Member order processing method and device, equipment and storage medium

Abstract

The invention relates to the technical field of order processing, and discloses a member order processing method, device and equipment and a storage medium, and the method comprises the steps: collecting multi-platform member information, constructing a member graph structure, and carrying out the node feature extraction, and obtaining a multi-dimensional member feature vector; performing feature fusion on the multi-dimensional member feature vector and the to-be-processed order data to obtain a member-order fusion feature vector; order classification is carried out to obtain an order classification result and a processing priority; performing multi-platform collaborative integer programming and solving to obtain a cross-platform order allocation scheme; performing member value evaluation to obtain a dynamic member value score and hierarchical division information; and inputting the dynamic member value score and the hierarchical division information into a dynamic right and interest distribution model based on the deep Q network for right and interest distribution strategy optimization to obtain a personalized member right and interest combination, and the method can comprehensively utilize multi-platform data to realize intelligent order processing and dynamic right and interest distribution.

Description

Technical Field

[0001] The present invention relates to the technical field of order processing, and in particular to a member order processing method, device, equipment and storage medium. Background Art

[0002] Traditional member order processing methods are often limited to a single platform and cannot fully utilize multi-platform data resources, resulting in inaccurate member value assessment, inefficient order allocation, and other issues. At the same time, existing order processing systems generally lack a deep integration of member characteristics and order characteristics, making it difficult to achieve accurate order classification and priority division.

[0003] In addition, in a multi-platform collaborative environment, how to effectively allocate order resources and balance the processing capabilities and expertise of each platform has become a problem that needs to be solved urgently. Traditional order allocation methods often use simple rule matching or random allocation, which cannot achieve the optimal allocation of resources and easily cause some platforms to be overloaded while other platforms have idle resources. On the other hand, the formulation of member rights allocation strategies usually relies on static rules, lacks flexibility and personalization, and cannot adjust the rights portfolio in time according to the dynamic value and preferences of members. This not only affects member satisfaction, but also limits the potential for companies to enhance member value through precision marketing. Summary of the invention

[0004] The present invention provides a member order processing method, device, equipment and storage medium, which can comprehensively utilize multi-platform data to realize intelligent order processing and dynamic equity distribution.

[0005] In a first aspect, the present invention provides a member order processing method, the member order processing method comprising: Collect member information from multiple platforms and construct a member graph structure to extract node features, and obtain a multi-dimensional member feature vector; Performing feature fusion on the multi-dimensional member feature vector and the order data to be processed to obtain a member-order fusion feature vector; Classify the orders based on the member-order fusion feature vector to obtain order classification results and processing priorities; Based on the order classification results and the processing priorities, multi-platform collaborative integer programming and solving are performed to obtain a cross-platform order allocation solution; Perform member value assessment based on the cross-platform order allocation plan and order processing results to obtain dynamic member value scores and hierarchical division information; The dynamic member value score and the hierarchical division information are input into a dynamic equity allocation model based on a deep Q network to optimize the equity allocation strategy and obtain a personalized member equity combination.

[0006] In a second aspect, the present invention provides a member order processing device, the member order processing device comprising: The collection module is used to collect member information from multiple platforms and construct a member graph structure to extract node features and obtain a multi-dimensional member feature vector; A fusion module, used to perform feature fusion on the multi-dimensional member feature vector and the order data to be processed to obtain a member-order fusion feature vector; A classification module, used to classify orders based on the member-order fusion feature vector, and obtain order classification results and processing priorities; A planning module, used for performing multi-platform collaborative integer programming and solving based on the order classification results and the processing priorities to obtain a cross-platform order allocation plan; An evaluation module, used to evaluate member value based on the cross-platform order allocation plan and order processing results, and obtain dynamic member value scores and hierarchical division information; The strategy optimization module is used to input the dynamic member value score and the hierarchical division information into a dynamic equity allocation model based on a deep Q network to optimize the equity allocation strategy and obtain a personalized member equity combination.

[0007] The third aspect of the present invention provides a member order processing device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the member order processing device executes the above-mentioned member order processing method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned member order processing method.

[0009] In the technical solution provided by the present invention, by adopting graph neural network and soft attention mechanism to extract member features, the complex relationship between member attributes can be effectively captured, and the accuracy and comprehensiveness of member feature representation can be improved. The adversarial bidirectional encoder and bidirectional LSTM network are used to fuse member-order features, and the deep integration of member information and order data is realized. The order classification method based on multi-layer perceptron network and support vector regression model can accurately identify order types and reasonably allocate processing priorities, thereby improving the efficiency and accuracy of order processing. The multi-platform collaborative integer programming model is used for order allocation, which fully considers the processing capabilities and expertise of each platform, realizes the optimal configuration of order resources, and improves the overall order processing efficiency. The entropy weight method and hierarchical analysis method are combined to evaluate member value, which not only considers the distribution characteristics of objective data, but also incorporates expert experience, making the member value evaluation results more comprehensive and reliable. The dynamic equity allocation model based on deep Q network is used to realize the personalized allocation and real-time optimization of member rights and interests, and can timely adjust the equity combination according to the dynamic value and preference of members, thereby improving member satisfaction and loyalty. The present invention improves the efficiency and accuracy of member management and order processing, and provides a comprehensive member order processing solution for enterprises.

[0010] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0011] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A schematic diagram of an embodiment of a method for processing member orders in an embodiment of the present invention; Figure 2 A schematic diagram of an embodiment of a member order processing device in an embodiment of the present invention; Figure 3 It is a schematic diagram of an embodiment of a member order processing device in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0015] To facilitate understanding of this embodiment, a member order processing method disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps: 101. Collect member information from multiple platforms and construct a member graph structure to extract node features, thereby obtaining a multi-dimensional member feature vector; It is understandable that the execution subject of the present invention may be a member order processing device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0016] Specifically, the member information of multiple platforms is cleaned and formatted in a unified manner, duplicate data is removed, missing values ​​are processed, and invalid data items are eliminated. At the same time, the data formats of each platform are unified to meet the unified data standards, and a standardized member data set is obtained. Based on the standardized member data set, a member graph structure is constructed. The construction of the graph structure takes members as nodes, and the relationship between member attributes is set as edges. Through the setting of nodes and edges, a graph representing members and their mutual relationships is formed. Each node represents a member, and the edge represents the association relationship between members, such as interaction frequency, same purchase behavior and other characteristics, thereby forming a member network graph structure reflecting multidimensional relationships. The member graph structure is represented by a sparse matrix, which effectively compresses the amount of information in the graph and improves the efficiency of storage and calculation. The member graph structure is encoded by the adjacency matrix and the feature matrix to obtain a graph structure encoding matrix. Among them, the adjacency matrix is ​​used to represent the connection relationship between nodes, and the feature matrix is ​​used to represent the attribute characteristics of each node. The two matrices are combined to form the overall encoding representation of the graph structure, so that the structured information of the graph is fully input into the subsequent neural network. The graph structure encoding matrix is ​​input into the self-attention gated graph neural network for processing. Graph neural networks can effectively extract and aggregate features of nodes in a graph. Through multi-layer graph convolution operations, the features of nodes are updated layer by layer, and the information of neighboring nodes is gradually aggregated to obtain an updated node feature matrix. In this process, the self-attention mechanism is used to enhance the ability to identify the importance of nodes in the graph convolution process, so that in the process of information aggregation, the network assigns different weights to different neighboring nodes, thereby focusing on important features. Based on the updated node feature matrix, the attention weights between nodes are calculated to capture the dynamic relationship between nodes. The node attention weight matrix is ​​obtained by calculating the attention weights between nodes based on the updated node feature matrix and normalizing the attention scores using the softmax function. The node attention weight matrix is ​​matrix multiplied with the updated node feature matrix to obtain a weighted node feature matrix. The weighted node feature matrix is ​​summed by column vectors and normalized by L2 norm to obtain a graph-level representation vector. Column vector summation aggregates node-level information to the overall level of the graph to obtain a feature vector that describes the entire graph. L2 norm normalization is used to standardize the vector so that the size of the vector is not affected by the feature dimension. The graph-level representation vector is connected with the basic attribute vector of the member, and feature fusion and dimensionality reduction are performed through the fully connected layer to obtain the final multi-dimensional member feature vector. Through the feature fusion of the fully connected layer, the graph structure information and the basic attributes of the member are combined to generate a comprehensive feature vector containing graph structure features and attribute features. At the same time, dimensionality reduction processing effectively reduces the dimension of the feature, reduces the computational complexity, and avoids overfitting of the model under high-dimensional data.

