A method for maximizing the influence of hypergraphs in smart home scenarios

Abstract

This invention relates to a method for maximizing the influence of a hypergraph in smart home scenarios, specifically as follows: Hypergraph data is generated by acquiring entity-to-entity interaction events in the smart home scenario; a function is defined to determine the number of entities and interaction events required for activation; the activation propagation process is modeled based on a hypergraph linear threshold model; a low-dimensional vector network embedding of the hypergraph is generated; a population is initialized, individuals are assigned to subpopulations, and a differentiated adaptive differential mutation strategy is defined; the performance of the DQN agent is evaluated through MDP and a reward mechanism; an experience replay pool is initialized to store interaction data; differential evolution is performed on the DQN individuals; the evaluation and evolution steps are repeated until no better individual is found; finally, the optimal DQN individual is selected, and k seed nodes are chosen to form a seed set for maximizing influence. This method can accurately characterize high-order relationships among multiple entities in a heterogeneous network of a smart home, quickly find the optimal seed node, efficiently trigger desired device interaction and user behavior, and achieve intelligent proactive optimization control in the home scenario.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and the Internet of Things, and in particular to a method for maximizing the influence of a hypergraph in smart home scenarios. Background Technology

[0002] With the rapid development of IoT and AI technologies, smart homes have evolved from remote control of single devices into a complex intelligent ecosystem characterized by multi-dimensional information perception, device linkage, and scene adaptation. The core objective of this system is to improve users' quality of life and achieve a more comfortable, energy-efficient, and convenient living experience. In a smart home environment, various intelligent devices, sensors, users, and environmental factors are interconnected and mutually influential, forming a dynamically changing and complex network.

[0003] In this network, one or more initial actions (e.g., a user executing a "home mode" command, or the system detecting that indoor light levels are below a threshold) can trigger a series of chain reactions, much like information spreading on a social network, thereby guiding the entire home environment to a specific "scene state." How to trigger a series of desired device interactions and user behaviors most efficiently with minimal initial intervention or at the lowest cost to achieve the globally optimal scene state is essentially an Influence Maximization (IM) problem.

[0004] Traditional research on maximizing influence has primarily focused on social networks, aiming to find a small number of "seed nodes" to maximize the breadth and depth of information dissemination. However, existing influence maximization algorithms face numerous challenges when applied to smart home scenarios. First, traditional methods are typically based on simple graph models where the relationships between nodes and edges are homogeneous (e.g., the "follow" relationship between users). A smart home, however, is a typical heterogeneous network containing various types of entities, including users, devices, sensors, and the spatial environment, with many-to-many, high-order, and complex interactions. For example, the action of "turning on the air conditioner" simultaneously affects entities in three different dimensions: "user's comfort," "indoor temperature," and "energy consumption." Traditional graph models struggle to accurately characterize these complex high-order relationships.

[0005] To more accurately model higher-order relationships in smart homes, some research has begun to introduce hypergraph theory, where a single hyperedge can connect multiple nodes, naturally describing group interactions among multiple entities. Meanwhile, to address dynamic decision-making problems, reinforcement learning, particularly deep Q-networks (DQNs), has been used to learn strategies for selecting optimal behavior in dynamic environments. However, within the vast state and action space of a smart home, single reinforcement learning algorithms still face challenges such as slow convergence and a tendency to get trapped in local optima when exploring optimal strategies.

[0006] Therefore, designing an efficient and fast solution algorithm to achieve intelligent and proactive optimization control of home scenarios is a pressing technical challenge in the field of smart homes. This invention proposes a method for maximizing the influence of a supergraph in smart home scenarios based on a DQN-evolutionary hybrid algorithm to solve the above problem. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention develops a method for maximizing the influence of hypergraphs in smart home scenarios. This invention and method can accurately characterize the high-order relationships of multiple entities in heterogeneous networks of smart homes, quickly find the optimal seed node, efficiently trigger the linkage of desired devices and user behaviors, and realize intelligent and proactive optimization control of home scenarios.

