CN116611731A - A scoring model training method, user push method and device - Google Patents

Abstract

The embodiment of this specification provides a scoring model training method, a user push method and a device. In the scoring model training method, construct the interactive relationship network between the user and the push object, estimate the node representation continuous distribution of the user and the push object through the variational graph reasoning network, and sample the continuous distribution of the node representation, and obtain as a comparison A first sampled representation and a second sampled representation of an object, and a node-level contrastive loss function is determined based on the contrastive object. At the same time, the continuous distribution of node representations is used to reconstruct the relationship network between users and push objects, so as to obtain the reconstruction loss function. Cluster users and push objects using node representation continuous distribution, and calculate a cluster-aware contrastive loss function based on the clustering results. Combine multiple loss functions for multi-task learning, and update the model until convergence. Use the trained scoring model to calculate the user's score on the push object, and push the user based on the score.

Description

A scoring model training method, user push method and device

technical field

One or more embodiments of this specification relate to the field of computer technology, and in particular to a scoring model training method, a user push method and a device.

Background technique

With the development of society and the advancement of technology, more and more service platforms have emerged to provide users with various services. Many service platforms can provide users with convenient online services through servers and clients. For example, e-commerce platforms can provide users with various commodity information for users to browse, select and purchase; content platforms can provide users with favorite e-books, articles, music and videos, etc. In order to provide richer services, the service platform will push richer push objects for users to choose from according to the user's historical behavior records on the premise of ensuring the user's data privacy and security after obtaining the user's authorization. To a certain extent, it provides convenience to users.

At present, it is hoped that there will be an improved solution that can provide users with push notifications more accurately and reasonably.

Contents of the invention

One or more embodiments of this specification describe a scoring model training method, user push method and device, which can provide push to users more accurately and reasonably. Concrete technical scheme is as follows.

In the first aspect, the embodiment provides a method for training a scoring model, the scoring model is used to predict the scores of several users to push objects to be pushed; the scoring model is based on a relationship network comprising a plurality of nodes For training, a plurality of nodes include nodes representing users and nodes representing push objects, and edges representing connection relationships between nodes; the parameters to be learned of the scoring model include node representations; the method includes:

Through a graph neural network, based on the node representations of multiple nodes and the original adjacency matrix of the relationship network, determine the aggregation representation of the node; wherein, the aggregation representation aggregates neighbor user features and/or neighbor push object features;

Based on the aggregate representation and random parameters of nodes, construct a continuous distribution of node representations, and by determining the value of the random parameters, sample the continuous distribution of node representations to obtain the first sampling representation and the second sampling representation as the comparison target. Sampling representation; wherein, the node representation continuous distribution includes: a node representation continuous distribution containing user characteristics and a node representation continuous distribution containing push object characteristics;

Determining a first loss based on the difference between the first sampled representation and the second sampled representation for the same node, and the difference between the first sampled representation and the second sampled representation for different nodes, based on said first loss determines a predicted loss;

The node representations of a plurality of nodes are updated in the direction of reducing the prediction loss.

In one embodiment, the graph neural network is implemented using a graph convolutional network; the graph convolutional network includes several convolutional layers; the step of determining the aggregation representation of a node includes:

Any convolutional layer outputs the intermediate representation of any node in the following way: based on the intermediate representation of the node's neighbor nodes output by the previous convolutional layer, determine the intermediate representation of the node in the convolutional layer;

For any node, based on the intermediate representations of the node output by several convolutional layers, the aggregated representation of the node is determined.

In one embodiment, the node represents a continuous distribution conforming to a Gaussian distribution;

The step of constructing a node to represent a continuous distribution includes:

Taking the aggregated representation as a mean value; determining a corresponding variance based on the mean value;

Based on the mean value, the variance and the random parameter, a continuous distribution of node representation conforming to the Gaussian distribution is constructed.

In one embodiment, the step of constructing nodes conforming to Gaussian distribution to represent continuous distribution includes:

Using the variance as the coefficient of the random parameter, constructing the first term of the node representing the continuous distribution;

Based on the sum of the first term and the mean value, the node representation continuous distribution is constructed.

In one embodiment, the random parameter is random Gaussian noise with a mean of 0 and a variance of 1.

In one embodiment, the step of determining the predicted loss based on the first loss includes:

determining a second loss based on the difference between the mean and the variance and the mean and variance of a preset Gaussian distribution, respectively;

A predicted loss is determined based on the first loss and the second loss.

In one embodiment, after constructing the node representation continuous distribution, the method further includes:

Predicting the probability of interconnection between multiple nodes based on the node representation continuous distribution;

determining a reconstructed adjacency matrix of the relational network based on the probability;

The step of determining a predicted loss based on the first loss includes:

determining a third loss based on the difference between the reconstructed adjacency matrix and the original adjacency matrix;

A predicted loss is determined based on the first loss and the third loss.

