Recommendation Method, Device, System and Medium Based on Graph Neural Network Architecture Search

Abstract

The present invention discloses a recommendation method, device, system and medium based on graph neural network architecture search. The method includes: designing a corresponding search space for graph neural network architecture search; using the particle swarm algorithm as the search strategy for graph neural network architecture search, converting the positions of each particle into corresponding graph neural network models according to the search space, and selecting the historical optimal value of the particle swarm as the optimal graph neural network architecture; wherein, the particle position corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses the graph convolution operation in the decoupled mode to aggregate neighbor node information; training the optimal graph neural network architecture, and generating a recommendation list by using the trained optimal graph neural network architecture. In the present invention, the graph neural network in the decoupled mode deepens the number of layers of information propagation and improves the recommendation performance; the architecture search technology is used to automatically find an optimal graph neural network for different data sets.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation systems, and particularly relates to a recommendation method, device, system, terminal device, and storage medium based on graph neural network architecture search. Background Art

[0002] In the field of recommendation systems, due to the widespread existence of graph-structured data, graph neural network technology is used to effectively extract valuable information from the target dataset. Traditional graph neural network-based recommendation systems, such as Pinsage, NGCF, etc., all adopt non-decoupled graph convolution. Its characteristic is that after aggregating information from adjacent nodes each time, it needs to pass through a fully connected MLP network. According to the research in "Towards Deeper Graph Neural Networks", this non-decoupled graph convolution is usually the main cause of the performance degradation of deep graph neural networks. Summary of the Invention

[0003] To solve the above deficiencies of the prior art, the present invention provides a recommendation method, device, system, terminal device, and storage medium based on graph neural network architecture search. By adopting a decoupled graph convolution, the graph convolution process is divided into two processes: information propagation and feature transformation. First, information propagation is performed to aggregate the information of K-order neighbor nodes, and then feature transformation is performed to learn the final feature representation. The decoupled graph convolution is used to implement a deep graph neural network, enabling the recommendation system to utilize more information in the target dataset, thereby improving the recommendation performance. At the same time, neural network architecture search technology is adopted to find a network structure suitable for the target dataset in the deep network space, so that it is not necessary to manually adjust a network structure for different datasets, but an optimal graph neural network can be automatically found for different datasets. The method of the present invention combines a recommendation method based on collaborative filtering and uses the interaction data between users and items to achieve effective and accurate recommendations.

[0004] The first object of the present invention is to provide a recommendation method based on graph neural network architecture search.

[0005] The first object of the present invention is to provide a recommendation device based on graph neural network architecture search.

[0006] The third object of the present invention is to provide a recommendation system based on graph neural network architecture search.

[0007] The fourth object of the present invention is to provide a terminal device.

[0008] The fifth object of the present invention is to provide a storage medium.

[0009] The first object of the present invention can be achieved by adopting the following technical solutions:

[0010] A recommendation method based on graph neural network architecture search, the method comprising:

[0011] Design a corresponding search space for graph neural network architecture search;

[0012] Adopt the particle swarm algorithm as the search strategy for graph neural network architecture search, convert the positions of each particle into corresponding graph neural network models according to the search space, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture; wherein, the position of the particle corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses the decoupled mode of graph convolution operation to aggregate neighbor node information;

[0013] Train the optimal graph neural network architecture, and generate a recommendation list by using the trained optimal graph neural network architecture.

[0014] Further, the converting the positions of each particle into corresponding graph neural network models according to the search space and selecting the historical optimal value of the particle swarm as the optimal graph neural network architecture includes:

[0015] S1: Initialize the particle swarm, including the position and velocity of the particles;

[0016] S2: Convert the positions of each particle into corresponding graph neural network architectures according to the search space; train the graph neural network architectures, and calculate performance evaluation indicators by using the trained graph neural network architectures;

[0017] S3: According to the performance evaluation indicators, select the historical optimal value pbest of each particle and the historical optimal value gbest of the population, and then update the position and velocity of each particle;

[0018] Repeat steps S2 and S3 for multiple rounds of search iteration, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture after the iteration ends.