[0008] The technical solution of this invention is a method for maximizing the influence of a hypergraph in smart home scenarios, comprising the following steps:

[0009] S1. Obtain entities and their interaction events in the smart home scenario, generate hypergraph data, and define function P to determine the number of entities and interaction events required to activate the hypergraph data.

[0010] S2. A hypergraph-based linear threshold model is used to model the activation and propagation process of entities and interconnected events in a smart home scenario, simulate the influence diffusion law, and define the activation principle.

[0011] S3. Using the NHNE embedding algorithm to process hypergraph data. Perform network embedding to generate a hypergraph Network embedding of low-dimensional vectors ;

[0012] S4. Initialize a population of X individuals, each represented by a deep DQN network. The network weights of the DQNs are randomized. Each DQN serves as a strategy for selecting seed nodes. A high-influence seed node is selected in a Markov Decision Process (MDP) with k discrete time steps.

[0013] S5. Distribute X individuals evenly into 4 subpopulations, each subpopulation containing X / 4 individuals. Define four adaptive differential mutation strategies with differentiating parameters for each of the four subpopulations.

[0014] S6. Based on MDP and reward mechanism, evaluate the performance of each DQN agent in the population, calculate the state-action value and select the optimal action through a greedy strategy.

[0015] S7. Initialize the experience replay pool and store the 4-tuple of state transitions generated by each DQN individual's interaction with the environment.

[0016] S8. Perform differential evolution operations on each subpopulation of DQN individuals after evaluation, including direct inheritance of elite individuals based on the elite retention ratio, adaptive strategy selection combined with success counters, and mutation calculation of the corresponding strategy.

[0017] S9. Repeat steps S6 to S8, cycling through G rounds until no better individual is generated;

[0018] S10. After completing G rounds of iterative training, select the optimal DQN individual and use this optimal DQN individual to select k seed nodes in the hypergraph data to form the final seed set for maximizing influence.

[0019] S1 is as follows:

[0020] The generated hypergraph data is represented as ,in, For a set of nodes, , For the first individual entities and , For a set of superedges, , For the first A series of linked events and , and ;

[0021] Define a function P to determine the number of nodes and hyperedges required to activate the hypergraph. Specifically, calculate the threshold number of nodes and hyperedges using function P, and determine the threshold number of each node based on the number of hyperedges it is connected to. The threshold number of a hyperedge is determined based on the number of nodes contained in each hyperedge. A node is activated when its threshold number is less than or equal to the number of activated hyperedges it connects to; conversely, a hyperedge is activated when its threshold number is less than or equal to the number of activated nodes it contains. This process is used to determine the optimal set of seed nodes. To maximize the impact of propagation, the seed node set The size is .

[0022] S2 is as follows:

[0023] The activation rule for the linear threshold model of a hypergraph is:

[0024] When node Number of active superedges connected At that time, node Activated nodes are marked with a status of 1, while inactive nodes are marked with a status of 0.

[0025] When the super-edge Number of activated nodes included At that time, the super-border Activated;

[0026] The influence spreads from the seed node set, iteratively propagating between superedges and nodes until no new nodes can be activated.

[0027] S3 is as follows:

[0028] (1) Read Hypergraph data ,use Random walk strategy generates node sequence. Through hyperparameters Guiding the direction of movement, among which... and Whether to prioritize depth-first or breadth-first search for navigation. To guide the swimmer to move towards nodes with larger or smaller degrees;

[0029] Input the generated node sequence into The skipping word model predicts context nodes by using a sliding window to predict the center node of the sequence and learns a low-dimensional vector for each node by optimizing the co-occurrence probability.

[0030] (2) Construct the original hypergraph dual hypergraph, primal hypergraph Each hyperedge corresponds to a node in the dual hypergraph, the original hypergraph. Each node corresponds to a hyperedge of the dual hypergraph; using Algorithm combination The model learns on the dual hypergraph to obtain a low-dimensional vector for each hyperedge in the original hypergraph;

[0031] (3) Input the low-dimensional vector of each node in step (1) and the low-dimensional vector of each hyperedge obtained in step (2) into a 1D convolutional neural network for processing, concatenate the results, and then jointly optimize them through the tuple similarity function to generate a hypergraph. Network Embedding .

