Training method of collaborative filtering recommendation model based on graph neural network

Abstract

The invention belongs to the field of collaborative filtering recommendation, and particularly relates to a training method of a collaborative filtering recommendation model based on a graph neural network. The method comprises the steps that the user side embedded feature relevancy and the object side embedded feature relevancy output by a neural network layer of each graph are determined, and the embedded feature relevancy is obtained by calculating the average feature relevancy of a node embedded matrix; when the relevancy of the user-side embedded features exceeds a threshold value, a user-side penalty weight inversely proportional to the relevancy is distributed, and when the relevancy of the object-side embedded features exceeds the threshold value, an object-side penalty weight inversely proportional to the object-side embedded features is distributed, and the threshold value is determined by scene attributes of a business scene; determining total decoupling loss according to embedded feature relevancy and penalty weight, adding the total decoupling loss to an original loss function of the collaborative filtering recommendation model to generate a total loss function, and training the collaborative filtering recommendation model based on the total loss function.

Description

Technical Field

[0001] The present disclosure belongs to the field of collaborative filtering recommendation, and specifically relates to a training method for a collaborative filtering recommendation model based on graph neural network. Background Art

[0002] Recommender systems, as a key technology for addressing information overload and providing personalized services, are crucial in various online applications. Collaborative filtering has become a mainstream recommendation technology because it can leverage collective intelligence to discover subtle connections of interest.

[0003] Graph neural networks are increasingly being used in collaborative filtering recommendations. In traditional GNN-based collaborative filtering recommendation models, a single neural network layer primarily considers the immediate neighbors of users and items, limiting their ability to capture deep collaborative signals. To address this, current GNN-based collaborative filtering recommendation models are stacking multiple GNN layers to expand their receptive domain. However, as the number of stacked layers increases, GNN-based collaborative filtering recommendation models experience performance degradation.

[0004] The increase in the number of stacking layers will cause the node representations to tend to be similar, resulting in an over-smoothing problem. At the same time, it will also cause the correlation between the feature dimensions of the node representations to gradually increase, resulting in an over-correlation problem.

[0005] To address the problem that the performance of collaborative filtering recommendation models based on graph neural networks decreases as the number of stacked layers increases, the current mainstream approach is to try to alleviate the over-smoothing problem through various attempts, such as simplifying the message passing process or performing feature propagation within segmented subgraphs. However, the results are not good. Summary of the Invention

[0006] The disclosed embodiments propose a collaborative filtering recommendation method based on graph neural networks to solve the problem of poor performance of existing methods due to failure to comprehensively consider over-smoothing and over-correlation.

[0007] A first aspect of the present disclosure provides a method for training a collaborative filtering recommendation model based on a graph neural network, comprising:

[0008] Determining the correlation between the user-side embedded features and the item-side embedded features output by each graph neural network layer, wherein the embedded feature correlation is obtained by calculating the average feature correlation of the node embedding matrix;

[0009] When the correlation of the user-side embedded features exceeds a threshold, a user-side penalty weight inversely proportional to the correlation is assigned; when the correlation of the item-side embedded features exceeds the threshold, an item-side penalty weight inversely proportional to the correlation is assigned, wherein the threshold is determined by the scenario attributes of the business scenario;

[0010] A total decoupling loss is determined according to the embedding feature correlation and the penalty weight, the total decoupling loss is added to the original loss function of the collaborative filtering recommendation model to generate a total loss function, and the collaborative filtering recommendation model is trained based on the total loss function.

[0011] In some embodiments of the present disclosure, the embedding feature correlation is obtained by calculating the average feature correlation of the node embedding matrix, including:

[0012] The embedding feature correlation is calculated based on a feature correlation calculation function for the node embedding matrix, wherein the feature correlation calculation function is:

[0013]

[0014] Among them, r(Φ) is the feature correlation calculation function, Φ is the node embedding matrix, Φ *i and Φ *j are any two columns of the node embedding matrix, γ(·,·) is the Pearson correlation coefficient, and the larger the γ is, the more serious the dimensional coupling is.