[0017] 102. Perform feature fusion on the multi-dimensional member feature vector and the pending order data to obtain a member-order fusion feature vector; Specifically, the multi-dimensional member feature vector is dimensionally expanded. By copying and zero-filling operations, the multi-dimensional member feature vector is expanded to the same length as the order data sequence to be processed, and the expanded member feature vector is obtained. The expanded member feature vector is concatenated with the order data to be processed in the feature dimension to obtain the initial fused feature sequence. In the fused feature sequence, each time step contains the member's features and the corresponding order features. The initial fused feature sequence is input into a bidirectional long short-term memory network (LSTM). The bidirectional LSTM includes a forward LSTM and a reverse LSTM. The LSTM network in each direction consists of two layers, each layer contains 128 hidden units, and the tanh activation function is used to introduce nonlinear characteristics. Through forward and reverse bidirectional processing, the temporal dependency between the front and back directions in the fused feature sequence is captured. The forward LSTM feature sequence and the reverse LSTM feature sequence start from two time directions respectively to extract the potential temporal correlation between the member features and the order features. The forward LSTM feature sequence and the reverse LSTM feature sequence are concatenated at the output of the last hidden layer to obtain a bidirectional LSTM feature sequence, which not only contains the global temporal information of the input sequence, but also effectively combines the forward and reverse contextual information. The bidirectional LSTM feature sequence is input into the multi-head self-attention layer. The multi-head self-attention layer can mine important information in the feature sequence. Eight attention heads are set, each with a dimension of 64, and query, key, and value matrices are generated through linear transformation. Then, the self-attention feature sequence is obtained by calculating the scaled dot product attention and performing softmax normalization. The multi-head attention mechanism can capture the hidden feature relationships in the sequence from different angles and scales, enabling the model to pay more attention to the correlation between different time steps. Through the parallel calculation of multiple attention heads, multiple information patterns in the input sequence are captured at the same time, enhancing the feature expression ability of the model. The self-attention feature sequence is input into a two-layer feedforward neural network for feature processing. The first layer uses the ReLU activation function to introduce nonlinearity to enhance the expression ability of the model; the second layer uses the linear activation function to linearly transform the features. The dimension of the hidden layer is set to 512 to ensure that the network can process rich enough information and perform complex feature transformations to obtain the encoder output feature sequence. The encoder's feedforward network fuses and transforms the features extracted by LSTM and self-attention to improve the abstractness and effectiveness of the features. In order to enhance the discriminative ability of the encoder to extract features, the encoder output feature sequence is input into the adversarial discriminator for optimization. The adversarial discriminator consists of three layers of one-dimensional convolutional layers with kernel sizes of 3, 4, and 5, respectively, to capture local features at different time steps. The number of output channels of each convolution layer is 100, ensuring that the convolution layer has sufficient capacity to extract rich feature patterns. The discriminator is followed by a fully connected layer to output the binary classification results.Through adversarial training, the encoder and the discriminator compete with each other, so that the encoder can learn more discriminative and generalized features, and obtain the feature sequence after adversarial optimization. The feature sequence after adversarial optimization is subjected to maximum pooling operation in the time dimension, and the member-order fusion feature vector is obtained by selecting the maximum value in each feature dimension.

[0018] 103. Classify orders based on the member-order fusion feature vector to obtain order classification results and processing priorities; Specifically, the member-order fusion feature vector is input into the multilayer perceptron network. The multilayer perceptron network consists of three hidden layers, the first layer contains 512 neurons, the second layer contains 256 neurons, and the third layer contains 128 neurons. The ReLU activation function is used between each layer to introduce nonlinear capabilities, so that the model can learn more complex feature patterns. The last layer uses a linear activation function to maintain the continuity of the features and obtain the representation of the order features. The multilayer structure of the multilayer perceptron effectively abstracts the input fusion feature vector layer by layer, extracts the deep features representing the order, and reflects the correlation between the order and the member and the various attributes of the order. The order feature representation is L2 regularized. The L2 norm of the order feature vector is calculated, and the original feature vector is divided by the L2 norm to obtain the normalized order feature vector. The L2 regularization process can ensure the consistency of the scale of the feature vector and prevent the subsequent classification from being adversely affected by the scale difference of different feature values. At the same time, the normalization operation can avoid overfitting of the model to a certain extent and improve the generalization ability of the model. The normalized order feature vector is input into the softmax classifier. The softmax classifier calculates the probability of each category through the softmax function to obtain the category probability distribution of the order. The softmax function maps the feature vector to a probability space so that the output is the probability value of each possible category, and the sum of the probabilities is 1. Based on the probability distribution of order categories, the argmax function is used to find the index of the category with the highest probability value, and the category corresponding to the index is used as the classification result of the order. The entropy of the order category probability distribution is calculated to evaluate the uncertainty of the classification result. Entropy is an important concept in information theory, which is used to measure the degree of uncertainty of probability distribution. For the probability distribution of order categories, a scalar value representing the uncertainty of classification is obtained by calculating its entropy value. The scalar entropy value is compared with the predefined threshold. If the entropy value is less than the threshold, it indicates that the classification is highly certain and the order is considered to belong to the determined category; on the contrary, if the entropy value is greater than the threshold, it indicates that the model has a high uncertainty in the classification of the order and it is classified as an uncertain category. For orders with determined classification, the preset priority mapping table is queried based on the order classification result. The priority mapping table is a mapping rule that maps each order category to an integer priority from 1 to 5, and predefines the processing priority of each category. For example, certain important categories of orders are given higher priorities to ensure that these orders are processed first. The preliminary processing priority of the order is obtained by querying the priority mapping table. For orders with uncertain classification, a more refined priority determination method is adopted. The normalized order feature vector is input into a pre-trained support vector regression model. The support vector regression model uses the RBF kernel function to capture the complex nonlinear relationship of order features in the feature space.Through the support vector regression model, a continuous priority score is obtained to reflect the importance of the order. The continuous score is converted into a discrete processing priority from 1 to 5 by setting interval boundaries, providing a more flexible and sophisticated priority determination strategy for orders with uncertain classification.

[0019] 104. Perform multi-platform collaborative integer programming and solving based on order classification results and processing priorities to obtain a cross-platform order allocation plan; Specifically, a weight coefficient is assigned to each order according to the classification result and processing priority of the order. The calculation formula of the weight coefficient is: weight = α* classification score + β* priority, where α and β are preset balance parameters, which are used to balance the impact of the classification score and processing priority of the order on the final weight. By calculating the weight coefficient, the relative importance of each order is quantified, and the order weight list is obtained. The platform order fitness matrix is ​​constructed according to the processing capabilities and expertise of each platform. The element aij of the matrix represents the fitness of platform i to process order j, and the fitness range is between 0 and 1, which is used to quantify the matching degree between the platform and the order. For example, if a platform is good at processing a certain type of order, its fitness will be higher, while the fitness for other orders that it is not good at will be lower. Based on the order weight list and the platform order fitness matrix, the objective function of order allocation is constructed, and three constraints are set. The construction of the objective function aims to maximize the overall order processing efficiency, which is usually defined as the sum of the weighted scores of all orders on each platform. The weighted score of each order is obtained by multiplying its weight and the platform adaptability. By maximizing the weighted score, important orders are ensured to be assigned to the most suitable platform for processing. While constructing the objective function, three constraints are set to ensure the rationality of order allocation. The first constraint is that each order can only be assigned to one platform to ensure the independence of order allocation; the second constraint is that the order processing volume of each platform cannot exceed the maximum processing capacity of the platform to avoid overload; the third constraint is that the cross-platform allocation ratio of preset type orders cannot exceed the preset threshold. This condition is used to ensure that the allocation of certain specific types of orders meets business needs and policy requirements. The objective function and the three constraints are combined to form a multi-platform collaborative integer programming model. The core of the multi-platform collaborative integer programming model is the decision variable, which represents the weighted score of each order on each platform. The goal of the model is to maximize the benefits of overall order processing by appropriately allocating orders. The design of the multi-platform collaborative integer programming model can comprehensively consider the order weight, platform adaptability, processing capacity of each platform and order type allocation requirements to achieve the global optimal solution for cross-platform order allocation. The multi-platform collaborative integer programming model is input into the interactive optimization solver for solving. Before inputting the solver, some key parameters of the solution are set, including the maximum number of iterations, convergence threshold, and time limit. The maximum number of iterations is used to control the number of iterations of the optimization algorithm to prevent the solution process from entering an infinite loop; the convergence threshold is used to determine whether the solution has reached the optimal solution or is close enough to the optimal solution; the time limit is used to limit the solution time to prevent the overall performance of the system from being affected when the solution time is too long. After setting the parameters, start the interactive optimization solver to start the solution process. During the solution process, the interactive optimization solver continuously adjusts the distribution of orders among various platforms to maximize the value of the objective function while ensuring that all constraints are met.When the solver completes the solution, the solution result is obtained from it, and the optimal value of the decision variable is extracted. The order-platform pairs with a decision variable of 1 are extracted and combined into a cross-platform order allocation plan. The final cross-platform order allocation plan can achieve reasonable and efficient order allocation, so that the processing capacity of each platform can be fully utilized, while ensuring the overall benefit of order processing is maximized.