[0032] The specific method for initializing X DQNs in S4 is as follows:

[0033] Each DQN is a neural network containing two linear layers. All weights in the network are randomly initialized, forming X weight sequences for each DQN. , Indicates the first The weights of each DQN.

[0034] The parameter initialization for the four adaptive differential mutation strategies in S5 is as follows:

[0035] The first subpopulation is highly exploratory, the second is standard, the third is highly efficient, and the fourth is balanced. A scaling factor is set for each subpopulation. Crossover probability and the proportion of elites retained Each subpopulation has a different scaling factor and crossover probability, and the elite retention rate is the same for all four subpopulations.

[0036] The specific process for evaluating DQN agents in S6 is as follows:

[0037] (1) Calculate the state-action function value ,in These are the weight parameters of DQN. Indicates the current state Next action Value, action The process of selecting nodes;

[0038] state Embedded by Hypergraph Network Impact estimation characteristics and node state sequence It is formed by fusion. ;

[0039] To influence the estimation characteristics, , Represents a node The number of activated hyperedges is calculated using the following formula:

[0040] ,

[0041] in, It is with nodes Connected hyperedge sets, Represents the set of seed nodes The set of active nodes below To be with the super-edge The set of connected nodes, where δ is the Kronecker indicator function; and , Represents a node Seed node, Represents a node Non-seed nodes;

[0042] (2) Use a greedy strategy to select the optimal action: at time step Determine the optimal value And select the optimal action. , This represents a function operation that retrieves the maximum value; the action performed. Then, update the node state o{ }=1 and = ,based on Extract seed node set Update impact estimation features = Generate the state for the next time step. ;

[0043] (3) Calculate the reward value ,in, Representing state The set of active nodes below, through Quantify execution actions The resulting increase in influence;

[0044] The quadruples generated after each DQN individual, representing a different strategy, interacts with the environment are represented as follows: .

[0045] The differential evolution operation in S8 is as follows:

[0046] (1) The DQN individuals in each subpopulation are sorted in descending order of evaluation performance, and the elite retention rate of each subpopulation is determined. , will the first in the subpopulation × The optimal individual in the subpopulation is directly replicated into the offspring population;

[0047] (2) Initialize the success counter =[3,3,3,3] is used to record the number of times the four subpopulations generate a better DQN. A better DQN refers to the seed node that can activate more nodes under the hypergraph linear threshold model. The more successful the subpopulations are, the greater the probability that their corresponding mutation strategy will be selected.

[0048] (3) Perform mutation calculations according to the selected mutation strategy:

[0049] 1) When selecting the first subpopulation strategy, select 3 DQN individuals from the first subpopulation. Generate new individuals

[0050] 2) When selecting the second subpopulation strategy, select 3 random DQN individuals from the second subpopulation. and the optimal individual Generate new individuals

[0051] 3) When selecting the third subpopulation strategy, target individuals are selected from the third subpopulation. 2 random individuals and the optimal individual Generate new individuals ;

[0052] 4) When selecting the fourth subpopulation strategy, select 5 random DQN individuals from the fourth subpopulation. Generate new individuals

[0053] (4) Evaluate the performance of offspring individuals. If an offspring individual is superior to its parent individual, then the corresponding subpopulation... Increment the count by 1.

[0054] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0055] This invention discloses a method for maximizing the influence of a hypergraph in smart home scenarios. It employs a hypergraph for modeling, utilizing the characteristic of hyperedges connecting multiple nodes to naturally describe many-to-many group interactions. This overcomes the limitations of traditional binary graph models, thus more accurately simulating higher-order relationships between multiple entities in a smart home. Furthermore, addressing the low policy search efficiency and susceptibility to local optima caused by the vast state and action space in smart homes, this invention combines a deep DQN network with an evolutionary algorithm. Through population iteration, elite retention, crossover, and mutation operations, it enhances global optimization capabilities, accelerates convergence, and improves solution quality. In addition, to further enhance the robustness and adaptability of the algorithm, this invention introduces an adaptive mutation mechanism based on sliding window monitoring. By dynamically adjusting the exploration intensity and maintaining population diversity, it effectively avoids premature convergence, enabling the algorithm to adapt to the needs of different optimization stages.