[0015] In some embodiments of the present disclosure, the embedded feature correlation is calculated based on the deformed feature correlation calculation function, wherein the feature correlation calculation function is rewritten as:

[0016]

[0017] Among them, r(Φ) is the feature correlation calculation function, is the feature correlation coefficient matrix, is the feature covariance matrix, Iκ is the identity matrix, and diag creates a diagonal matrix.

[0018] In some embodiments of the present disclosure, during the training process of the collaborative filtering recommendation model, a skip probability is set for each graph neural network layer, and a Bernoulli random variable is sampled based on the skip probability in each training round;

[0019] When the Bernoulli random variable takes the value of 1, the decoupling loss calculation of the layer is skipped;

[0020] The skip probability is 0.9 in the initial training round, decreases linearly with each training round, and is 0 in the tenth training round.

[0021] In some embodiments of the present disclosure, if the Bernoulli random variable of the graph neural network layer does not take a value of 1 and the user-side feature correlation exceeds the threshold, a penalty weight is assigned to the user side according to the following formula:

[0022] and / or

[0023] If the Bernoulli random variable of the graph neural network layer is not 1 and the item-side feature correlation exceeds the threshold, a penalty weight is assigned to the item side according to the following formula:

[0024]

[0025] Among them, 1 represents the indicator function, S represents the total number of layers, s represents the current layer number, q t Indicates whether to skip the decoupling calculation of this layer. Indicates the penalty weight assigned to the user side of the current layer, Indicates the penalty weight assigned to the item side of the current layer, represents the user embedding representation matrix of the current layer, Represents the user embedding representation matrix of each layer, Represents the item embedding representation matrix of the current layer, represents the item embedding representation matrix of each layer, τ(·) represents the target correlation threshold calculation function, and r(·) represents the feature correlation calculation function.

[0026] In some embodiments of the present disclosure, the method further includes:

[0027] When the user side satisfies the first rule, the penalty weight assigned to the user side of the current layer is 0; and / or

[0028] When the item side satisfies the second rule, the penalty weight allocated to the item side in the current layer is 0.

[0029] In some embodiments of the present disclosure, determining the total decoupling loss according to the feature correlation and the penalty weight includes:

[0030] For each graph neural network layer that needs to calculate the decoupling loss, its decoupling loss is calculated by the following formula:

[0031] Layer decoupling loss = (user-side feature relevance × user-side penalty weight) + (item-side feature relevance × item-side penalty weight);

[0032] The layer decoupling losses of all the graph neural network layers that need to calculate the decoupling loss are added together to obtain the total decoupling loss.

[0033] In some embodiments of the present disclosure, adding the total decoupling loss to the original loss function of the collaborative filtering recommendation model to generate a total loss function includes:

[0034] The total decoupling loss is added to the original loss function of the collaborative filtering recommendation model based on the following formula:

[0035] L=L CF +βLAFD

[0036] Among them, L is the total loss function, L CF is the original loss function, L AFD is the total decoupling loss and β is the balance coefficient.

[0037] In some embodiments of the present disclosure, the training of the collaborative filtering recommendation model based on the total loss function further includes:

[0038] The balance coefficient is dynamically adjusted based on the performance indicator of the collaborative filtering recommendation model on the validation set.

[0039] In some embodiments of the present disclosure, dynamically adjusting the balance coefficient based on the performance indicator of the collaborative filtering recommendation model on the validation set includes:

[0040] If the performance indicator improvement exceeds a first preset threshold, increase β;

[0041] If the performance indicator decreases by more than a second preset threshold, reducing β;

[0042] If the performance indicator changes within a preset range, the current β value remains unchanged.

[0043] In summary, the training methods for collaborative filtering recommendation models based on graph neural networks provided by various embodiments of the present disclosure, by explicitly calculating and penalizing the correlation between feature dimensions in the embedding representation, force the model to learn embedding representations that are more independent between dimensions and have less information redundancy, thereby alleviating the over-correlation problem and improving the quality and efficiency of the representation. At the same time, by assigning lower penalty weights to layers with higher correlation (typically deep layers), the present disclosure suppresses excessive correlation while avoiding excessively damaging the smoothness of deep layer representations, which is beneficial for collaborative filtering. This adaptive mechanism finds a better balance between decorrelation and maintaining the necessary smoothness, thereby more effectively alleviating the performance degradation caused solely by oversmoothing or overcorrelation. Furthermore, the present disclosure adaptively adjusts the relevance threshold based on scenario attributes, significantly increasing the adaptability of the recommendation system in different scenarios. The user-item asynchronous decoupling strategy avoids the impact of decoupling loss in the cold start phase on model training convergence, thereby maintaining stable gain in both high-density and cold start scenarios. Furthermore, layered self-hopping frequency control significantly reduces computational overhead while maintaining the effectiveness of decoupling. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The features and advantages of the present disclosure will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present disclosure in any way. In the accompanying drawings:

[0045] Figure 1is a flowchart of a method for training a collaborative filtering recommendation model based on a graph neural network according to some embodiments of the present disclosure;

[0046] Figure 2 The performance comparison of applying this method to the graph collaborative filtering baseline model and using the baseline model alone is shown.

[0047] Figure 3 We compared the training speed of various baseline models before and after applying this method, including the time per round and the number of rounds required.

[0048] Figure 4 This is a schematic diagram of a training device for a collaborative filtering recommendation model based on a graph neural network according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0049] In the detailed description that follows, many specific details of the present disclosure are set forth by way of example in order to provide a thorough understanding of the relevant disclosure. However, it will be apparent to one of ordinary skill in the art that the present disclosure can be implemented without these details. It should be understood that the use of the terms "system," "device," "unit," and / or "module" in the present disclosure is a method for distinguishing between different parts, elements, parts, or assemblies at different levels in a sequential arrangement. However, these terms may be replaced by other expressions if they can achieve the same purpose.

[0050] It should be understood that when a device, unit, or module is referred to as being "on," "connected to," or "coupled to" another device, unit, or module, it may be directly on, connected to, coupled to, or in communication with the other device, unit, or module, or there may be intervening devices, units, or modules, unless the context clearly indicates an exception. For example, the term "and / or" as used in this disclosure includes any and all combinations of one or more of the associated listed items.

[0051] The terms used in this disclosure are only for describing specific embodiments and are not intended to limit the scope of this disclosure. As shown in the specification and claims of this disclosure, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified features, wholes, steps, operations, elements and / or components, and such expressions do not constitute an exclusive list, and other features, wholes, steps, operations, elements and / or components may also be included.

[0052] These and other features and characteristics of the present disclosure, as well as the methods of operation, the functions of the related elements of the structure, the combination of parts, and the economy of manufacture may be better understood with reference to the following description and accompanying drawings, which form a part of this specification. However, it is to be expressly understood that the drawings are for illustration and description purposes only and are not intended to limit the scope of protection of the present disclosure. It is to be understood that the drawings are not drawn to scale.

[0053] Various structural diagrams are used in this disclosure to illustrate various variations of the embodiments of the present disclosure. It should be understood that the preceding or following structures are not intended to limit the present disclosure. The scope of protection of the present disclosure is subject to the claims.

[0054] Despite the remarkable achievements of graph neural networks in recommender systems, significant challenges remain in robustness and accuracy, such as over-correlation and over-smoothing of embeddings. In graph neural network-based recommendation models, a single neural network layer primarily considers the immediate neighbors of users and items, limiting its ability to capture deep collaborative signals. To address this limitation, traditional graph neural network-based recommendation models stack multiple graph neural network layers to expand their receptive domain. However, performance degradation is observed with increasing numbers of stacked layers. This performance degradation is generally believed to be due to over-smoothing, where node representations tend to become more similar with increasing layers. However, it is important to note that as the number of graph neural network layers increases, the correlation between the feature dimensions of node representations increases, which can negatively impact the quality of learned embeddings. While over-correlation and over-smoothing are somewhat related, their primary difference lies in their focus on different relationships: over-smoothing addresses the relationships between node representations (in the row direction of the representation matrix), while over-correlation addresses the relationships between the feature dimensions of representations (in the column direction). It is worth noting that over-smoothing does not necessarily negatively impact recommender systems. This is because the smoothness of embeddings plays a key role in the effectiveness of graph neural network-based collaborative filtering models.

