A crowd-sourcing logistics overlapping community detection method based on graph convolution network and probability decoding

Abstract

This invention discloses a crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding, applied to the field of overlapping community partitioning in complex networks. The method includes: obtaining a graph adjacency matrix and a worker node attribute matrix based on the relationships between worker nodes; performing feature encoding on worker nodes using an improved graph convolutional network to obtain a worker node-community membership matrix; introducing learnable community weight parameters to define the structural association strength between node pairs; and decoding the structural association strength as the probability of edges existing between node pairs using a Bernoulli-Poisson random graph generation mechanism; constructing a structural reconstruction loss function using a set of positive sample edges and a set of negative sample non-edges sampled from the graph adjacency matrix; and introducing community membership constraints and attribute consistency constraints, which are then weighted and combined to form the overall optimization objective function of the model. This invention achieves accurate overlapping community partitioning of worker nodes, effectively improving the accuracy and practicality of the partitioning.

Description

Technical Field

[0001] This invention relates to the field of overlapping community partitioning in complex networks, and more specifically to a crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding. Background Technology

[0002] With the development of the platform economy, crowdsourced logistics, as a new model that utilizes idle social transportation capacity to complete the delivery of goods, has been widely applied in various fields such as retail, e-commerce, and local services. In actual operation, crowdsourced workers often spontaneously form dynamic and overlapping collaborative groups due to shared order areas, time slot coordination, or task preferences. However, such group structures are usually not explicitly modeled or utilized by existing scheduling systems. How to effectively identify and utilize this overlapping collaborative structure to improve the task allocation efficiency and service robustness of crowdsourced logistics systems has become a pressing technical problem to be solved in this field.

[0003] In recent years, research on crowdsourced logistics has mainly focused on individual-level optimization issues such as task allocation, worker selection, and incentive mechanisms, with less attention paid to exploring the collaborative structure of groups. Among them, Yu Haiyan et al. (Yu Haiyan, Jiang Renlian. Research on real-time delivery order allocation and path optimization based on crowdsourcing platform for food delivery [J]. Industrial Engineering and Management, 2022, 27(02):146-152.) used insertion heuristic strategies such as greedy proximity and minimum path increment for order allocation in the food delivery crowdsourcing scenario. Although it takes into account both delivery cost and timeliness, it treats riders as independent units and plans their paths separately, without considering the implicit collaborative relationships formed by them in shared areas. To address this issue, Jiang et al. (Jiang J, An B, Jiang Y, et al. Group-oriented task allocation for crowdsourcing in social networks[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2019, 51(7): 4417-4432.) proposed a group-oriented task allocation method. This method assesses the ability of the main group to collaborate with its socially connected groups to complete tasks by evaluating the "contextual crowdsourcing value". It also designs semi-supervised and fully supervised algorithms to select the main group and assistant groups, effectively reducing the team formation cost. However, its group structure relies on a predefined social network, uses only binary connections, and does not integrate worker attributes. To address this, Hu Linbo et al. (Hu Linbo, Ni Zhiwei, Cheng Jiale, et al. Collaborative Crowdsourcing Task Allocation Method Based on Community Detection [J]. Computer Applications, 2025, 45(02): 534-545.) constructed a weighted collaboration graph by integrating workers' social attributes, cooperation history, and communication frequency, and generated candidate communities through Louvain community detection to achieve data-driven group discovery; however, it adopted non-overlapping partitioning and forced workers to belong to a single entity, which could not characterize overlapping collaborative roles in multiple scenarios.