[0056] In summary, the present invention and method can accurately characterize the high-order relationships of multiple entities in a heterogeneous smart home network, quickly find the optimal seed node, efficiently trigger the linkage of desired devices and user behavior, and realize intelligent proactive optimization control of the home scene. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0058] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0059] Figure 2 This is a line graph showing the influence expansion of the method of this invention and existing methods on the Music dataset at different seed sizes.

[0060] Figure 3 This is a line graph showing the influence spread of the method of this invention and existing methods on the NDC dataset at different seed sizes.

[0061] Figure 4 This is a line graph showing the influence spread of the method of this invention and existing methods on the Restaurant dataset at different seed sizes.

[0062] Figure 5 This is a line graph showing the training time of the method of this invention and existing methods on the Music dataset at different seed sizes.

[0063] Figure 6 This is a line graph showing the training time of the method of this invention and existing methods on the NDC dataset under different seed sizes.

[0064] Figure 7 This is a line graph showing the training time of the method of this invention and existing methods on the Restaurant dataset under different seed sizes. Detailed Implementation

[0065] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0066] Example 1

[0067] like Figure 1 As shown, a method for maximizing the influence of a hypergraph in a smart home scenario includes the following steps:

[0068] S1. Obtain entities and their interaction events in the smart home scenario, generate hypergraph data, and define function P to determine the number of entities and interaction events required to activate the hypergraph data.

[0069] S2. A hypergraph-based linear threshold model is used to model the activation and propagation process of entities and interconnected events in a smart home scenario, simulate the influence diffusion law, and define the activation principle.

[0070] S3. Using the NHNE embedding algorithm to process hypergraph data. Perform network embedding to generate a hypergraph Network embedding of low-dimensional vectors ;

[0071] S4. Initialize a population of X individuals, each represented by a deep DQN network. The network weights of the DQNs are randomized. Each DQN serves as a strategy for selecting seed nodes. A high-influence seed node is selected in a Markov Decision Process (MDP) with k discrete time steps.

[0072] S5. Distribute X individuals evenly into 4 subpopulations, each subpopulation containing X / 4 individuals. Define four adaptive differential mutation strategies with differentiating parameters for each of the four subpopulations.

[0073] S6. Based on MDP and reward mechanism, evaluate the performance of each DQN agent in the population, calculate the state-action value and select the optimal action through a greedy strategy.

[0074] S7. Initialize the experience replay pool and store the 4-tuple of state transitions generated by each DQN individual's interaction with the environment.

[0075] S8. Perform differential evolution operations on each subpopulation of DQN individuals after evaluation, including direct inheritance of elite individuals based on the elite retention ratio, adaptive strategy selection combined with success counters, and mutation calculation of the corresponding strategy.

[0076] S9. Repeat steps S6 to S8, cycling through G rounds until no better individual is generated;

[0077] S10. After completing G rounds of iterative training, select the optimal DQN individual and use this optimal DQN individual to select k seed nodes in the hypergraph data to form the final seed set for maximizing influence.

[0078] In a specific implementation, S1 is as follows:

[0079] The generated hypergraph data is represented as ,in, For a set of nodes, , For the first individual entities and , For a set of superedges, , For the first A series of linked events and , and ;

[0080] Define a function P to determine the number of nodes and hyperedges required to activate the hypergraph. Specifically, calculate the threshold number of nodes and hyperedges using function P, and determine the threshold number of each node based on the number of hyperedges it is connected to. The threshold number of a hyperedge is determined based on the number of nodes contained in each hyperedge. A node is activated when its threshold number is less than or equal to the number of activated hyperedges it connects to; conversely, a hyperedge is activated when its threshold number is less than or equal to the number of activated nodes it contains. This process is used to determine the optimal set of seed nodes. To maximize the impact of propagation, the seed node set The size is ;

[0081] For nodes , , Represents a node The number of connected superedges, Represents a node The number of thresholds;

[0082] For hyperedge , , Indicates the superedge The number of nodes included. Indicates the superedge The number of thresholds.

