A method for constructing a collaborative device group in a social Internet of Things

Abstract

The present invention relates to a method for constructing a collaborative device group in a social Internet of Things, and belongs to the field of the Internet of Things. The method is to first construct a heterogeneous social graph, a task set, social relationships between SIoT objects, and the relationship between each SIoT object and the task; at the same time, according to different actual needs, different constraints are applied to construct a device group selection method based on reducing communication losses or increasing robustness. The present invention solves the problem of querying the best smart device group when a service request is received in a SIoT environment. For different scenarios, it considers reducing communication losses or increasing robustness and customizes different strategies, that is, by considering the interaction between different devices in the social Internet of Things to find the best device group for certain tasks.

Description

Technical Field

[0001] The present invention belongs to the field of Internet of Things, and relates to a method for constructing a collaborative device group in a social Internet of Things. Background Art

[0002] As IoT technology matures and becomes more popular, it is widely believed that IoT represents the next paradigm shift. The future IoT will consist of a huge number of objects that can be controlled and have the ability to provide valuable information. In addition, since these objects will be able to interact with each other, they can collaborate with other related objects to provide services to end users, such as those related to environmental monitoring, surveillance, smart homes, healthcare, and product management. At the same time, by incorporating the concept of social networks into the IoT, some researchers have proposed the concept of social Internet of Things (SIoT) to support new applications and network services of the IoT in a more efficient way. Compared with traditional sensor data monitoring systems, SIoT can enhance the scalability and service discovery capabilities of the network. In addition, social relationships help improve resource discovery and can solve specific tasks. However, current research usually focuses on the architecture and protocol design of SIoT under specific scenarios. However, how to effectively use the collaborative capabilities of SIoT to complete complex tasks has remained largely unexplored.

[0003] Community search in social networks is an active research area. For example, Wen et al. enumerated k-vertex connected components by partitioning the input graph into subgraphs, which significantly improved the efficiency, while Conte et al. used an effective strategy to identify k-plexes with small diameters to improve the efficiency. To extract quasi-completely connected subgraphs from networks, SaneiMehri et al. used the kernel (i.e., extremely dense subgraphs) to efficiently enumerate the top-k quasi-completely connected subgraphs. Chang studied the maximum clique problem in sparse networks and designed a branch-and-result bound algorithm to solve this problem efficiently. In addition, Yang et al. adopted a new UCF index that can extract the kernel of an uncertain graph in linear time. Different from discovering communities in social networks, a recent series of studies explored the idea of ​​discovering anti-communities in social networks, that is, extracting sparse subgraphs in social networks, which can be used in the formation of mental health treatment groups.

[0004] There are also researchers thinking about how to extract dense subgraphs in multi-layer social networks and heterogeneous social networks. For example, Zhang et al. explored the problem of dense subgraph extraction by enumerating spatial clusters in two-dimensional space. The authors used geometric properties to improve the efficiency of spatial cluster enumeration. In addition, considering the social network, the location of each user, and a set of intersections, Ghosh et al. proposed Top-k FSSGQ to identify the top k groups and corresponding intersections, so that each group satisfies mutual familiarity and group size restrictions. On the other hand, SDSQ considers real-time multi-stream scenarios in social networks, which combines the social closeness of users, the diversity of multi-stream channels, and the preferences between users and channels.

[0005] The idea of ​​crowdsourcing is particularly helpful for tasks that contain subtasks that may be easier for humans to solve. For example, SLADE decomposes crowdsourcing tasks to achieve the minimum cost, and MAPS studies the pricing strategy of spatial crowdsourcing. In addition, Zheng et al. conducted an in-depth study to compare a large number of algorithms to solve reasoning problems, while Xu et al. developed a blockchain-based method to protect privacy in crowdsourcing environments.

[0006] On the other hand, the formation of expert teams has also attracted extensive research interest. To form an expert team is to find a group of experts with the required skills, and the communication cost between the selected experts needs to be minimized so that the team members can communicate efficiently. For this reason, researchers have considered several communication costs under different considerations. For example, Lapas et al. found a team that covers all the required skills and minimizes the social diameter of the team or the total edge weight of the spanning tree within the team. In addition, Kargar et al. suggested selecting a leader for each skill and minimizing the social distance of the skill members from each skill leader. In the paper, the authors further considered the spatial proximity and skill requirements for finding a fast response team.