[0020] 105. Evaluate member value based on the cross-platform order allocation plan and order processing results to obtain dynamic member value scores and tier division information; Specifically, based on the cross-platform order allocation scheme and the order processing results of pending order data, various indicators of member value evaluation are extracted, including consumption frequency, consumption amount, order completion rate, platform activity and customer service demand, reflecting the consumption behavior, activity level and interaction status of members with the platform, so as to construct an original indicator matrix for evaluating member value. The original indicator matrix is ​​normalized to eliminate the dimensional differences between indicators so that each indicator can be compared and calculated on the same scale. The range transformation method is used to transform indicators of different dimensions and convert them into dimensionless values ​​to obtain a standardized indicator matrix. Based on the standardized indicator matrix, the information entropy of each indicator is calculated. The information entropy is used to measure the uncertainty and information size of each indicator in the entire evaluation process. The entropy weight method is used to calculate the objective weight of each indicator according to the information entropy of each indicator to obtain an objective weight vector. The entropy weight method measures the importance of each indicator through objective data, which can reflect the degree of contribution to the overall value in actual member behavior. A hierarchical analysis structure model is constructed to divide the member value evaluation indicators into three levels: target level, criterion level and indicator level. In the goal layer, the ultimate goal is to evaluate the value of members; in the criterion layer, the main aspects that affect member value are divided, such as consumption behavior, platform activity and customer service needs; and in the indicator layer, specific evaluation indicators are included, such as consumption frequency, consumption amount, order completion rate, etc. A judgment matrix is ​​established to compare the relative importance of each indicator. The judgment matrix is ​​constructed by expert scoring and represents the pairwise comparison results between each indicator. The judgment matrix is ​​tested for consistency to ensure the logical consistency of the judgment matrix. The maximum eigenvalue λmax of the judgment matrix is ​​calculated, and the consistency index CI is calculated according to the formula CI=(λmax-n) / (n-1), where n is the order of the judgment matrix. The consistency ratio CR=CI / RI is calculated, where RI is the random consistency index. If the consistency ratio CR is less than 0.1, it is considered that the judgment matrix passes the consistency test, indicating that the judgment matrix has good logical consistency in subjective judgment; if CR≥0.1, the value in the judgment matrix needs to be adjusted until it passes the consistency test. Based on the judgment matrix that passes the consistency test, the eigenvalue method is used to calculate the subjective weight of each indicator to obtain the subjective weight vector. Subjective weights reflect the subjective judgment of experts on the relative importance of each indicator, and can introduce human experience and domain knowledge. In order to comprehensively consider objective data and expert judgment, the objective weight vector and the subjective weight vector are weighted averaged to obtain a comprehensive weight vector. Based on the comprehensive weight vector and the standardized indicator matrix, the dynamic value score of each member is calculated. The comprehensive weight vector is weighted and summed with the indicator data of each member in the standardized indicator matrix to obtain the total value score of each member, reflecting the comprehensive performance of members in multiple aspects such as consumption and activity. The member value scores are clustered to obtain hierarchical division information.Through clustering algorithms (such as K-means clustering), members are grouped according to their value scores and divided into different levels.

[0021] 106. The dynamic member value score and hierarchical division information are input into the dynamic equity allocation model based on the deep Q network to optimize the equity allocation strategy and obtain a personalized member equity combination.

[0022] Specifically, the dynamic member value score and the hierarchical division information are combined into a state vector S. The state vector S contains the member's value score, member level, and historical equity usage rate. The state vector S is input into a deep Q network, which contains 3 hidden layers, each with 256, 128, and 64 neurons, respectively, and a ReLU activation function is used between each layer. The ReLU activation function introduces nonlinear capabilities, enabling the model to better learn the relationship between complex states and actions. Through the deep structure, the deep Q network extracts high-dimensional features from the state vector and outputs a Q value matrix. Each element in the Q value matrix represents the expected return that can be obtained by taking a certain equity distribution action under the current state. Based on the Q value matrix, the ε-greedy strategy is used to select the equity distribution action. The ε-greedy strategy is an exploration and exploitation strategy. When selecting an action, it randomly selects an action with a probability of ε, and selects the action with the largest current Q value with a probability of (1-ε). In order to better meet the needs of different member levels, the ε value is dynamically adjusted according to the level of the member. For higher-level members, the ε value will be smaller so that the action with the largest expected return can be selected more frequently, thereby improving the accuracy of equity distribution; while for lower-level members, the ε value is relatively large so that more equity distribution combinations can be explored and potential optimal solutions can be tapped. In this way, the initial equity distribution plan is obtained. The feasibility of the initial equity distribution plan is checked to ensure that the allocated equity meets the preset equity distribution rules and resource constraints. The rules and constraints include the maximum number of equity restrictions, platform resource restrictions, and the types of equity that each member can obtain within a certain period of time. By iteratively adjusting the equity items in the plan, the initial plan is optimized on the premise of meeting all rules and constraints to obtain an effective equity distribution plan. The effective equity distribution plan is applied to specific members, and feedback from members on the use of the allocated equity is collected within a fixed time window. The actual use of the equity by members is an important basis for evaluating the effectiveness of the equity distribution strategy, and the instant reward value is calculated based on the preset reward function. The reward function is designed based on factors such as the frequency of members using the equity, satisfaction, and further consumption behavior on the platform. The instant reward value reflects the effectiveness of the current equity distribution plan and the satisfaction of members. The cumulative reward is calculated based on the instant reward value and the preset discount factor. The discount factor is used to balance the relationship between current rewards and future rewards, thereby optimizing the member equity distribution strategy in the long term. Each experience sample (i.e., the four-tuple of state, action, reward, and next state) is stored in an experience replay pool with a capacity of 10,000. The experience replay pool is used to store historical experience samples. By random sampling, the training process of the model does not depend on the order of samples, reduces the correlation between samples, and improves the stability and generalization ability of the model. A small batch of experience data of size 64 is randomly sampled from the experience replay pool to update the parameters of the deep Q network. The dual Q learning algorithm is used to update the parameters of the target Q network.Dual Q learning introduces two independent Q networks to estimate the value of actions and select actions respectively, thereby effectively reducing the overestimation problem in Q value estimation and improving the stability and convergence speed of the model. After each parameter update, the updated target Q network parameters are copied to the evaluation Q network in a soft update manner. The soft update sets a smaller learning rate so that the target network gradually approaches the evaluation network, ensuring the stability of model training and the effect of gradual convergence. When the preset convergence conditions are met, a personalized member rights portfolio is output. The personalized rights portfolio is based on the member's dynamic value score and hierarchical division information, combined with historical rights usage feedback, and continuously optimized through deep reinforcement learning to maximize member satisfaction with the use of rights and improve the overall operational efficiency of the platform.

[0023] In the embodiment of the present invention, by adopting graph neural network and soft attention mechanism to extract member features, the complex relationship between member attributes can be effectively captured, and the accuracy and comprehensiveness of member feature representation are improved. The adversarial bidirectional encoder and bidirectional LSTM network are used to fuse member-order features, and the deep integration of member information and order data is realized. The order classification method based on multi-layer perceptron network and support vector regression model can accurately identify the order type and reasonably allocate processing priority, thereby improving the efficiency and accuracy of order processing. The multi-platform collaborative integer programming model is used for order allocation, which fully considers the processing capabilities and expertise of each platform, realizes the optimal configuration of order resources, and improves the overall order processing efficiency. The entropy weight method and hierarchical analysis method are combined to evaluate member value, which not only considers the distribution characteristics of objective data, but also incorporates expert experience, making the member value evaluation results more comprehensive and reliable. The dynamic equity allocation model based on deep Q network is used to realize the personalized allocation and real-time optimization of member rights and interests, and can timely adjust the equity combination according to the dynamic value and preference of members, thereby improving member satisfaction and loyalty. The present invention improves the efficiency and accuracy of member management and order processing, and provides a comprehensive member order processing solution for enterprises.

[0024] In a specific embodiment, the process of executing step 101 may specifically include the following steps: Perform data cleaning and format unification on member information on multiple platforms to obtain a standardized member data set. Based on the standardized member data set, member attributes are set as nodes, and the relationships between attributes are set as edges to construct a member graph structure. The member graph structure is represented by a sparse matrix, and the member graph structure information is encoded through the adjacency matrix and the feature matrix to obtain the graph structure encoding matrix, and the graph structure encoding matrix is ​​input into the self-attention gated graph neural network. The node features are updated through multi-layer graph convolution operations to obtain the updated node feature matrix; The attention weights between nodes are calculated based on the updated node feature matrix, and the attention scores are normalized using the softmax function to obtain the node attention weight matrix. The node attention weight matrix is ​​then multiplied by the updated node feature matrix to obtain the weighted node feature matrix. The column vector summation operation is performed on the weighted node feature matrix, and the L2 norm normalization processing is performed to obtain the graph-level representation vector. The graph-level representation vector is connected to the member information attribute vector, and feature fusion and dimensionality reduction processing are performed through the fully connected layer to obtain a multi-dimensional member feature vector.