[0083] In a specific implementation, S2 is as follows:

[0084] The activation rule for the linear threshold model of a hypergraph is:

[0085] When node Number of active superedges connected At that time, node Activated nodes are marked with a status of 1, while inactive nodes are marked with a status of 0.

[0086] When the super-edge Number of activated nodes included At that time, the super-border Activated;

[0087] The influence spreads from the seed node set, iteratively propagating between superedges and nodes until no new nodes can be activated.

[0088] In a specific implementation, S3 is as follows:

[0089] (1) Read Hypergraph data ,use Random walk strategy generates node sequence. Through hyperparameters Guiding the direction of movement, among which... and Whether to prioritize depth-first or breadth-first search for navigation. To guide the swimmer to move towards nodes with larger or smaller degrees;

[0090] Input the generated node sequence into The skipping word model predicts context nodes by using a sliding window to predict the center node of the sequence and learns a low-dimensional vector for each node by optimizing the co-occurrence probability.

[0091] (2) Construct the original hypergraph dual hypergraph, primal hypergraph Each hyperedge corresponds to a node in the dual hypergraph, the original hypergraph. Each node corresponds to a hyperedge of the dual hypergraph; using Algorithm combination The model learns on the dual hypergraph to obtain a low-dimensional vector for each hyperedge in the original hypergraph;

[0092] (3) Input the low-dimensional vector of each node in step (1) and the low-dimensional vector of each hyperedge obtained in step (2) into a 1D convolutional neural network for processing, concatenate the results, and then jointly optimize them through the tuple similarity function to generate a hypergraph. Network Embedding .

[0093] In a specific implementation, the method for initializing X DQNs in S4 is as follows:

[0094] Each DQN is a neural network containing two linear layers. All weights in the network are randomly initialized, forming X weight sequences for each DQN. , Indicates the first The weights of each DQN.

[0095] In a specific implementation, the parameter initialization of the four adaptive differential mutation strategies in S5 is as follows:

[0096] The first subpopulation is highly exploratory, the second is standard, the third is highly efficient, and the fourth is balanced. A scaling factor is set for each subpopulation. Crossover probability and the proportion of elites retained Each subpopulation has a different scaling factor and crossover probability, and the elite retention rate is the same for all four subpopulations.

[0097] Set scaling factor in the first subpopulation Crossover probability ;

[0098] Set the scaling factor in the second subpopulation Crossover probability ;

[0099] Set the scaling factor in the third subpopulation Crossover probability ;

[0100] In the fourth subpopulation, the scaling factor F = 0.6 and the crossover probability CR = 0.6 are set.

[0101] Elite retention rates of the four subpopulations All are 0.1.

[0102] In a specific implementation, the evaluation process of the DQN agent in S6 is as follows:

[0103] (1) Calculate the state-action function value ,in These are the weight parameters of DQN. Indicates the current state Next action Value, action The process of selecting nodes;

[0104] state Embedded by Hypergraph Network Impact estimation characteristics and node state sequence It is formed by fusion. ;

[0105] To influence the estimation characteristics, , Represents a node The number of activated hyperedges is calculated using the following formula:

[0106] ,

[0107] in, It is with nodes Connected hyperedge sets, Represents the set of seed nodes The set of active nodes below To be with the super-edge The set of connected nodes, where δ is the Kronecker indicator function; and , Represents a node Seed node, Represents a node Non-seed nodes;

[0108] (2) Use a greedy strategy to select the optimal action: at time step Determine the optimal value And select the optimal action. , This represents a function operation that retrieves the maximum value; the action performed. Then, update the node status. and = ,based on Extract seed node set Update impact estimation features = Generate the state for the next time step. ;

[0109] (3) Calculate the reward value ,in, Representing state The set of active nodes below, through Quantify execution actions The resulting increase in influence;

[0110] The quadruples generated after each DQN individual, representing a different strategy, interacts with the environment are represented as follows: .