[0019] Further, the training of the graph neural network architecture includes:

[0020] The graph neural network model first randomly initializes the feature representations of users and items, and then uses the information propagation operation in the decoupled mode of graph convolution operation to aggregate neighbor node information and learn the feature representations of users and items; wherein, the information propagation operation is only executed once during the initialization process of the graph neural network model;

[0021] Calculate the inner product of the user feature vector and the item feature vector according to the feature representations of the user and the item; sort the calculation results of the inner product from largest to smallest, and take the top K from the corresponding item list as the final recommendation list; where K is a positive integer greater than or equal to 1.

[0022] The particle swarm algorithm uses the BPR loss function to train the graph neural network model in a gradient descent manner.

[0023] Further, calculate performance evaluation metrics using the trained graph neural network architecture, including:

[0024] Input the data in the test dataset into the trained graph neural network architecture to obtain a recommendation list.

[0025] Calculate performance evaluation metrics according to the recommendation list and the actual preference list of users in the test set.

[0026] Further, select the historical best value pbest of each particle and the historical best value gbest of the population according to the performance evaluation metrics, and then update the position and velocity of each particle, including:

[0027] Select the historical best value pbest of each particle and the historical best value gbest of the population according to the performance evaluation metrics.

[0028] According to pbest and gbest, use the velocity and position update formulas of the particle swarm to modify the information of the particles. The formulas are as follows:

[0029] v i+1 (t + 1) = v i (t) + c1 × rand × (pbest i (t) - x i (t)) + c2 × rand × (gbest - xi(t))x i+1 (t + 1) = x i (t) + v i+1 (t + 1)

[0030] To ensure that the position coordinates of the particles are integers after each movement, the following constraints are made:

[0031]

[0032] Among them, x i represents the position of the particle at time t, corresponding to the encoding of the graph neural network architecture; v i represents the movement speed of the particle at time t, v max is the maximum speed of the particle.; pbest iIt represents the historical optimal value of the particle individual, and gbest represents the historical optimal value of the entire particle swarm; both c1 and c2 are predefined hyperparameters used to control the search direction of the particles.

[0033] Further, in order to improve the efficiency of architecture search, in step S2, part of the data in the training dataset is used to train the graph neural network architecture;

[0034] So that the trainable parameters in the graph neural network architecture can fully fit the target dataset, and thus the best recommendation effect is obtained. Then, all the data in the training dataset is used to train the optimal graph neural network architecture.

[0035] Further, the corresponding search space is designed for the graph neural network architecture search, specifically:

[0036] The search space is decomposed into an information propagation search space, a feature transformation search space, and a hyperparameter search space.

[0037] The second object of the present invention can be achieved by adopting the following technical solutions:

[0038] A recommendation device based on graph neural network architecture search, the recommendation device includes:

[0039] A search space design module for designing a corresponding search space for graph neural network architecture search;

[0040] An optimal graph neural network architecture selection module, which uses the particle swarm algorithm as the search strategy for graph neural network architecture search, converts the positions of each particle into corresponding graph neural network models according to the search space, and selects the historical optimal value of the particle swarm as the optimal graph neural network architecture; among them, the position of the particle corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses the decoupled graph convolution operation to aggregate neighbor node information;

[0041] A recommendation list generation module for training the optimal graph neural network architecture and generating a recommendation list using the trained optimal graph neural network architecture.

[0042] The third object of the present invention can be achieved by adopting the following technical solutions:

[0043] A recommendation system based on graph neural network architecture search, the recommendation system includes a front-end page and a back-end service. The front-end page is used to input the interaction information between users and items and display the recommendation list returned by the back-end service; the back-end service generates a recommendation list according to the interaction information between users and items through the trained optimal graph neural network architecture obtained by running the above-mentioned recommendation method.

