Sequence recommendation method based on edge-enhanced global decoupling graph neural network

Abstract

The application discloses a sequence recommendation method based on an edge-enhanced global decoupling graph neural network, comprising: an input layer: all items in the training data are taken as nodes, and the interaction order between the user and the item is taken as an edge to construct a global link graph and input into the graph neural network; a decoupling learning layer: the influence probability of a preset number of factors from all neighbor items of the item and the influence probability of a preset number of factors from the target user's items in the front according to the interaction time sequence; a prediction layer: the decoupled item representation of the global level and the local level is accumulated, multiplied by the initial embedding representation of the candidate item, and the probability that the target user appears as the next interaction item v i The application has the beneficial effects of perceiving the transfer of the user's intention and more accurately recommending the product to the user.

Description

Technical Field

[0001] This invention relates to the field of sequence recommendation technology. More specifically, this invention relates to a sequence recommendation method based on an edge-enhanced globally decoupled graph neural network. Background Technology

[0002] Recommender systems play a crucial role in the rapidly evolving internet era, and sequence recommendation is a vital component. Sequence recommendation models user behavior as a sequence of items rather than a set of items. Markov chains are a classic approach that models short-term item transitions and predicts the next item a user might like. With the development of deep learning networks, recurrent neural networks (RNNs) have achieved success in sequence recommendation. For example, Long Short-Term Memory (LSTM) networks are a common variant of RNNs used to enhance the model's ability to retain sequential information through memory units. GRU4REC applies gated recurrent units in parallel mini-batches for session-based recommendation. However, RNN-based methods face the challenge of retaining long-range information, and self-attention networks have recently been applied to sequence recommendation to capture both long-term and short-term dependencies.

[0003] However, previous work modeled user intent through historical sequential interactions, neglecting the dynamic underlying relationships between items. Edges linking pairs of items contain rich semantic information, explaining why and how users choose one item after another. These underlying factors relate to real-world concepts, and a particular factor often dominates in a single situation. For example, suppose there are two users interacting with six items, such as... Figure 2 As shown in the diagram, the linking graph shows that item 2 is adjacent to all five other items. However, these links are intuitively driven by different factors. Item 2 is linked to items 1 and 4 because they are the same color, while it is linked to items 5 and 6 because they are both short-sleeved. Item 3 is connected to item 2 because it can be used as a jacket over a T-shirt. These different factors reveal the shift in user intent and also expose the shared characteristics of paired items. Therefore, it is worthwhile to explore how to identify and differentiate these underlying factors.

[0004] Decoupling representation learning is very popular in many fields, including computer vision, and has recently been applied to recommender systems. The general goal of decoupling representation learning is to separate the uniqueness and informative factors of data variation, where each unit is associated with a single concept in the real world. A single change in one factor will lead to a change in the associated unit. The most common networks for learning decoupling representations are β-VAE and InfoGAN. In recommender systems, Macrid-VAE infers high-level concepts of user intent at the macro level and applies VAEs at the micro level to enhance decoupling. The authors also proposed a self-supervised seq2seq training strategy for sequence recommendation, which utilizes an intent decoupling encoder to generate subsequences and compares user intent between pairwise subsequences. However, these studies do not consider the linking patterns of items and fail to distinguish the different user intents behind the sequences. Therefore, the learned sequence models will be sensitive to noisy data and lack interpretability. Summary of the Invention

[0005] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0006] Another objective of this invention is to provide a sequence recommendation method based on edge-enhanced global decoupled graph neural networks, which can identify the potential intent of user interaction item conversion, thereby accurately recommending products to users and improving transaction speed and success rate.