In one embodiment, the step of predicting the probability of interconnection between multiple nodes includes:

For any first node and second node, the similarity between the node representation continuous distribution of the first node and the node representation continuous distribution of the second node is used as the input parameter of the activation function, and the obtained output The value is used as the probability of interconnection between the first node and the second node.

In one embodiment, after constructing the node representation continuous distribution, the method further includes:

Based on the continuous distribution of the node representation, clustering the multiple nodes to obtain the clusters to which the multiple nodes belong respectively;

The step of determining a predicted loss based on the first loss includes:

Determining a fourth loss based on a difference between the first sampled representation and the second sampled representation of nodes in a homogeneous cluster and a difference between the first sampled representation and the second sampled representation of nodes in a heterogeneous cluster ;

A predicted loss is determined based on the first loss and the fourth loss.

In one embodiment, the step of determining the predicted loss based on the first loss includes:

Based on the first loss, the prediction loss is determined using a contrastive loss function.

In the second aspect, the embodiment provides a method of using a scoring model to push users, and the scoring model is trained using the method provided in the first aspect; the method includes:

Through the scoring model, based on the continuous distribution of the node representation of the user node and the continuous distribution of the node representation of the push object node, the similarity between the corresponding user and the push object is determined;

Through the scoring model, the similarity is input into an activation function as an input parameter, and the obtained output value is used as the user's rating of the pushed object;

Based on the user's ratings on several push objects, select the several push objects, and push the user based on the selection result.

In a third aspect, the embodiment provides a scoring model training device, the scoring model is used to predict the scores of several users to push objects to be pushed respectively; the scoring model is based on a relationship network comprising a plurality of nodes For training, a plurality of nodes include a node representing a user and a node representing a push object, and an edge representing a connection relationship between nodes; the parameters to be learned of the scoring model include a node representation; the device includes:

The aggregation module is configured to determine an aggregated representation of a node based on the node representations of multiple nodes and the original adjacency matrix of the relationship network through a graph neural network; wherein the aggregated representation aggregates neighbor user features and/or neighbors push object characteristics;

A building module configured to construct a continuous distribution of node representations based on the aggregate representation and random parameters of nodes, and sample the continuous distribution of node representations by determining the value of the random parameters to obtain the first A sampling representation and a second sampling representation; wherein, the node representation continuous distribution includes: a node representation continuous distribution containing user characteristics and a node representation continuous distribution containing push object characteristics;

a loss module configured to determine based on a difference between the first sampled representation and the second sampled representation of the same node, and a difference between the first sampled representation and the second sampled representation of a different node a first loss, determining a predicted loss based on the first loss;

An update module configured to update the node representations of a plurality of nodes in a direction to reduce the prediction loss.

In the fourth aspect, the embodiment provides a device for pushing users by using a scoring model, and the scoring model is trained using the method provided in the first aspect; the device includes:

The determination module is configured to determine the similarity between the corresponding user and the push object based on the continuous distribution of the node representation of the user node and the continuous distribution of the node representation of the push object node through the scoring model;

The scoring module is configured to input the similarity as an input parameter into an activation function through the scoring model, and use the obtained output value as the user's score on the pushed object;

The push module is configured to select the several push objects based on the user's ratings on the several push objects, and push the user based on the selection result.

In a fifth aspect, the embodiment provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed in a computer, the computer is instructed to execute any one of the first aspect to the second aspect. Methods.

In a sixth aspect, the embodiment provides a computing device, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, the first aspect to the second aspect are implemented. any one of the methods described.

In the method and device provided in the embodiments of this specification, the scoring model is trained using a relational network. During the training process, the continuous distribution of node representations is constructed based on the aggregate representation of nodes and random parameters, and the continuous distribution of node representations is sampled to obtain the first sampling representation and the second sampling representation as the comparison target. Based on the constructed comparison target, determine Prediction loss can use the sparse connection relationship in the network to construct comparison targets, and realize unsupervised learning of users and push objects, so that the scoring model can extract the deep valuable features of users and push objects. When pushing, it can provide users with push services more accurately and reasonably.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings that need to be used in the description of the embodiments. Apparently, the drawings in the following description are only some embodiments of the present invention, and those skilled in the art can obtain other drawings according to these drawings without creative efforts.

FIG. 1 is a schematic diagram of an implementation scenario of an embodiment disclosed in this specification;

Fig. 2 is a schematic flow chart of a scoring model training method provided by an embodiment;

Fig. 3 is a schematic flowchart of a method for pushing users by using a scoring model provided by an embodiment;

Fig. 4 is a schematic block diagram of a scoring model training device provided by an embodiment;

Fig. 5 is a schematic block diagram of an apparatus for pushing users by using a scoring model provided by an embodiment.

Detailed ways

The solutions provided in this specification will be described below in conjunction with the accompanying drawings.