[0055] Currently, the mainstream attempt to solve the problem that the performance of collaborative filtering recommendation models based on graph neural networks decreases as the number of stacked layers increases is to alleviate the transition smoothing problem. The first type of method aims to simplify the message passing process. For example, LightGCN removes feature transformation and nonlinear activation in graph neural networks, while GCCF adopts a residual network structure and removes nonlinear activation. UltraGCN skips the explicit message passing of infinite layers. SVD-GCN replaces graph convolution with truncated singular value decomposition. The second type of method performs feature propagation within segmented subgraphs. For example, IMP-GCN performs graph convolution within a subgraph consisting of users with similar interests and the items they interact with. However, the performance of the above methods has not met expectations.

[0056] In light of this, this paper provides an adaptive feature-decoupled graph collaborative filtering recommendation method for multi-scenario recommendations. By combining it with any graph neural network-based collaborative filtering model, it alleviates the over-correlation and over-smoothing issues of the original model while finding a balance between over-correlation and representation smoothness, thereby improving the recommendation performance of the original model. Thanks to the adaptive threshold of scene attributes, this method can maintain stable gain in both high-density and cold-start scenarios.

[0057] Figure 1 This is a flowchart of a method for training a collaborative filtering recommendation model based on a graph neural network according to some embodiments of the present disclosure. In some embodiments, the method for training a collaborative filtering recommendation model based on a graph neural network is performed by a training server, which is used to train the collaborative filtering recommendation model based on a graph neural network. The method includes the following steps:

[0058] S110, determining the user-side embedded feature correlation and the item-side embedded feature correlation output by each graph neural network layer, wherein the embedded feature correlation is obtained by calculating the average feature correlation of the node embedding matrix.

[0059] First, the embedding feature correlation is calculated based on the node embedding matrix for the user side and item side of each graph neural network layer.

[0060]

[0061] Among them, r(Φ) is the feature correlation calculation function, Φ is the node embedding matrix, Φ *i and Φ *j are any two columns of the node embedding matrix, γ(·,·) is the Pearson correlation coefficient, and the larger the γ is, the more serious the dimensional coupling is.

[0062] In some embodiments of the present disclosure, to facilitate fast calculation and gradient transfer, r(Φ) can be rewritten as the Frobenius norm form:

[0063]

[0064] in, is the feature correlation coefficient matrix, is the feature covariance matrix, I κ is the identity matrix, and diag creates a diagonal matrix.

[0065] S120, when the correlation of the user-side embedded features exceeds a threshold, a user-side penalty weight inversely proportional to the user-side feature is assigned; when the correlation of the item-side embedded features exceeds the threshold, an item-side penalty weight inversely proportional to the item-side feature is assigned, wherein the threshold is determined by the scenario attributes of the business scenario.

[0066] In some embodiments of the present disclosure, in order to automatically adjust the penalty intensity according to the business scenario, a scenario attribute vector ζ (including information such as time period, region, and session length) is introduced, and a target relevance threshold is generated for each layer s through the function g(·):

[0067] τ s =g(ζ,s)

[0068] When τ(Φ (s) )≤τ s When , the correlation of this layer is considered to be within an acceptable range and no decoupling penalty is imposed.

[0069] During the training process, a skip probability p is set for each layer that decreases with the rounds. s , and sample the Bernoulli random variable based on the skip probability:

[0070] q s ~Bernoulli(p s )

[0071] When the Bernoulli random variable takes the value of 1, the decoupling calculation of this layer is skipped. s It takes the value 0.1 in the initial round and increases linearly with the number of training rounds, reaching 1 at round 10.

[0072] In other embodiments of the present disclosure, if the Bernoulli random variable of the graph neural network layer does not take a value of 1 and the user-side feature correlation exceeds the threshold, a penalty weight is assigned to the user side according to the following formula:

[0073]

[0074] If the Bernoulli gate value of the graph neural network layer is not 1 and the item-side feature relevance exceeds the threshold, a penalty weight is assigned to the item side according to the following formula:

[0075]

[0076] Among them, 1 represents the indicator function, S represents the total number of layers, s represents the current layer number, q t Indicates whether to skip the decoupling calculation of this layer. Indicates the penalty weight assigned to the user side of the current layer, Indicates the penalty weight assigned to the item side of the current layer, represents the user embedding representation matrix of the current layer, Represents the user embedding representation matrix of each layer, Represents the item embedding representation matrix of the current layer, represents the item embedding representation matrix of each layer, τ(·) represents the target correlation threshold calculation function, and r(·) represents the feature correlation calculation function.