[0007] To achieve these objectives and other advantages according to the present invention, a sequence recommendation method based on an edge-enhanced globally decoupled graph neural network is provided, comprising: Input layer: Using all items in the training data as nodes and the interaction order between the user and the items as edges, a global link graph is constructed and input into the graph neural network; Decoupled Learning Layer: The global layer employs a graph neural network-based decoupled representation method to aggregate data from various projects. v i The probability of influence of a preset number of factors on all neighboring projects is obtained. v i A global-level decoupled project representation; The local layer first preprocesses the target user's item interaction sequence using a sequence neural network, and then aggregates the items from the target user using a decoupled representation method based on a graph neural network. v i Based on the influence probability of a preset number of factors for the first L-1 items in the interaction time sequence, the items for the target user are obtained. v i The decoupling of projects at a local level; Prediction layer: Accumulated items v iDecoupling project representation at the global level from the target user's project v i The decoupled item representation at the local level is multiplied by the initial embedding representation of each candidate item to obtain the target user's item representation. v i The probability of it appearing as the next interactive item; Iterative training: Set training objectives, train the graph neural network, update item representations and internal parameters, and obtain the trained graph neural network; The outputs of the prediction layer of the trained graph neural network are arranged in descending order, and the item with the highest probability is the item recommended to the target user.

[0008] Preferably, the decoupling representation method at the global level includes: preset items v i Benefiting from its neighbor's project v j The number of influencing factors is K, that is, from the project v i To the neighbor's project v j With K channels, each channel corresponding to an influencing factor, the neighboring items are calculated based on a channel-aware mechanism. v j K influencing factors on the project v i The probability of impact, and aggregated from the project v i All Neighbor Projects v j The probability of impact, to update the project. v i The project indicated that it received .

[0009] Preferably, neighboring projects v j The kth influencing factor on the project v i The probability of influence is calculated using Formula 1: Formula 1 in, Indicates project v i and v j It is an adjacent link, and the project v j Influencing factors k The project v i The probability of influence, Indicates channelk The parameters, W k′ Indicates channel k’ The parameters, d in d represents the dimension of the input item embedding. channel This represents the dimension of the embedding representation for each channel. h i Indicates project v i The initial embedding representation, h j Indicates project v j The initial embedding representation, σ (·) denotes a non-linear activation function. W k T and W k′ T Representing the matrix W k and W k′ Take the transpose.

[0010] Preferably, the items are accumulated using Formula 2. v i The probability of the influence of the k-th influencing factor in the neighborhood of the project v i middle: Formula 2 in, Indicates cumulative items v i The probability of the influence of the kth influencing factor in the neighborhood of the project v i The project indicates that, Indicates project v i The neighborhood of a, that is, the set of all neighboring items; Formula 3 is used to convert K channels Accumulation: Formula 3 in, Indicates project v i The item representation after summing the influence probabilities of K channels, i.e., the item v i A global-level decoupled representation.

[0011] Preferably, the decoupling representation method at the local level includes position embedding of the item sequence, specifically as follows: Select target users u Projects that interact at a given moment v i and Project v i The most recent L-1 interacting items form an item sequence, and each item in the item sequence is embedded in position according to the sequence order. The item sequence is then input into a self-attention encoder for encoding, so that the position information is aggregated and updated into the item representation of the item sequence. Formula 4 is used to represent the position-embedded item sequence: Formula 4 in, p 1 ~ p L This represents the embedding representation at L positions. h 1 ~ h L This represents the initial embedding of L items. H This represents a sequence of items with positional embeddings. The output of a single-head self-attention encoder is calculated using Formula 5: Formula 5 Where D represents the output of a single-head self-attention network. W Q , W Key , W V All represent the parameters of the self-attention network, and softmax is the normalized exponential function; Formula 6 is used to calculate the item sequence representation after passing through the single-head self-attention network, i.e. h s : Formula 6.

[0012] Preferably, a variational autoencoder is used to process the item sequence with location information. The specific method is as follows: The item sequence output by the self-attention encoder h s The input items are fed into a variational autoencoder for encoding, and the output is a representation of the item sequence that follows a normal distribution. Calculate using formulas 7, 8, and 9: Formula 7 Formula 8 Formula 9 in, l 1 and l 2 represents the linear transformation function, z v This represents the item sequence representation after passing through a variational autoencoder network. μ v This represents the mean of a Gaussian distribution. σ v This represents the variance of the Gaussian distribution. .

[0013] Preferably, the k-th influencing factor of each project in the project sequence has a significant impact on the project. v i The aggregation of influence probabilities is calculated using Formula 10: Formula 10 in, This represents the item representation after summing the probability of the kth influencing factor of L-1 items; Formula 11 is used to convert K channels Accumulation: Formula 11 in, z i l Indicates project v i The combination of the cumulative influence probabilities of K channels, i.e., the project v i Decoupling representation at the local level.