Fig. 1 is a schematic diagram of an implementation scene of an embodiment disclosed in this specification. Figure 1 is also a schematic diagram of the training process of the scoring model. Among them, the relationship network contains multiple nodes and edges between nodes. The plurality of nodes includes a user node representing a user and a push object node representing a push object. In the example of the relationship network shown in FIG. 1 , circles represent user nodes, and squares represent push object nodes. The node representations of multiple nodes and the connection relationship between nodes are input into the graph neural network, and the aggregate representations of multiple nodes can be determined through the graph neural network. A continuous distribution of node representations can be constructed by using aggregated representations and random parameters, and the continuous distribution of node representations is sampled to obtain the first sampling representation and the second sampling representation. Based on the first sampling representation and the second sampling representation, a prediction loss can be obtained through comparative learning, and the node representation is updated using the prediction loss. Model training is performed iteratively until convergence. The relationship network shown in FIG. 1 includes the connection relationship between user nodes and user nodes, and the connection relationship between user nodes and push object nodes. The schematic diagram of the relationship network is only an example. In practical applications, the relationship network may only include the connection relationship between user nodes and push object nodes.

Wherein, the above-mentioned scoring model is used to predict the scores of several users on several pushing objects to be pushed respectively. To be pushed means to be pushed to the user, and the object to be pushed is the object waiting or preparing to be pushed to the user, which may include content such as commodities, articles, videos, or music. The user's rating on the push object reflects the user's interest in the push object. The service platform can provide users with push services in a server-client manner. On the user side, users can view the push objects pushed by the service platform through the client. The user involved in the embodiments of this specification refers to a user device, or a user device that has logged in a user account, that is, a user device where a client that has logged in a user account is located.

A relational network includes multiple nodes and multiple edges. Nodes in the relational network include user nodes and push object nodes, as well as edges between nodes. The edges between nodes reflect the connection relationship between nodes, including the edges between user nodes and user nodes, the edges between user nodes and push objects, etc. The connection relationship reflects the association relationship between different types of nodes. For example, the association relationship between user nodes may include family and friend relationship, like relationship and loan relationship, etc. The association relationship between the user node and the push object node may include click, purchase, view, distribution, affiliation, and so on. For example, when user a has clicked on commodity v, a connection edge may be formed between the user node corresponding to user a and the commodity node corresponding to commodity v. User-related data such as clicks, purchases, and viewings between push objects are used after obtaining user authorization, and the privacy of user-related data will not be leaked during use.

Node representations can also be called node features, which include user features or push object features. The node representation of the user node includes user features, and the node representation of the push object includes the push object features. User features can be understood as abstract features extracted from user-related data. Push object features can be understood as abstract features extracted from push object features. A node representation can be represented by a vector, and multiple node representations can be represented by a matrix composed of multiple vectors.

The users in the relationship network are the users of the service platform, and the push object may be the push object provided by the service platform. The number of users and push objects of the service platform is generally very large. The relationship network includes the association relationship between some users and some push objects. In order to predict the ratings of more users on more push objects, a scoring model can be trained based on information such as nodes with associated relationships in the relationship network, and the scoring model can be used to predict the ratings between all users and all push objects in the relationship network. Data such as node representations are used as parameters to be learned in the scoring model, and through multiple rounds of iterative training, the scoring model can learn node representations from the relational network.

In order to learn deeper node representations in order to predict more reasonable ratings, an embodiment of this specification provides a scoring model training method, including the following steps: Step S210, through a graph neural network, based on node representations of multiple nodes and the original adjacency matrix of the relationship network to determine the aggregation representation of nodes; Step S220, based on the aggregation representation of nodes and random parameters, construct a continuous distribution of node representations, and sample the continuous distribution of node representations by determining the value of random parameters, to obtain The first sampling representation and the second sampling representation as the comparison target; step S230, based on the difference between the first sampling representation and the second sampling representation of the same node, and the difference between the first sampling representation and the second sampling representation of different nodes difference, determine the first loss, and determine the prediction loss based on the first loss; step S240, update the node representations of multiple nodes in the direction of reducing the prediction loss.

In this embodiment, the scoring model can build a continuous distribution of node representations, obtain comparison targets through sampling, and update node representations through comparison learning, thereby extracting valuable deep-level features of users and push objects, so that the scoring model can be used more Provide users with push services accurately and more reasonably.

The present embodiment will be described in detail below with reference to FIG. 2 .

Fig. 2 is a schematic flow chart of a scoring model training method provided in an embodiment. The method can be executed through the service platform. The service platform can be realized by any device, device, platform, device cluster, etc. that have computing and processing capabilities. The scoring model described above is trained on a relational network. Wherein, the multiple nodes include a node representing a user and a node representing a push object, and an edge representing a connection relationship between the nodes. That is, the nodes include user nodes and push object nodes. When referring to a node, it includes a user node and/or a pushed object node. The parameters to be learned of the scoring model include node representations.