[0004] To accurately identify multiple affiliations of delivery personnel, it is necessary to introduce a community detection method that supports node overlap. VanLierde et al. (Van Lierde H, Chow TWS, Chen G. Scalable spectral clustering for overlapping community detection in large-scale networks[J]. IEEE Transactions on Knowledge and Data Engineering, 2019, 32(4): 754-767.) proposed the VRONcut algorithm based on spectral clustering, which achieves efficient overlapping community detection by extending the normalized cut model. Although this method can introduce node weights as priors, the weights still originate from the graph topology and fail to utilize the node's own attributes independent of the network topology, thus limiting the semantic consistency of the discovered communities. Liu et al. (He C, Zheng Y, Cheng J, et al. Semi-supervised overlapping community detection in attributed graph with graph convolutional autoencoder[J]. Information Sciences, 2022, 608: 1464-1479.) proposed a semi-supervised graph convolutional autoencoder (SSGCAE), which achieves effective overlapping community detection on attributed networks by jointly learning node attributes and graph topology and combining modularity optimization. However, the research and validation of this method are limited to general social networks and have not yet been explored for group structure modeling in new collaborative scenarios such as crowdsourced logistics.

[0005] Therefore, how to provide a crowdsourcing logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding that can integrate the attributes and collaborative relationship topology of crowdsourcing worker nodes, effectively identify their multiple affiliations, and is applicable to crowdsourcing logistics scenarios is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding, characterized by comprising: Step 1: Obtain the graph adjacency matrix and worker node attribute matrix based on the relationships between worker nodes, and perform feature encoding on worker nodes based on the improved graph convolutional network to obtain the worker node-community membership matrix used to characterize the degree of worker node affiliation to multiple communities. Step 2: Based on the worker node-community membership matrix, a learnable community weight parameter is introduced to define the structural association strength between node pairs. The structural association strength is then decoded into the probability that there is an edge between node pairs through the Bernoulli-Poisson random graph generation mechanism. Step 3: Based on the probability that there are edges between node pairs, construct a structural reconstruction loss function using the set of positive edge samples and the set of negative non-edge samples obtained from the graph adjacency matrix. Introduce community membership constraints and attribute consistency constraints, and combine them in a weighted manner to form the overall optimization objective function of the model. Finally, based on the worker node-community membership matrix obtained from training, complete the multi-community affiliation division of worker nodes based on a preset threshold and output the overlapping community division results.

[0008] Optionally, in step 1, the method for obtaining the graph adjacency matrix and worker node attribute matrix based on the relationships between worker nodes is as follows: Construct a worker relations network diagram, denoted as ;in, For worker nodes; It is a set of relationships between nodes; Further construct the graph adjacency matrix ,like ,but Otherwise and order To remove self-loops; Obtain the attribute vector of each worker node and form an attribute matrix. ;in, The total number of rider nodes in the crowdsourced logistics network; For attribute dimensions; Represents a node The attribute feature vector.

[0009] Optionally, in step 1, the worker nodes are feature-encoded based on an improved graph convolutional network to obtain a worker node-community membership matrix that characterizes the degree of worker node affiliation to multiple communities, specifically: First, a self-loop is added to the graph adjacency matrix. Then, the degree matrix corresponding to the graph adjacency matrix is ​​calculated, and a symmetric normalized graph propagation operator is constructed as follows:

[0010] in, The graph adjacency matrix after adding self-loops; The graph adjacency matrix; It is the identity matrix, i.e., the self-loop that has been added; The degree matrix is ​​the first Line 1 The value in the column represents the number of columns. The degree of each node, while the off-diagonal values ​​in the degree matrix are all 0; The total number of rider nodes in the crowdsourced logistics network; for The Middle Line 1 The column values ​​represent the nodes. and nodes The true connection relationship; It is a symmetric normalized adjacency matrix; It is a degree matrix; Then the first The output feature matrix of the layer is calculated as follows:

[0011] in, For the first The layer's output feature matrix; initial input This is the original worker node attribute matrix; Transform the weight matrix for the node's own features; The neighborhood aggregation transformation weight matrix; It is the bias vector; In the intermediate layer, the output is further mapped using the nonlinear activation function tanh; after multi-layer graph convolution propagation, the original score matrix is ​​finally output. ;in, The preset number of communities; Introducing temperature parameters The score matrix is ​​then normalized using the Sigmoid function, mapping the original scores output by the graph convolutional network to probabilistic community membership degrees. The mapping process is as follows:

[0012] in, For the Sigmoid function; This is the final worker node-community membership matrix; Indicates the first The worker belongs to the first The probability of each community.