[0014] Preferably, Formula 12 is used to calculate the target user's performance in the project. v i The probability of it appearing as the next interactive item: Formula 12 in, Indicates the target users in the project v i The probability of it appearing as the next interactive item. z L T express z L transpose, h i This represents the initial embedding representation of the candidate item. This represents the final embedding representation of the most recent item in the sequence, where softmax is the normalized exponential function.

[0015] Preferably, the lower bound of evidence is used as the training target, and formula 13 is used for calculation: Formula 13 in, Indicates reconstruction error. Indicates KL divergence; in, , , in, Indicates project v i The probability of it appearing as the next interactive item in the target user's sequence. y i Project demonstration v i Does it actually appear as the next interactive item in the target user's sequence? y i If the condition actually occurs, the value is 1; otherwise, the value is 0.

[0016] Preferably, it also includes the use of L2 Normalization and .

[0017] This invention offers at least the following advantages: It proposes an edge-enhanced global decoupled graph neural network model to capture item link information. Item representation and user intent are modeled at both global and local levels. Firstly, a global item link graph is constructed across all sequences. A channel-aware mechanism is applied to decompose item edges into multiple channels, each corresponding to an influencing factor. Channels extract specific factor features from item neighbors and aggregate different factors to the target item. Secondly, the decoupled sequence representation is modeled on the current sequence. First, latent variables are inferred to be Gaussian distributions for decoupled representation learning from a variational autoencoder statistical perspective. Then, a graph neural network is applied to aggregate item information with previous items and model user intent. This allows for the identification of potential intent shifts in user interaction item transitions, thereby accurately recommending products to users and improving transaction speed and success rate.

[0018] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0019] Figure 1 This is an architecture diagram of the sequence recommendation model according to one of the technical solutions of the present invention; Figure 2 This is an example of a portion of the global link diagram of one of the technical solutions of the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0021] It should be noted that, unless otherwise specified, the experimental methods described in the following embodiments are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified. In the description of this invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this invention and simplifying the description. It does not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0022] like Figures 1-2 As shown, this invention provides a sequence recommendation method based on an edge-enhanced globally decoupled graph neural network, comprising: Input layer: Using all items in the training data as nodes and the interaction order between the user and the items as edges, a global link graph is constructed and input into the graph neural network; Decoupled Learning Layer: The global layer employs a graph neural network-based decoupled representation method to aggregate data from various projects. v i The probability of influence of a preset number of factors on all neighboring projects is obtained. v i A global-level decoupled project representation; The local layer first preprocesses the target user's item interaction sequence using a sequence neural network, and then aggregates the items from the target user using a decoupled representation method based on a graph neural network. v i Based on the influence probability of a preset number of factors for the first L-1 items in the interaction time sequence, the items for the target user are obtained. v i The decoupling of projects at a local level; Prediction layer: Accumulated items v i Decoupling project representation at the global level from the target user's project v i The decoupled item representation at the local level is multiplied by the initial embedding representation of each candidate item to obtain the target user's item representation. v i The probability of it appearing as the next interactive item; Iterative training: Set training objectives, train the graph neural network, update item representations and internal parameters, and obtain the trained graph neural network; The outputs of the prediction layer of the trained graph neural network are arranged in descending order, and the item with the highest probability is the item recommended to the target user.

[0023] In the above technical solution, by establishing connections between users and items in the training data, and based on the interaction sequence, the global link graph can represent items and determine their neighborhood information, providing an information foundation for subsequent mining of interaction intent. Then, a decoupled learning method is used to learn the interaction intent between items from two levels. Firstly, by mining the neighborhood of an item in the global link graph, the probability of influence from multiple factors from all neighboring items is aggregated, i.e., the magnitude of the influence of a certain factor in the neighborhood on the item, thus identifying the most influential factor. Secondly, a sequence neural network is used to associate information about the target user's item interaction sequence within a certain time period, and then a decoupled representation method of the graph neural network is used to learn the target user's recent item interaction association information, thereby aggregating the factors influencing the target user's recent interactive items and identifying the most influential factor.

[0024] Based on the mining of potential intentions in the project interaction sequence of the above two steps, the next most likely interaction item can be predicted and recommended to the target user. The accuracy of the recommendation and the likelihood of promoting the transaction are significantly increased.

