Cross-domain recommendation optimization method, device and equipment based on graph neural architecture search

Abstract

The invention provides a cross-domain recommendation optimization method, device and equipment based on graph neural architecture search, relates to the technical field of data processing, and aims to solve the problems of poor adaptability and negative migration of a cross-domain recommendation model so as to improve the cross-domain recommendation performance. The method comprises the following steps: constructing a cross-domain customized super network; according to the cross-domain customized super network, the click rate of a source domain behavior and the click rate of a target domain behavior are predicted, and the click rate of each behavior represents the possibility that a user clicks a commodity; determining an importance weight for the source domain behavior through a behavior importance perceptron, wherein the importance weight represents the contribution degree of the source domain behavior to improvement of the performance of the cross-domain customized super network; determining a first loss according to the importance weight, the click rate of the source domain behavior and the click rate of the target domain behavior; and optimizing the cross-domain customized super network according to the first loss, and taking the cross-domain customized super network meeting a convergence condition as a final cross-domain recommendation model.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to a cross-domain recommendation optimization method, device, and equipment based on graph neural architecture search. Background Art

[0002] With the rapid development of Internet technology, the importance of recommendation systems has become increasingly prominent in fields such as e-commerce and social media. However, data sparsity and cold start problems seriously affect the recommendation effect. Cross-domain recommendation technology has become an effective means to solve these problems by leveraging source domain information to enhance the recommendation effect in the target domain.

[0003] Cross-domain recommendation technology based on graph neural networks (GNNs) has received extensive attention because it can capture complex interaction patterns between users and items. However, existing methods have two major limitations: one is the dependence on manually designed network architectures, which may lead to suboptimal choices and increase design costs; the other is the failure to fully consider the impact of user behavior, that is, the difference in the contribution of different behaviors in the source domain to model optimization, which limits the recommendation performance. Therefore, there is an urgent need for a cross-domain recommendation method that can automatically optimize the network architecture and fully integrate the impact of user behavior to further improve the performance of the recommendation system. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide a cross-domain recommendation optimization method, device, and equipment based on graph neural architecture search to overcome or at least partially solve the above problems.

[0005] In the first aspect of the embodiments of this application, a cross-domain recommendation optimization method based on graph neural architecture search is disclosed. The method includes: Construct a cross-domain customized hypernetwork, where the cross-domain customized hypernetwork includes multiple graph neural network architectures, and the weights of each graph neural network architecture are optimized during training; According to the cross-domain customized hypernetwork, predict the click-through rate of source domain behaviors and the click-through rate of target domain behaviors, where the click-through rate of each behavior represents the likelihood that a user clicks on a product; Determine importance weights for source domain behaviors through a behavior importance perceptron, where the importance weights represent the contribution of source domain behaviors to improving the performance of the cross-domain customized hypernetwork; Determine a first loss according to the importance weights, the click-through rate of the source domain behaviors, and the click-through rate of the target domain behaviors; Optimize the cross-domain customized hypernetwork according to the first loss, and use the cross-domain customized hypernetwork that meets the convergence condition as the final cross-domain recommendation model.

[0006] Optionally, optimizing the cross-domain customized hypernetwork according to the first loss includes: Optimize the cross - domain customized hyper - network a preset number of times based on the first loss; Determine a second loss according to the click - through rate of the target - domain behavior generated by the cross - domain customized hyper - network that has been optimized a preset number of times; Optimize the behavior importance perceptron a preset number of times based on the second loss, where the behavior importance perceptron that has been optimized a preset number of times is used to: determine importance weights for source - domain behaviors during the next preset number of optimization processes of the cross - domain customized hyper - network; Perform multiple rounds of alternating optimization according to the above steps until the cross - domain customized hyper - network meets the convergence condition.

[0007] Optionally, the method further includes: Construct an interaction heterogeneous graph with source - domain commodities as the first nodes, target - domain commodities as the second nodes, users as the third nodes, and the interaction behaviors between users and commodities as labeled edges, where the label represents the interaction type between users and commodities, and the interaction types include click and non - click; Predict the click - through rate of source - domain behaviors and the click - through rate of target - domain behaviors according to the cross - domain customized hyper - network, including: Input the interaction heterogeneous graph into the cross - domain customized hyper - network for click - through rate prediction processing to obtain the click - through rate of source - domain behaviors and the click - through rate of target - domain behaviors.

[0008] Optionally, inputting the interaction heterogeneous graph into the cross - domain customized hyper - network for click - through rate prediction processing to obtain the click - through rate of source - domain behaviors and the click - through rate of target - domain behaviors includes: Map each node in the interaction heterogeneous graph to a feature space to obtain source - domain commodity embedding features, target - domain commodity embedding features, and user embedding features; Perform multi - layer graph convolution and cross - domain information transfer processing on the source - domain commodity embedding features, the target - domain commodity embedding features, and the user embedding features through the cross - domain customized hyper - network to obtain final source - domain commodity embedding features, final target - domain commodity embedding features, and final user embedding features; Generate the click - through rate of the source - domain behavior according to the final source - domain commodity embedding features and the final user embedding features, and generate the click - through rate of the target - domain behavior according to the final target - domain commodity embedding features and the final user embedding features.

[0009] Optionally, determining importance weights for source - domain behaviors through a behavior importance perceptron includes: Determine domain importance weights through the behavior importance perceptron, where the domain importance weights represent the contribution degree of the source domain to improving the performance of the cross - domain customized hyper - network; Determine the global importance weight of the commodity through the behavior importance perception device, where the global importance weight of the commodity represents the contribution degree of the commodity in the source domain behavior to improving the performance of the cross-domain customized hypernetwork; Determine the user-specific importance weight through the behavior importance perception device, where the user-specific importance weight represents the contribution degree of the commodity in the source domain behavior to improving the performance of the cross-domain customized hypernetwork for a specific user in the target domain, and the specific user is the user in the source domain behavior; Obtain the importance weight according to the domain importance weight, the global importance weight of the commodity, and the user-specific importance weight.