[0044] The fourth object of the present invention can be achieved by adopting the following technical solutions:

[0045] A terminal device includes a processor and a memory for storing programs executable by the processor. When the processor executes the programs stored in the memory, the above-mentioned recommendation method is implemented.

[0046] The fifth object of the present invention can be achieved by adopting the following technical solutions:

[0047] A storage medium stores a program. When the program is executed by a processor, the above-mentioned recommendation method is implemented.

[0048] The present invention has the following beneficial effects compared with the prior art:

[0049] The method provided by the present invention deepens the number of layers of information propagation in the graph neural network by adopting graph convolution in a decoupled mode, and can utilize more information in the target dataset to improve the recommendation performance by aggregating information to high-order neighbor nodes; by adopting the neural network architecture search technology, it avoids the repeated parameter tuning process of manually designing the network, and under the guidance of the search space of excellent designs, the architecture search can often find a network structure with better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the structures shown in these drawings without creative efforts.

[0051] Figure 1 It is a flowchart of the recommendation method based on graph neural network architecture search in Embodiment 1 of the present invention.

[0052] Figure 2 It is a flowchart of the particle swarm algorithm in Embodiment 1 of the present invention.

[0053] Figure 3 It is a schematic diagram of particle coding in Embodiment 1 of the present invention.

[0054] Figure 4 It is a schematic diagram of an iterative process in the particle swarm algorithm in Embodiment 1 of the present invention.

[0055] Figure 5 It is a framework diagram of the recommendation system based on graph neural network architecture search in Embodiment 1 of the present invention.

[0056] Figure 6This is the structural framework diagram of the recommendation device based on graph neural network architecture search in Embodiment 2 of the present invention.

[0057] Figure 7 This is the structural framework diagram of the terminal device in Embodiment 3 of the present invention. Detailed implementation manners

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. It should be understood that the specific embodiments described are only used to explain the present application and are not used to limit the present application.

[0059] Embodiment 1:

[0060] As Figure 1 shown, this embodiment provides a recommendation method based on graph neural network architecture search. The applicable scenarios of this recommendation method can be the recommendation of items such as movies, music, and books. The recommendation method includes the following steps:

[0061] S101. Design a corresponding search space for graph neural network architecture search.

[0062] In this embodiment, the search space is decomposed into an information propagation, a feature transformation, and a hyperparameter search space. The specific search space design is as follows:

[0063] Table 1 Information Propagation Search Space

[0064]

[0065] Table 2 Feature Transformation Search Space

[0066]

[0067] Table 3 Hyperparameter Search Space

[0068] Dropout rate 0.1,0.2,0.3,0.4,0.5,0.6,0.7 Learning rate 5e-4, 1e-3, 5e-3, 1e-2, 1e-1 Weight decay rate 0, 5e-4, 8e-4, 1e-3, 4e-3

[0069] S102. Use the particle swarm optimization algorithm as the search strategy for graph neural network architecture search, convert the positions of each particle into corresponding graph neural network models according to the search space, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture.

[0070] Further, as Figure 2 shown, step S102 includes:

[0071] (1) First, initialize the particle swarm, including the positions and velocities of the particles.

[0072] The default initial value of the particle velocity is 0, and the position coordinates of each particle represent the encoding relative to the graph neural network structure. The algorithm randomly initializes the positions of the particles at the beginning. Figure 3 Shows the meanings of each part in the particle encoding, where:

[0073] G1: Represents the type of skip connection of the network structure, and the value range is [0, 4].

[0074] G2: Represents the dimension of the hidden layer, and the value range is [0, 2].

[0075] G3: Represents the type of activation function, and the value range is [0, 7].

[0076] G4: Represents the number of transformation layers, and the value range is [0, 5].

[0077] G5: Represents the number of propagation layers, and the value range is [0, 5].