[0007] Machine learning can be used to solve combinatorial optimization problems that have been mainly solved by algorithmic methods in the past few decades, such as the traveling salesman problem, the maximum independent set problem, the minimum vertex cover, and mixed integer programming. He et al. applied machine learning methods to learn the node search order problem of branch and bound methods. Several other works solve combinatorial optimization problems as sequence transformation problems, that is, generating answers sequentially from given problem inputs.

[0008] Reinforcement Learning (RL) has the potential to effectively solve graph optimization problems, such as the satisfiability problem, vehicle routing problem, traveling salesman problem, and maximum independent set. There are two reasons for this: on the one hand, RL can learn to explore potential combinations and use experience at the same time, effectively avoiding falling into local optimality. On the other hand, RL uses goals as rewards without the need to derive optimal solutions as supervision. For example, Cappart et al. improved decision graph-based optimization by using RL to identify variable ordering, which can be used to solve the MIS problem and the maximum cut problem. In addition, to solve problems that require supervision of pointer networks, Bello et al. used RL to optimize the policy modeled by the pointer network to solve the TSP and knapsack problems.

[0009] To complete a complex task in the SIoT environment, the basic solution is to specify all the required functions of the complex task and execute these functions on the corresponding SIoT objects. However, due to the huge number of SIoT objects, executing these functions on all compatible SIoT objects will cause great redundancy and inefficiency. At the same time, users also pay usage costs in the form of rental fees or required data costs based on usage, so it is necessary to select an appropriate number of devices to perform tasks. In addition, since the SIoT network needs to ensure reliability, the selected smart objects need to be tightly coupled.

[0010] Therefore, there is an urgent need for a method that can solve the problem of querying the best group of smart devices when receiving a service request in a SIoT environment. Summary of the invention

[0011] In view of this, the purpose of the present invention is to provide a method for constructing a collaborative device group in a social Internet of Things that reduces communication loss or increases robustness, and finds the best device group for certain tasks by considering the interaction between different devices in the social Internet of Things, thereby solving the problem of querying the best smart device group when a service request is received in a SIoT environment.

[0012] In order to achieve the above object, the present invention provides the following technical solutions:

[0013] A method for constructing a collaborative device group in a social Internet of Things is firstly to construct a heterogeneous social graph, a task set, social relations between SIoT objects and the relationship between each SIoT object and the task; at the same time, according to different actual needs, different constraints are applied to construct a device group selection method based on reducing communication loss or increasing robustness.

[0014] Furthermore, the method specifically includes: given a heterogeneous graph G = (T, S, E, R), where T is a task set, that is, a set of tasks that can be implemented by SIoT devices; S represents a set of SIoT devices; E represents a social relationship between devices, (u, v) ∈ E represents that device u and device v can communicate; R is a task precision edge set, where each precision edge r = [t, v] connects a task vertex t ∈ T and a device vertex v ∈ S, that is, the precision of v executing a certain task t is used as the edge weight;

[0015] Given a heterogeneous graph G and a task query group The goal of the problem to be solved is to find a target device group with exactly p number of devices. to optimize the accuracy of the selected task in Y; constraint p represents how many devices are planned to be controlled or carried according to the application scenario; the sum of the precision edge weights of each vertex in Y is used to measure the quality of the solution of the found device group, let I F (t) represents the sum of the task accuracy edge weights of the target device group F for task t∈Y, that is, I F (t)=∑ v∈F ω[t,v]; the sum of the event weights of all tasks t in the task query group Y to the target device group F is used to represent the aggregation quality corresponding to the task query group Y to the target device group F, that is, the objective function is defined as Ω(F) = ∑ t∈Y I F (t); the optimization goal is to maximize the objective function Ω(F); in addition, the problem contains an accuracy constraint τ, which is used to ensure the performance of the target device group in the worst case; at the same time, different constraints are applied to Y according to different actual needs to reduce communication loss or increase the robustness of the device group selected in F.

[0016] Furthermore, a device group selection method based on reducing communication loss is constructed, which specifically includes: in addition to optimizing the accuracy of the selected task in Y, the communication loss between different devices is also considered, that is, an upper limit constraint is set for the jump distance between devices; this constraint requires that the number of jumps between each pair of vertices in F is at most h, that is, Aiming at the problem that the device group with the maximum task accuracy does not always meet the hop count constraint due to the interaction between two different edge sets E and R, a selection method based on communication loss is adopted. The specific steps are as follows:

[0017] First, a preprocessing step is performed to ensure that the accuracy edge weight of all tasks for each device in S is at least τ; for a device v in the target device set F, any device u∈F must satisfy That is, for each device v∈S, a candidate device set S is constructed v, which only contains vertices within h hops; for device u ∈ S, α(u) is expressed as the sum of the task precision edge weights linked from u to Y, that is, α(u) = ∑ s∈Y w[u, s]; then select p vertices with the largest α(u) from the candidate set of v to construct the candidate solution of v, repeat the above steps to construct different candidate solutions, and return the solution with the largest Ω(F) as the target device group F; since all vertices in S need to be scanned, it will cause a large computational overhead. Considering that if vertices are searched in a certain predefined order, some vertices may not need to be verified because it is doomed that no better solution can be combined from their candidate vertices. Therefore, a vertex access sorting and search strategy can be formulated to avoid unnecessary searches.

[0018] Further, the vertex access sorting and search strategy specifically includes: accessing each device, that is, vertex v ∈ S, in descending order of α(u), which can better evaluate the solution quality in each candidate group to avoid redundant checks; specifically, associate a list L with each vertex v ∈ S v , which is used to store the top-p vertices with the largest α(·) in S v ; each time vertex v is checked and S v is constructed in descending order of α(v), at the same time, insert v into the list L of each vertex u u where |L u | < p; then adopt a pruning strategy. Before constructing the vertex set within h hops for vertex v ∈ S, first check L v to determine whether a better solution than the current solution will be generated. If not, directly skip it.

[0019] Further, construct a device group selection method based on increasing robustness, which specifically includes: in addition to optimizing the accuracy of the selected tasks in Y, it is also required that each device in F has at least k adjacent devices to successfully transmit messages; that is, each device, that is, vertex, must also have at least k adjacent vertices in F; according to the edge set E in the subgraph , represent the in-degree of vertex v as which is the number of vertices u ∈ H such that (, v) ∈ E;

[0020] Due to the interaction between accuracy and communication robustness, devices with high accuracy may not have robust communication capabilities, while devices with robust communication capabilities may not always have optimal accuracy. To achieve a better balance between the quality and efficiency of the solution, adopt the method of constructing a partial solution S sub to gradually construct the complete solution, which specifically includes the following steps:

[0021] First, the filtering strategy is applied to remove the devices u∈S that do not meet the accuracy constraints and node degree constraints from G; secondly, the candidate nodes are sorted according to the accuracy. When adding a new node to the candidate solution C, its robustness is first calculated to see if it meets the conditions. If it does not meet the conditions, the node will not be included in the partial solution S. sub ; Then prune according to accuracy and robustness respectively. Assuming that there are m nodes in the current partial solution, the accuracy-based pruning strategy is that if the sum of the α(v) of the m nodes and the sum of the maximum α(u) of the (pm) candidate nodes is still less than the sum of the α(v) of the current candidate complete solution, that is, ∑ v∈S α(v)+(pm)·max u∈C α(u)≤Ω(S * ), this part of the solution can be deleted, where Ω(S * ) represents the sum of α(v) of the current candidate complete solution; the robustness-based pruning strategy is to delete the partial solution if the sum of the minimum node degrees of the remaining (pm) candidate nodes and m nodes is still less than k, or the sum of the node degrees of the remaining (pm) candidate nodes is less than k(pm); however, choosing a suitable strategy requires domain knowledge and a time-consuming trial-and-error process. Considering the advantages of reinforcement learning in solving graph optimization problems, it is proposed to use reinforcement learning to explore the optimization strategy.

[0022] Furthermore, the optimization strategy using reinforcement learning is specifically as follows: given a heterogeneous graph G = (T, S, E, R), a task query group Y∈T, a degree constraint k, a size constraint p and an accuracy constraint τ, the goal is to extract a target device group according to a learning function H(ξ) using reinforcement learning where ξ is the set of hyperparameters such that:

[0023] 1) |F| = p;

[0024] 2)

[0025] 3)

[0026] 4) Maximize the target value Ω(F);

[0027] When constructing partial solutions, reinforcement learning needs to evaluate candidate devices based on their discriminant features, i.e., their states. Feature information should include global and local information about the graph structure, and should also consider how the partial solutions satisfy the constraints to form usable solutions. Since the quality of the solution is highly correlated with the node degree, it is necessary to extract features based on the average degree:

[0028] 1) The average degree of partial solutions;

[0029] 2) Partial solution S sub The normalized average degree of is the average degree normalized by the total degree;

[0030] 3) The average degree divided by k indicates the density of the partial solutions;

[0031] When the density is not high enough, the basis for selecting nodes may be to increase the density rather than to increase the target value Ω(F). Therefore, the following three features are further extracted to compare the partial solution with the entire graph:

[0032] 4) Density ratio, the ratio of the density of the partial solution to the density of the entire graph, indicating the tendency of forming a dense subgraph in the current input graph;

[0033] 5) Minimum degree ratio, that is, the ratio of the minimum degree in the partial solution to the minimum degree of the entire graph, which is used to measure the relative minimum degree;

[0034] 6) Minimum density ratio, the ratio of the density of the partial solution to the minimum density of the partial solution. All feature information is spliced ​​into a multidimensional feature vector F sub , and linearly transformed by the projection matrix W1, that is, V sub =W1F sub . V sub It contains the subgraph information of partial solutions and the constraint information of the problem;

[0035] After extracting the subgraph features of the partial solution, the graph convolutional network (GCN) is used to effectively integrate the node features and subgraph features of the graph structure; first, the enhanced features are input into the node evaluation function Q with the learning parameter Θ to evaluate the addition of node v to the current partial solution S sub The performance in Q(S sub ,v;Θ); then aggregate node features layer by layer; let α(v) and x v Respectively represent the sum of the task accuracy edge weights of v and whether v is in a partial solution; by aggregating α(v), x v and the neighboring node features from the previous layer Get the node features of v at the (l+1) layer, and use Indicates that

[0036]

[0037] Where W2∈R d ,3∈R d ,4∈R d×d is a learnable parameter, ReLU is the activation function; It can be initialized by the existing graph embedding method; after aggregating the features through L layers, the embedding of each node v can be obtained, that is,

[0038] The evaluation function Q evaluates the features of vertex v by the enhanced features of v Adjacent nodes pooled sum And the feature embedding V of the subgraph sub constitute; that is

[0039]

[0040] Among them, W6∈R d ,7∈R d ,5∈R d×d is a learnable parameter; if the target value of vertex v is high, it should be added to the partial solution first; if vertex v has been added to the partial solution, its neighbor node N(v) should be ranked higher than other vertices;

[0041] Compared with the number of training sets required, real-world data is relatively sparse and cannot meet the needs. Graph generators can be used to synthesize examples for learning, but the characteristics of real graphs are different from those of graph generators. Therefore, the idea of ​​meta-learning is proposed to explore models that can quickly adapt to new environments.

[0042] Furthermore, the idea of ​​using meta-learning to explore the model architecture that can quickly adapt to new environments is divided into an encoder-decoder architecture, where: encoder Θ e Contains {W1,W2,W3,W4,W5}, the decoder parameters Θ d ={W6,W7}; The goal of meta-learning is to make the encoder parameters Θ e It can be transferred to the new graph distribution; during training, the encoder parameters and decoder parameters are optimized alternately according to the gradient descent function. During the test process, the encoder parameters Θ are fixed e And optimize the decoder parameters Θ d .

[0043] The beneficial effect of the present invention is that the present invention solves the problem of querying the best smart device group when receiving a service request in a SIoT environment. The present invention finds the best device combination that can complete a given task set in a task pool by giving a heterogeneous social graph, a task set, social relationships between SIoT objects, and the relationship between each SIoT object and the task. For different scenarios, reducing communication loss or increasing robustness is considered and different strategies are customized, that is, the best device group is found for certain tasks by considering the interaction between different devices in the social Internet of Things.

[0044] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0046] Figure 1 The present invention is a flowchart of a method for building a collaborative device group in a social Internet of Things. DETAILED DESCRIPTION

[0047] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0048] See also Figure 1 ,The present invention provides a method for building a collaborative device group in a social Internet of Things. ,First, a heterogeneous social graph, a task set, social relationships between SIoT objects, and the relationship between each SIoT object and the task is constructed; ,At the same time, according to different actual needs, different constraints are applied to ,construct a device group selection method based on reducing communication loss or increasing robustness.

[0049] Specifically, given a heterogeneous graph G = (T, S, E, R), T is the task set, that is, the set of tasks that can be implemented by SIoT devices; S represents the set of SIoT devices; E represents the social relationship between devices, (u, v) ∈ E means that device u and device v can communicate; R is the task precision edge set, where each precision edge r = [t, v] connects a task vertex t∈T and a device vertex v∈S, that is, the precision of v executing a task t is used as the edge weight.