[0010] Optionally, obtaining the importance weight according to the domain importance weight, the global importance weight of the commodity, and the user-specific importance weight includes: Normalize the product of the global importance weight of the commodity and the user-specific importance weight, and multiply the normalization result by the domain importance weight to obtain the importance weight.

[0011] Optionally, determining the first loss according to the importance weight, the click-through rate of the source domain behavior, and the click-through rate of the target domain behavior includes: Obtain the target domain loss according to the cross-entropy loss between the click-through rate of the target domain behavior and the label of the target domain behavior; Obtain the source domain loss according to the cross-entropy loss between the click-through rate of the source domain behavior and the label of the source domain behavior, and the importance weight; Obtain the first loss according to the target domain loss and the source domain loss.

[0012] Optionally, optimizing the behavior importance perception device a preset number of times based on the second loss includes: For each optimization, calculate and determine the implicit gradient of the second loss using the chain rule, and optimize the behavior importance perception device based on the implicit gradient.

[0013] In the second aspect of the embodiments of the present application, a cross-domain recommendation optimization device based on graph neural architecture search is disclosed, and the device includes: A network construction module for constructing a cross-domain customized hypernetwork, where the cross-domain customized hypernetwork includes multiple graph neural network architectures, and the weights of each graph neural network architecture are optimized during training; A behavior generation module for predicting the click-through rate of the source domain behavior and the click-through rate of the target domain behavior according to the cross-domain customized hypernetwork, where the click-through rate of each behavior represents the possibility of a user clicking on a commodity; A weight determination module, configured to determine an importance weight for a source domain behavior through a behavior importance perceptron, where the importance weight characterizes the contribution of the source domain behavior to improving the performance of the cross-domain customized hypernetwork; A loss determination module, configured to determine a first loss according to the importance weight, the click-through rate of the source domain behavior, and the click-through rate of the target domain behavior; A model optimization module, configured to optimize the cross-domain customized hypernetwork according to the first loss, and use the cross-domain customized hypernetwork that meets the convergence condition as the final cross-domain recommendation model.

[0014] In a third aspect of the embodiments of the present application, an electronic device is disclosed, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the cross-domain recommendation optimization method based on graph neural architecture search described in the first aspect of the embodiments of the present application are implemented.

[0015] In a fourth aspect of the embodiments of the present application, a computer-readable storage medium is disclosed, on which a computer program is stored. When the computer program is executed by a processor, the steps of the cross-domain recommendation optimization method based on graph neural architecture search described in the first aspect of the embodiments of the present application are implemented.

[0016] In a fifth aspect of the embodiments of the present application, a computer program product is disclosed, including a computer program. When the computer program is executed by a processor, the steps of the cross-domain recommendation optimization method based on graph neural architecture search described in the first aspect of the embodiments of the present application are implemented.

[0017] The embodiments of the present application have the following advantages: In the embodiments of the present application, cross-domain recommendation optimization is performed based on behavior importance perception and graph neural architecture search. By constructing a cross-domain customized hypernetwork, which includes multiple graph neural network architectures and the weights of each graph neural network architecture are optimized during training, the optimal graph neural network architecture can be found in one search based on this cross-domain customized hypernetwork, avoiding the need for repeated training. For source domain behaviors, the importance of each source domain behavior is dynamically evaluated through a behavior importance perceptron to determine an importance weight for the source domain behavior, and a first loss for optimizing the cross-domain customized hypernetwork is determined based on the importance weight to guide the optimization of the cross-domain customized hypernetwork, thereby improving the recommendation quality of the target domain. In this way, the method jointly optimizes the graph neural network architecture and data importance, automatically searches for the optimal architecture, and dynamically evaluates the impact of source domain behaviors, thereby effectively alleviating the problems of poor adaptability and negative transfer of the cross-domain recommendation model and significantly improving the cross-domain recommendation performance. Description of the Drawings

[0018] To more clearly illustrate the technical solutions of the embodiments of the present application, the accompanying drawings required for the description of the embodiments of the present application will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a flowchart of the steps of a cross-domain recommendation optimization method based on graph neural architecture search provided by an embodiment of the present application; Figure 2 It is an architecture diagram of a cross-domain recommendation optimization method based on graph neural architecture search provided by an embodiment of the present application; Figure 3 It is a schematic structural diagram of a cross-domain recommendation optimization device based on graph neural architecture search provided by an embodiment of the present application; Figure 4 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0020] To make the above objects, features, and advantages of the present application more obvious and understandable, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present application belong to the scope of protection of the present application.

[0021] Existing cross - domain recommendation methods usually rely on manually designed neural network architectures, as follows: One type of method uses matrix factorization to integrate features between different domains. However, due to the limited ability of this type of method in modeling complex user preferences, its performance improvement space is also limited. With the progress of deep learning technology, some cross - domain recommendation models achieve interaction between cross - domain data by sharing projection matrices, thus significantly improving the performance of the recommendation system. However, most of these deep - learning - based methods focus on using known user - item interaction information and are not good at capturing higher - order and implicit relationship patterns. The emergence of graph neural networks (GNNs) provides a new way to overcome the above limitations. For example, some recommendation methods construct heterogeneous graphs and combine global and domain - specific sub - graphs for click - through rate prediction, improving the model's ability to understand complex relationships. Another type of recommendation method proposes variational bipartite graph encoders and mutual information regularizers to achieve the decoupling of domain - shared and domain - specific user representations, enhancing the flexibility and adaptability of the model. In addition, some recommendation methods optimize the cross - domain recommendation effect by introducing decoupled embeddings and domain alignment strategies, enabling the model to better transfer useful information between different domains.