[0078] G6: Represents the type of aggregation function, and the value range is [0, 3].

[0079] G7: Represents the type of graph normalization method, and the value range is [0, 3].

[0080] G8: Represents the dropout rate, and the value range is [0, 6].

[0081] G9: Represents the learning rate, and the value range is [0, 4].

[0082] G10: Represents the weight decay rate, and the value range is [0, 4].

[0083] (2) Convert the positions of each particle (i.e., the network structure encoding) into the corresponding graph neural network model according to the search space, and train the graph neural network model; use the trained graph neural network to calculate the performance evaluation index.

[0084] Use part of the data in the training dataset to train the graph neural network model; input the data in the test dataset into the trained graph neural network model, compare the obtained recommendation list with the actual preference list of users in the test set, and obtain the performance evaluation index.

[0085] Obtain public datasets on the network as the dataset, such as Amazon - book, Gowalla. Divide the dataset into a training dataset and a test dataset.

[0086] (2 - 1) Train the graph neural network model.

[0087] Generate the corresponding graph neural network architecture according to the position encoding of the particles, and use this architecture to perform TOP-K recommendation on the target dataset.

[0088] Further, step (2-1) specifically includes the following process:

[0089] a) The graph neural network model first randomly initializes the feature representations of users and items, and uses the information propagation operation in the graph convolution operation in the decoupled mode to aggregate the information of neighbor nodes and learn the feature representations of users and items. The information propagation operation is executed once during the initialization process of the model, and only the feature transformation operation is performed during the subsequent gradient training process of the model, so the training efficiency of the model is greatly improved;

[0090] b) According to the feature representations of users and items, calculate the inner product of the user feature vector and the item feature vector, sort the item list from largest to smallest according to the calculation result of the inner product, and take the top K as the final TOP-K recommendation list. Among them, the result of the vector inner product represents the degree of association between the user and the item, and K is a positive integer greater than 0;

[0091] c) The particle swarm algorithm uses the BPR loss function to train the model in the way of gradient descent. Among them is the inner product of the user representation and the positive sample item representation, is the inner product of the user representation and the negative sample item representation. A positive sample means that this sample appears in the user's interest list in the training set.

[0092]

[0093] In order to improve the efficiency of architecture search, part of the data in the training dataset is used to train the graph neural network model in this step.

[0094] (2-2) Use the trained graph neural network model to calculate the performance evaluation index.

[0095] Use the test dataset to predict the trained model, and by comparing the obtained recommendation list with the actual preference list of users in the test set, obtain the ndcg@k index as the performance index for evaluating the network model.

[0096]

[0097] The above formula shows the calculation process of the ndcg@k index, where rel indicates whether the recommendation result is in the user's preference list, and the value is {0, 1}. IDCG is the maximum DCG value in the ideal case, and REL means sorting the list according to the rel value from largest to smallest.

[0098] (3) Select the historical optimal value pbest of each particle and the historical optimal value gbest of the population according to the performance evaluation index, and then update the position and velocity of each particle.

[0099] Use the ndcg@k index as the evaluation of the particle, select the pbest and gbest indexes of the particle swarm in this round of iteration, and then update the position and velocity of each particle.

[0100] Use the ndcg@k index calculated in step (2) as the evaluation of the particle, and select the pbest and gbest indexes of the particle swarm in this round of iteration; according to the μbest and gbest indexes, use the velocity and position update formulas of the particle swarm algorithm to modify the particle information. The formulas are as follows:

[0101] v i+1 = v i + c1 × rand × (pbest i - x i ) + c2 × rand × (gbest - x i )

[0102] x i+1 = x i + v i+1

[0103] To ensure that the position coordinates of the particle after each move are integers, the algorithm makes the following constraints:

[0104]

[0105] Among them, x i is the abbreviation of x i (t), representing the position of the particle at time t in the algorithm, which is the encoding of the actual network structure. v i is the abbreviation of v i (t), representing the movement speed of the particle at time t in the algorithm. v max is the maximum speed of the particle. pbest i represents the historical optimal value of this particle individual, and gbest represents the historical optimal value of the entire particle swarm. c1 and c2 are predefined hyperparameters used to control the search direction of the particle.