In one embodiment, the scoring model may include a graph neural network and a variational encoder. Among them, the graph neural network is used to determine the aggregate representation of nodes based on the node representations of multiple nodes and the original adjacency matrix of the relational network. The variational encoder is used to construct node representation continuous distribution based on node aggregation representation and random parameters, and sample the node representation continuous distribution by determining the value of random parameters to obtain the first sampling representation and the second sampling as the comparison target characterization. In this case, the parameters to be learned of the scoring model may also include parameters to be learned in the graph neural network and/or parameters to be learned in the variational encoder. Graph neural networks and variational encoders can be collectively referred to as variational graph inference networks.

In one embodiment, the scoring model may include a variational encoder instead of a graph neural network. Graph neural networks can perform joint training with scoring models.

The scoring model can be trained through several model iteration processes, and any one model iteration process can include steps S210-S240 shown in FIG. 2 .

In step S210, the aggregate representation μ of the node is determined based on the node representations of multiple nodes and the original adjacency matrix A of the relational network G through the graph neural network. Wherein, the aggregation representation aggregates features of neighbor users and/or features of objects pushed by neighbors. The aggregation representation of the user node includes user characteristics, and the aggregation representation of the push object node includes the push object characteristics. The determined aggregation representation can be understood as determining the aggregation representation vector of each node, or determining the aggregation representation matrix of all nodes.

The following takes the push object as an example for illustration. Let U represent the user set, and U={u ₁ ,…,u _a ,…,u _b ,…,u _M }, u _a represents the a-th user, u _b represents the b-th user, and M represents the total number of users , 1≤a, b≤M. Let V represent the commodity set, and V={v ₁ ,…,v _i ,…,v _j ,…,v _N }, v _i represents the i-th commodity, v _j represents the j-th commodity, and N represents the total number of commodities , 1≤i, j≤N. Let r _ai represent the scoring value of the ath user u _a on the ith commodity v _i , then the user's scoring matrix R={ra _ai } _M×N for the commodity. If the a-th user u _a has behavior data on the i-th commodity v _i , then r _ai =1, otherwise r _ai =0. Among them, behavioral data includes watching, purchasing and clicking.

According to the above raw data, the user-to-product interaction graph G, that is, the relationship network G, can be constructed according to the following formula (1):

Among them, A is the original adjacency matrix of the relational network G, and U∪V represents the node set of the relational network G.

Initially, the node representation matrix E of each node in the relational network G can be randomly determined. When the relationship network contains M+N nodes, E can be represented by a (M+N)×d-dimensional matrix, that is, E={e ₁ ,…,e _a ,…,e _M ,…e _i ,…, e _M+N }. Among them, e _a represents the d-dimensional node representation vector of the ath user node. The determined aggregation representation μ can also be represented by a (M+N)×d-dimensional matrix, that is, μ={μ _i }, where μ _i represents the aggregation representation vector of the i-th node. The i-th node here can be a user node or a product node.

The node representation and the original adjacency matrix A are input into the graph neural network, and the graph neural network can output the aggregated representation μ of the node. Graph neural networks can be implemented using a variety of network models, such as graph convolutional networks, graph recurrent networks, or graph attention networks. The following uses the graph convolutional network as an example to illustrate the implementation of the graph neural network. When using the graph convolutional network to determine the aggregation representation of nodes, there may be multiple implementations, and the specific implementations are described below.

A graph convolutional network consists of L convolutional layers. Each convolutional layer can determine the intermediate representation of each node based on the output of the previous convolutional layer and the original adjacency matrix A. Based on the intermediate representations output by multiple convolutional layers, an aggregated representation of a node can be determined.

For any convolutional layer l+1, the intermediate representation of any node (such as the i-th node) can be output in the following way: Based on the intermediate representation of the neighbor node of the i-th node output by the previous convolutional layer l, determine The i-th node is represented in the middle of this convolutional layer l+1. Neighbor nodes of the i-th node may include nodes that have a connection relationship with the i-th node. Specifically, the weighted average of the intermediate representations of multiple neighboring nodes may be used as the intermediate representation of the i-th node. For example, the following formula (2) can be used to determine the intermediate representation μ _i l+1 of the i-th node output by the ^l+1th convolutional layer:

Among them, S _i represents the number of nodes in the node set that has a connection relationship with the i-th node. j takes a value in S _i , and S _j represents the number of nodes in the node set that has a connection relationship with the jth node. S _i and S _j can be determined based on the original adjacency matrix A. μ _j ^l is the intermediate representation output by the i-th node in the l-th convolutional layer. The above formula (2) is only an implementation manner, and it is feasible to modify the formula, such as removing or modifying the weighted value.

For any node, such as the i-th node, based on the intermediate representations of the i-th node output by several convolutional layers, the aggregated representation of the node is determined. Specifically, the mean or weighted mean of the intermediate representations of the i-th node output by several convolutional layers may be determined as the aggregated representation of the i-th node. For example, the following formula (3) can be used to obtain the aggregation representation output by the i-th node through the graph convolutional network:

Among them, μ _i represents the aggregation representation of the i-th node, and the dimension of the vector is d. For the aggregated representation matrix μ formed by the aggregated representations of all nodes, its matrix dimension is (M+N)×d.