[0022] Existing cross - domain recommendation methods have the following problems: On the one hand, traditional cross - domain recommendation methods based on user representations have limited expressive power and are difficult to fully model user preferences and high - order interaction patterns, resulting in limited recommendation performance. On the other hand, although models that combine user representations and neural network structures for knowledge transfer improve the cross - domain information modeling ability by introducing graph neural networks, these methods usually rely on fixed manually designed architectures, have poor adaptability, and are vulnerable to the negative transfer problem. Especially when the source - domain data is sparse, the attention mechanism may over - emphasize noise features, leading to sub - optimal recommendation effects. In addition, existing methods often only use the simple sum of source loss and target loss when optimizing the model, ignoring the impact differences of different behaviors in the source domain on the recommendation effect, which further limits the performance of the model.

[0023] To overcome the limitations of related technologies, the embodiments of this application provide a cross - domain recommendation optimization method based on graph neural architecture search. This method optimizes cross - domain recommendations based on behavior importance perception and graph neural architecture search. It automatically searches for the optimal graph neural network architecture through a cross - domain customized hyper - network, dynamically evaluates the impact of source - domain behaviors on the target - domain recommendation effect through a behavior importance perceptron, and assigns importance weights to each source - domain behavior to guide model optimization. In this way, by jointly optimizing the graph neural network architecture and data importance, automatically searching for the optimal architecture and dynamically evaluating the impact of source - domain behaviors, it effectively alleviates the problems of poor model adaptability and negative transfer, and significantly improves cross - domain recommendation performance.

[0024] The following will describe the cross - domain recommendation optimization method based on graph neural architecture search in the embodiments of the present application with reference to the accompanying drawings.

[0025] Referring to Figure 1 as shown, Figure 1 is a flowchart of the steps of a cross - domain recommendation optimization method based on graph neural architecture search provided by an embodiment of the present application. As Figure 1 shown, the cross - domain recommendation optimization method based on graph neural architecture search may include steps S110 to S150: Step S110: Construct a cross - domain customized hyper - network, where the cross - domain customized hyper - network includes multiple graph neural network architectures, and the weights of each graph neural network architecture are optimized during the training process.

[0026] Among them, the multiple graph neural network architectures refer to all possible graph neural network architectures. Each graph neural network architecture is used to process commodity - user pairs in different domains, that is, to predict the possibility that a user clicks on a commodity. Each graph neural network architecture is composed of multiple operators, each operator has a weight, and the weight of each graph neural network architecture is determined according to the weights of the multiple operators it includes.

[0027] During the optimization process (training process) of the cross - domain customized hyper - network, the weights of each operator can be optimized through gradient descent to achieve the optimization of the weights of each graph neural network architecture. Therefore, based on this cross - domain customized hyper - network, the optimal graph neural network architecture can be found in one search, avoiding the need for repeated training.

[0028] Step S120: According to the cross - domain customized hyper - network, predict the click - through rate of source - domain behaviors and the click - through rate of target - domain behaviors. The click - through rate of each behavior characterizes the possibility that a user clicks on a commodity.

[0029] In the embodiments of the present application, each behavior represents a commodity - user pair. That is, the source - domain behavior refers to the source - domain commodity - user pair, and the target - domain behavior refers to the target - domain commodity - user pair. A user clicking on a commodity refers to interactive behaviors such as a user's purchase or rating of the commodity. For each source - domain behavior and each target - domain behavior, the cross - domain customized hyper - network can predict the click - through rate of this behavior. Among them, the click - through rate is a number between 0 and 1. In some embodiments, a click - through rate threshold can be set. When the click - through rate of a behavior is greater than the click - through rate threshold, it indicates that the interaction type between the user and the commodity is a click. When the click - through rate is not greater than the click - through rate threshold, it indicates that the interaction type between the user and the commodity is not a click.

[0030] In a specific embodiment, before performing step S120, the following steps are further included: taking the source domain product as the first node, the target domain product as the second node, the user as the third node, and taking the interaction behavior between the user and the product as a labeled edge, constructing an interaction heterogeneous graph, where the label represents the interaction type between the user and the product, and the interaction type includes click and non - click.

[0031] In the embodiments of the present application, the source domain behavior and the target domain behavior are constructed based on the existing user behavior data. The existing user - product interaction data is stored in the form of single - interaction records. According to the existing user - product interaction data, it is organized into a heterogeneous graph structure, where the user and the product are taken as nodes, and the interaction behavior between the user and the product is taken as a labeled edge.

[0032] For each source domain behavior in the interaction heterogeneous graph, the label of the corresponding edge is the label of the source domain behavior. If the interaction type between the user and the product is click, the label of this source domain behavior is click (the label can be represented as 1). If the interaction type between the user and the product is non - click, the label of this source domain behavior is non - click (the label can be represented as 0). Similarly, for each target domain behavior in the interaction heterogeneous graph, the label of the corresponding edge is the label of the target domain behavior. If the interaction type between the user and the product is click, the label of this target domain behavior is click (the label can be represented as 1). If the interaction type between the user and the product is non - click, the label of this target domain behavior is non - click (the label can be represented as 0).

[0033] It should be noted that for different source domains and different target domains, different interaction heterogeneous graphs can be constructed respectively. For example, for source domain 1, target domain 1, and target domain 2, based on the interaction data between the user and the products in source domain 1, and the interaction data between the user and the products in target domain 1, an interaction heterogeneous Figure 1 graph can be constructed; based on the interaction data between the user and the products in source domain 1, and the interaction data between the user and the products in target domain 2, an interaction heterogeneous Figure 2 graph can be constructed.