[0050] Given the above heterogeneous graph G and task query group The goal of the problem to be solved is to find a target device group with exactly p number of devices. To optimize the accuracy of the selected task in Y. Constraint p represents how many devices are planned to be controlled or carried according to the application scenario. The quality of the solution of the found device group is measured by the sum of the precision edge weights of each vertex in Y. Let I F (t) represents the sum of the task accuracy edge weights of the target device group F for task t∈Y, that is, I F (t) = ∑ v∈Fω[t,v]. The aggregate quality corresponding to the task query group Y to the target device group F is represented by the sum of the event weights of all tasks t in the task query group Y to the target device group F, that is, the objective function is defined as Ω(F) = ∑ t∈Y I F (t). The goal of the present invention is to maximize the objective function Ω(F). In addition, an accuracy constraint τ is included in the problem, which can ensure the performance of the target device group in the worst case. At the same time, different constraints are applied to Y according to different actual needs to reduce communication losses or increase the robustness of the device group selected in F. Based on different constraints, the main contents of the present invention are as follows.

[0051] 1. A device group selection method suitable for reducing communication loss in social Internet of Things.

[0052] Reducing the problem of communication loss: For the problem of communication loss, in addition to optimizing the accuracy of the selected task in Y, the communication loss between different devices should also be considered, that is, an upper limit is set for the jump distance between devices to limit the number of forwarded messages. The constraint is that the number of jumps between each pair of vertices in F is at most h, that is, To reduce potential communication losses. Due to the interaction between two different edge sets E and R, the device group with the maximum task accuracy does not always meet the hop count constraint. To address this problem, a selection method based on communication loss is proposed. The specific steps are as follows:

[0053] First, a preprocessing step is performed to ensure that the accuracy edge weight of all tasks for each device in S is at least τ. For a device v in the target device set F, any device u∈F must satisfy That is, for each device v∈S, a candidate device set S is constructed v , which only contains vertices within h hops. For a device u∈S, α(u) is expressed as the sum of the edge weights of the task precisions from u to Y, that is, α(u) = ∑ s∈Y w[u,s]. Then select p vertices with the largest α(u) from v's candidate set, construct a candidate solution for v, repeat the above steps, construct different candidate solutions, and return the solution with the largest Ω(F) as the target device group F. Since all vertices in S need to be scanned, a large computational overhead will result. Considering that if vertices are searched in a predefined order, some vertices may not need to be verified because their candidate vertices are destined not to be combined into a better solution. Therefore, a vertex access sorting and search strategy can be formulated to avoid unnecessary searches. For example, visiting each vertex v∈S in descending order of α(u) can better evaluate the quality of the solution in each candidate group to avoid redundant checks. Specifically, a list L is associated with each vertex v∈S. v , which is used to store S vThe top-p vertices with the largest α(·). Each time a vertex v is examined and the corresponding S is constructed in descending order of α(v) v while inserting v into the list L of each vertex u u such that |L u |< p. Then, a pruning strategy is adopted. Before constructing the set of vertices within h hops for the vertex v ∈ S, first check L v to determine whether a solution better than the current solution will be generated. If not, directly skip it.

[0054] 2. A method for selecting a device group suitable for increasing robustness in the social Internet of Things scenario.

[0055] The problem of increasing robustness:

[0056] The robustness problem mainly considers the number of different message transmission paths, that is, a device can transmit or back up its data through multiple different adjacent devices. Therefore, in addition to optimizing the accuracy of the selected tasks in Y, it is also required that each device in F has at least k adjacent devices to successfully transmit messages. That is to say, each vertex must also have at least k adjacent vertices in F. According to the edge set E in the subgraph the in-degree of the vertex v is represented as which is the number of vertices u ∈ H such that (u, v) ∈ E.

[0057] Due to the interaction between accuracy and communication robustness, devices with high accuracy may not have robust communication capabilities, and devices with robust communication capabilities may not always have optimal accuracy. To achieve a better balance between the quality and efficiency of the solution, a partial solution S sub is constructed step by step to obtain the complete solution.

[0058] First, apply a filtering strategy to remove the device u ∈ S from G that does not meet the accuracy constraint and the node degree constraint. Second, sort the candidate nodes according to accuracy. When adding a new node to the candidate solution C, first calculate whether its robustness meets the conditions. If not, the node is not included in the partial solution S sub . Then, pruning is performed according to accuracy and robustness respectively. Suppose there are m nodes in the current partial solution. The pruning strategy based on accuracy is that if the sum of α(v) of the m nodes and the sum of the maximum value of α(u) among the (p - m) candidate nodes is still less than the sum of α(v) of the current candidate complete solution, that is, ∑ v∈S α(v)+(p - m)·max u∈C α(u) ≤ Ω(S *), this part of the solution can be deleted; the robustness-based pruning strategy is to delete this part of the solution if the sum of the remaining (pm) candidate nodes and the minimum node degree of m nodes is still less than k, or the sum of the node degrees of the remaining (pm) candidate nodes is less than k(pm). However, choosing a suitable strategy requires domain knowledge and a time-consuming trial-and-error process. Considering the advantages of reinforcement learning in solving graph optimization problems, reinforcement learning is proposed to explore optimization strategies.