Through the execution of step S210, based on the association relationship between users in the relationship network and the association relationship between users and push objects, the characteristics of its neighbor nodes are aggregated in the aggregation representation of user nodes, including neighbor user characteristics and and/or neighbor push object features; aggregate features of its neighbor nodes in the aggregate representation of push object nodes, including neighbor user features and/or neighbor push object features, so that the aggregate representation aggregates useful information in the relationship network.

In step S220, based on the aggregate representation of nodes μ and the random parameter ∈, a continuous distribution of node representations is constructed, and by determining the value of the random parameter ∈, the continuous distribution of node representations is sampled to obtain the first sampling representation Z as the comparison target ' and the second sample characterizes Z".

Wherein, the continuous distribution of node representations includes: the continuous distribution of node representations including user features and the continuous distribution of node representations including push object features. The continuous distribution of node representation here can be a vector of each node, or a matrix composed of vectors of all nodes. The first sample representation Z′ and the second sample representation Z″ obtained through sampling also include user features and/or push object features.

Step S220 utilizes the reparameterization technique to reconstruct the continuous distribution of node representations. This reparameterization can reconstruct the aggregated representation of the input, and exploit the randomness of random parameters to generate meaningful data, providing more samples for comparative learning. Continuous distributions can be Gaussian or other distributions such as uniform. The implementation process of step S220 will be described below by taking the continuous distribution of node representation conforming to the Gaussian distribution as an example.

When constructing the Gaussian distribution of node representation, the aggregate representation can be used as the mean value, and the corresponding variance can be determined based on the mean value; then, based on the mean value, variance and random parameters, a continuous distribution of node representation conforming to the Gaussian distribution can be constructed. Various implementations can be included when constructing node representation continuous distributions based on mean, variance, and random parameters. For example, the variance can be used as the coefficient of the random parameter to construct a node representing the first item of the continuous distribution, and based on the sum of the first item and the mean value, the node is constructed to represent the continuous distribution.

When constructing a Gaussian distribution, for each node's node representation, a node representation continuous distribution corresponding to the node can be constructed. In the following description, when the mean, variance and random parameters are used for calculation, the mean, variance and random parameters of a certain node are used, and the three correspond to each other. A pair of mean and variance defines a Gaussian distribution.

When determining the corresponding variance based on the mean value, it may be determined by referring to an existing conversion formula between the mean value and the variance. For example, the following formula (4) can be used to determine the variance:

σ＝exp(μW+b) (4)

Among them, W and b are parameters to be learned. exp is an exponential function with a natural constant e as the base. μ is the mean and σ is the variance. Considering that there is exp in the density function of Gaussian distribution, in order to eliminate the exp in the calculation result of formula (4), the commonly used conversion formula between variance and mean is appropriately modified. In formula (4), when μ is the mean value vector of a certain node, σ is the variance vector of the node. When μ is the mean matrix formed by the mean vectors of all nodes, σ is the variance matrix formed by the variance vectors of all nodes.

When constructing a node to represent a continuous distribution, it can be carried out according to the following formula (5):

Z＝μ+σ˙∈ (5)

Among them, Z represents the continuous distribution of nodes, and σ˙∈ is the first term. ∈ is a random parameter. When μ is the mean value vector of a certain node and σ is the variance vector of the node, Z is the node representation continuous distribution vector of the node. When μ is the mean matrix formed by the mean vectors of all nodes, and σ is the variance matrix formed by the variance vectors of all nodes, Z is the matrix formed by the node representation continuous distribution vectors of all nodes.

In order to make the calculation simple, the continuous distribution of the node representation can be assumed to be a Gaussian distribution with a mean of 0 and a variance of 1. Correspondingly, the random parameter ∈ can also be random Gaussian noise with a mean of 0 and a variance of 1.

When determining the value of the random parameter ∈, the random parameter ∈ can be generated according to the corresponding rules. When two sampling representations need to be generated, two random parameters ∈ can be obtained. For example, when two random parameters ∈' and ∈" are generated, the comparison targets Z' and Z" in the following formula (6) can be obtained based on the above formula (5):

Z′=μ+σ˙∈′, Z″=μ+σ˙∈″ (6)

Wherein, Z' may be the first sampling representation, and Z" may be the second sampling representation.

When constructing a comparison target, two sampling representations may be obtained, but not limited to, or more sampling representations may be obtained.

When the Gaussian distribution is used for the continuous distribution of the node representation, before step S210, the Gaussian distribution may be used to randomly initialize the node representation, which can speed up the iterative process of the scoring model.

The foregoing illustrates an implementation of refactoring node representations to represent continuous distributions using the reparameterization technique. Some improvements can be made to these embodiments, and more embodiments can also be obtained. For example, the combination of the variance and the random parameter is not limited to multiplying the two to obtain the first item, and may also include many ways.