[0034] Further, in step S120, "predicting the click - through rate of the source domain behavior and the click - through rate of the target domain behavior according to the cross - domain customized hyper - network" specifically includes: inputting the interaction heterogeneous graph into the cross - domain customized hyper - network for click - through rate prediction processing to obtain the click - through rate of the source domain behavior and the click - through rate of the target domain behavior.

[0035] In the embodiments of the present application, the interaction heterogeneous graph includes the source domain product, the target domain product, and the user, as well as the interaction behavior between the user and the product. The cross - domain customized hyper - network can predict the click - through rate of the behavior (source domain behavior or target domain behavior) composed of the user node and the product node by learning the interaction behavior.

[0036] Specifically, input the interactive heterogeneous graph into the cross-domain customized hypernetwork for click-through rate prediction processing to obtain the click-through rate of source-domain behaviors and the click-through rate of target-domain behaviors, including sub-steps A1 to A3: Step A1: Map each node in the interactive heterogeneous graph to a feature space to obtain source-domain commodity embedding features, target-domain commodity embedding features, and user embedding features.

[0037] Among them, in order to enable the cross-domain customized hypernetwork to perform calculations, each node in the interactive heterogeneous graph is mapped to a feature space, that is, the first node is mapped to source-domain commodity embedding features, the second node is mapped to target-domain commodity embedding features, and the third node is mapped to user embedding features.

[0038] Step A2: Through the cross-domain customized hypernetwork, perform multi-layer graph convolution and cross-domain information transfer processing on the source-domain commodity embedding features, the target-domain commodity embedding features, and the user embedding features to obtain final source-domain commodity embedding features, final target-domain commodity embedding features, and final user embedding features.

[0039] For each embedding feature, after the cross-domain customized hypernetwork performs multi-layer graph convolution and cross-domain information transfer processing, multi-layer embedding features are obtained respectively. After splicing the multi-layer embedding features, the final embedding feature is obtained. For example, for a target-domain commodity embedding feature, after multi-layer graph convolution and cross-domain information transfer processing of the cross-domain customized hypernetwork, multi-layer source-domain commodity embedding features are obtained, and the multi-layer source-domain commodity embedding features are spliced to obtain the final source-domain commodity embedding feature.

[0040] Step A3: Generate the click-through rate of the source-domain behavior based on the final source-domain commodity embedding features and the final user embedding features, and generate the click-through rate of the target-domain behavior based on the final target-domain commodity embedding features and the final user embedding features.

[0041] Specifically, the similarity between the final source-domain commodity embedding features and the final user embedding features is used as the click-through rate of the source-domain behavior, and the similarity between the final target-domain commodity embedding features and the final user embedding features is used as the click-through rate of the target-domain behavior. Among them, the similarity can be the cosine similarity between the two final embedding features, or the dot product between the two final embedding features (the larger the dot product, the more similar the two final embedding features).

[0042] In this way, the cross-domain customized hypernetwork predicts the click-through rate of source-domain behaviors and the click-through rate of target-domain behaviors through the constructed interactive heterogeneous graph. The interactive heterogeneous graph contains source-domain commodities, target-domain commodities, and the interaction behaviors between users and commodities, and ensures the accuracy of behaviors by making full use of graph structure information and cross-domain information.

[0043] Step S130: Determine the importance weight for the source domain behavior through the behavior importance perceptron. The importance weight represents the contribution degree of the source domain behavior to improving the performance of the cross-domain customized hypernetwork.

[0044] In the embodiments of the present application, the behavior importance perceptron dynamically evaluates the importance of the source domain behavior through auxiliary learning, and determines an importance weight for each source domain behavior to guide the optimization of the model (cross-domain customized hypernetwork). When determining the importance weight for the source domain behavior through the behavior importance perceptron, two aspects can be considered, that is, an overall weight can be determined for the source domain, and a corresponding weight can be determined for each source domain behavior. Based on the overall weight and the source domain behavior, the corresponding weight is determined to obtain the importance weight of the source domain behavior.

[0045] In a specific implementation manner, "determine the importance weight for the source domain behavior through the behavior importance perceptron" in step S130 specifically includes sub-steps S130-1 to step S130-4: Step S130-1: Determine the domain importance weight through the behavior importance perceptron. The domain importance weight represents the contribution degree of the source domain to improving the performance of the cross-domain customized hypernetwork.

[0046] Among them, the domain importance weight refers to the overall weight of the source domain, and the domain importance weights corresponding to different source domains may be different. For example, for source domain 1 and source domain 2, the behavior importance perceptron can determine the respective domain importance weights for source domain 1 and source domain 2. In some embodiments, for different source domains, the domain importance weight can be comprehensively obtained by combining the commodity information of the source domain and all interaction data of the user in the source domain.

[0047] Step S130-2: Determine the global commodity importance weight through the behavior importance perceptron. The global commodity importance weight represents the contribution degree of the commodity in the source domain behavior to improving the performance of the cross-domain customized hypernetwork.

[0048] Among them, the behavior importance perceptron includes a multi-layer perceptron (MLP) inside. Determining the global commodity importance weight through the behavior importance perceptron can be obtained by processing the source domain commodity embedding representation through the multi-layer perceptron. It can be understood that the source domain commodity embedding representation mentioned here refers to: the embedding representation corresponding to the commodity in a certain source domain behavior.

[0049] Step S130-3: Determine the user-specific importance weight through the behavior importance perceptron. The user-specific importance weight represents the contribution degree of the commodity in the source domain behavior to improving the performance of the cross-domain customized hypernetwork for a specific user in the target domain. The specific user is the user in the source domain behavior.