[0025] The sequential recommendation method of this invention is particularly effective in product recommendations on shopping websites, significantly improving both the speed and success rate of transactions.

[0026] In another technical solution, the global-level decoupling representation method includes: preset items. v i Benefiting from its neighbor's project v j The number of influencing factors is K, that is, from the project v i To the neighbor's project v j With K channels, each channel corresponding to an influencing factor, the neighboring items are calculated based on a channel-aware mechanism. v j K influencing factors on the project v i The probability of impact, and aggregated from the project v i All Neighbor Projects v j The probability of impact, to update the project. v i The project indicated that it received .

[0027] In the above technical solution, since it is unclear how many influencing factors there are for the given training data, but in fact there are a certain number of influencing factors, the number of influencing factors is set and iterative training is performed. If the training result is not good, the number of influencing factors can be changed. This process is repeated many times until the number of influencing factors suitable for the given training data is obtained.

[0028] In another technical solution, neighbor projects v j The kth influencing factor on the project v i The probability of influence is calculated using Formula 1: Formula 1 in, Indicates project v i and v j It is an adjacent link, and the project v j Influencing factors k The project v i The probability of influence, Indicates channel k The parameters, W k′ Indicates channel k′ The parameters, d in d represents the dimension of the input item embedding. channel This represents the dimension of the embedding representation for each channel. h i Indicates project v i The initial embedding representation, h j Indicates project v j The initial embedding representation, σ (·) denotes a non-linear activation function. W k T and W k′ T Representing the matrix W k and W k′ Take the transpose.

[0029] In the above technical solution, the relationship between each neighboring project and the project can be calculated using Formula 1. v i The probability of influence of each influencing factor is used to quantify the influencing factors.

[0030] In another technical solution, formula 2 is used to accumulate the items. v i The probability of the influence of the k-th influencing factor in the neighborhood of the project v i middle: Formula 2 in, Indicates cumulative items v i The probability of the influence of the kth influencing factor in the neighborhood of the project v i The project indicates that, Indicates project v i The neighborhood of a, that is, the set of all neighboring items; Formula 3 is used to convert K channels Accumulation: Formula 3 in, Indicates project v i The item representation after summing the influence probabilities of K channels, i.e., the item v i A global-level decoupled representation.

[0031] In the above technical solution, formulas 2 and 3 can aggregate the impact of multiple influencing factors of all neighboring projects to the project. v i In the middle, for the project v i This achieves the goal of globally decoupled learning.

[0032] In another technical solution, the decoupling representation method at the local level includes position embedding of the item sequence, specifically as follows: Select target users u Projects that interact at a given moment v i and Project v i The most recent L-1 interacting items form an item sequence, and each item in the item sequence is embedded in position according to the sequence order. The item sequence is then input into a self-attention encoder for encoding, so that the position information is aggregated and updated into the item representation of the item sequence. Formula 4 is used to represent the position-embedded item sequence: Formula 4 in, p 1~ p L This represents the embedding representation at L positions. h 1 ~ h L This represents the initial embedding of L items. H This represents a sequence of items with positional embeddings. The output of a single-head self-attention encoder is calculated using Formula 5: Formula 5 Where D represents the output of a single-head self-attention network. W Q , W Key , W V All represent the parameters of the self-attention network, and softmax is the normalized exponential function; Formula 6 is used to calculate the item sequence representation after passing through the single-head self-attention network, i.e. h s : Formula 6.

[0033] In the above technical solution, the recent project interaction information of the target user is represented in the current project by the self-attention encoder, and the position embedding representation is combined to represent the information of the target user's project interaction sequence, so as to provide an information basis for subsequent decoupling learning.

[0034] In another technical solution, a variational autoencoder is used to process the item sequence with location information. The specific method is as follows: The item sequence output by the self-attention encoder h s The input items are fed into a variational autoencoder for encoding, and the output is a representation of the item sequence that follows a normal distribution. Calculate using formulas 7, 8, and 9: Formula 7 Formula 8 Formula 9 in, l 1 and l 2 represents the linear transformation function, z v This represents the item sequence representation after passing through a variational autoencoder network. μ v This represents the mean of a Gaussian distribution. σ v This represents the variance of the Gaussian distribution. .