[0106] Figure 4 This is a complete iteration process in the particle swarm algorithm. After performing multiple iterations as above, the optimal network structure is obtained. The algorithm first converts the particle encoding into the target neural network structure according to the division of the search space, trains the network model on the training dataset, and then performs Top-K recommendation. The evaluation value of the particle is calculated for the recommendation result.

[0107] (4) Before reaching the maximum number of iterations, steps (2) and (3) are repeatedly executed for multiple rounds of search iteration. According to the idea of the particle swarm algorithm, the particle swarm will randomly move in the entire optimization space and adjust its position in the next round based on the evaluation results after each move. After the iteration ends, the algorithm will select the historical optimal value of the particle swarm as the optimal graph neural network architecture.

[0108] In this embodiment, when the number of iterations reaches the specified number, the search iteration ends.

[0109] S103. Train the optimal graph neural network architecture and generate a recommendation list using the trained graph neural network architecture.

[0110] Use the dataset to fully train the optimal graph neural network architecture so that the trainable parameters in the network fully fit the target dataset to obtain the best recommendation effect.

[0111] To improve the efficiency of architecture search, only part of the data in the dataset is used for model training during the search process in step S102 to speed up, but this will also cause the performance of the trained model to deviate. Therefore, in this step, a complete training is performed using all the data in the dataset.

[0112] As Figure 5 shown, this embodiment also provides a recommendation system based on graph neural network architecture search. The system includes a front-end page, a back-end service, and a database, where:

[0113] The front-end page is mainly used to display the actual recommendation effect of the above method. In this embodiment, music recommendation is used as the actual application scenario. After the user logs in to the system and enters the music recommendation list page, the system backend will return the music recommendation list for this user, and the front-end page receives and displays this data;

[0114] The back-end service is used to generate a music recommendation list according to the information input by the user and run the model obtained by architecture search (the trained graph neural network architecture) and return it to the front-end page. In this embodiment, Node.Js technology is adopted, and the express framework is used to quickly implement a Node.Js Web application. child_process is also a functional package of Node.Js, and the system uses it to call and execute the network model generated by architecture search.

[0115] The database is used to store various information of users and items. In this embodiment, MySQL is used to store data.

[0116] Those skilled in the art can understand that all or part of the steps in the method of implementing the above embodiment can be completed by instructing relevant hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.

[0117] It should be noted that although the method operations of the above embodiments are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the depicted steps can be changed in the order of execution. Additionally or alternatively, certain steps can be omitted, multiple steps can be combined into one step for execution, and / or one step can be decomposed into multiple steps for execution.

[0118] Embodiment 2:

[0119] As Figure 6 shown, this embodiment provides a recommendation device based on graph neural network architecture search. The device includes a search space design module 601, an optimal graph neural network architecture selection module 602, and a recommendation list generation module 603, where:

[0120] The search space design module 601 is used to design a corresponding search space for graph neural network architecture search;

[0121] The optimal graph neural network architecture selection module 602 is used to adopt the particle swarm algorithm as the search strategy for graph neural network architecture search, convert the positions of each particle into corresponding graph neural network models according to the search space, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture; wherein, the position of the particle corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses decoupled graph convolution operations to aggregate neighbor node information;

[0122] The recommendation list generation module 603 is used to train the optimal graph neural network architecture and generate a recommendation list using the trained optimal graph neural network architecture.