[0025] Specifically, member data from different platforms are cleaned and formatted in a unified manner. Through data cleaning, noise in the data is removed, missing values ​​are filled, and duplicate data is merged to ensure that the data on all platforms are consistent in semantics and format. Format unification standardizes data from different platforms into the same format, such as converting data from different units into consistent units, and mapping all member information into a common structured data table to form a standardized member data set. Based on the standardized member data set, a member graph structure is constructed. In the graph structure, member attributes are set as nodes, and the relationship between attributes is set as edges. For example, there are common purchase records or similar browsing behaviors between member A and member B, which are represented as an edge in the graph; member A's attributes (such as purchase frequency, browsing habits) are used as nodes in the graph. In the member graph structure, each node represents a member, and the existence of an edge indicates that there is some kind of relationship between these members, such as common interests, similar consumption behaviors, etc. The member graph structure is represented by a sparse matrix, which effectively compresses the amount of information in the graph and improves the efficiency of storage and calculation. Using the adjacency matrix represents the connectivity of the graph, where Representation Node and nodes There are edges between them, and This means that there is no direct relationship between the two. At the same time, through the feature matrix Represents the attribute characteristics of each node, such as consumption frequency, purchase amount, etc., thus obtaining a matrix containing node feature information. The graph structure encoding matrix is ​​combined with the adjacency matrix and the feature matrix It can be expressed as: ; in, represents the initial graph structure encoding matrix, is the adjacency matrix, and is the feature matrix. The graph structure encoding matrix The input is fed into the self-attention gated graph neural network, and the node features are updated through multi-layer graph convolution operations. Graph convolution aggregates the features of neighboring nodes and updates the features of the node itself. The application of the self-attention mechanism in graph convolution assigns different neighbor weights to each node, focusing on the neighbor nodes that are most important for updating node features. Through multi-layer graph convolution operations, the information of distant nodes is gradually aggregated to each node, and the updated node feature matrix is ​​obtained, which is recorded as ,in Indicates the number of graph convolution layers. Based on the updated node feature matrix , calculate the attention weights between nodes, and quantify the importance of each node to other nodes during the feature update process. The attention scores between each pair of nodes are calculated through the self-attention mechanism, and the softmax function is used to normalize the attention scores to obtain the node attention weight matrix The soft attention mechanism ensures that the sum of the attention scores of all neighbor nodes is 1 through normalization, generating a probability distribution. and the updated node feature matrix Perform matrix multiplication to obtain the weighted node feature matrix: ; in, Represents the weighted node feature matrix, which aggregates the features of neighboring nodes through attention weights to reflect the importance and dependency between nodes. Perform column vector summation to aggregate all node features into a graph-level representation vector. Column vector summation aggregates node-level information to the overall level of the graph, and obtains a feature vector describing the entire graph, recorded as In order to eliminate the deviation caused by the inconsistency of eigenvalues, Perform L2 norm normalization to obtain the standardized graph-level representation vector: ; in, Representation vector The L2 norm of is the normalized graph-level representation vector. Connect with the basic attribute vector of the member to obtain a comprehensive feature vector containing graph structure information and basic member information. The comprehensive feature vector is passed through a fully connected layer for feature fusion and dimensionality reduction. The fully connected layer can perform linear transformation on the input high-dimensional features, introduce nonlinearity through the activation function, extract more meaningful features, and map the high-dimensional feature vector to a lower-dimensional space, finally obtaining a multi-dimensional member feature vector.

[0026] In a specific embodiment, the process of executing step 102 may specifically include the following steps: Performing dimension expansion processing on the multi-dimensional member feature vector, expanding the multi-dimensional member feature vector to the same length as the order data sequence to be processed through replication and zero-filling operations, and obtaining an expanded member feature vector; The extended member feature vector is concatenated with the pending order data in the feature dimension to obtain an initial fused feature sequence, where each time step contains the member feature and the corresponding order feature; The initial fused feature sequence is input into the bidirectional LSTM network, where the forward LSTM and the reverse LSTM each contain two layers, each layer contains 128 hidden units, and the tanh activation function is used to obtain the forward LSTM feature sequence and the reverse LSTM feature sequence; The outputs of the forward LSTM feature sequence and the reverse LSTM feature sequence in the last hidden layer are concatenated to obtain a bidirectional LSTM feature sequence, which contains temporal dependency information captured from two directions. Input the bidirectional LSTM feature sequence into the multi-head self-attention layer, set 8 attention heads, each with a dimension of 64, generate query, key and value matrices through linear transformation, calculate the scaled dot product attention and perform softmax normalization to obtain the self-attention feature sequence; The self-attention feature sequence is input into a two-layer feedforward neural network. The first layer uses the ReLU activation function, the second layer uses the linear activation function, and the hidden layer dimension is 512. Nonlinear transformation is performed to obtain the encoder output feature sequence; The feature sequence output by the encoder is input into the adversarial discriminator. The adversarial discriminator contains three one-dimensional convolutional layers with convolution kernel sizes of 3, 4, and 5 respectively. The number of output channels of each layer is 100. Finally, a fully connected layer is connected to output the binary classification result. The encoder parameters are optimized through adversarial training to obtain the feature sequence after adversarial optimization. Perform the maximum pooling operation on the time dimension of the feature sequence after adversarial optimization, select the maximum value on each feature dimension, and obtain the member-order fusion feature vector.

[0027] Specifically, the multi-dimensional member feature vector is dimensionally expanded, and the multi-dimensional member feature vector is expanded to the same length as the order data sequence to be processed through replication and zero padding operations, so that it has the same time dimension as the order data, and the member and order features are aligned. The expanded member feature vector is concatenated with the order data to be processed on the feature dimension to obtain the initial fused feature sequence. Assume that the expanded member feature vector is , the order data is ,in is the length of the order sequence, is the dimension of member characteristics, is the dimension of the order feature. After concatenating the two in the feature dimension, we get the initial fusion feature sequence , where the features of each time step include the corresponding member features and order features. Input into the bidirectional LSTM network. The bidirectional LSTM network contains forward LSTM and reverse LSTM. Each LSTM has two layers, and each layer contains 128 hidden units. LSTM is a recurrent neural network used to process time series data and can capture the time dependency in the data. The forward LSTM starts from the time step arrive Processing the input sequence, the reverse LSTM starts from the time step arrive The input sequence is processed in reverse order to capture the temporal features in the input sequence from two directions. Each layer of the LSTM network uses the tanh activation function to introduce nonlinear features. After bidirectional LSTM processing, the forward LSTM feature sequence is obtained. and reverse LSTM feature sequence The output of the forward and reverse LSTM feature sequences at the last hidden layer are concatenated to obtain a bidirectional LSTM feature sequence The bidirectional LSTM feature sequence contains the temporal dependency information captured from two directions, which can reflect the relationship between the order and member features at each time step. Input to the multi-head self-attention layer. The self-attention layer is used to calculate the dependencies between different time steps. 8 attention heads are set, and the dimension of each head is 64. Generate the query matrix through linear transformation , key matrix Sum Matrix ,in: ; in, is a linear transformation matrix, which is used to generate query, key and value matrices respectively. By calculating the scaled dot product attention and performing softmax normalization, the attention weight matrix is ​​obtained and multiplied with the value matrix to obtain the self-attention feature sequence: ; in, is the dimension of the key matrix, and the softmax function is used to normalize the attention scores so that the relationship between each time step can be effectively measured. Through the multi-head attention mechanism, the information pattern in the sequence is captured from different angles at the same time, and the self-attention feature sequence is obtained. . The self-attention feature sequence The input is fed into a two-layer feedforward neural network. The first layer uses the ReLU activation function to introduce nonlinearity to enhance the expressiveness of features. The dimension of the hidden layer is 512. The second layer uses a linear activation function to maintain the continuity of the output. Through the nonlinear transformation of the feedforward neural network, the deep features in the sequence are extracted to obtain the encoder output feature sequence. In order to enhance the discriminative ability of feature sequences, the encoder output feature sequence Input into the adversarial discriminator. The adversarial discriminator contains three one-dimensional convolutional layers, with convolution kernel sizes of 3, 4, and 5 respectively, and the number of output channels of each layer is 100. The convolution operation extracts local patterns in the feature sequence, and finally outputs the binary classification results through a fully connected layer to judge the authenticity or quality of the feature sequence. Through adversarial training, the encoder and the discriminator compete with each other, and the encoder is forced to generate more discriminative features, thereby optimizing its parameters and obtaining the feature sequence after adversarial optimization. The maximum pooling operation is performed on the feature sequence after adversarial optimization in the time dimension, and the maximum value is selected in each feature dimension to retain the most significant information in the feature sequence. The dimension of the feature is effectively reduced, and the most representative features are extracted at the same time, and finally the member-order fusion feature vector is obtained.