[0035] In the above technical solution, the item representation is inferred as a Gaussian distribution by a variational autoencoder, thereby introducing uncertainty modeling into the system and further improving the independence between the dimensions of the item representation in decoupled representation learning.

[0036] In another technical solution, the k-th influencing factor of each project in the project sequence affects the project. v i The aggregation of influence probabilities is calculated using Formula 10: Formula 10 in, This represents the item representation after summing the probability of the kth influencing factor of L-1 items; Formula 11 is used to convert K channels Accumulation: Formula 11 in, z i l Indicates project v i The combination of the cumulative influence probabilities of K channels, i.e., the project v i Decoupling representation at the local level.

[0037] In the above technical solution, the decoupling learning method can be used to perform local intent mining for target users. It decouples the impact of multiple influencing factors by analyzing the items recently interacted by the target user in the order of interaction and aggregates them into the item. v i In the middle, for the project v i This achieves the goal of localized decoupling learning for the target user.

[0038] In another technical solution, Formula 12 is used to calculate the target user's performance in the project. v i The probability of it appearing as the next interactive item: Formula 12 in, Indicates the target users in the project v i The probability of it appearing as the next interactive item. z L T express z L transpose, h iThis represents the initial embedding representation of the candidate item. This represents the final embedding representation of the most recent item in the sequence, where softmax is the normalized exponential function.

[0039] In the above technical solution, the candidate items are actually all the items in the training data. Formula 12 can be used to calculate the next possible interaction item for a certain target user, and the item with the highest probability is the most likely next interaction item.

[0040] In another technical approach, the lower bound of evidence is used as the training target, and formula 13 is employed for calculation: Formula 13 in, Indicates reconstruction error. Indicates KL divergence; in, , , in, Indicates project v i The probability of it appearing as the next interactive item in the target user's sequence. y i Project demonstration v i Does it actually appear as the next interactive item in the target user's sequence? y i If the condition actually occurs, the value is 1; otherwise, the value is 0.

[0041] In the above technical solution, using the lower bound of evidence as the training target can train the graph neural network faster and more effectively, avoid overfitting, and improve prediction accuracy.

[0042] Another technical solution also includes using L2 Normalization and Using regularization to process data can ensure numerical stability.

[0043] Specifically: Figure 1 The diagram shows the model architecture of the entire edge-enhanced globally decoupled graph neural network.

[0044] 1. Problem Definition Given M users and N items, we represent the user set as follows: Representing a project set as For each user, Indicates user u and vThe sequential behavior of interactions. Given in t The historical sequence of time points, the sequence recommendation model aims to predict when a user will... t+1 The next project that might be of interest at any time.

[0045] In the inventive method, a global-level link graph is provided to capture project link transformation information. The global link graph is defined as follows: ,in V It is the collection of all items in the training data. E It is a set of edges. Each edge Indicates that users are in order v i After that and v j Interact with it. Indicates project v i The neighborhood of, i.e., the sequence containing, v i Adjacent projects.

[0046] 2. Channel sensing mechanism Given a historical sequence ,definition Indicates user u In relation to the project Potential intentions during interaction. Assume there exists an underlying intent related to the user's intentions. K One factor will represent the potential representation Divided into K One channel, namely . No. k Each channel independently corresponds to the first... k One factor. For each pair of adjacent items, and The correlation between items v i and projects v j Influencing factors k The similarity between the two projects was analyzed, and it was revealed why and how the two projects are connected.

[0047] 3. Global-level decoupled representation learning Assume there exists something related to user intent. K This concept means there are K Several potential factors need to be decoupled. Given a global graph constructed based on training data... Nodes (i.e., projects) are divided in the latent space into K Each component is represented by a link edge, which is then correspondingly divided into components. K One channel. The first... k The first component and the user intentk The first factor is related, while the second factor is related. k Each channel represents a factor. k How to affect the linking of paired projects.

[0048] For a single node in the graph v i The goal is to aggregate data from its neighborhood. The information. First, the project representation is broken down into... K Each component, and calculate the factors. k From his neighbor Impact on Project v i The probability of: Revealing why the project v i and v j It is an adjacent link, and the project v j How to factor k Projects with impact v i Then you can start from the project. v i Accumulate and update information from neighbors v i The project indicates: To ensure numerical stability, use L2 Normalization ,Will writing .