[0123] For the specific implementation of each module in this embodiment, reference can be made to Embodiment 1 above, which will not be elaborated here one by one; it should be noted that the device provided in this embodiment is only illustrated by the above division of each functional module. In practical applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0124] Embodiment 3:

[0125] This embodiment provides a terminal device, which can be a computer, such as Figure 7As shown in the figure, it includes a processor 702, a memory, an input device 703, a display 704, and a network interface 705 connected through a system bus 701. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium 706 and an internal memory 707. The non-volatile storage medium 706 stores an operating system, a computer program, and a database. The internal memory 707 provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the processor 702 executes the computer program stored in the memory, the recommended method of the above-mentioned Embodiment 1 is implemented as follows:

[0126] Design a corresponding search space for graph neural network architecture search;

[0127] Adopt the particle swarm algorithm as the search strategy for graph neural network architecture search. Convert the positions of each particle into corresponding graph neural network models according to the search space, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture; among them, the position of the particle corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses decoupled graph convolution operations to aggregate neighbor node information;

[0128] Train the optimal graph neural network architecture, and use the trained optimal graph neural network architecture to generate a recommendation list.

[0129] Embodiment 4:

[0130] This embodiment provides a storage medium, which is a computer-readable storage medium. It stores a computer program. When the computer program is executed by a processor, the recommended method of the above-mentioned Embodiment 1 is implemented as follows:

[0131] Design a corresponding search space for graph neural network architecture search;

[0132] Adopt the particle swarm algorithm as the search strategy for graph neural network architecture search. Convert the positions of each particle into corresponding graph neural network models according to the search space, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture; among them, the position of the particle corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses decoupled graph convolution operations to aggregate neighbor node information;

[0133] Train the optimal graph neural network architecture, and use the trained optimal graph neural network architecture to generate a recommendation list.

[0134] It should be noted that the computer-readable storage medium of this embodiment can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0135] As mentioned above, the above is only a preferred embodiment of the present invention patent, but the protection scope of the present invention patent is not limited thereto. Any person skilled in the art within the scope disclosed by the present invention patent, according to the technical solution and inventive concept of the present invention patent, makes equivalent substitutions or changes, all belong to the protection scope of the present invention patent.

Claims

1. A recommendation method based on graph neural network architecture search, characterized in that The method includes: For the search of graph neural network architectures, decomposing the search space into an information propagation search space, a feature transformation search space, and a hyperparameter search space; Adopting the particle swarm optimization algorithm as the search strategy for graph neural network architecture search, transforming the positions of each particle into corresponding graph neural network models according to the search space, and selecting the historical optimal value of the particle swarm as the optimal graph neural network architecture; wherein, the position of the particle corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses decoupled graph convolutional operations to aggregate neighbor node information; Training the optimal graph neural network architecture, and generating a recommendation list by using the trained optimal graph neural network architecture; Wherein, the transforming the positions of each particle into corresponding graph neural network models according to the search space and selecting the historical optimal value of the particle swarm as the optimal graph neural network architecture includes: S1: Initializing the particle swarm, including the positions and velocities of the particles; S2: Transforming the positions of each particle into corresponding graph neural network architectures according to the search space; training the graph neural network architectures, and calculating performance evaluation metrics by using the trained graph neural network architectures; S3: According to the performance evaluation metrics, selecting the historical optimal value pbest of each particle and the historical optimal value gbest of the population, and then updating the positions and velocities of each particle; Repeatedly execute steps S2 and S3 for multiple rounds of search iteration, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture after the iteration ends.

2. The recommendation method according to claim 1, characterized in that The training of the graph neural network architecture includes: The graph neural network model first randomly initializes the feature representations of users and items, and then uses the information propagation operation in the decoupled graph convolutional operation to aggregate neighbor node information and learn the feature representations of users and items; wherein, the information propagation operation is only executed once during the initialization process of the graph neural network model; According to the feature representations of users and items, calculating the inner product of the user feature vector and the item feature vector; sorting the calculation results of the inner product from large to small, and taking the top K in the corresponding item list as the final recommendation list; wherein, K is a positive integer greater than or equal to 1; The particle swarm optimization algorithm uses the BPR loss function to train the graph neural network model in a gradient descent manner.

