of action locations and their related attributes in the global user cross-action location graph at each moment.

5. The method for intelligent recommendation of action planning based on neighborhood enhancement according to claim 1, characterized in that: In step S3, each local action sequence graph is converted into a local action sequence graph vector containing local action transfer features by the following method: S31: using the embedding layer to convert the elements of each user-generated node and virtual neighborhood node in the given local action sequence graph into embedding vectors respectively, and concatenate the embedding vectors to form the potential representation of the node; S32: adding the potential representation of the node to the node representation consistent with its action location in the global user cross-action location graph, and fusing the time information to serve as the initial state of the node in the local action sequence graph; S33: constructing a pairwise connection matrix of a local action sequence graph, the pairwise connection matrix of the local action sequence graph includes an in-degree matrix of the local action sequence graph and an out-degree matrix of the local action sequence graph, and then assigning values ​​to two adjacent nodes, two non-adjacent nodes, and virtual neighboring nodes in the action sequence of the local action sequence graph in the in-degree matrix of the local action sequence graph and the out-degree matrix of the local action sequence graph, and normalizing the in-degree matrix of the local action sequence graph and the out-degree matrix of the local action sequence graph after the assignment by row; S34: Updating the node states in the local action sequence graph through a gated graph neural network based on long short-term memory units, and transforming each local action sequence graph into a local action sequence graph vector containing local action transfer features.

6. The method for intelligent recommendation of action planning based on neighborhood enhancement according to claim 5, characterized in that: In step S34, the node status in the local action sequence diagram is updated according to formula (7): （7）； in: express The input features of the local action sequence graph nodes at each moment, represents the weight matrix for the linear transformation of the in-degree matrix of the local action sequence graph, represents the in-degree matrix of the local action sequence graph, represents the bias vector for the linear transformation of the in-degree matrix of the local action sequence graph, represents the weight matrix for the linear transformation of the out-degree matrix of the local action sequence graph, represents the out-degree matrix of the local action sequence graph, represents the bias vector for the linear transformation of the out-degree matrix of the local action sequence graph, express The value of the input gate of the long short-term memory network of the local action sequence graph at the moment, represents the weight matrix for the current input in the input gate, represents the weight matrix in the input gate for the state at the previous time step, represents the bias vector of the input gate in the process of updating the node state in the local action sequence graph, Represents a local action sequence diagram The value of the forget gate of the short-term memory network at time length, represents the weight matrix of the forget gate of the long short-term memory network of the local action sequence graph, represents the weight matrix for the node state at the previous time step in the forget gate of the long short-term memory network of the local action sequence graph, represents the bias vector of the forget gate of the long short-term memory network of the local action sequence graph, Represents a local action sequence diagram The candidate memory unit state of the short-term memory network at time length, The weight matrix representing the candidate memory cell states of the local action sequence graph long short-term memory network, Represents the weight matrix for the node state at the previous time step in the candidate memory unit of the long short-term memory network of the local action sequence graph, The bias vector representing the candidate memory unit of the LSTM network of the local action sequence graph, Represents a local action sequence diagram The state of the memory unit of the short-term memory network, Represents a local action sequence diagram The state of the memory unit of the short-term memory network, Indicates the node status after the global transfer relationship is exchanged, Represents a local action sequence diagram The value of the output gate of the short-term memory network at time length, represents the weight matrix of the output gate of the LSTM network of the local action sequence graph, represents the weight matrix for the node state at the previous time step in the output gate of the LSTM network of the local action sequence graph, represents the bias vector of the output gate of the LSTM network of the local action sequence graph, express The state of the local action sequence diagram node at the moment, express The state of the local action sequence diagram node at the moment, Represents the sigmoid function.