[0028] In a specific embodiment, the process of executing step 103 may specifically include the following steps: The member-order fusion feature vector is input into the multi-layer perceptron network. The multi-layer perceptron network contains three hidden layers. The first layer has 512 neurons, the second layer has 256 neurons, and the third layer has 128 neurons. The ReLU activation function is used between each layer, and the last layer uses the linear activation function to obtain the order feature representation; Perform L2 regularization on the order feature representation, calculate the L2 norm of the feature vector, and divide the feature vector by the L2 norm to obtain the normalized order feature vector; Input the normalized order feature vector into the softmax classifier, use the softmax function to calculate the probability of each category, and obtain the order category probability distribution. Based on the order category probability distribution, use the argmax function to find the index with the highest probability, and use the category corresponding to the index as the order classification result. The entropy of the probability distribution of order categories is calculated to obtain a scalar entropy value representing classification uncertainty, and the scalar entropy value representing classification uncertainty is compared with a predefined threshold. If the scalar entropy value representing classification uncertainty is less than the threshold, the order is classified into a certain category, otherwise it is classified into an uncertain category. For orders with certain classification, the preset priority mapping table is queried based on the order classification results. The priority mapping table maps each order category to an integer priority of 1-5 to obtain a preliminary processing priority. For orders with uncertain classification, the normalized order feature vector is input into the pre-trained support vector regression model. The pre-trained support vector regression model uses the RBF kernel function to obtain a continuous priority score, and the continuous score is converted into a discrete processing priority of 1-5 by setting the interval boundary.

[0029] Specifically, the member-order fusion feature vector is input into the multi-layer perceptron network. The multi-layer perceptron network contains three hidden layers. The first layer contains 512 neurons, the second layer contains 256 neurons, and the third layer contains 128 neurons. The ReLU activation function is used between each layer. By introducing the nonlinear activation function ReLU, the model captures complex feature relationships and increases its representation ability. After each layer of neurons and ReLU activation, a linear activation function is used in the last layer to maintain the continuity of the output and obtain the feature representation of the order. The input feature vector is , No. The output of the layer is represented as: ; in, It is The weight matrix of the layer, It is The bias vector of the layer, represents the output of the previous layer, is the output after ReLU activation. After passing through all hidden layers, the order feature representation is obtained. In order to ensure the scale consistency and numerical stability of the order feature representation, the order feature representation is L2 regularized. The L2 norm of the feature vector is calculated, and then the feature vector is divided by the L2 norm to obtain the normalized order feature vector. The calculation formula of the L2 norm is: ; in, represents the L2 norm of the feature vector, is the order feature representation Quantity, is the dimension of the feature vector. The normalized order feature vector It is expressed as: ; Normalization ensures that the length of the feature vector is 1, eliminating the impact of inconsistent feature values ​​on model training and prediction. The normalized order feature vector is input into the softmax classifier to classify the orders. The softmax classifier uses the softmax function to calculate the probability of each category. The expression of the softmax function is: ; in, Indicates that orders are classified into categories The probability of and The categories are The weight vector and bias of is the total number of categories. By calculating the probability of each category, the probability distribution of the order category is obtained. Use the argmax function to find the index with the highest probability, and use the category corresponding to the index as the classification result of the order. In this way, each order is classified to determine the category to which it belongs. In order to evaluate the certainty of the classification result, the entropy of the probability distribution of the order category is calculated to measure the uncertainty of the classification result. The entropy calculation formula is: ; in, is the entropy value of the probability distribution of order categories, Yes Category The probability of . Entropy value The larger the entropy value is, the higher the uncertainty of the classification is; the smaller the entropy value is, the clearer the classification result is. Comparing with the predefined threshold, if the entropy value is less than the threshold, it means that the classification result has a high certainty, and the order is classified into a certain classification; if the entropy value is greater than the threshold, it means that the classification result is uncertain, and the order is classified into an uncertain classification. For orders with a certain classification, the preset priority mapping table is queried based on the order classification result. The priority mapping table maps each order category to an integer priority of 1-5 to obtain the preliminary processing priority of the order. For example, if the order is classified as a high-value category, it is mapped to a priority of 1, indicating that it needs to be processed first. For orders with uncertain classification, a more detailed priority determination method is adopted. The normalized order feature vector is input into a pre-trained support vector regression model to obtain a continuous priority score for the order. The support vector regression model uses the RBF kernel function to capture the complex nonlinear relationship between order features and priority. The output of the support vector regression is a continuous score, which represents the priority of the order. In order to convert the continuous priority score into a discrete priority, the method of setting interval boundaries is adopted, such as dividing the score into an integer priority interval of 1-5, so that the priority of the order is comparable with other orders, and finally the processing priority of each order is obtained.

[0030] In a specific embodiment, the process of executing step 104 may specifically include the following steps: Based on the order classification results and processing priority, a weight coefficient is assigned to each order. The weight coefficient calculation formula is: weight = α*classification score + β*priority, where α and β are preset balance parameters, and the order weight list is obtained; According to the processing capabilities and expertise of each platform, a platform order suitability matrix is ​​constructed. The matrix element aij represents the suitability of platform i to process order j, and the suitability range is 0-1. Based on the order weight list and the platform order fitness matrix, the objective function is constructed and three constraints are set: each order can only be assigned to one platform, the order processing volume of each platform does not exceed the maximum processing capacity of the platform, and the cross-platform allocation ratio of preset type orders does not exceed the preset threshold; The objective function and three constraints are combined to construct a multi-platform collaborative integer programming model. The decision variables of the multi-platform collaborative integer programming model are the weighted scores of each order on each platform, and the goal is to maximize the overall order processing benefits. The multi-platform collaborative integer programming model is input into the interactive optimization solver, and the solution parameters including the maximum number of iterations, convergence threshold and time limit are set. The solution process is started, and the solution results are obtained from the interactive optimization solver. The optimal values ​​of the decision variables are extracted, and the order-platform pairs with the decision variable being 1 are combined into a cross-platform order allocation plan.

[0031] Specifically, a weight coefficient is assigned to each order according to the classification result and processing priority of the order. The calculation formula of the weight coefficient is: Classification score Priority, where and It is a preset balance parameter used to adjust the relative importance of classification score and priority in the final weight. The weight of each order is obtained through the calculation formula to form an order weight list. According to the processing capabilities and expertise of each platform, the platform order fitness matrix is ​​constructed. The elements of the platform order fitness matrix represent the platform Processing orders The degree of fit, The fitness value ranges from 0 to 1, which is used to indicate the degree of match between the platform and the order. The higher the fitness, the better the platform is at handling this type of order. Based on the order weight list and the platform order fitness matrix, the objective function of order allocation is constructed and related constraints are set. The core of the objective function is to maximize the overall order processing efficiency, which is defined as the sum of the weighted scores of all orders on each platform. The weighted score is expressed as follows: ; in, represents the overall order processing benefit, is a decision variable, representing the order Is it assigned to the platform? , if assigned then , otherwise 0; Indicates order The weight of Indicates platform For orders The degree of fit; and Represent the number of platforms and orders respectively. Objective function The goal is to maximize the weighted score, so that important orders are assigned to the most appropriate platform and the processing capacity of the platform is fully utilized. While constructing the objective function, three constraints are set to ensure the rationality and operability of order allocation. The first constraint is that each order can only be assigned to one platform, which is expressed by the following constraint formula: ; This formula means that each order It can only be assigned to one platform, thus ensuring the independence and clarity of order allocation. The second constraint is that the order processing volume of each platform cannot exceed the maximum processing capacity of the platform, which is expressed as the following formula: ; in, Indicates platform The maximum processing capacity of the platform is the constraint condition that ensures that the platform will not be overloaded due to excessive order volume and maintains the processing stability of the platform. The third constraint condition is that the cross-platform allocation ratio of preset type orders cannot exceed the preset threshold, ensuring that the cross-platform allocation of certain specific types of orders meets business needs and strategies. For example, certain high-value orders are only allowed to be allocated to specific platforms. The above objective function and three constraints are combined to construct a multi-platform collaborative integer programming model. The decision variable of the model is the weighted score of each order on each platform, and the goal is to maximize the overall order processing efficiency. The form of integer programming enables the system to find the optimal solution in the discrete space and determine how to allocate orders to maximize the efficiency while satisfying the constraints. The multi-platform collaborative integer programming model is input into the interactive optimization solver for solution. Before solving, some key parameters of the solution are set, including the maximum number of iterations, convergence threshold and time limit. The maximum number of iterations is used to control the amount of optimization calculation to ensure that the solution process will not go on endlessly; the convergence threshold is used to determine whether the solution has reached the optimal solution or is close enough to the optimal solution; the time limit is used to limit the solution time to avoid affecting the real-time performance of the system when the solution time is too long. By setting these parameters, the solver can find a better order allocation solution in a reasonable time. After the solver starts the solution process, the optimization algorithm continuously adjusts the distribution of orders between platforms to maximize the value of the objective function and ensure that all constraints are met. The interactive optimization solver gradually approaches the optimal solution by iteratively updating the decision variables. When the preset stop condition is reached, the final solution result is obtained from the solver and the optimal value of the decision variable is extracted. For each order and Platform , if the decision variable , it means the order Assigned to the platform . The order-platform pairs with decision variables equal to 1 are extracted and combined into the final cross-platform order allocation plan.