[0013] Optionally, in step 2, based on the worker node-community membership matrix, a learnable community weight parameter is introduced to define the structural association strength between node pairs, specifically: Introducing a learnable community weight vector ; where each weight component From the corresponding original learnable parameters The result is obtained through positive value mapping, as follows:

[0014] in, For the first The weight parameter of the community, i.e., the _th_ community The intensity of each community's contribution to edge generation; For any pair of nodes Define the structural association strength as follows:

[0015] in, For node pairs Structural correlation strength; For smoothing terms; The preset number of communities; , worker nodes Worker nodes For the first The degree of affiliation of each community.

[0016] Optionally, in step 2, the structural association strength is decoded into the probability that there is an edge between node pairs using the Bernoulli-Poisson random graph generation mechanism, as follows:

[0017] in, For node pairs There is a probability that there are connected edges between them; For node pairs The structural correlation strength.

[0018] Optionally, in step 3, the method for obtaining the set of positive sample edges and the set of negative sample non-edges from the graph adjacency matrix is ​​as follows: Training samples are constructed using a random sampling method, which involves uniformly sampling with replacement from the set of real edges in the graph adjacency matrix. The edges are used as positive samples, and candidate node pairs are randomly generated while real edges and self-loops are filtered out, thus obtaining... Non-connected edges are used as negative samples.

[0019] Optionally, in step 3, based on the probability that there is an edge between node pairs, a structural reconstruction loss function is constructed using the set of positive sample edges and the set of negative sample non-edges obtained from the graph adjacency matrix, as follows:

[0020] in, It is a set of relationships between nodes; This represents the number of worker node pairs connected by edges in the adjacency matrix; This represents the number of non-connected worker node pairs in the adjacency matrix. For node pairs There is a probability that there are connected edges between them; For structural reconstruction loss; , These are the loss terms for connected node pairs and the loss terms for non-connected node pairs, respectively.

[0021] Optionally, in step 3, the community membership constraint term is constructed as follows: Calculate the average row sum of the membership degrees of all nodes and compare it with the preset target value. Alignment, construct the following constraint terms:

[0022] in, As a constraint on community participation, it serves as a constraint on community affiliation. The total number of rider nodes in the crowdsourced logistics network; The preset number of communities; To represent nodes Belongs to the community The probability of membership; The target is the average community engagement parameter.

[0023] Optionally, in step 3, the method for constructing the attribute consistency constraint is as follows: Based on the membership matrix Calculate the first The attribute center of each community Then, the weighted distance between the node attribute and the corresponding community center is calculated to form the attribute consistency loss, as follows:

[0024] in, The attribute consistency loss is used as an attribute consistency constraint term. The total number of rider nodes in the crowdsourced logistics network; To represent nodes Belongs to the community The probability of membership; Represents a node Attribute vectors; It is a smoothing constant; The preset number of communities.

[0025] Optionally, in step 3, based on the structural reconstruction loss function, and by introducing community membership constraints and attribute consistency constraints, a weighted combination is formed to create the overall optimization objective function of the model, which also includes: Introducing a community deduplication constraint, the final overall optimization objective function of the model is as follows:

[0026] in, The ultimate goal is to optimize the overall model. , , , These are the loss weighting coefficients; For structural reconstruction loss; As a constraint on community participation, it serves as a constraint on community affiliation. The attribute consistency loss is used as an attribute consistency constraint term. Remove duplicate constraints from the community.