[0050] Among them, the behavior importance perceptron internally includes a Graph Neural Network (GraphSAGE). The user-specific importance weight can be determined through the behavior importance perceptron, which can be obtained by processing the interaction heterogeneous graph through the graph neural network.

[0051] Step S130-4: Obtain the importance weight according to the domain importance weight, the global commodity importance weight, and the user-specific importance weight.

[0052] Among them, the importance weight can be the product of the domain importance weight, the global commodity importance weight, and the user-specific importance weight.

[0053] In this way, the importance of the source domain behavior is evaluated from three perspectives: domain, commodity, and user, and the importance weight of each source domain behavior is obtained to guide the optimization of the cross-domain customized hypernetwork through this importance weight.

[0054] Furthermore, obtaining the importance weight according to the domain importance weight, the global commodity importance weight, and the user-specific importance weight includes: normalizing the product of the global commodity importance weight and the user-specific importance weight, and multiplying the result of the normalization process by the domain importance weight to obtain the importance weight.

[0055] Exemplarily, the importance weight can be expressed as:

[0056] Among them, represents the importance weight, represents the domain importance weight, represents the global commodity importance weight, represents the user-specific importance weight, represents the normalization function.

[0057] In the embodiments of the present application, since the global commodity importance weight and the user-specific importance weight are directly related to the behavior, while the domain importance weight is the overall weight of the source domain itself, the weights related to the behavior are normalized to reflect the differences of all source domain behaviors, and then the importance weight of each source domain behavior is accurately determined.

[0058] Step S140: Determine the first loss according to the importance weight, the click-through rate of the source domain behavior, and the click-through rate of the target domain behavior.

[0059] In the embodiments of the present application, after obtaining the importance weights of the source domain behaviors, the first loss is determined according to the importance weights, the click-through rates of the source domain behaviors, and the click-through rates of the target domain behaviors, so as to guide the optimization of the cross-domain customized hypernetwork through the importance weights. The information of different source domain behaviors is introduced in the optimization of the cross-domain customized hypernetwork, enabling the cross-domain customized hypernetwork to learn the source domain behaviors with high importance, thereby improving the recommendation quality of the target domain.

[0060] In a specific implementation manner, determining the first loss according to the importance weights, the click-through rates of the source domain behaviors, and the click-through rates of the target domain behaviors includes: obtaining the target domain loss according to the cross-entropy loss between the click-through rate of the target domain behavior and the label of the target domain behavior; obtaining the source domain loss according to the cross-entropy loss between the click-through rate of the source domain behavior and the label of the source domain behavior, and the importance weights; and obtaining the first loss according to the target domain loss and the source domain loss.

[0061] Exemplarily, the first loss is expressed as:

[0062] where is the binary cross-entropy loss function; represents the click-through rate of the target domain behavior; represents the label of the target domain behavior. The label of the target domain behavior is 1 or 0, where 1 indicates that the interaction behavior type between the user and the target domain commodity is a click, and 0 indicates that the interaction behavior type between the user and the target domain commodity is not a click; represents the click-through rate of the source domain behavior, represents the label of the source domain behavior. The label of the source domain behavior is 1 or 0, where 1 indicates that the interaction behavior type between the user and the source domain commodity is a click, and 0 indicates that the interaction behavior type between the user and the source domain commodity is not a click.

[0063] In this way, the importance weights are used as the weights of the source domain loss in the first loss, enabling the cross-domain customized hypernetwork to learn the source domain behaviors with high importance, thereby improving the recommendation quality of the target domain.

[0064] Step S150: Optimize the cross-domain customized hypernetwork according to the first loss, and use the cross-domain customized hypernetwork that meets the convergence condition as the final cross-domain recommendation model.

[0065] In the embodiments of the present application, the cross-domain customized hypernetwork is optimized with the goal of minimizing the first loss. Specifically, the gradient descent method can be used to update the parameters of the cross-domain customized hypernetwork (including updating the weights of each operator in the cross-domain customized hypernetwork) until the cross-domain customized hypernetwork converges, and the cross-domain customized hypernetwork that meets the convergence condition is used as the final cross-domain recommendation model; where the convergence condition can be that the first loss value is less than a preset threshold, or the number of training rounds reaches a preset upper limit.

[0066] In a specific embodiment, in order to jointly optimize the cross-domain customized hypernetwork and the behavior importance perceptron, a two-layer optimization algorithm is proposed to alternately perform bottom-layer optimization and top-layer optimization on the cross-domain customized hypernetwork and the behavior importance perceptron. Specifically, "optimizing the cross-domain customized hypernetwork according to the first loss" in the above step S150 includes the following steps S150-1 to step S150-4: Step S150-1: Optimize the cross-domain customized hypernetwork a preset number of times based on the first loss.

[0067] Step S150-2: Determine the second loss according to the click-through rate of the target domain behavior generated by the cross-domain customized hypernetwork that has been optimized a preset number of times.

[0068] Step S150-3: Optimize the behavior importance perceptron a preset number of times based on the second loss, where the behavior importance perceptron that has been optimized a preset number of times is used to: determine the importance weight for each source domain behavior in the next preset number of optimization processes of the cross-domain customized hypernetwork.

[0069] Step S150-4: Perform multiple rounds of alternating optimization according to the above steps until the cross-domain customized hypernetwork meets the convergence condition.

[0070] In the embodiments of the present application, first perform the bottom-layer optimization of step S150-1, fixing the parameters of the behavior perception module and optimize the cross-domain customized hypernetwork a preset number of times with the goal of minimizing the first loss. After optimizing the cross-domain customized hypernetwork a preset number of times, perform the top-layer optimization of steps S150-2 to S150-3, fixing the parameters of the cross-domain customized hypernetwork and optimize the behavior importance perceptron a preset number of times with the goal of minimizing the second loss.