[0032] In a specific embodiment, the process of executing step 105 may specifically include the following steps: Based on the cross-platform order allocation plan and the order processing results of pending order data, member value evaluation indicators are extracted, including consumption frequency, consumption amount, order completion rate, platform activity and customer service demand, and the original indicator matrix is ​​constructed; The original indicator matrix is ​​normalized, and the indicators of different dimensions are converted into dimensionless values ​​using the range transformation method to obtain a standardized indicator matrix. Based on the standardized indicator matrix, the information entropy of each indicator is calculated, and the objective weight of each indicator is calculated using the entropy weight method to obtain an objective weight vector. Construct a hierarchical analysis structure model, divide the member value evaluation indicators into three levels: target level, criterion level and indicator level, and establish a judgment matrix; Perform consistency check on the judgment matrix, calculate the maximum eigenvalue λmax and consistency index CI=(λmax-n) / (n-1), where n is the matrix order, calculate the consistency ratio CR=CI / RI, where RI is the random consistency index. If the consistency ratio CR<0.1, the judgment matrix passes the consistency check; otherwise, adjust the value in the judgment matrix until it passes the consistency check; Based on the judgment matrix that has passed the consistency test, the eigenvalue method is used to calculate the subjective weight of each indicator to obtain the subjective weight vector, and the objective weight vector and the subjective weight vector are weighted averaged to obtain the comprehensive weight vector; Based on the comprehensive weight vector and the standardized indicator matrix, the value score of each member is calculated, and the member value scores are clustered to obtain the hierarchical division information.

[0033] Specifically, data on multiple important dimensions related to each member are collected, including consumption frequency, consumption amount, order completion rate, platform activity, and customer service demand. Each indicator represents a certain aspect of a member's performance on the platform, where consumption frequency measures the number of times a member consumes within a certain period of time, consumption amount is the total consumption amount of the member, order completion rate indicates the proportion of orders completed by members, platform activity reflects the interaction between members and the platform, such as the frequency of browsing and leaving messages, and customer service demand measures the frequency of members' demand for customer service help. By extracting this information from the cross-platform order allocation plan and order processing results, an original indicator matrix is ​​formed, in which each row represents a member and each column represents an indicator. The original indicator matrix is ​​normalized. Using the range transformation method, indicators of different dimensions are converted into dimensionless values ​​to obtain a standardized indicator matrix. The normalization formula of the range transformation method is: ; in, is the original indicator value, and are the minimum and maximum values ​​of the indicator respectively. is the normalized indicator value. Through normalization, all indicators are converted to the range of [0,1], thereby eliminating the dimensional differences between different indicators and making the values ​​of each indicator weighted and compared under the same standard. Based on the normalized indicator matrix, the information entropy of each indicator is calculated to evaluate the importance and information content of the indicator. Information entropy The calculation formula is: ; in, For the The information entropy of the indicator, is a positive constant, usually taken as , used to ensure that the entropy value range is between [0,1], For the Among the indicators The proportion of the standardized value of each member to the total value of the indicator. is the number of members. By calculating the information entropy of each indicator, the amount of information contained in the indicator in the entire data set is measured. The smaller the entropy value of the indicator, the stronger its distinguishing ability and the greater its importance. The objective weight of each indicator is calculated using the entropy weight method. The calculation formula for the objective weight is: ; in, For the The objective weight of an indicator, For the The information entropy of the indicator, is the total number of indicators. The objective weight of each indicator is calculated through this formula, reflecting the relative importance of each indicator to the member value. A hierarchical analysis structure model is constructed to divide the member value evaluation indicators into three levels: target level, criterion level and indicator level. The target level is the member value evaluation, the criterion level includes the main aspects that affect the member value, such as consumption behavior, activity and service demand, and the indicator level includes specific quantitative indicators, such as consumption frequency, consumption amount, etc. Then a judgment matrix is ​​established to compare the relative importance of each indicator in pairs. The elements of the judgment matrix are Indicates The indicator is relative to The importance of each indicator is scored by experts based on their experience. The judgment matrix is ​​checked for consistency to ensure its logical consistency. The maximum eigenvalue of the judgment matrix is ​​calculated. , and then calculate the consistency index ; in, is the maximum eigenvalue of the judgment matrix, is the order of the judgment matrix. Calculate the consistency ratio ; in, is a random consistency index, which depends on the order of the judgment matrix. When , the judgment matrix is ​​considered to have passed the consistency test, indicating that the logical consistency of the matrix is ​​good; otherwise, the value in the judgment matrix needs to be adjusted until it passes the consistency test. After passing the consistency test, the eigenvalue method is used to calculate the subjective weight of each indicator. The eigenvalue method calculates the eigenvector corresponding to the maximum eigenvalue of the judgment matrix and normalizes it to obtain the subjective weight vector of each indicator. In order to consider both objective and subjective weights, the objective weight vector and the subjective weight vector are weighted averaged to obtain a comprehensive weight vector: ; in, It is a weighing factor with a value range of [0,1], which is used to balance the influence of subjective weight and objective weight. Based on the comprehensive weight vector and the standardized indicator matrix, the comprehensive value score of each member is calculated. The calculation formula is: ; in, For the The comprehensive value score of each member, For the The comprehensive weight of the indicators, For the Members in The standardized value of each indicator. The final value score of each member is obtained by weighted summation to evaluate the overall value and contribution of the member. The comprehensive value scores of members are clustered to divide members into different levels. For example, using the K-means clustering algorithm, members are divided into three categories: high value, medium value, and low value. The clustered hierarchical division information helps the platform formulate targeted marketing and service strategies, so as to better manage member resources and improve member satisfaction and loyalty.

[0034] In a specific embodiment, the process of executing step 106 may specifically include the following steps: Combine the dynamic member value score and hierarchical division information into a state vector S, which includes the member value score, member level, and historical equity utilization rate. Input the state vector S into a deep Q network, which contains three hidden layers, each with 256, 128, and 64 neurons, respectively. Use the ReLU activation function to obtain the Q value matrix. Based on the Q-value matrix, the ε-greedy strategy is used to select the equity distribution action, where the ε value is dynamically adjusted according to the membership level to obtain the initial equity distribution plan; Conduct a feasibility check on the initial equity distribution plan to ensure that it meets the preset equity distribution rules and resource constraints, and obtain an effective equity distribution plan by iteratively adjusting the equity items in the plan; Apply the effective equity distribution plan to members, collect feedback on the use of equity by members within a fixed time window, and calculate the instant reward value based on the preset reward function; Based on the instant reward value and the preset discount factor, the cumulative reward is calculated and the experience tuple is stored in the experience replay pool with a capacity of 10,000; A small batch of experience data of size 64 is randomly sampled from the experience replay pool, and the parameters of the target Q network are updated using the dual Q learning algorithm. The updated target Q network parameters are copied to the evaluation Q network in a soft update manner until the preset convergence conditions are reached, and a personalized membership benefit combination is output.