[0077] In other embodiments of the present disclosure, if one side enters the cold start phase (for the user side, the average click traffic is less than 0.1% of the number of items; for the item side, the average shelf time of the items is less than or equal to 3 days), it can be temporarily placed. (or ), only punishing the other side.

[0078] S130, determining a total decoupling loss based on the embedding feature relevance and the penalty weight, adding the total decoupling loss to the original loss function of the collaborative filtering recommendation model to generate a total loss function, and training the collaborative filtering recommendation model based on the total loss function.

[0079] The total decoupling loss in this disclosure is obtained by multiplying the correlation of all valid layers with the weights and summing them up.

[0080]

[0081] Among them, L AFD is the total decoupling loss, and are the penalty weights assigned to the user side and item side of the sth layer respectively.

[0082] The final training target of the collaborative filtering recommendation model based on graph neural network in this disclosure is the original collaborative filtering loss L CF The weighted sum of the total decoupling loss, where the balance coefficient β can be dynamically adjusted by the performance of the validation set:

[0083] L=L CF +βL AFD ·

[0084] Among them, L is the total loss function, L CF is the original loss function, L AFD is the total decoupling loss and β is the balance coefficient.

[0085] In some embodiments of the present disclosure, the specific adjustment rules are as follows: based on the performance indicators on the validation set (such as NDCG@K, Recall@K, etc.), the β value is dynamically adjusted:

[0086] Performance improvement stage: If the validation set indicator improves by more than the preset threshold (e.g., 2%), β is increased (e.g., by 10%) to strengthen the optimization of the cold start problem.

[0087] Performance degradation phase: If the validation set metric drops by more than a preset threshold (e.g., 1%), reduce β (e.g., by 20%) to avoid over-focusing on cold starts and damaging the overall recommendation quality.

[0088] Performance stabilization stage: If the validation set indicators do not change significantly, the current β value remains unchanged.

[0089] One embodiment of the present disclosure tests the Figure 1 The effect of the method described in S110-S130 (hereinafter referred to as this method) on a graph collaborative filtering baseline model. The graph collaborative filtering baseline models include: NGCF, GCCF, HMLET, and LightGCN. The public datasets include MovieLens, Yelp, Gowalla, and Amazon-book. Each dataset is divided into training / validation / test sets with an 8:1:1 ratio. Yelp / Amazon-book retains ≥ 15 interactions, and Gowalla retains ≥ 10 interactions to ensure data validity.

[0090] The experimental configuration is:

[0091] 1. Optimizer: Adam, learning rate 1e–3, batch size 4096; embedding dimension / hidden dimension 128;

[0092] Initialize Xavier parameters and use early stopping to avoid overfitting (monitor NDCG@10 and stop if no improvement is seen).

[0093] 2. Baseline model setup: Use 3-layer propagation (3-layer graph neural network) for all graph collaborative filtering models

[0094] Figure 2 The performance comparison between this method applied to a graph collaborative filtering baseline model and the baseline model alone is shown. The performance metrics compared are Recal l, NDCG, and MAP.

[0095] Depend on Figure 2 It can be seen that after applying this method, Recal l, NDCG, and MAP are all steadily improved.

[0096] Taking the Recal l@10 indicator of the Yelp dataset as an example, the application of this method improved it by 12.20%, 11.73%, and 9.52% respectively compared with the original model alone.

[0097] It can be seen that this method effectively alleviates the problems of over-correlation and over-smoothing, and brings significant performance gains on datasets of various sizes and various graph collaborative filtering baseline models.

[0098] Figure 3 The training speed (time per round and number of rounds required) of various baseline models before and after applying this method is compared. The baseline models compared include HMLET and LightGCN. Figure 3In [1], epoch (ep) refers to the training round. Models starting with the prefix AFD- are the baseline models to which this method is applied.

[0099] Depend on Figure 3 It can be seen that:

[0100] 1.AFD-HMLET only needs 112–205 rounds to reach the optimal point, which is approximately 33%–41% less than the original HMLET. Although the time per round is slightly increased, the overall training time is significantly reduced.