[0071] Among them, the second loss can be expressed as:

[0072] Among them, Denotes the click-through rate of the target domain behavior generated by the cross-domain customized hypernetwork after a preset number of optimizations. Characterizes the label of the target domain behavior. In this way, by constructing the second loss based on the target domain behavior, the behavior importance perceptron can be enabled to be based on the learned source domain behavior, making the cross-domain customized hypernetwork perform better in the target domain.

[0073] Specifically, performing a preset number of optimizations on the behavior importance perceptron based on the second loss includes: for each optimization, calculating and determining the implicit gradient of the second loss using the chain rule, and optimizing the behavior importance perceptron based on the implicit gradient.

[0074] In the embodiments of the present application, since depends on the parameters of the cross-domain customized hypernetwork and in turn depends on the parameters of the behavior perception module , so the implicit gradient is used to update :

[0075] where represents the implicit gradient of the second loss, which is calculated by the chain rule and uses the K-truncated Neumann series to approximate the Hessian inverse matrix.

[0076] In this way, by alternately performing bottom-layer optimization and top-layer optimization on the cross-domain customized hypernetwork and the behavior importance perceptron according to the above steps, the joint optimization of the graph neural network architecture and the user behavior importance is achieved.

[0077] As Figure 2 shown, Figure 2 is an architecture diagram of a cross-domain recommendation optimization method based on graph neural architecture search provided by the embodiments of the present application. The specific implementation process of the cross-domain recommendation optimization method based on graph neural architecture search is as follows: taking the source domain commodities as the first nodes, the target domain commodities as the second nodes, the users as the third nodes, and taking the interaction behaviors between the users and the commodities as the labeled edges, constructing an interaction heterogeneous graph, where the label characterizes the interaction type between the user and the commodity, and the interaction type includes click and non-click.

[0078] Furthermore, map each node in the interactive heterogeneous graph to the representation space to obtain the embedded representations (i.e., the source domain commodity embedded representation, the target domain commodity embedded representation, and the user embedded representation). Perform multi-layer graph convolution and cross-domain information transfer processing on each embedded representation through the cross-domain customized hypernetwork to obtain the final representations (i.e., the final source domain commodity embedded representation, the final target domain commodity embedded representation, and the final user embedded representation). And generate the click-through rate of the source domain behavior according to the final source domain commodity embedded representation and the final user embedded representation, and generate the click-through rate of the target domain behavior according to the final target domain commodity embedded representation and the final user embedded representation.

[0079] Next, determine the importance weight for the source domain behavior through the behavior importance perceptron, and determine the first loss according to the importance weight, the click-through rate of the source domain behavior, and the click-through rate of the target domain behavior; perform a preset number of optimizations on the cross-domain customized hypernetwork based on the first loss; after performing a preset number of optimizations on the cross-domain customized hypernetwork, determine the second loss according to the click-through rate of the target domain behavior generated by the cross-domain customized hypernetwork after a preset number of optimizations, and perform a preset number of optimizations on the behavior importance perceptron based on the second loss. In this way, perform multiple rounds of alternating optimizations according to the steps until the cross-domain customized hypernetwork meets the convergence condition, and use the cross-domain customized hypernetwork that meets the convergence condition as the final cross-domain recommendation model.

[0080] Adopt the technical solution of the embodiment of the present application to optimize cross-domain recommendation based on behavior importance perception and graph neural architecture search. By constructing a cross-domain customized hypernetwork, the cross-domain customized hypernetwork includes multiple graph neural network architectures, and the weights of each graph neural network architecture are optimized during the training process. Therefore, based on this cross-domain customized hypernetwork, the optimal graph neural network architecture can be found in one search, avoiding the need for repeated training. For the source domain behavior, dynamically evaluate the importance of each source domain behavior through the behavior importance perceptron to determine an importance weight for the source domain behavior, and determine the first loss used to optimize the cross-domain customized hypernetwork based on the importance weight to guide the optimization of the cross-domain customized hypernetwork, thereby improving the recommendation quality of the target domain. In this way, this method effectively alleviates the problems of poor adaptability and negative transfer of the cross-domain recommendation model by jointly optimizing the graph neural network architecture and data importance, and significantly improves the cross-domain recommendation performance by automatically searching for the optimal architecture and dynamically evaluating the impact of the source domain behavior.

[0081] In addition, the method is made sufficiently generalizable through importance perception and graph neural architecture search, enabling it to automatically adapt to different cross-domain recommendation models and datasets. By jointly optimizing the graph neural network architecture and behavior importance, the method performs excellently in multiple cross-domain recommendation tasks, significantly outperforming existing state-of-the-art methods. Compared with existing cross-domain recommendation methods, since existing cross-domain recommendation methods mainly use manually designed neural network architectures and suffer from the problem of insufficient architecture flexibility, while this solution can adaptively generate efficient recommendation models for different domain characteristics, improving the performance of cross-domain recommendation.