[0035] Specifically, the member's dynamic value score, hierarchical classification information, and historical rights usage are combined to form a state vector The state vector As a description of the member, it contains three parts of information: the member's value score , used to reflect the overall contribution of members to the platform; member level information , used to indicate the member's level in the platform; and the utilization rate of historical benefits , describes the frequency of members' use of the rights and interests assigned to them in history. This information helps to judge members' interests and needs for different types of rights and interests. As input, it is input into the deep Q network. The deep Q network contains 3 hidden layers, each with 256, 128 and 64 neurons respectively, and the ReLU activation function is used between the hidden layers. The ReLU activation function is a nonlinear function defined as , which can introduce nonlinearity and thus enhance the learning ability of the model. After passing through each layer of the neural network, the deep Q network learns the complex patterns and features in the state vector and outputs a Q value matrix. Each element in the Q value matrix represents the expected return that can be obtained by taking a certain equity allocation action under a given state, where Indicates the status, Represents the equity allocation actions that can be taken. Based on the Q-value matrix, the ϵ-greedy strategy is used to select the equity allocation action. The ϵ-greedy strategy is a strategy that strikes a balance between exploration and utilization, where ϵ represents the probability of exploration. In each decision step, there is a probability of ϵ to select a random action in order to explore more possible equity allocation methods, and there is a probability of 1−ϵ to select the action with the current maximum Q value to utilize the best strategy that has been learned. In order to better meet the needs of different membership levels, the ϵ value is dynamically adjusted according to the membership level. For example, for higher-level members, the ϵ value is smaller so that more high-return actions can be selected to ensure the refinement and personalization of equity allocation; while for lower-level members, the ϵ value is larger so that more different equity allocation combinations can be explored to dig out the possible optimal strategy. In this way, an initial equity allocation plan is obtained. The feasibility of the initial equity allocation plan is checked to ensure that the allocated equity meets the equity allocation rules and resource constraints preset by the platform. The preset rights and interests distribution rules include the limit on the number of rights and interests that members can obtain, the total resource limit of the platform, etc. The feasibility of the initial plan is verified, and the distribution plan is improved by iteratively adjusting the rights and interests items when necessary. Assuming that the resources of a certain rights and interests item are limited and cannot meet the needs of all members, the initial plan is adjusted, such as reducing the number of certain rights and interests items or replacing them with other alternative rights and interests, to obtain an effective rights and interests distribution plan that meets all constraints. Apply the effective rights and interests distribution plan to members, and collect feedback from members on the use of the allocated rights and interests within a fixed time window. By tracking the actual use of rights and interests by members during this period, the satisfaction and frequency of use of the members for the allocated rights and interests can be obtained. Based on the usage feedback, the preset reward function is used to calculate the instant reward value The reward function is designed based on factors such as the frequency of members using their rights, satisfaction, and further consumption behavior on the platform. It is usually in the form of: ; in, For the The instantaneous reward value of time steps, and are the reward coefficients, which represent the importance of equity usage and consumption behavior, is the utilization rate of equity, It is the amount of consumption generated by members on the platform after the equity distribution. Based on the instant reward value and a preset discount factor , calculate the cumulative reward ; in, is a discount factor, usually between [0,1], which is used to balance the relationship between immediate rewards and future rewards, so that the model not only pays attention to the current equity distribution effect, but also pays attention to the long-term benefits brought by the future. The calculated cumulative reward is used to measure the overall effect of a certain equity distribution scheme on the entire time series. The experience tuple (i.e., the four-tuple of state, action, immediate reward, and next state) is stored in an experience replay pool with a capacity of 10,000. The experience replay pool stores past experience samples and randomly extracts these samples during the training process to update the model parameters, effectively reducing the correlation between samples and improving the generalization ability and training stability of the model. Each time, a small batch of experience data of size 64 is randomly sampled from the experience replay pool, and these data are used to update the parameters of the deep Q network. When updating the parameters, the dual Q learning algorithm is used to avoid the problem of overestimation of Q values. Dual Q learning introduces two independent Q networks: evaluation Q network and target Q network. The evaluation Q network is used to select actions, while the target Q network is used to evaluate the value of actions. During the training process, the parameters of the target Q network are updated using the following formula: ; in, Indicates that the target Q network is in state Next select action The Q value, For instant rewards, For the next state, Indicates the Q value of selecting the optimal action in the next state. Gradually approach the optimal Q value function to improve the decision-making ability of the model. After each update, the updated target Q network parameters are copied to the evaluation Q network in a soft update manner. The soft update is implemented by the following formula: ; in, and are the parameters of the evaluation Q network and the target Q network, respectively. It is a soft update coefficient, usually a small value (such as 0.01) is taken to make the target network gradually approach the parameters of the evaluation network. Ensure that the model gradually stabilizes during the training process and eventually reaches the preset convergence conditions. When the deep Q network converges, it outputs a personalized member benefits combination.

[0036] The member order processing method in the embodiment of the present invention is described above. The member order processing device in the embodiment of the present invention is described below. Figure 2 In one embodiment of the present invention, a member order processing device includes: The collection module 201 is used to collect member information from multiple platforms and construct a member graph structure to extract node features to obtain a multi-dimensional member feature vector; Fusion module 202, used to perform feature fusion on the multi-dimensional member feature vector and the order data to be processed to obtain a member-order fusion feature vector; Classification module 203, used to classify orders based on member-order fusion feature vectors, and obtain order classification results and processing priorities; Planning module 204, used to perform multi-platform collaborative integer programming and solving based on order classification results and processing priorities to obtain a cross-platform order allocation plan; Evaluation module 205, used to evaluate member value based on the cross-platform order allocation plan and order processing results, and obtain dynamic member value scores and level division information; The strategy optimization module 206 is used to input the dynamic member value score and the level division information into the dynamic equity allocation model based on the deep Q network to optimize the equity allocation strategy and obtain a personalized member equity combination.

[0037] Through the collaborative cooperation of the above-mentioned components, by using graph neural network and soft attention mechanism to extract member features, the complex relationship between member attributes can be effectively captured, and the accuracy and comprehensiveness of member feature representation can be improved. The adversarial bidirectional encoder and bidirectional LSTM network are used to fuse member-order features, and the deep integration of member information and order data is realized. The order classification method based on multi-layer perceptron network and support vector regression model can accurately identify order types and reasonably allocate processing priorities, thereby improving the efficiency and accuracy of order processing. The multi-platform collaborative integer programming model is used for order allocation, which fully considers the processing capabilities and expertise of each platform, realizes the optimal configuration of order resources, and improves the overall order processing efficiency. The entropy weight method and hierarchical analysis method are combined to evaluate member value, which not only considers the distribution characteristics of objective data, but also incorporates expert experience, making the member value evaluation results more comprehensive and reliable. The dynamic equity allocation model based on deep Q network is used to realize the personalized allocation and real-time optimization of member rights and interests, and can timely adjust the equity combination according to the dynamic value and preference of members, thereby improving member satisfaction and loyalty. The present invention improves the efficiency and accuracy of member management and order processing, and provides a comprehensive member order processing solution for enterprises.

[0038] above Figure 2 The member order processing device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The member order processing device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0039] Figure 33 is a structural diagram of a member order processing device provided by an embodiment of the present invention. The member order processing device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 may be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the member order processing device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the member order processing device 300 to implement the steps of the above-mentioned member order processing method.

[0040] The member order processing device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The member order processing device structure shown does not constitute a limitation on the member order processing device provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0041] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the member order processing method.

[0042] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0043] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0044] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A member order processing method, characterized in that: The method comprises: Collect member information from multiple platforms and construct a member graph structure to extract node features, and obtain a multi-dimensional member feature vector; Performing feature fusion on the multi-dimensional member feature vector and the order data to be processed to obtain a member-order fusion feature vector; Classify the orders based on the member-order fusion feature vector to obtain order classification results and processing priorities; Based on the order classification results and the processing priorities, multi-platform collaborative integer programming and solving are performed to obtain a cross-platform order allocation solution; Perform member value assessment based on the cross-platform order allocation plan and order processing results to obtain dynamic member value scores and hierarchical division information; The dynamic member value score and the hierarchical division information are input into a dynamic equity allocation model based on a deep Q network to optimize the equity allocation strategy and obtain a personalized member equity combination.

2. The member order processing method according to claim 1, characterized in that: The method collects member information from multiple platforms and constructs a member graph structure to extract node features, thereby obtaining a multi-dimensional member feature vector, including: Perform data cleaning and format unification processing on multi-platform member information to obtain a standardized member data set, and based on the standardized member data set, set member attributes as nodes, set relationships between attributes as edges, and construct a member graph structure; The member graph structure is represented by a sparse matrix, the member graph structure information is encoded by an adjacency matrix and a feature matrix to obtain a graph structure encoding matrix, and the graph structure encoding matrix is ​​input into a self-attention gated graph neural network, and node features are updated by multi-layer graph convolution operations to obtain an updated node feature matrix; Calculating the attention weights between nodes based on the updated node feature matrix, normalizing the attention scores using a softmax function to obtain a node attention weight matrix, and performing matrix multiplication operation on the node attention weight matrix and the updated node feature matrix to obtain a weighted node feature matrix; The weighted node feature matrix is ​​subjected to column vector summation and L2 norm normalization to obtain a graph-level representation vector, the graph-level representation vector is connected to the member information attribute vector, and feature fusion and dimensionality reduction are performed through a fully connected layer to obtain a multi-dimensional member feature vector.