[0059] In reinforcement learning, the research problem of this invention is stated as follows: Given a heterogeneous graph G = (T, S, E, R), a task query set Y∈T, a degree constraint k, a size constraint p, and an accuracy constraint τ, the goal is to extract the target device set according to the learning function H(ξ) using reinforcement learning where ξ is the set of hyperparameters such that:

[0060] 1) |F| = p;

[0061] 2)

[0062] 3)

[0063] 4) Maximize the target value Ω(F).

[0064] When constructing partial solutions, reinforcement learning needs to evaluate based on the discriminant features (i.e., states) of candidate devices. Feature information should contain global and local information of the graph structure, and should also consider how the partial solution satisfies the constraints to form a usable solution. Since the quality of the solution is highly correlated with the node degree, it is necessary to extract features based on the average degree:

[0065] 1) The average degree of partial solutions;

[0066] 2) Partial solution S sub The normalized average degree of is the average degree normalized by the total degree;

[0067] 3) The average degree divided by k indicates the density of the partial solutions.

[0068] When the density is not high enough, the basis for selecting nodes may be to increase the density rather than to increase the target value Ω(F). Therefore, the following three features are further extracted to compare the partial solution with the entire graph:

[0069] 4) Density ratio, the ratio of the density of the partial solution to the density of the entire graph, indicating the tendency of forming a dense subgraph in the current input graph;

[0070] 5) Minimum degree ratio, that is, the ratio of the minimum degree in the partial solution to the minimum degree of the entire graph, which is used to measure the relative minimum degree;

[0071] 6) Minimum density ratio, the ratio of the density of the partial solution to the minimum density of the partial solution. All feature information is spliced ​​into a multidimensional feature vector F sub , and linearly transformed by the projection matrix W1, that is, V sub =W1F sub . V sub It contains the subgraph information of partial solutions and the constraint information of the problem.

[0072] After extracting the subgraph features of the partial solution, the graph convolutional network (GCN) is used to effectively integrate the node features and subgraph features of the graph structure. First, the enhanced features are input into the node evaluation function Q with the learning parameter Θ to evaluate the addition of node v to the current partial solution S. sub The performance in Q(S sub ,v;Θ). Then aggregate the node features layer by layer. Let α(v) and x v They represent the sum of the task accuracy edge weights of v and whether v is in a partial solution. By aggregating α(v), x v and the neighboring node features from the previous layer Get the node features of v at the (l+1) layer, and use Indicates that

[0073]

[0074] where W2∈R d ,W3∈R d ,W4∈R d×d is a learnable parameter and ReLU is the activation function. It can be initialized by existing graph embedding methods. After L layers of aggregation features, the embedding of each node v can be obtained, that is,

[0075] The evaluation function Q evaluates the features of vertex v by the enhanced features of v , the sum of adjacent node pooling And the feature embedding V of the subgraph sub Composition.

[0076]

[0077] where W6∈R d ,W7∈R d ,W5∈R d×d is a learnable parameter. If the target value of vertex v is high, it should be added to the partial solution first. If vertex v has been added to the partial solution, its neighbor node N(v) should be ranked higher than other vertices.

[0078] Compared with the number of training sets required, real-world data is relatively sparse and cannot meet the needs. Graph generators can be used to synthesize examples for learning, but the characteristics of real graphs are different from those of graph generators. Therefore, the idea of ​​meta-learning is proposed to explore models that can quickly adapt to new environments.

[0079] The model architecture is divided into an encoder-decoder architecture, where: encoder Θ e Contains {W1,W2,W3,W4,W5}, the decoder parameters Θ d ={W6,W7}. The goal of meta-learning is to make the encoder parameters Θ e can be transferred to the new graph distribution. During training, the encoder parameters and decoder parameters are optimized alternately according to the gradient descent function. During testing, the encoder parameters Θ are fixed e And optimize the decoder parameters Θ d .