7. The method for intelligently recommending action plans based on neighborhood enhancement according to claim 5, characterized in that: In step S4, the following method is used to obtain the local action sequence diagram after the semantic representation ability of the virtual neighborhood feature vector is enhanced: S41: Input each local action sequence graph vector containing local action transfer features as input of action planning recommendation into the contrastive learning framework for contrastive learning; S42: For each local action sequence graph, compare the action sequence representation with the target virtual neighborhood node of the last single-point action and the source virtual neighborhood node of the next single-point action, and calculate the total contrast loss according to formula (8): （8）； in: represents the loss of the target virtual neighborhood node, Represents the source virtual neighborhood node vector of the next single-point action after linear transformation, represents the positive sample local action sequence graph vector, represents the negative sample local action sequence graph vector, represents the loss of the source virtual neighborhood node, represents the target virtual neighborhood node vector of the last single-point action after linear transformation, represents the total contrast loss, represents the similarity function used for pairwise sample comparison, Represents the sigmoid function; S43: updating the learning parameters through the back propagation algorithm according to the total contrast loss, and obtaining an updated matrix for linear transformation of the source virtual neighborhood node vector of the next single-point action and a matrix for linear transformation of the target virtual neighborhood node vector of the last single-point action; S44: Using the updated matrix for linear transformation of the source virtual neighborhood node vector for the next single-point action, the source virtual neighborhood node vector of the next single-point action after linear transformation is updated; using the updated matrix for linear transformation of the target virtual neighborhood node vector of the last single-point action, the target virtual neighborhood node vector of the last single-point action after linear transformation is updated to obtain a local action sequence diagram with enhanced semantic representation capability of the virtual neighborhood feature vector.

8. The method for intelligently recommending action plans based on neighborhood enhancement according to claim 7, characterized in that: In step S5, the method for training the feature representation of the local action sequence graph with enhanced semantic representation capability of the virtual neighborhood feature vector using a multi-task learning method is as follows: S51: Calculate the embedding vector of the user-generated candidate action location according to formula (9): （9）； in: represents the embedding vector of the user-generated candidate action location, represents the embedding vector of the candidate action location generated by the user before fusing related attributes, The location category embedding vector representing the candidate action location generated by the user, The geographic cluster embedding vector representing the candidate action locations generated by the user, represents the embedding of user-generated candidate action locations and their related attributes, Represents the global user cross-action location graph user-generated node candidate vector, Represents the user-generated node candidate vector of the global transition graph; S52: Based on the embedding vector of the candidate action location generated by the user, the prediction score of the subsequent action location generated by the user is calculated according to formula (10): （10）； in: represents the prediction score of the next action location generated by the user, represents the normalized exponential function, Indicates feature vector of user-generated nodes, Represents the linear transformation matrix after node vector aggregation, represents a set of indices associated with user-generated nodes in the corresponding local action sequence graph; S53: Based on the prediction score of the subsequent action location generated by the user, the prediction loss of the subsequent action location generated by the user is calculated according to formula (11): （11）； in: represents the predicted loss of the user-generated subsequent action location, A one-hot encoding vector representing the location of the subsequent user-generated real action; S54: Calculate the embedding vector of the candidate action location of the source virtual neighborhood node according to formula (12): （12）； in: represents the embedding vector of the candidate action location corresponding to the source virtual neighborhood node, represents the embedding vector of the candidate action location corresponding to the source virtual neighborhood node before fusing related attributes, represents the location category embedding vector of the candidate action location corresponding to the source virtual neighborhood node, represents the geographic cluster embedding vector of the candidate action locations corresponding to the source virtual neighborhood node, represents the candidate action locations and their related attribute embeddings corresponding to the source virtual neighborhood nodes, Represents the candidate vector of virtual neighborhood nodes of the global user cross-action location graph source, Represents the candidate vector of source virtual neighborhood nodes in the global transition graph; S55: Based on the embedding vector of the candidate action location of the source virtual neighbor node, the prediction score of the action location corresponding to the source virtual neighbor node is calculated according to formula (13): （13）； in: represents the predicted score of the action location corresponding to the source virtual neighborhood node, Indicates The feature vector of the source virtual neighbor nodes, Represents the linear transformation matrix after the source virtual neighborhood node vector aggregation, represents a set of indices associated with source virtual neighborhood nodes in the corresponding local action sequence graph; S56: Based on the prediction score of the action location corresponding to the source virtual neighbor node, the prediction loss of the action location corresponding to the source virtual neighbor node is calculated according to formula (14): （14）； in: represents the prediction loss of the action location corresponding to the source virtual neighborhood node, A one-hot encoding vector representing the real action location corresponding to the source virtual neighborhood node; S57: Calculate the embedding vector of the candidate action location of the target virtual neighborhood node according to formula (15): （15）； in: The embedding vector representing the candidate action locations of the target virtual neighborhood node, represents the embedding vector of the candidate action location corresponding to the target virtual neighborhood node before fusing related attributes, represents the location category embedding vector of the candidate action location corresponding to the virtual neighborhood node of the day target, The geographic cluster embedding vector representing the candidate action locations corresponding to the target virtual neighborhood node, represents the candidate action locations and their related attribute embeddings corresponding to the target virtual neighborhood nodes, Represents the candidate vector of the target virtual neighborhood node in the global user cross-action location graph, Represents the candidate vector of the target virtual neighborhood node in the global transition graph; S58: Based on the embedding vector of the candidate action location of the target virtual neighborhood node, the prediction score of the action location corresponding to the target virtual neighborhood node is calculated according to formula (16): （16）； in: represents the predicted score of the action location corresponding to the target virtual neighborhood node, Indicates The feature vector of the target virtual neighborhood node, represents the set of indices associated with the target virtual neighborhood nodes in the corresponding local action sequence graph, Represents the linear transformation matrix after the target virtual neighborhood node vector is aggregated; S59: Based on the prediction score of the action location corresponding to the target virtual neighborhood node, the prediction loss of the action location corresponding to the target virtual neighborhood node is calculated according to formula (17): （17）； in: represents the prediction loss of the action location corresponding to the target virtual neighborhood node, A one-hot encoding vector representing the real action location corresponding to the target virtual neighborhood node; S510: Calculate the comprehensive loss function according to formula (18): （18）； in: represents the comprehensive loss function, It represents the coefficient used to control the influence of total contrast loss on comprehensive loss; S511: With the goal of minimizing the comprehensive loss function, the prediction process and the back propagation process are repeatedly iterated during the training process until the prediction results converge, completing the training of the feature representation method of the local action sequence graph.