[0082] Based on the same inventive concept, an embodiment of the present application further provides a cross-domain recommendation optimization device based on graph neural architecture search. Referring to Figure 3 as shown in Figure 3 FIG. is a schematic structural diagram of a cross-domain recommendation optimization device based on graph neural architecture search provided by an embodiment of the present application. The device includes: A network construction module 310, configured to construct a cross-domain customized hypernetwork, where the cross-domain customized hypernetwork includes multiple graph neural network architectures, and the weights of each graph neural network architecture are optimized during the training process; A behavior generation module 320, configured to predict the click-through rate of source domain behaviors and the click-through rate of target domain behaviors according to the cross-domain customized hypernetwork, where the click-through rate of each behavior represents the likelihood of a user clicking on a product; A weight determination module 330, configured to determine importance weights for source domain behaviors through a behavior importance perceptron, where the importance weights represent the contribution degree of source domain behaviors to improving the performance of the cross-domain customized hypernetwork; A loss determination module 340, configured to determine a first loss according to the importance weights, the click-through rate of the source domain behaviors, and the click-through rate of the target domain behaviors; A model optimization module 350, configured to optimize the cross-domain customized hypernetwork according to the first loss, and use the cross-domain customized hypernetwork that meets the convergence condition as the final cross-domain recommendation model.

[0083] In some optional embodiments, the model optimization module is specifically configured to: perform a preset number of optimizations on the cross-domain customized hypernetwork based on the first loss; determine a second loss according to the click-through rate of the target domain behaviors generated by the cross-domain customized hypernetwork that has undergone a preset number of optimizations; perform a preset number of optimizations on the behavior importance perceptron based on the second loss, where the behavior importance perceptron that has undergone a preset number of optimizations is used to: determine importance weights for each source domain behavior during the next preset number of optimizations of the cross-domain customized hypernetwork; perform multiple rounds of alternating optimizations according to the above steps until the cross-domain customized hypernetwork meets the convergence condition.

[0084] In some alternative embodiments, it further includes: An interactive heterogeneous graph construction module, configured to construct an interactive heterogeneous graph with source domain commodities as the first nodes, target domain commodities as the second nodes, users as the third nodes, and the interaction behaviors between users and commodities as labeled edges, where the labels represent the interaction types between users and commodities, and the interaction types include clicks and non-clicks; The behavior generation module is specifically configured to: input the interactive heterogeneous graph into the cross-domain customized hypernetwork for click-through rate prediction processing to obtain the click-through rate of source domain behaviors and the click-through rate of target domain behaviors.

[0085] In some alternative embodiments, the behavior generation module includes: A feature mapping module, configured to map each node in the interactive heterogeneous graph to a feature space to obtain source domain commodity embedded features, target domain commodity embedded features, and user embedded features; A feature processing module, configured to perform multi-layer graph convolution and cross-domain information transfer processing on the source domain commodity embedded features, the target domain commodity embedded features, and the user embedded features through the cross-domain customized hypernetwork to obtain final source domain commodity embedded features, final target domain commodity embedded features, and final user embedded features; A generation sub-module, configured to generate the click-through rate of the source domain behavior according to the final source domain commodity embedded features and the final user embedded features, and generate the click-through rate of the target domain behavior according to the final target domain commodity embedded features and the final user embedded features.

[0086] In some alternative embodiments, the weight determination module includes: A first determination sub-module, configured to determine a domain importance weight through the behavior importance perceptron, where the domain importance weight represents the contribution degree of the source domain to improving the performance of the cross-domain customized hypernetwork; A second determination sub-module, configured to determine a commodity global importance weight through the behavior importance perceptron, where the commodity global importance weight represents the contribution degree of commodities in the source domain behavior to improving the performance of the cross-domain customized hypernetwork; A third determination sub-module, configured to determine a user-specific importance weight through the behavior importance perceptron, where the user-specific importance weight represents the contribution degree of commodities in the source domain behavior to improving the performance of the cross-domain customized hypernetwork for a specific user in the target domain, and the specific user is the user in the source domain behavior; A fourth determination sub-module, configured to obtain an importance weight according to the domain importance weight, the commodity global importance weight, and the user-specific importance weight.

[0087] In some alternative embodiments, the fourth determination sub-module is specifically configured to: normalize the product of the global importance weight of the commodity and the user-specific importance weight, and multiply the result of the normalization process by the domain importance weight to obtain the importance weight.

[0088] In some alternative embodiments, the loss determination module is specifically configured to: obtain a target domain loss according to the cross-entropy loss between the click-through rate of the target domain behavior and the label of the target domain behavior; obtain a source domain loss according to the cross-entropy loss between the click-through rate of the source domain behavior and the label of the source domain behavior, and the importance weight; and obtain the first loss according to the target domain loss and the source domain loss.

[0089] An embodiment of the present application further provides an electronic device. Refer to Figure 4 , Figure 4 FIG. is a schematic structural diagram of an electronic device provided by an embodiment of the present application. As Figure 4 shown, the electronic device 400 includes: a memory 410 and a processor 420. The memory 410 is communicatively connected to the processor 420 through a bus. A computer program is stored in the memory 410, and the computer program can run on the processor 420, thereby implementing the steps of the cross-domain recommendation optimization method based on graph neural architecture search described in the embodiments of the present application.

[0090] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the cross-domain recommendation optimization method based on graph neural architecture search described in the embodiments of the present application are implemented.

[0091] An embodiment of the present application further provides a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the cross-domain recommendation optimization method based on graph neural architecture search described in the embodiments of the present application are implemented.

[0092] Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other.

[0093] Embodiments of the present application are described with reference to the flowcharts and / or block diagrams of methods and devices according to embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal devices generate a device for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0094] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to generate a computer-implemented process. Therefore, the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or multiple blocks.

[0096] Although the preferred embodiments of the embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they know the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present application.

[0097] Finally, it should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or terminal device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or terminal device comprising the said element.