[0101] 2.AFD-LightGCN also reduces the best rounds from 309–742 rounds to 114–360 rounds, an acceleration of more than 50%.

[0102] The above results show that this method can greatly accelerate model convergence and improve training efficiency.

[0103] Figure 4 Schematic diagram of a training device for a collaborative filtering recommendation model based on a graph neural network according to some embodiments of the present disclosure. Figure 4 As shown, the training device 400 for the collaborative filtering recommendation model based on a graph neural network includes a determination module 410, an allocation module 420, and a training module 430. In the present disclosure, the training function of the collaborative filtering recommendation model based on a graph neural network is performed by a training server, which is used to train the collaborative filtering recommendation model based on a graph neural network, wherein:

[0104] Determination module 410, configured to determine the correlation between the user-side embedded features and the item-side embedded features output by each graph neural network layer, wherein the embedded feature correlation is obtained by calculating the average feature correlation of the node embedding matrix;

[0105] an allocating module 420 configured to allocate a user-side penalty weight inversely proportional to a threshold when the correlation of the user-side embedded features exceeds a threshold, and allocate an item-side penalty weight inversely proportional to the threshold when the correlation of the item-side embedded features exceeds the threshold, wherein the threshold is determined by a scenario attribute of the business scenario;

[0106] The training module 430 is used to determine the total decoupling loss based on the embedding feature correlation and the penalty weight, add the total decoupling loss to the original loss function of the collaborative filtering recommendation model to generate a total loss function, and train the collaborative filtering recommendation model based on the total loss function.

[0107] In summary, the training methods for collaborative filtering recommendation models based on graph neural networks provided by various embodiments of the present disclosure, by explicitly calculating and penalizing the correlation between feature dimensions in the embedding representation, force the model to learn embedding representations that are more independent between dimensions and have less information redundancy, thereby alleviating the over-correlation problem and improving the quality and efficiency of the representation. At the same time, by assigning lower penalty weights to layers with higher correlation (typically deep layers), the present disclosure suppresses excessive correlation while avoiding excessively damaging the smoothness of deep layer representations, which is beneficial for collaborative filtering. This adaptive mechanism finds a better balance between decorrelation and maintaining the necessary smoothness, thereby more effectively alleviating the performance degradation caused solely by oversmoothing or overcorrelation. Furthermore, the present disclosure adaptively adjusts the relevance threshold based on scenario attributes, significantly increasing the adaptability of the recommendation system in different scenarios. The user-item asynchronous decoupling strategy avoids the impact of decoupling loss in the cold start phase on model training convergence, thereby maintaining stable gain in both high-density and cold start scenarios. Furthermore, layered self-hopping frequency control significantly reduces computational overhead while maintaining the effectiveness of decoupling.

[0108] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding descriptions in the aforementioned device embodiments and will not be repeated here.

[0109] Although the subject matter described herein is provided in the general context of being executed in conjunction with the execution of an operating system and application programs on a computer system, those skilled in the art will recognize that other implementations may also be performed in conjunction with other types of program modules. Generally speaking, program modules include routines, programs, components, data structures, and other types of structures that perform specific tasks or implement specific abstract data types. Those skilled in the art will appreciate that the subject matter described herein may be practiced using other computer system configurations, including handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframe computers, and the like, and may also be used in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

[0110] Those skilled in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0111] It should be understood that the above-described specific embodiments of the present disclosure are merely illustrative of or explanation of the principles of the present disclosure and do not constitute limitations on the present disclosure. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present disclosure shall be included within the scope of protection of the present disclosure. In addition, the claims appended to the present disclosure are intended to cover all variations and modifications that fall within the scope and metes and bounds of the appended claims, or equivalents of such scope and metes and bounds.

Claims

1. A training method for a collaborative filtering recommendation model based on graph neural network, characterized in that: include: Determining the correlation between the user-side embedded features and the item-side embedded features output by each graph neural network layer, wherein the embedded feature correlation is obtained by calculating the average feature correlation of the node embedding matrix; When the correlation of the user-side embedded features exceeds a threshold, a user-side penalty weight inversely proportional to the correlation is assigned; when the correlation of the item-side embedded features exceeds the threshold, an item-side penalty weight inversely proportional to the correlation is assigned, wherein the threshold is determined by the scenario attributes of the business scenario; A total decoupling loss is determined according to the embedding feature correlation and the penalty weight, the total decoupling loss is added to the original loss function of the collaborative filtering recommendation model to generate a total loss function, and the collaborative filtering recommendation model is trained based on the total loss function.