[0111] In a specific implementation, the differential evolution operation in S8 is as follows:

[0112] (1) The DQN individuals in each subpopulation are sorted in descending order of evaluation performance, and the elite retention rate of each subpopulation is determined. , will the first in the subpopulation × The optimal individual in the subpopulation is directly replicated into the offspring population;

[0113] (2) Initialize the success counter =[3,3,3,3] is used to record the number of times the four subpopulations generate a better DQN. A better DQN refers to the seed node that can activate more nodes under the hypergraph linear threshold model. The more successful the subpopulations are, the greater the probability that their corresponding mutation strategy will be selected.

[0114] (3) Perform mutation calculations according to the selected mutation strategy:

[0115] 1) When selecting the first subpopulation strategy, select 3 DQN individuals from the first subpopulation. Generate new individuals

[0116] 2) When selecting the second subpopulation strategy, select 3 random DQN individuals from the second subpopulation. and the optimal individual Generate new individuals

[0117] 3) When selecting the third subpopulation strategy, target individuals are selected from the third subpopulation. 2 random individuals and the optimal individual Generate new individuals ;

[0118] 4) When selecting the fourth subpopulation strategy, select 5 random DQN individuals from the fourth subpopulation. Generate new individuals

[0119] (4) Evaluate the performance of offspring individuals. If an offspring individual is superior to its parent individual, then the corresponding subpopulation... Increment the count by 1.

[0120] Example 2

[0121] To demonstrate the beneficial effects of the method of the present invention, the method of the present invention is compared with existing methods. Specifically, three hypergraphs from the real world are used as experimental datasets, namely MUSIC, Restaurant and NDC. The higher-order relationships between nodes are clearly marked. MUSIC and NDC have similar numbers of nodes and similar numbers of hyperedges. The network connectivity of the three datasets is sorted from strong to weak as follows: MUSIC, Restaurant, NDC. The specific details of the three datasets are shown in Table 1.

[0122] Table 1 shows the three datasets used from the real world.

[0123]

[0124] The existing methods are HHD, HD, CELF, Greedy, and HEDRL-IM;

[0125] HHD is a heuristic algorithm that sorts nodes according to their degree and then selects the top k nodes as seed nodes.

[0126] HD is a heuristic algorithm that sorts nodes according to their degree and then selects the node with the highest degree at each step.

[0127] The CELF method calculates the initial influence score of each node in the network, selects the node with the highest score to add to the seed set, and then iteratively updates the influence scores of the relevant nodes, repeating the above process until a predetermined number of seed nodes are selected.

[0128] In the Greedy method, the number of inactive hyperedges is continuously updated at each propagation step of the hypergraph, and the seed node is dynamically selected as the node with the most inactive hyperedges.

[0129] HEDRL-IM is an evolutionary deep reinforcement learning algorithm for hypergraphs that selects seed nodes by combining a genetic algorithm with a deep Q-network.

[0130] As shown in Table 2, Figures 2 to 4 As shown in Table 3, the influence expansion of the method of this invention is compared with that of existing methods on different datasets under different seed sizes K. Figures 5 to 7 As shown, the training time of the method of this invention and the HEDRL-IM method are compared on different datasets with different seed sizes K.

[0131] Table 2 Comparison Results of Influence Expansion

[0132]

[0133] Table 3 Comparison of Training Time

[0134]

[0135] The experiment compared the traditional hypergraph-based influence maximization algorithm with the influence maximization algorithm based on evolutionary deep reinforcement learning, where a larger influence expansion indicates better results. The comparison, as shown in the figures and tables, demonstrates that the method of this invention improves performance across different datasets, significantly outperforming traditional algorithms and slightly outperforming the HEDRL-IM algorithm. It also outperforms the HEDRL-IM algorithm in terms of training time. This indicates that using the method of this invention can more efficiently and quickly derive the optimal seed set.

[0136] Example 3

[0137] To illustrate how the method of this invention can be applied in a smart home scenario, let's take a "good morning" scenario as an example. The user's home is a network containing entities such as the user, mobile phone alarm clock, curtains, lights, and coffee machine. The system using this method learns that the action of "getting up" is associated with the alarm clock, curtains, and lights, while "preparing breakfast" is associated with the user and coffee machine. Its core task is to find out which action is the "key first step" that can trigger the entire ideal chain reaction.