[0098] The above has introduced in detail a cross-domain recommendation optimization method, device and equipment provided by the present application. Specific examples are used in this text to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A cross-domain recommendation optimization method based on graph neural architecture search, characterized in that: The method comprises: Constructing a cross-domain customized super network, wherein the cross-domain customized super network includes multiple graph neural network architectures, and the weight of each graph neural network architecture is optimized during the training process; Predicting the click-through rate of source domain behaviors and target domain behaviors according to the cross-domain customized hypernetwork, wherein the click-through rate of each behavior represents the possibility of a user clicking on a product; Determining an importance weight for a source domain behavior through a behavior importance sensor, wherein the importance weight represents a contribution of the source domain behavior to improving the performance of the cross-domain customized hypernetwork; determining a first loss according to the importance weight, the click-through rate of the source domain behavior, and the click-through rate of the target domain behavior; The cross-domain customized super network is optimized according to the first loss, and the cross-domain customized super network that meets the convergence condition is used as the final cross-domain recommendation model.

2. The method according to claim 1, characterized in that Optimizing the cross-domain customized super network according to the first loss includes: Optimizing the cross-domain customized super network a preset number of times based on the first loss; determining a second loss according to a click-through rate of the target domain behavior generated by the cross-domain customized super network optimized a preset number of times; Optimizing the behavior importance sensor a preset number of times based on the second loss, wherein the behavior importance sensor optimized a preset number of times is used to: determine the importance weight for the source domain behavior in the next round of optimization of the cross-domain customized super network a preset number of times; According to the above steps, multiple rounds of alternating optimization are performed until the cross-domain customized super network meets the convergence condition.

3. The method according to claim 1, characterized in that The method further comprises: The source domain product is used as the first node, the target domain product is used as the second node, the user is used as the third node, and the interaction behavior between the user and the product is used as the edge with a label to construct an interaction heterogeneous graph, wherein the label represents the interaction type between the user and the product, and the interaction type includes click and no click; According to the cross-domain customized hypernetwork, predicting the click-through rate of the source domain behavior and the click-through rate of the target domain behavior includes: The interaction heterogeneous graph is input into the cross-domain customized super network for click-through rate prediction processing to obtain the click-through rate of the source domain behavior and the click-through rate of the target domain behavior.

4. The method according to claim 3, characterized in that Inputting the interaction heterogeneous graph into the cross-domain customized hypernetwork to perform click-through rate prediction processing to obtain the click-through rate of the source domain behavior and the click-through rate of the target domain behavior, including: Mapping each node in the interactive heterogeneous graph to a representation space to obtain a source domain product embedding representation, a target domain product embedding representation, and a user embedding representation; Performing multi-layer graph convolution and cross-domain information transfer processing on the source domain product embedding representation, the target domain product embedding representation, and the user embedding representation through the cross-domain customized super network to obtain a final source domain product embedding representation, a final target domain product embedding representation, and a final user embedding representation; The click rate of the source domain behavior is generated according to the final source domain product embedding representation and the final user embedding representation, and the click rate of the target domain behavior is generated according to the final target domain product embedding representation and the final user embedding representation.

5. The method according to claim 1, characterized in that The importance weight of the source domain behavior is determined by the behavior importance sensor, including: Determining a domain importance weight by the behavior importance sensor, wherein the domain importance weight represents a contribution of a source domain to improving the performance of the cross-domain customized supernetwork; Determining a global importance weight of a product through the behavior importance sensor, wherein the global importance weight of the product represents a contribution of the product in the source domain behavior to improving the performance of the cross-domain customized super network; Determining a user-specific importance weight through the behavior importance sensor, wherein the user-specific importance weight represents a contribution of the commodity in the source domain behavior to improving the performance of the cross-domain customized supernetwork for a specific user in the target domain, wherein the specific user is a user in the source domain behavior; An importance weight is obtained according to the domain importance weight, the commodity global importance weight and the user specific importance weight.

6. The method according to claim 5, characterized in that Obtaining importance weights according to the domain importance weight, the commodity global importance weight, and the user-specific importance weight includes: The product of the global importance weight of the commodity and the user-specific importance weight is normalized, and the normalized result is multiplied by the domain importance weight to obtain the importance weight.

7. The method according to claim 1, characterized in that Determining a first loss according to the importance weight, the click rate of the source domain behavior, and the click rate of the target domain behavior includes: Obtaining a target domain loss according to a cross entropy loss between a click rate of the target domain behavior and a label of the target domain behavior; Obtaining a source domain loss according to a cross entropy loss between a click rate of the source domain behavior and a label of the source domain behavior, and the importance weight; The first loss is obtained according to the target domain loss and the source domain loss.

8. The method according to claim 2, characterized in that: Optimizing the behavior importance sensor a preset number of times based on the second loss includes: For each optimization, the implicit gradient of the second loss is calculated and determined using the chain rule, and the behavior importance perceptron is optimized based on the implicit gradient.

9. A cross-domain recommendation optimization device based on graph neural architecture search, characterized in that: The device comprises: A network construction module, used to construct a cross-domain customized super network, wherein the cross-domain customized super network includes multiple graph neural network architectures, and the weight of each graph neural network architecture is optimized during the training process; A behavior generation module, used to predict the click-through rate of source domain behavior and the click-through rate of target domain behavior according to the cross-domain customized super network, wherein the click-through rate of each behavior represents the possibility of a user clicking on a product; A weight determination module, used to determine an importance weight for a source domain behavior through a behavior importance sensor, wherein the importance weight represents a contribution of the source domain behavior to improving the performance of the cross-domain customized super network; a loss determination module, configured to determine a first loss according to the importance weight, the click-through rate of the source domain behavior, and the click-through rate of the target domain behavior; A model optimization module is used to optimize the cross-domain customized super network according to the first loss, and use the cross-domain customized super network that meets the convergence condition as the final cross-domain recommendation model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the cross-domain recommendation optimization method based on graph neural architecture search described in any one of claims 1-8 are implemented.

Images