3. The member order processing method according to claim 2, characterized in that: The step of fusing the multi-dimensional member feature vector and the order data to be processed to obtain a member-order fusion feature vector includes: Performing dimension expansion processing on the multi-dimensional member feature vector, expanding the multi-dimensional member feature vector to a length equal to the order data sequence to be processed by duplication and zero-filling operations, to obtain an expanded member feature vector; The extended member feature vector is concatenated with the to-be-processed order data in the feature dimension to obtain an initial fused feature sequence, wherein each time step contains a member feature and a corresponding order feature; Input the initial fusion feature sequence into a bidirectional LSTM network, wherein the forward LSTM and the reverse LSTM each include two layers, each layer includes 128 hidden units, and a tanh activation function is used to obtain a forward LSTM feature sequence and a reverse LSTM feature sequence; The outputs of the forward LSTM feature sequence and the reverse LSTM feature sequence in the last hidden layer are concatenated to obtain a bidirectional LSTM feature sequence, which contains temporal dependency information captured from two directions; Input the bidirectional LSTM feature sequence into a multi-head self-attention layer, set 8 attention heads, each with a dimension of 64, generate query, key and value matrices through linear transformation, calculate the scaled dot product attention and perform softmax normalization to obtain a self-attention feature sequence; The self-attention feature sequence is input into a two-layer feedforward neural network, the first layer uses a ReLU activation function, the second layer uses a linear activation function, the hidden layer dimension is 512, and a nonlinear transformation is performed to obtain an encoder output feature sequence; The feature sequence output by the encoder is input into an adversarial discriminator, wherein the adversarial discriminator comprises three one-dimensional convolutional layers, the convolution kernel sizes are 3, 4, and 5 respectively, the number of output channels of each layer is 100, and finally a fully connected layer is connected to output a binary classification result, and the encoder parameters are optimized through adversarial training to obtain an adversarially optimized feature sequence; The maximum pooling operation is performed on the feature sequence after the adversarial optimization in the time dimension, and the maximum value in each feature dimension is selected to obtain the member-order fusion feature vector.

4. The member order processing method according to claim 3, characterized in that: The order classification is performed based on the member-order fusion feature vector to obtain the order classification result and processing priority, including: Input the member-order fusion feature vector into a multilayer perceptron network, wherein the multilayer perceptron network comprises three hidden layers, the first layer has 512 neurons, the second layer has 256 neurons, the third layer has 128 neurons, a ReLU activation function is used between each layer, and a linear activation function is used in the last layer to obtain order feature representation; Performing L2 regularization processing on the order feature representation, calculating the L2 norm of the feature vector, and dividing the feature vector by the L2 norm to obtain a normalized order feature vector; Input the normalized order feature vector into the softmax classifier, use the softmax function to calculate the probability of each category, obtain the order category probability distribution, and based on the order category probability distribution, use the argmax function to find the index with the highest probability, and use the category corresponding to the index as the order classification result; Performing entropy calculation on the probability distribution of the order category to obtain a scalar entropy value representing classification uncertainty, and comparing the scalar entropy value representing classification uncertainty with a predefined threshold value, if the scalar entropy value representing classification uncertainty is less than the threshold value, classifying the order into a definite category, otherwise classifying the order into an uncertain category; For orders with certain classification, a preset priority mapping table is queried based on the order classification results. The priority mapping table maps each order category to an integer priority of 1-5 to obtain a preliminary processing priority. For orders with uncertain classification, the normalized order feature vector is input into a pre-trained support vector regression model. The pre-trained support vector regression model uses an RBF kernel function to obtain a continuous priority score, and the continuous score is converted into a discrete processing priority of 1-5 by setting interval boundaries.

5. The member order processing method according to claim 4, characterized in that: The multi-platform collaborative integer programming and solving based on the order classification results and the processing priorities to obtain a cross-platform order allocation solution includes: Based on the order classification result and the processing priority, a weight coefficient is assigned to each order. The weight coefficient calculation formula is: weight = α*classification score + β*priority, where α and β are preset balance parameters, and an order weight list is obtained; According to the processing capabilities and expertise of each platform, a platform order suitability matrix is ​​constructed. The matrix element aij represents the suitability of platform i to process order j, and the suitability range is 0-1. Based on the order weight list and the platform order fitness matrix, an objective function is constructed, and three constraints are set: each order can only be assigned to one platform, the order processing volume of each platform does not exceed the maximum processing capacity of the platform, and the cross-platform allocation ratio of preset type orders does not exceed the preset threshold; Combining the objective function and the three constraints to construct a multi-platform collaborative integer programming model, wherein the decision variables of the multi-platform collaborative integer programming model are the weighted scores of each order on each platform, and the goal is to maximize the overall order processing benefit; The multi-platform collaborative integer programming model is input into an interactive optimization solver, solution parameters including a maximum number of iterations, a convergence threshold and a time limit are set, a solution process is started, solution results are obtained from the interactive optimization solver, optimal values ​​of decision variables are extracted, and order-platform pairs with a decision variable of 1 are combined into a cross-platform order allocation plan.

6. The member order processing method according to claim 5, characterized in that: The member value evaluation is performed based on the cross-platform order allocation scheme and the order processing result to obtain a dynamic member value score and hierarchical division information, including: Based on the cross-platform order allocation scheme and the order processing results of the pending order data, member value evaluation indicators are extracted, including consumption frequency, consumption amount, order completion rate, platform activity and customer service demand, and an original indicator matrix is ​​constructed; The original indicator matrix is ​​normalized, and the indicators of different dimensions are converted into dimensionless values ​​by using the range transformation method to obtain a standardized indicator matrix, and based on the standardized indicator matrix, the information entropy of each indicator is calculated, and the objective weight of each indicator is calculated by using the entropy weight method to obtain an objective weight vector; Construct a hierarchical analysis structure model, divide the member value evaluation indicators into three levels: target level, criterion level and indicator level, and establish a judgment matrix; Perform a consistency check on the judgment matrix, calculate the maximum eigenvalue λmax and the consistency index CI=(λmax-n) / (n-1), where n is the matrix order, calculate the consistency ratio CR=CI / RI, where RI is the random consistency index, and if the consistency ratio CR<0.1, the judgment matrix passes the consistency check; otherwise, adjust the value in the judgment matrix until it passes the consistency check; Based on the judgment matrix that has passed the consistency test, the subjective weight of each indicator is calculated by the eigenvalue method to obtain a subjective weight vector, and the objective weight vector and the subjective weight vector are weighted averaged to obtain a comprehensive weight vector; Based on the comprehensive weight vector and the standardized indicator matrix, the value score of each member is calculated, and the member value scores are clustered to obtain hierarchical division information.

7. The member order processing method according to claim 6, characterized in that: The step of inputting the dynamic member value score and the hierarchical division information into a dynamic equity allocation model based on a deep Q network to optimize the equity allocation strategy and obtain a personalized member equity combination includes: The dynamic member value score and the hierarchical division information are combined into a state vector S, which includes the member value score, the member level, and the historical equity utilization rate. The state vector S is input into a deep Q network, which includes three hidden layers, each with 256, 128, and 64 neurons, respectively, and a ReLU activation function is used to obtain a Q value matrix; Based on the Q value matrix, an ε-greedy strategy is used to select the equity distribution action, where the ε value is dynamically adjusted according to the membership level to obtain an initial equity distribution plan; Conduct a feasibility check on the initial equity distribution plan to ensure that it meets the preset equity distribution rules and resource constraints, and obtain an effective equity distribution plan by iteratively adjusting the equity items in the plan; Apply the effective equity distribution plan to members, collect members' feedback on the use of equity within a fixed time window, and calculate the instant reward value based on a preset reward function; Based on the instant reward value and the preset discount factor, the cumulative reward is calculated, and the experience tuple is stored in the experience replay pool with a capacity of 10,000; A small batch of experience data with a size of 64 is randomly sampled from the experience replay pool, the parameters of the target Q network are updated using the dual Q learning algorithm, and the updated target Q network parameters are copied to the evaluation Q network in a soft update manner until the preset convergence conditions are reached, and a personalized member benefit combination is output.

8. A member order processing device, characterized in that: The device is used to execute the member order processing method according to any one of claims 1 to 7, and comprises: The collection module is used to collect member information from multiple platforms and construct a member graph structure to extract node features and obtain a multi-dimensional member feature vector; A fusion module, used to perform feature fusion on the multi-dimensional member feature vector and the order data to be processed to obtain a member-order fusion feature vector; A classification module, used to classify orders based on the member-order fusion feature vector, and obtain order classification results and processing priorities; A planning module, used for performing multi-platform collaborative integer programming and solving based on the order classification results and the processing priorities to obtain a cross-platform order allocation plan; An evaluation module, used to evaluate member value based on the cross-platform order allocation plan and order processing results, and obtain dynamic member value scores and hierarchical division information; The strategy optimization module is used to input the dynamic member value score and the hierarchical division information into a dynamic equity allocation model based on a deep Q network to optimize the equity allocation strategy and obtain a personalized member equity combination.

9. A member order processing device, characterized in that: The member order processing device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the member order processing device to execute the member order processing method according to any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the member order processing method according to any one of claims 1 to 7 is implemented.

Images