[0027] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding, which achieves the following beneficial effects: This invention employs a dual-weight structure with separated self-feature transformation parameters and neighborhood aggregation parameters in the graph convolutional encoding stage to model the features of worker nodes. It then constructs a Bernoulli-Poisson probabilistic decoding model based on the node-community membership matrix to achieve a probabilistic mapping between structural association strength and edge probability. Furthermore, it introduces a balanced structural reconstruction loss with a positive-negative sample ratio adjustment mechanism to correct for the imbalance between the number of connected and non-connected edges in sparse graphs. Simultaneously, it combines attribute consistency constraints based on community attribute centers and a constraint mechanism that adjusts the similarity between the row sum of the membership matrix and the column similarity of the community, creating a synergistic constraint between structural reconstruction, attribute aggregation, and community differentiation. Through this multi-mechanism joint optimization, this invention can alleviate degradation phenomena such as community collapse, duplicate column generation, and excessive overlap during overlapping community detection, maintaining a stable balance between structural tightness and attribute consistency in the obtained community partitioning results. This invention provides technical support for modeling potential collaborative structures in crowdsourced logistics systems, helping to utilize group structure information in subsequent task scheduling or resource allocation processes to improve the system's adaptability and robustness in complex dynamic environments. In the future, it can be further expanded to real-time collaborative relationship analysis scenarios by combining time-series behavioral data or dynamic network evolution information. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of the method flow provided by the present invention.

[0030] Figure 2 This is a schematic diagram of the crowdsourced logistics overlapping community detection model provided by the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1: Embodiment 1 of this invention discloses a crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding, such as... Figure 1 As shown, it includes: Step 1: Obtain the graph adjacency matrix and worker node attribute matrix based on the relationships between worker nodes, and perform feature encoding on worker nodes based on the improved graph convolutional network to obtain the worker node-community membership matrix used to characterize the degree of worker node affiliation to multiple communities.

[0033] The method for obtaining the graph adjacency matrix and worker node attribute matrix based on the relationships between worker nodes is as follows: Construct a worker relations network diagram, denoted as ;in, For worker nodes; It is a set of relationships between nodes; Further construct the graph adjacency matrix ,like ,but Otherwise and order To remove self-loops; Obtain the attribute vector of each worker node and form an attribute matrix. ;in, The total number of rider nodes in the crowdsourced logistics network; For attribute dimensions; Represents a node The attribute feature vector.

[0034] By performing feature encoding on worker nodes using an improved graph convolutional network, a worker node-community membership matrix is ​​obtained to characterize the degree to which worker nodes belong to multiple communities. Specifically: To enhance the ability of nodes to express their own features, a self-loop is first added to the graph adjacency matrix. Then, the degree matrix corresponding to the graph adjacency matrix is ​​calculated, and a symmetric normalized graph propagation operator is further constructed to eliminate the influence of node degree differences on feature propagation, as follows:

[0035] in, The graph adjacency matrix after adding self-loops; The graph adjacency matrix; It is the identity matrix, i.e., the self-loop that has been added; The degree matrix is ​​the first Line 1 The value in the column represents the number of columns. The degree of each node, while the off-diagonal values ​​in the degree matrix are all 0; The total number of rider nodes in the crowdsourced logistics network; for The Middle Line 1 The column values ​​represent the nodes. and nodes The true connection relationship; It is a symmetric normalized adjacency matrix; It is a degree matrix; In this embodiment of the invention, the feature dimensions of each graph convolutional layer can be set to 64, 128, and 256 respectively, and the transformation of the node's own features and the aggregation of neighborhood information are separated and modeled in the graph convolutional layer. Then the... The output feature matrix of the layer is calculated as follows:

[0036] in, For the first The layer's output feature matrix; initial input This is the original worker node attribute matrix; Transform the weight matrix for the node's own features; The neighborhood aggregation transformation weight matrix; It is the bias vector; In the intermediate layer, the output is further mapped using the nonlinear activation function tanh; after multi-layer graph convolution propagation, the original score matrix is ​​finally output. ;in, The preset number of communities; Introducing temperature parameters The score matrix is ​​then normalized using the Sigmoid function, mapping the original scores output by the graph convolutional network to probabilistic community membership degrees. The mapping process is as follows:

[0037] in, For the Sigmoid function; This is the final worker node-community membership matrix; Indicates the first The worker belongs to the first The probability of each community.