2. The method according to claim 1, characterized in that The embedding feature correlation is obtained by calculating the average feature correlation of the node embedding matrix, including: The embedding feature correlation is calculated based on a feature correlation calculation function for the node embedding matrix, wherein the feature correlation calculation function is: Among them, r(Φ) is the feature correlation calculation function, Φ is the node embedding matrix, Φ *i and Φ *j are any two columns of the node embedding matrix, γ(·,·) is the Pearson correlation coefficient, and the larger the γ, the more serious the dimensional coupling.

3. The method according to claim 2, wherein: The embedded feature correlation is calculated based on the deformed feature correlation calculation function, wherein the feature correlation calculation function is rewritten as: Among them, r(Φ) is the feature correlation calculation function, is the feature correlation coefficient matrix, is the feature covariance matrix, I κ is the identity matrix, and diag creates a diagonal matrix.

4. The method according to claim 1, characterized in that The method further comprises: During the training process of the collaborative filtering recommendation model, a skip probability is set for each graph neural network layer, and a Bernoulli random variable is sampled based on the skip probability in each training round; When the Bernoulli random variable takes the value of 1, the decoupling loss calculation of the layer is skipped; The skip probability is 0.9 in the initial training round, decreases linearly with each training round, and is 0 in the tenth training round.

5. The method according to claim 4, characterized in that: If the Bernoulli random variable of the graph neural network layer does not take a value of 1 and the user-side feature correlation exceeds the threshold, a penalty weight is assigned to the user side according to the following formula: and / or If the Bernoulli random variable of the graph neural network layer does not take a value of 1 and the item-side feature correlation exceeds the threshold, a penalty weight is assigned to the item side according to the following formula: Among them, 1 represents the indicator function, S represents the total number of layers, s represents the current layer number, q t Indicates whether to skip the decoupling calculation of this layer. Indicates the penalty weight assigned to the user side of the current layer, Indicates the penalty weight assigned to the item side of the current layer, represents the user embedding representation matrix of the current layer, Represents the user embedding representation matrix of each layer, Represents the item embedding representation matrix of the current layer, represents the item embedding representation matrix of each layer, τ(·) represents the target correlation threshold calculation function, and r(·) represents the feature correlation calculation function.

6. The method according to claim 5, characterized in that The method further comprises: When the user side satisfies the first rule, the penalty weight assigned to the user side of the current layer is 0; and / or When the item side satisfies the second rule, the penalty weight allocated to the item side in the current layer is 0.

7. The method according to claim 1, characterized in that: Determining the total decoupling loss according to the feature correlation and the penalty weight includes: For each graph neural network layer that needs to calculate the decoupling loss, its decoupling loss is calculated by the following formula: Layer decoupling loss = (user-side feature relevance × user-side penalty weight) + (item-side feature relevance × item-side penalty weight); The layer decoupling losses of all the graph neural network layers that need to calculate the decoupling loss are added together to obtain the total decoupling loss.

8. The method according to claim 1, characterized in that: Adding the total decoupling loss to the original loss function of the collaborative filtering recommendation model to generate a total loss function includes: The total decoupling loss is added to the original loss function of the collaborative filtering recommendation model based on the following formula: L=L CF +βL AFD Among them, L is the total loss function, L CF is the original loss function, L AFD is the total decoupling loss and β is the balance coefficient.

9. The method according to claim 8, characterized in that The training of the collaborative filtering recommendation model based on the total loss function further includes: The balance coefficient is dynamically adjusted based on the performance indicator of the collaborative filtering recommendation model on the validation set.

10. The method according to claim 9, characterized in that: The dynamically adjusting the balance coefficient based on the performance indicator of the collaborative filtering recommendation model on the validation set includes: If the performance indicator improvement exceeds a first preset threshold, increase β; If the performance indicator decreases by more than a second preset threshold, reducing β; If the performance indicator changes within a preset range, the current β value remains unchanged.

Images