In the constructed node representation continuous distribution, the random parameter ∈ is randomly generated, and the variance is used as the coefficient of the random parameter ∈, which amplifies this randomness and makes the randomness of the node representation continuous distribution controllable . The random parameter ∈ brings randomness, and the variance determines the scale of this randomness, thus making this randomness controllable. Moreover, as the model iterates, the variance is gradually learned, which makes the magnitude of the noise also learned. Moreover, the variances of different nodes are different, that is, the variances of node representations of different user nodes are different, and the variances of node representations of different push object nodes are different, and the variances carry user characteristics and push object characteristics. When the variance is different, the size of the random parameter ∈ in the corresponding node representation continuous distribution is also different. The random parameter ∈ is affected by the variance of the node and becomes related to the node, that is, related to the user and the push object. When the variance is large, the corresponding random noise is large. When calculating the loss of the node, the difference value of the node will be relatively large, and the adjustment of the node representation of the node will be correspondingly large. In this way, the unique properties of users and push objects are considered, and the continuous distribution of generated node representations also takes into account the unique properties of users and push objects, which makes the comparison target constructed more efficient and reasonable.

In step S230, based on the difference between the first sampling representation and the second sampling representation of the same node, and the difference between the first sampling representation and the second sampling representation of different nodes, the first loss is determined, based on the first loss Determine the prediction loss. In step S240, node representations of multiple nodes are updated in the direction of reducing the prediction loss.

The first sampling representation Z' and the second sampling representation Z" of the same node should be as close as possible, or as similar as possible. The first sampling representation Z' and the second sampling representation Z" of different nodes should be as far away as possible, or as dissimilar as possible. The first sampled representation Z' and the second sampled representation Z" of the same node can be used as positive samples, and the first sampled representation Z' and the second sampled representation Z" of different nodes can be used as negative samples. The nodes here can be represented by node identifiers. Judging whether the two nodes are the same node may include judging whether the node identifiers are the same. For example, in the first sampled representation matrix and the second sampled representation matrix, the two vectors corresponding to the i-th node can be used as positive samples, and the vector of the i-th node and the vector of the j-th node can be used as negative samples, where i and j Are not the same.

When using the first loss calculated in step S230, it can be determined by using a contrastive loss function, and the first loss can be used as a part of the prediction loss. When determining the predicted loss based on the first loss, the first loss can be directly used as the predicted loss for model update, or the first loss can be used as a part of the predicted loss, and the first loss and losses determined by other methods can be used as the predicted loss.

The manner of calculating the prediction loss in step S230 may be called node-level contrastive loss. Nodes include user nodes and push object nodes. The following formulas (7) and (8) can be used to calculate the user-level comparative learning loss function L _N ^U and the push object-level comparative loss function L _N ^I :

where _τ1 is the contrast temperature, which is a pre-set hyperparameter. B _u is a user in a batch of nodes, and B _i is a push object in a batch of nodes. T is the matrix transpose notation. z _a ′ is the first sampling characterization vector of the ath user, z _a ″ is the second sampling characterization vector of the ath user. _{z i} ′ is the first sampling characterization vector of the i push object, z _i ″ is The second sampling representation vector of the i-th pushed object.

When determining the prediction loss, the second loss may also be determined based on the difference between the above mean and variance and the mean and variance of the preset Gaussian distribution, and the prediction loss may be determined based on the first loss and the second loss. When determining the predicted loss based on the first loss and the second loss, the sum of the first loss and the second loss may be used as the predicted loss, or the first loss, the second loss and losses determined in other ways may be used as the predicted loss.

Other ways of determining the prediction loss continue below. After constructing the node representation continuous distribution Z, the probability of interconnection between multiple nodes can also be predicted based on the node representation continuous distribution Z; based on this probability, the reconstructed adjacency matrix A' of the relationship network is determined.

When determining the prediction loss, the third loss can be determined based on the difference between the reconstructed adjacency matrix A' and the original adjacency matrix A, and the prediction loss can be determined based on the first loss and the third loss.

When predicting the probability of interconnection between multiple nodes, for any first node and second node, the similarity between the node representation continuous distribution of the first node and the node representation continuous distribution of the second node can be taken as The input parameter of the activation function, and the obtained output value is used as the probability of interconnection between the first node and the second node.

When predicting the probability of interconnection among multiple nodes, the third sampling representation Z ₃ can be obtained by sampling the continuous distribution of node representations by determining the value of the random parameter ∈. Based on the third sampled representation Z ₃ of the plurality of nodes, the probability of interconnection among the plurality of nodes is predicted. The third sampling representation Z ₃ may also use any one of the first sampling representation Z′ and the second sampling representation Z″, or a new third sampling representation Z ₃ may be obtained by randomly generating ∈.

For any first node and second node, the similarity between the third sampling characterization vector of the first node and the third sampling characterization vector of the second node can be used as the input parameter of the activation function, and the output value obtained will be As the probability of interconnection between the first node and the second node.