[0038] Step 2: Based on the worker node-community membership matrix, a learnable community weight parameter is introduced to weight and superimpose the membership probability of a node in each community, define the structural association strength between node pairs, and decode the structural association strength into the probability that there is an edge between node pairs through the Bernoulli-Poisson (BerPo) random graph generation mechanism, thereby forming a graph structure generation distribution model.

[0039] Based on the worker node-community membership matrix, a learnable community weight parameter is introduced to define the structural association strength between node pairs, specifically: Introducing a learnable community weight vector ; where each weight component From the corresponding original learnable parameters The result is obtained through positive value mapping, as follows:

[0040] in, For the first The weight parameter of the community, i.e., the _th_ community The intensity of each community's contribution to edge generation; For any pair of nodes Define the structural association strength as follows:

[0041] in, For node pairs Structural correlation strength; For smoothing terms; The preset number of communities; , worker nodes Worker nodes For the first The degree of affiliation of each community.

[0042] Using the Bernoulli-Poisson random graph generation mechanism, the structural association strength is decoded as the probability that there is an edge between node pairs, as follows:

[0043] in, For node pairs There is a probability that there are connected edges between them; For node pairs The structural correlation strength.

[0044] Step 3: Based on the probability that there are edges between node pairs, construct a structural reconstruction loss function using the positive sample edge set and negative sample non-edge set obtained from the graph adjacency matrix. Introduce community membership constraint term and attribute consistency constraint term to enable the model to capture network structure features and node attribute features simultaneously. The weighted combination forms the overall optimization objective function of the model. Finally, based on the worker node-community membership matrix obtained from training, complete the multi-community classification of worker nodes and output the overlapping community classification results based on the preset threshold (when the membership exceeds the threshold, the corresponding node is determined to belong to the community).

[0045] The method for obtaining the set of positive edge samples and the set of negative non-edge samples from the graph adjacency matrix is ​​as follows: Training samples are constructed using a random sampling method, which involves uniformly sampling with replacement from the set of real edges in the graph adjacency matrix. The edges are used as positive samples, and candidate node pairs are randomly generated while real edges and self-loops are filtered out, thus obtaining... Non-connected edges are used as negative samples.

[0046] Based on the probability that there is an edge between node pairs, a structural reconstruction loss function is constructed using the set of positive sample edges and the set of negative sample non-edges obtained from the graph adjacency matrix. This loss is composed of the sum of the mean of the negative log-likelihood of the positive sample edges and the mean of the negative log-likelihood of the negative sample non-edges, as follows:

[0047] in, It is a set of relationships between nodes; This represents the number of worker node pairs connected by edges in the adjacency matrix; This represents the number of non-connected worker node pairs in the adjacency matrix. For node pairs There is a probability that there are connected edges between them; For structural reconstruction loss; , These are the loss terms for connected node pairs and the loss terms for non-connected node pairs, respectively.

[0048] To suppress excessive dispersion or concentration of node membership, a community participation constraint (represented as the community membership constraint term in this invention) is constructed to control the average overlap of nodes. The community membership constraint term is constructed as follows: Calculate the average row sum of the membership degrees of all nodes and compare it with the preset target value. Alignment, construct the following constraint terms:

[0049] in, As a constraint on community participation, it serves as a constraint on community affiliation. The total number of rider nodes in the crowdsourced logistics network; The preset number of communities; To represent nodes Belongs to the community The probability of membership; The target is the average community engagement parameter.

[0050] To enhance the consistency of node attributes within the same community, based on the node attribute matrix Construct attribute consistency constraints. The specific method for constructing attribute consistency constraints is as follows: Based on the membership matrix Calculate the first The attribute center of each community Then, the weighted distance between the node attribute and the corresponding community center is calculated to form the attribute consistency loss, as follows:

[0051] in, The attribute consistency loss is used as an attribute consistency constraint term. The total number of rider nodes in the crowdsourced logistics network; To represent nodes Belongs to the community The probability of membership; Represents a node Attribute vectors; It is a smoothing constant; The preset number of communities.