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for constructing a collaborative device group in a social Internet of Things, characterized in that: Construct a heterogeneous social graph, a set of tasks, social relations between SIoT objects, and the relationship between each SIoT object and task; where SIoT stands for Social Internet of Things; Specifically, it includes: given a heterogeneous graph G = (T, S, E, R), where T is a task set, that is, a set of tasks that can be implemented by SIoT devices; S represents a set of SIoT devices; E represents the social relationship between devices, (u, v) ∈ E represents that devices u and v can communicate; R is a task precision edge set, where each precision edge r = [t, v] connects a task vertex t ∈ T and a device vertex v ∈ S, that is, the precision of v executing a certain task t is used as the edge weight; Given a heterogeneous graph G and a task query group The goal of the problem to be solved is to find a target device group with exactly p number of devices. to optimize the accuracy of the selected task in Y; constraint p represents how many devices are planned to be controlled or carried according to the application scenario; the sum of the precision edge weights of each vertex in Y is used to measure the quality of the solution of the found device group, let I F (t) represents the sum of the task accuracy edge weights of the target device group F for task t∈Y, that is, I F (t)=∑ v∈F ω[t,v]; the sum of the event weights of all tasks t in the task query group Y to the target device group F is used to represent the aggregation quality corresponding to the task query group Y to the target device group F, that is, the objective function is defined as Ω(F) = ∑ t∈Y I F (t); The optimization goal is to maximize the objective function Ω(F); In addition, the problem contains an accuracy constraint τ, which is used to ensure the performance of the target device group in the worst case; At the same time, according to different actual needs, different constraints are applied to Y to reduce communication loss or increase the robustness of the device group selected in F; Aiming at the problem that the device group with the maximum task accuracy does not always satisfy the hop count constraint due to the interaction of two different edge sets E and R, a selection method based on communication loss is adopted.

2. The method for constructing a collaborative device group in a social Internet of Things according to claim 1, characterized in that: Construct a device group selection method based on reducing communication loss, which specifically includes: in addition to optimizing the accuracy of the selected task in Y, the communication loss between different devices is also considered, that is, an upper limit constraint is set for the jump distance between devices; this constraint requires that the number of jumps between each pair of vertices in F is at most h, that is, Aiming at the problem that the device group with the maximum task accuracy does not always meet the hop count constraint due to the interaction between two different edge sets E and R, a selection method based on communication loss is adopted. The specific steps are as follows: First, a preprocessing step is performed to ensure that the accuracy edge weight of all tasks for each device in S is at least τ; for a device v in the target device set F, any device u∈F must satisfy That is, for each device v∈S, a candidate device set S is constructed v , which only contains vertices within h hops; for device u∈S, α(u) is expressed as the sum of the edge weights of the task precisions from u to Y, that is, α(u) = ∑ s∈Y w[u,s]; then select p vertices with the largest α(u) from the candidate set of v, construct a candidate solution for v, repeat the above steps, construct different candidate solutions, and return the solution with the largest Ω(F) as the target device group F; formulate vertex access sorting and search strategies to avoid unnecessary searches.

3. The method for constructing a collaborative device group in a social Internet of Things according to claim 2, characterized in that: The vertex access sorting and search strategy specifically includes: accessing each device, i.e., vertex v ∈ S, in descending order of α(u), which can better evaluate the solution quality in each candidate group to avoid redundant checks; associating a list L with each vertex v ∈ S v , which is used to store the top-p vertices with the largest α(·) in S v ; each time vertex v is checked and the corresponding S is constructed in descending order of α(v) v , v is inserted into the list L of each vertex u at the same time u ; |L u | < p; then, a pruning strategy is adopted. Before constructing the vertex set within h hops for vertex v ∈ S, first check L v to determine whether a solution better than the current solution will be generated. If not, directly skip it.