For example, the following formula (9) can be used to determine the probability p that any i-th node is connected to the j-th node:

Among them, z _i is the third sampling characterization vector of the i-th node, z _j is the third sampling characterization vector of the j-th node, and sigmoid is the activation function. For the third sampling representation matrix, z _i and z _j are the i-th vector and the j-th vector in the third sampling representation matrix, respectively. The value of p is between 0 and 1.

In one embodiment, the following formula (10) can be used to determine the difference between the reconstructed adjacency matrix A' and the original adjacency matrix A, and the mean and variance based on the above mean and variance respectively and the preset Gaussian distribution The difference between the resulting graph reconstruction loss function L _ELBO :

Among them, δ is the activation function. D _a is the training data of user a. For example, for user a, if user a has a connection relationship with 1, 2 and 3 among the five commodities 1 to 5, then user a and commodity 1, 2 and 3 constitute positive samples respectively, and user a and commodity 4 and 5 constitute negative samples. The first term on the right side of equation (10) is the loss function between adjacency matrices, and the second term is the loss function between Gaussian distributions.

In order to make the determined prediction loss more comprehensive and improve the iteration speed of the model, after constructing the node representation continuous distribution, the nodes can also be clustered to determine the basic comparative loss function of the cluster.

Specifically, multiple nodes can be clustered based on the continuous distribution Z of the node representation, and clusters to which multiple nodes belong can be obtained.

When determining the prediction loss, it can be based on the difference between the first sampled representation Z' and the second sampled representation Z" of the same cluster nodes, and the difference between the first sampled representation Z' and the second sampled representation Z" of the heterogeneous cluster nodes , determine the fourth loss, determine the prediction loss based on the first loss and the fourth loss, etc.

When clustering multiple nodes using the node to represent the continuous distribution Z, you can use the node to represent any value in the continuous distribution Z, which can be the first sampling representation Z' or the second sampling representation Z", or resampling after the value.

When clustering, various existing clustering algorithms can be used, such as K-Means algorithm or mean shift algorithm. The following uses the K-Means algorithm as an example to illustrate the clustering process of nodes. During clustering, user nodes and push object nodes can be clustered separately.

Initially, the number of clusters of user nodes can be defined as K _u , and the number of clusters of push objects can be defined as K _i . Use the K-Means algorithm to determine the clustering prototype of user nodes as C ^u ＝{c _k ^u } _Ku , and the clustering prototype of push objects as C ⁱ ＝{c _k ⁱ } _Ki , and then get the clustering distribution of user nodes ∏ _a=1 ^M ∏ _k=1 ^Ku p(c _k ^u |z _a ), and the cluster distribution of push object nodes ∏ _i=M+1 ^M+N ∏ _k=1 ^Ki p(c _k ⁱ |z _i ). The clustering prototype is the clustering center, and one clustering prototype corresponds to one cluster. The cluster distribution includes the cluster distribution of user nodes and the cluster distribution of push object nodes. Cluster distributions can be represented in vector or matrix form. In one embodiment, the dimension of the cluster distribution vector is the total number of categories of the cluster, which category the node belongs to, the corresponding elements in the vector can take a value of 1, and the other elements can take a value of 0. The cluster distribution matrix is a matrix composed of the cluster distribution vectors of all nodes. For example, the rows in the matrix can be the cluster distribution vectors of users or push objects, and the columns are the dimensions of the cluster distribution vectors.

In one embodiment, after determining the cluster distribution of nodes, the probability that any two nodes belong to the same cluster can be determined based on the cluster distribution of multiple nodes, and the comparative loss at the cluster level can be determined based on the probability function.

For example, the following formula (11) can be used to determine the probability that the ath user and the bth user belong to the same cluster prototype, that is, the probability p(a,b) of belonging to the same cluster:

Among them, z _a is the vector of the node representing the ath user in the continuous distribution, and z _b is the vector of the node representing the bth user in the continuous distribution.

Similarly, the following formula (12) can be used to determine the probability that the i-th push object and the j-th push object belong to the same cluster prototype, that is, the probability p(i,j) of the same cluster:

Among them, z _i is the vector representing the i-th pushing object in the continuous distribution, and z _j is the vector representing the j-th pushing object in the continuous distribution.

The above formulas (11) and (12) for calculating probabilities are just examples. Appropriate modifications are made to these formulas, such as adding coefficients or weights, etc., to obtain new implementations.