[0052] To avoid high similarity or duplication among column vectors of membership degrees in different communities, which could lead to community representation collapse, a structural reconstruction loss function is used. Community membership constraints and attribute consistency constraints are introduced and weighted together to form the overall optimization objective function of the model. This also includes: Introducing a community deduplication constraint, the final overall optimization objective function of the model is as follows:

[0053] in, The ultimate goal is to optimize the overall model. , , , These are the loss weighting coefficients; For structural reconstruction loss; As a constraint on community participation, it serves as a constraint on community affiliation. The attribute consistency loss is used as an attribute consistency constraint term. The community deduplication constraint is removed. Finally, the parameters of the graph convolutional network encoder and probabilistic decoder are jointly optimized through backpropagation.

[0054] like Figure 2 The diagram shown is a structural diagram of the crowdsourced logistics overlapping community detection model of the present invention.

[0055] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0056] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting a crowd-sourced logistics overlapping community based on a graph convolution network and a probabilistic decoding, characterized in that, include: Step 1: Obtain the graph adjacency matrix and worker node attribute matrix based on the relationships between worker nodes, and perform feature encoding on worker nodes based on the improved graph convolutional network to obtain the worker node-community membership matrix used to characterize the degree of worker node affiliation to multiple communities. Step 2: Based on the worker node-community membership matrix, a learnable community weight parameter is introduced to define the structural association strength between node pairs. The structural association strength is then decoded into the probability that there is an edge between node pairs through the Bernoulli-Poisson random graph generation mechanism. Step 3: Based on the probability that there are edges between node pairs, construct a structural reconstruction loss function using the positive sample edge set and negative sample non-edge set obtained from the graph adjacency matrix. Introduce community membership constraint term and attribute consistency constraint term, and combine them in a weighted manner to form the overall optimization objective function of the model. Finally, based on the worker node-community membership matrix obtained from training, complete the multi-community affiliation division of worker nodes based on a preset threshold and output the overlapping community division results.

2. The method of claim 1, wherein, In step 1, the method for obtaining the graph adjacency matrix and worker node attribute matrix based on the relationships between worker nodes is as follows: A worker relationship network graph is constructed, denoted as ; wherein, is a worker node set; is a node relationship set; Further construct the graph adjacency matrix , if , then , otherwise let , and let to remove self-loops; Obtain the attribute vector of each worker node and compose the attribute matrix ; wherein, is the total number of rider nodes in the crowd-sourced logistics network; is the attribute dimension; represents the attribute feature vector of the node .

3. The method of claim 1, wherein, In step 1, the worker nodes are feature-encoded using an improved graph convolutional network to obtain a worker node-community membership matrix that characterizes the degree to which worker nodes belong to multiple communities. Specifically: First, a self-loop is added to the graph adjacency matrix. Then, the degree matrix corresponding to the graph adjacency matrix is ​​calculated, and a symmetric normalized graph propagation operator is further constructed as follows: wherein, is the graph adjacency matrix after self-loop is added; is the graph adjacency matrix; is the identity matrix, i.e. the added self-loop; is the value in the degree matrix at the row and the column, representing the degree of the th node, while the non-diagonal values in the degree matrix are all 0; is the total number of rider nodes in the crowd-sourced logistics network; is the value in the row and the column, representing the real connection relationship between the node and the node ; is the symmetric normalized adjacency matrix; is the degree matrix; Then the The output feature matrix of the layer is calculated as follows: wherein, is the output feature matrix of the layer; the initial input is the output feature matrix of the layer; the initial input is the original worker node attribute matrix; is the node self-feature transformation weight matrix; is the neighborhood aggregation transformation weight matrix; is the bias vector; In the intermediate layer, the output is further mapped by a nonlinear activation function tanh; after propagation through multiple layers of graph convolution, the final output is the original score matrix ; wherein, is a preset number of communities; Introducing temperature parameters The score matrix is ​​then normalized using the Sigmoid function, mapping the original scores output by the graph convolutional network to probabilistic community membership degrees. The mapping process is as follows: in, For the Sigmoid function; This is the final worker node-community membership matrix; Indicates the first The worker belongs to the first The probability of each community.