[0138] Through simulation, the system will find that if only the coffee machine starts, the impact is small; but if the mobile phone alarm rings, it will activate the "user", thereby triggering the "wake up" event, causing the curtains to open automatically, the lights to turn on slowly, and then the "prepare breakfast" event to start the coffee machine.

[0139] Ultimately, users can seamlessly trigger a comfortable morning scene with just the simple action of being woken up by the alarm clock, perfectly demonstrating how this method can achieve proactive and automated services for smart homes by finding the most influential "seed nodes".

[0140] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for maximizing the influence of a hypergraph in smart home scenarios, characterized by: Includes the following steps: S1. Obtain entities and their interaction events in the smart home scenario, generate hypergraph data, and define function P to determine the number of entities and interaction events required to activate the hypergraph data. In the smart home scenario, entities include users, devices, sensors, and spatial environment, and the linkage events between entities are many-to-many interaction relationships between entities; S2. A hypergraph-based linear threshold model is used to model the activation and propagation process of entities and interconnected events in a smart home scenario, simulate the influence diffusion law, and define the activation principle. S3. Using the NHNE embedding algorithm to process hypergraph data. Perform network embedding to generate a hypergraph Network embedding of low-dimensional vectors ; S3 is as follows: (1) Read Hypergraph data ,use Random walk strategy generates node sequence. Through hyperparameters Guiding the direction of movement, among which... and Whether to prioritize depth-first or breadth-first search for navigation. To guide the swimmer to move towards nodes with larger or smaller degrees; Input the generated node sequence into The skipping word model predicts context nodes by using a sliding window to predict the center node of the sequence and learns a low-dimensional vector for each node by optimizing the co-occurrence probability. (2) Construct the original hypergraph dual hypergraph, primal hypergraph Each hyperedge corresponds to a node in the dual hypergraph, the original hypergraph. Each node corresponds to a hyperedge of the dual hypergraph; using Algorithm combination The model learns on the dual hypergraph to obtain a low-dimensional vector for each hyperedge in the original hypergraph; (3) Input the low-dimensional vector of each node in step (1) and the low-dimensional vector of each hyperedge obtained in step (2) into a 1D convolutional neural network for processing, concatenate the results, and then jointly optimize them through the tuple similarity function to generate a hypergraph. Network Embedding ; S4. Initialize a population of X individuals, each represented by a deep DQN network. The network weights of the DQNs are randomized. Each DQN serves as a strategy for selecting seed nodes. A high-influence seed node is selected in a Markov Decision Process (MDP) with k discrete time steps. S5. Distribute X individuals evenly into 4 subpopulations, each subpopulation containing X / 4 individuals. Define four adaptive differential mutation strategies with differentiating parameters for each of the four subpopulations. S6. Based on MDP and reward mechanism, evaluate the performance of each DQN agent in the population, calculate the state-action value and select the optimal action through a greedy strategy. The specific process for evaluating DQN agents is as follows: (1) Calculate the state-action function value ,in These are the weight parameters of DQN. Indicates the current state Next action Value, action The process of selecting nodes; state Embedded by Hypergraph Network Influence estimation characteristics and node state sequence It is formed by fusion. ; To influence the estimation characteristics, , Represents a node The number of activated hyperedges is calculated using the following formula: ， in, It is with nodes Connected hyperedge sets, Represents the set of seed nodes The set of active nodes below To be with the super-edge The set of connected nodes, where δ is the Kronecker indicator function; and , Represents a node Seed node, Represents a node Non-seed nodes; (2) Use a greedy strategy to select the optimal action: at time step Determine the optimal value And select the optimal action. , This represents a function operation that retrieves the maximum value; the action performed. Then, update the node state o{ }=1 and = ,based on Extract seed node set Update impact estimation features = Generate the state for the next time step. ; (3) Calculate the reward value ,in, Representing state The set of active nodes below, through Quantify execution actions The resulting increase in influence; The quadruples generated after each DQN individual, representing a different strategy, interacts with the environment are represented as follows: ; S7. Initialize the experience replay pool and store the 4-tuple of state transitions generated by each DQN individual's interaction with the environment. S8. Perform differential evolution operations on each subpopulation of DQN individuals after evaluation, including direct inheritance of elite individuals based on the elite retention ratio, adaptive strategy selection combined with success counters, and mutation calculation of the corresponding strategy. S9. Repeat steps S6 to S8, cycling through G rounds until no better individual is generated; S10. After completing G rounds of iterative training, select the optimal DQN individual and use this optimal DQN individual to select k seed nodes in the hypergraph data to form the final seed set for maximizing influence.