4. The method for constructing a collaborative device group in a social Internet of Things according to claim 1, characterized in that: Construct a device group selection method based on increasing robustness, which includes: in addition to optimizing the accuracy of the selected task in Y, it is also required that each device in F has at least k adjacent devices to successfully transmit messages; each device, i.e., vertex, must also have at least k adjacent vertices in F; according to the subgraph The edge set E in the graph represents the intrinsic degree of vertex v as It is the number of vertices u∈H such that (u,v)∈E; In order to strike a balance between quality and efficiency, a partial solution is constructed. sub The complete solution is gradually constructed in this way, which includes the following steps: First, the filtering strategy is applied to remove the devices u∈S that do not meet the accuracy constraints and node degree constraints from G; secondly, the candidate nodes are sorted according to the accuracy. When adding a new node to the candidate solution C, its robustness is first calculated to see if it meets the conditions. If it does not meet the conditions, the node will not be included in the partial solution S. sub ; Then prune according to accuracy and robustness respectively; Assuming that there are m nodes in the current partial solution, the accuracy-based pruning strategy is that if the sum of α(v) of m nodes and the sum of the maximum α(u) of (pm) candidate nodes is still less than the sum of α(v) of the current candidate complete solution, that is, ∑ v∈S α(v)+(pm)·max u∈C α(u)≤Ω(S * ), this part of the solution can be deleted, where Ω(S * ) represents the sum of α(v) of the current candidate complete solution; the robustness-based pruning strategy is to delete the partial solution if the sum of the minimum node degrees of the remaining (pm) candidate nodes and m nodes is still less than k, or the sum of the node degrees of the remaining (pm) candidate nodes is less than k(pm); selecting a suitable strategy requires domain knowledge and a time-consuming trial-and-error process, and reinforcement learning is used to explore the optimization strategy.

5. The method for constructing a collaborative device group in a social Internet of Things according to claim 4, characterized in that: The reinforcement learning exploration optimization strategy specifically includes: given a heterogeneous graph G = (T, S, E, R), a task query group Y∈T, a degree constraint k, a size constraint p and an accuracy constraint τ, the goal is to extract the target device group according to the learning function H(ξ) using reinforcement learning where ξ is the set of hyperparameters such that: 1) |F| = p; 2) 3) 4) Maximize the target value Ω(F); When constructing partial solutions, reinforcement learning evaluates candidate devices based on their discriminant features, i.e., their states. Feature information includes global and local information about the graph structure, and also considers how the partial solutions satisfy the constraints to form usable solutions. Since the quality of the solution is highly correlated with the node degree, it is necessary to extract features based on the average degree: 1) The average degree of partial solutions; 2) Partial solution S sub The normalized average degree of , that is, the average degree normalized by the total degree; 3) The average degree divided by k indicates the density of the partial solutions; 4) Density ratio, the ratio of the density of the partial solution to the density of the entire graph, indicating the tendency of forming a dense subgraph in the current input graph; 5) Minimum degree ratio, that is, the ratio of the minimum degree in the partial solution to the minimum degree of the entire graph, which is used to measure the relative minimum degree; 6) Minimum density ratio, the ratio of the density of the partial solution to the minimum density of the partial solution; all feature information is spliced ​​into a multidimensional feature vector F sub , and linearly transformed by the projection matrix W1, that is, V sub =W1F sub ; V sub It contains the subgraph information of partial solutions and the constraint information of the problem; After extracting the subgraph features of the partial solution, the graph convolutional network is used to integrate the node features and subgraph features of the graph structure; first, the enhanced features are input into the node evaluation function Q with the learning parameter Θ to evaluate the addition of node v to the current partial solution S sub The performance in Q(S sub ,v;Θ); then aggregate node features layer by layer; let α(v) and x v Respectively represent the sum of the task accuracy edge weights of v and whether v is in a partial solution; by aggregating α(v), x v and the neighboring node features from the previous layer Get the node features of v at the (l+1) layer, and use Indicates that Where W2∈R d ,W3∈R d ,W4∈R d×d is a learnable parameter, ReLU is the activation function; Initialize it through the existing graph embedding method; after aggregating the features through L layers, get the embedding of each node v, that is, The evaluation function Q evaluates the features of vertex v by the enhanced features of v Adjacent nodes pooled sum And the feature embedding V of the subgraph sub constitute; that is Among them, W6∈R d ,W7∈R d ,W5∈R d×d is a learnable parameter; if the target value of vertex v is high, it should be added to the partial solution first; if vertex v has been added to the partial solution, its neighbor node N(v) should be ranked higher than other vertices; Use the idea of ​​meta-learning to explore models that can quickly adapt to new environments.

6. The method for constructing a collaborative device group in a social Internet of Things according to claim 5, characterized in that: The idea of ​​using meta-learning to explore the model architecture that can quickly adapt to new environments is divided into an encoder-decoder architecture, where: encoder Θ e Contains {W1,W2,W3,W4,W5}, the decoder parameters Θ d ={W6,W7}; The goal of meta-learning is to make the encoder parameters Θ e Can be transferred to new graph distributions; during training, encoder parameters and decoder parameters are optimized alternately according to the gradient descent function; During the test, the encoder parameter Θ is fixed e And optimize the decoder parameters Θ d .

Images