9. The method for intelligently recommending action plans based on neighborhood enhancement according to claim 8, characterized in that: In step S5, the action plan prediction score in the plan recommendation list is obtained according to formula (19): （19）； in: The prediction score corresponding to the target virtual neighborhood node representing the last single-point action in the self-action sequence is represented, Represents the representation of the target virtual neighborhood node corresponding to the last single-point action in the self-action sequence, Indicates the prediction score of the action plan in the plan recommendation list.

10. An action planning intelligent recommendation system based on neighborhood enhancement, characterized in that: Used to execute an action planning intelligent recommendation method based on neighborhood enhancement as described in any one of claims 1 to 9, which includes a local action sequence graph construction module, a global user cross-action location map construction module, a global transfer map construction module, a local action sequence graph feature representation module, a contrastive learning module, a multi-task learning module and an action planning recommendation module; The local action sequence graph construction module is used to add a corresponding virtual neighborhood action set to the action sequence composed of each single-point action set to construct a local action sequence graph; The global user cross action location map construction module is used to construct a global user cross action location map, so as to propagate the collaborative information between action locations; The global transfer graph construction module updates the nodes of the global transfer graph based on the gated graph neural network of the long short-term memory unit, thereby constructing the global transfer graph; The local action sequence graph feature representation module is used to merge the global user cross-action location graph features and the global transition graph features into the local action sequence graph, and update the node states of the local action sequence graph based on the gated graph neural network of the long short-term memory unit to obtain a local action sequence vector containing local action transition features; The contrastive learning module is used to enhance the semantic representation capability of the virtual neighborhood feature vector of the local action sequence graph vector, and obtain the local action sequence graph with enhanced semantic representation capability of the virtual neighborhood feature vector; The multi-task learning module is used to train the feature representation of the local action sequence graph after the semantic representation capability of each virtual neighborhood vector is enhanced; The action plan recommendation module predicts the subsequent action locations using the local action sequence graph obtained by the multi-task learning module, and selects the optimal location for recommendation based on the prediction score ranking.

Images