2. The method for maximizing hypergraph influence in smart home scenarios according to claim 1, characterized in that, S1 is as follows: The generated hypergraph data is represented as ,in, For a set of nodes, , For the first individual entities and , For a set of superedges, , For the first A series of linked events and , and ; Define a function P to determine the number of nodes and hyperedges required to activate the hypergraph. Specifically, calculate the threshold number of nodes and hyperedges using function P, and determine the threshold number of each node based on the number of hyperedges it is connected to. The threshold number of a hyperedge is determined based on the number of nodes contained in each hyperedge. A node is activated when its threshold number is less than or equal to the number of activated hyperedges it connects to; conversely, a hyperedge is activated when its threshold number is less than or equal to the number of activated nodes it contains. This process is used to determine the optimal set of seed nodes. To maximize the impact of propagation, the seed node set The size is .

3. The method for maximizing the influence of a hypergraph in a smart home scenario according to claim 2, characterized in that, S2 is as follows: The activation rule for the linear threshold model of a hypergraph is: When node Number of active superedges connected At that time, node Activated nodes are marked with a status of 1, while inactive nodes are marked with a status of 0. When the super-edge Number of activated nodes included At that time, the super-border Activated; The influence spreads from the seed node set, iteratively propagating between superedges and nodes until no new nodes can be activated.

4. The method for maximizing hypergraph influence in smart home scenarios according to claim 3, characterized in that, The specific method for initializing X DQNs in S4 is as follows: Each DQN is a neural network containing two linear layers. All weights in the network are randomly initialized, forming X weight sequences for each DQN. , Indicates the first The weights of each DQN.

5. The method for maximizing the influence of a hypergraph in a smart home scenario according to claim 4, characterized in that the parameter initialization of the four adaptive differential mutation strategies in S5 is as follows: The first subpopulation is highly exploratory, the second is standard, the third is highly efficient, and the fourth is balanced. A scaling factor is set for each subpopulation. Crossover probability and the proportion of elites retained Each subpopulation has a different scaling factor and crossover probability, and the elite retention rate is the same for all four subpopulations.

6. The method for maximizing hypergraph influence in smart home scenarios according to claim 5, characterized in that, The differential evolution operation in S8 is as follows: (1) The DQN individuals in each subpopulation are sorted in descending order of evaluation performance, and the elite retention rate of each subpopulation is determined. , will the first in the subpopulation × The optimal individual in the subpopulation is directly replicated into the offspring population; (2) Initialize the success counter =[3,3,3,3] is used to record the number of times the four subpopulations generate a better DQN. A better DQN refers to the seed node that can activate more nodes under the hypergraph linear threshold model. The more successful the subpopulations are, the greater the probability that their corresponding mutation strategy will be selected. (3) Perform mutation calculations according to the selected mutation strategy: 1) When selecting the first subpopulation strategy, select 3 DQN individuals from the first subpopulation. Generate new individuals 2) When selecting the second subpopulation strategy, select 3 random DQN individuals from the second subpopulation. and the optimal individual Generate new individuals 3) When selecting the third subpopulation strategy, target individuals are selected from the third subpopulation. 2 random individuals and the optimal individual Generate new individuals ; 4) When selecting the fourth subpopulation strategy, select 5 random DQN individuals from the fourth subpopulation. Generate new individuals (4) Evaluate the performance of offspring individuals. If an offspring individual is superior to its parent individual, then the corresponding subpopulation... Increment the count by 1.

Images