3. The recommendation method according to claim 2, characterized in that The calculating of the performance evaluation metrics by using the trained graph neural network architecture includes: Inputting the data in the test dataset into the trained graph neural network architecture to obtain a recommendation list; Calculating performance evaluation metrics according to the recommendation list and the actual preference list of users in the test set.

4. The recommendation method according to claim 1, wherein The selecting the historical optimal value pbest of each particle and the historical optimal value gbest of the population according to the performance evaluation metrics and then updating the positions and velocities of each particle includes: Selecting the historical optimal value pbest of each particle and the historical optimal value gbest of the population according to the performance evaluation metrics; According to pbest and gbest, using the velocity and position update formula of the particle swarm to modify the information of the particle, and the formula is as follows: v i+1 v(t + 1)= i v(t)+c1×rand×(pbest i (t)-x i (t))+c2×rand×(gbest - x i (t)) x i+1 (t + 1)= x i (t)+ v i+1 (t + 1) To ensure that the position coordinates of the particle are integers after each movement, the following constraints are made: where x i represents the position of the particle at time t, corresponding to the encoding of the graph neural network architecture; v i represents the velocity of the particle's motion at time t, and v max is the maximum velocity of the particle; pbest i represents the historical optimal value of the individual particle, and gbest represents the historical optimal value of the entire particle swarm; c1 and c2 are both predefined hyperparameters used to control the search direction of the particle.

5. The recommendation method according to claim 1, wherein To improve the efficiency of architecture search, in step S2, part of the data in the training dataset is used to train the graph neural network architecture; To make the trainable parameters in the graph neural network architecture fully fit the target dataset, and thus obtain the best recommendation effect, all the data in the training dataset is used to train the optimal graph neural network architecture.

6. A recommendation device based on graph neural network architecture search, characterized in that The recommendation device includes: A search space design module, which is used for graph neural network architecture search, and decomposes the search space into an information propagation search space, a feature transformation search space, and a hyperparameter search space; An optimal graph neural network architecture selection module, which uses the particle swarm algorithm as the search strategy for graph neural network architecture search, transforms the positions of each particle into corresponding graph neural network models according to the search space, and selects the historical optimal value of the particle swarm as the optimal graph neural network architecture; among them, the position of the particle corresponds to the encoding of the graph neural network architecture, and the graph neural network architecture uses the decoupled mode of graph convolution operation to aggregate neighbor node information; A recommendation list generation module, which is used to train the optimal graph neural network architecture, and generates a recommendation list by using the trained optimal graph neural network architecture; Among them, the process of transforming the positions of each particle into corresponding graph neural network models according to the search space and selecting the historical optimal value of the particle swarm as the optimal graph neural network architecture includes: S1: Initialize the particle swarm, including the position and velocity of the particles; S2: Transform the positions of each particle into corresponding graph neural network architectures according to the search space; train the graph neural network architecture, and calculate the performance evaluation index by using the trained graph neural network architecture; S3: According to the performance evaluation index, select the historical optimal value pbest of each particle and the historical optimal value gbest of the population, and then update the position and velocity of each particle; Repeat steps S2 and S3 for multiple rounds of search iteration, and select the historical optimal value of the particle swarm as the optimal graph neural network architecture after the iteration ends.

7. A recommendation system based on graph neural network architecture search, characterized in that, The recommendation system includes a front-end page and a back-end service. The front-end page is used to input the interaction information between the user and the item and display the recommendation list returned by the back-end service; the back-end service generates a recommendation list according to the interaction information between the user and the item through the trained optimal graph neural network architecture obtained by running the recommendation method according to any one of claims 1 to 5.

8. A storage medium stores a program, characterized in that, When the program is executed by a processor, it implements the recommendation method according to any one of claims 1 to 5.

Images