4. The crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding according to claim 1, characterized in that, In step 2, based on the worker node-community membership matrix, a learnable community weight parameter is introduced to define the structural association strength between node pairs, specifically: Introducing a learnable community weight vector ; where each weight component From the corresponding original learnable parameters The result is obtained through positive value mapping, as follows: in, For the first The weight parameter of the community, i.e., the _th_ community The intensity of each community's contribution to edge generation; For any pair of nodes Define the structural association strength as follows: in, For node pairs Structural correlation strength; For smoothing terms; The preset number of communities; , worker nodes Worker nodes For the The degree of affiliation of each community.

5. The crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding according to claim 1, characterized in that, In step 2, the structural association strength is decoded into the probability that there is an edge between node pairs using the Bernoulli-Poisson random graph generation mechanism, as follows: in, For node pairs There is a probability that there are connected edges between them; For node pairs The structural correlation strength.

6. The crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding according to claim 1, characterized in that, In step 3, the method for obtaining the set of positive sample edges and the set of negative sample non-edges from the graph adjacency matrix is ​​as follows: Training samples are constructed using a random sampling method, by uniformly sampling with replacement from the set of real edges in the graph adjacency matrix. The edges are used as positive samples, and candidate node pairs are randomly generated while real edges and self-loops are filtered out, thus obtaining... Non-connected edges are used as negative samples.

7. The crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding according to claim 1, characterized in that, In step 3, based on the probability that there is an edge between node pairs, a structural reconstruction loss function is constructed using the set of positive sample edges and the set of negative sample non-edges obtained from the graph adjacency matrix, as follows: in, It is a set of relationships between nodes; This represents the number of worker node pairs connected by edges in the adjacency matrix; This represents the number of non-connected worker node pairs in the adjacency matrix. For node pairs There is a probability that there are connected edges between them; For structural reconstruction loss; , These are the loss terms for connected node pairs and the loss terms for non-connected node pairs, respectively.

8. The crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding according to claim 1, characterized in that, In step 3, the community membership constraint term is constructed as follows: Calculate the average row sum of the membership degrees of all nodes and compare it with the preset target value. Alignment, construct the following constraint terms: in, As a constraint on community participation, it serves as a constraint on community affiliation. The total number of rider nodes in the crowdsourced logistics network; The preset number of communities; To represent nodes Belongs to the community The probability of membership; The target is the average community engagement parameter.

9. The crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding according to claim 1, characterized in that, In step 3, the construction method of the attribute consistency constraint item is as follows: Based on the membership matrix Calculate the first The attribute center of each community Then, the weighted distance between the node attribute and the corresponding community center is calculated to form the attribute consistency loss, as follows: in, The attribute consistency loss is used as an attribute consistency constraint term. The total number of rider nodes in the crowdsourced logistics network; To represent nodes Belongs to the community The probability of membership; Represents a node Attribute vectors; It is a smoothing constant; The preset number of communities.

10. The crowdsourced logistics overlapping community detection method based on graph convolutional networks and probabilistic decoding according to claim 1, characterized in that, In step 3, based on the structural reconstruction loss function, and introducing community membership constraints and attribute consistency constraints, a weighted combination is used to form the overall optimization objective function of the model, which also includes: Introducing a community deduplication constraint, the final overall optimization objective function of the model is as follows: in, The ultimate goal is to optimize the overall model. , , , These are the loss weighting coefficients; For structural reconstruction loss; As a constraint on community participation, it serves as a constraint on community affiliation. The attribute consistency loss is used as an attribute consistency constraint term. Remove duplicate constraints from the community.

Images