[0049] By projecting project representations onto different channels, project information can be aggregated from the perspectives of different conceptual factors. This results in a global-level project representation of the model. It can be written as K The combination of channels, i.e. .

[0050] 4. Local-level decoupling representation learning Given that items appearing in a sequence rarely repeat, local learning models are based on sequence methods rather than graph neural networks. This is given by a series of historical user behaviors. Choose the nearest one L One interactive item. For sequences shorter than... L For users, zero vectors are repeatedly added to the left side of the sequence. To distinguish item representations at different positions in the sequence, a learnable positional embedding is added to the initial item embedding. And use it as the final input to the learning layer: First, a self-attention network is applied to the learning model of this invention to leverage its ability to sequentially capture the long-term and short-term dependencies of items. Let... For the output of a single-head self-attention encoder, h s As the input of the variational autoencoder, where .

[0051] In variational autoencoders, the posterior distribution is inferred as a multinomial layer. The mean and variance vectors are computed based on the self-attention vector, as shown below: By using the "reparameterization technique" that follows the traditional variational autoencoder model, the output of the variational autoencoder layer is written as... ,in .

[0052] Following the self-attention variational autoencoder model, an item representation of the entire sequence following a normal distribution is obtained. Then, a channel-aware aggregation mechanism is applied for local decoupling learning.

[0053] To obtain user intent conversion, a channel-aware sliding window strategy based on a graph neural network was used. The sliding window length was set to... L This means that for each target project v i The information will be derived from its previous L - Aggregates from 1 project. Project v i and v j The probability between them is calculated based on the channel-aware mechanism as follows: , Then the sequence representation of the local layer is... K A combination of factors: .

[0054] 5. Prediction layer Based on representations learned from global and local layers z g , z l This yields the final sequence representation, written as: .

[0055] The final recommendation probability of candidate items can be estimated based on the current sequence embedding and the initial item embedding. Indicates project vi The probability of appearing as the next interaction in the current sequence: The training objective is calculated from the lower bound of evidence, and the reconstruction error is defined as the cross-entropy, as shown below.

[0056] 6. Iterative training Using the lower bound of evidence as the training objective, the entire edge-reinforced global decoupled graph neural network model is trained, resulting in a fully trained model. For a target user at the current time, the model predicts the next item that the user might be interested in at the next time step and recommends it to the target user.

[0057] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A sequence recommendation method based on edge-enhanced globally decoupled graph neural networks, characterized in that, The method is used for product recommendations on shopping websites, including: Input layer: Using all product items in the training data as nodes and the interaction order between users and product items as edges, a global link graph is constructed and input into the graph neural network; Decoupled Learning Layer: The global layer adopts a decoupled representation method based on graph neural networks to aggregate data from product projects. v i The probability of influence of a preset number of factors on all neighboring product items is obtained by calculating the product item's... v i The decoupling of product projects at the global level; The local layer first preprocesses the product item interaction sequence of the target user using a sequence neural network, and then aggregates the product items from the target user using a decoupled representation method based on a graph neural network. v i Based on the influence probability of a preset number of factors for the first L-1 product items in the interaction time sequence, the product items for the target user are obtained. v i This refers to the decoupling of product projects at a local level. Prediction layer: Accumulated product items v i Decoupling product project representation from target user product project at the global level v i The decoupled product item representation at the local level is multiplied by the initial embedding representation of each candidate product item to obtain the target user's representation in the product item. v i The probability of it appearing as the next interactive product project; Iterative training: Set training objectives, train the graph neural network, update the product item representation and internal parameters, and obtain the trained graph neural network; The outputs of the prediction layer of the trained graph neural network are arranged in descending order, and the product with the highest probability is the product recommended to the target user. Among them, the decoupling representation method at the global level includes: preset product items v i Benefiting from its neighbor's product projects v j The number of influencing factors is K, that is, from the product project v i To the neighbor's product project v j With K channels, each channel corresponding to an influencing factor, the calculation of each neighboring product item is performed based on a channel-aware mechanism. v j K influencing factors on product projects v i The probability of impact, and aggregated from product projects v i All Neighbor Product Projects v j The probability of impact, to update product items. v i The product project indicated that it received ; Neighbor Product Project v j The kth influencing factor on the product project v i The probability of influence is calculated using Formula 1: Official 1 in, Indicates product items v i and v j It consists of adjacent links and product items. v j Influencing factors k The above product projects v i The probability of influence, Indicates channel k The parameters, W k′ Indicates channel k′ The parameters, d in d represents the dimension of the embedded product item in the input. channel This represents the dimension of the embedding representation for each channel. h i Indicates product items v i The initial embedding representation, h j Indicates product items v j The initial embedding representation, σ (·) denotes a non-linear activation function. W k T and W k′ T Representing the matrix W k and W k′ Take the transpose; Use Formula 2 to accumulate product items v i The probability of the influence of the k-th influencing factor in the neighborhood to the product item v i middle: Official 2 in, Indicates the cumulative product items v i The probability of the influence of the kth influencing factor in the neighborhood of the product project v i The product project indicates that, Indicates product items v i The neighborhood, that is, the set of all neighboring product items; Formula 3 is used to convert K channels Accumulation: Official 3 in, Indicates product items v i The product item representation after summing the influence probabilities of K channels, i.e., the product item v i Decoupled representation at the global level; The decoupling representation method at the local level includes positional embedding of the product item sequence, specifically as follows: Select target users u A product project that interacts at a certain moment. v i and product projects v i The most recent L-1 interactive product items form a product item sequence, and each product item in the product item sequence is embedded in position according to the sequence order. The product item sequence is input into a self-attention encoder for encoding, so as to aggregate and update the position information into the product item representation of the product item sequence. Formula 4 is used to represent the sequence of product items with embedded positions: Official 4 in, p 1 ~ p L This represents the embedding representation at L positions. h 1 ~ h L This represents the initial embedding of L product items. H This represents a sequence of product items with location embedding. The output of a single-head self-attention encoder is calculated using Formula 5: Official 5 Where D represents the output of a single-head self-attention network. W Q , W Key , W V All represent the parameters of the self-attention network, and softmax is the normalized exponential function; Formula 6 is used to calculate the product item sequence representation after passing through a single-head self-attention network, i.e. h s : Official 6; A variational autoencoder is used to process product item sequences with location information. The specific method is as follows: Product item sequence based on the output of a self-attention encoder h s The input is fed into a variational autoencoder for encoding, and the output is a representation of the product item sequence that follows a normal distribution. Calculate using formulas 7, 8, and 9: Official 7 Official 8 Official 9 in, l 1 and l 2 represents the linear transformation function, z v This represents the sequence of product items after passing through a variational autoencoder network. μ v This represents the mean of a Gaussian distribution. σ v This represents the variance of the Gaussian distribution. ; The k-th influencing factor of each product item in the product item sequence affects the product item. v i The aggregation of influence probabilities is calculated using Formula 10: Official 10 in, This represents the product item representation after summing the probability of the influence of the kth influencing factor for L-1 product items; Formula 11 is used to convert K channels Accumulation: Official 11 in, z i l Indicates product items v i The combination of the cumulative influence probabilities of K channels, i.e., the product project. v i Decoupling representation at the local level.

2. The sequence recommendation method based on edge-enhanced global decoupled graph neural network as described in claim 1, characterized in that, Formula 12 is used to calculate the target user's performance in the product project. v i The probability of it appearing as the next interactive product project: Official 12 in, Indicates target users in product projects v i The probability of it appearing as the next interactive product project. z L T express z L transpose, h i This represents the initial embedding representation of the candidate product item. This represents the final embedding representation of the most recent product item in the sequence, where softmax is a normalized exponential function.

3. The sequence recommendation method based on edge-enhanced global decoupled graph neural network as described in claim 1, characterized in that, Using the lower bound of evidence as the training target, the calculation is performed using Formula 13: Official 13 in, Indicates reconstruction error. Indicates KL divergence; in, , ， in, Indicates product items v i The probability of it appearing as the next interactive product project in the target user sequence. y i Show product items v i Whether it will actually appear as the next interactive product project in the target user's sequence. y i If the condition actually occurs, the value is 1; otherwise, the value is 0.

4. The sequence recommendation method based on edge-enhanced global decoupled graph neural network as described in claim 1, characterized in that, It also includes the use of L2 Normalization and .

Images