The same cluster nodes include different user nodes belonging to the same cluster and different push object nodes belonging to the same cluster. The heterogeneous cluster nodes include different user nodes belonging to different clusters and different push object nodes belonging to different clusters. In the first sampling representation Z′ and the second sampling representation Z″, for different users a and b belonging to the same cluster, the first sampling representation vector of user a and the second sampling representation vector of user b belong to positive samples. For For different users a and b of different clusters, the first sampling characterization vector of user a and the second sampling characterization vector of user b belong to negative samples. For different pushing objects i and j belonging to the same cluster, the first sampling characterization vector of pushing object i The sampled characterization vector and the second sampled characterization vector of the push object j belong to the positive samples. For different push objects i and j belonging to different clusters, the first sample characterization vector of the push object i and the second sample characterization vector of the push object j belong to Negative samples. Based on the difference between the first sampled representation Z' and the second sampled representation Z" of the nodes of the same cluster, and the difference between the first sampled representation Z' and the second sampled representation Z" of the heterogeneous cluster nodes, When determining the fourth loss, a comparison loss function may be used for determination.

For example, the following formula (13) can be used to determine the comparison loss function L _C ^U of the user clustering level, and the formula (14) can be used to determine the comparison loss function L _C ^I of the push object clustering level:

where _τ2 is the contrast temperature, which is a pre-set hyperparameter. B _u is a user in a batch of nodes, and B _i is a push object in a batch of nodes. z _a ′ is the first sampling characterization vector of the a-th user, z _b ″ is the second sampling characterization vector of the b-th user. _{z i} ′ is the first sampling characterization vector of the i-th push object, and z _j ″ is The second sample representation vector of the jth pushed object. SP(a) is the sum of the number of positive samples of the molecular part in formula (13), SP(i) is the sum of the number of positive samples of the molecular part in formula (14), adding the reciprocal of these two parameters makes formula (13) and (14) are at the same node level as formulas (7) and (8). Formulas (13) and (14) are just one implementation, and different implementations can be obtained by modifying them.

In an implementation manner, based on the sum of the first loss, the second loss and the third loss, the total prediction loss may be determined, and then the node representation waiting learning parameters are updated according to the total prediction loss. For example, based on formulas (7) and (8) and formulas (13) and (14), the total contrastive loss function L _cl can be determined using the following formula (15):

L _cl ＝(L _N ^U +L _N ^I )+γ(L _C ^U +L _C ^I ) (15)

Wherein, γ is a preset weight coefficient, which can be set according to experience.

Based on formula (15) and formula (10), the following formula (16) can be used to establish a multi-task optimization goal, that is, to determine the total loss function L(θ):

L(θ)＝ L _ELBO +αL _cl +β‖E‖ ² (16)

Among them, E is the node representation, and α and β are used to control the weight of the comparison loss function and the regularization item, respectively. ‖E‖ ² is to find the L2 norm of E. θ=[E, W, b] is the parameter to be learned. The above optimization objective can be solved by the gradient descent method, and the parameter θ can be updated.

The above steps S210-S240 are an iterative process of the scoring model. In one iteration process, a batch of nodes in the relational network can be taken for training. When the prediction loss of the scoring model is less than the threshold, or the number of iterations reaches the threshold, the model converges.

The above is the training process of the scoring model. In the above embodiment, the node representation is reconstructed by using aggregated representation and random parameters, and the reconstructed parameters are sampled, and a comparison target is constructed for comparative learning, so as to learn node representation from the relational network. When building a node to represent a continuous distribution, using the variance as the coefficient of the random parameter can make the process of parameter construction fully take into account the individuality of the nodes in the relationship network, thereby constructing a comparison target with lossless semantics and individualized nodes, which effectively improves the push rate. Effect.

In the above-mentioned embodiment, the user's historical interaction data is deeply mined through the scoring model, the high-order potential preferences of the user and the push object are modeled by using the graph neural network, and the continuous distribution of node representations is accurately inferred. Moreover, in cluster-aware contrastive learning, the similarity between nodes is calculated in an unsupervised manner, which further improves contrastive learning from the perspective of cluster distribution and effectively alleviates the problem of sparse interaction data in relational networks.

Fig. 3 is a schematic flowchart of a method for pushing users by using a scoring model provided by an embodiment. The scoring model is trained using the method provided in the embodiment shown in FIG. 2 . The method includes the following steps.

Step S310, through the scoring model, based on the continuous distribution of node representations of user nodes and the continuous distribution of node representations of push object nodes, determine the similarity between the corresponding user and the push object.

Step S320, through the scoring model, the similarity is input into the activation function as an input parameter, and the obtained output value is used as the user's rating on the pushed object.

For example, the following formula (17) can be used to predict the rating value r _ai ′ of the a-th user on the i-th pushed object:

r _ai ′=sigmoid(z _a ^T z _i ) (17)

Among them, z _a ^T z _i is the similarity between the ath object and the ith push object, and sigmoid is the activation function. Through the above formula, the prediction score matrix R'={r _ai '} _M×N can be obtained.

Step S330, based on the user's ratings on several push objects, select several push objects, and push the user based on the selection result. When selecting, several push objects with the highest scores may be pushed to the user.

In this embodiment, when determining the similarity between the user and the pushed object, the continuous distribution of node representation is used as the node vector to participate in score prediction. Compared with node representation, the continuous distribution of node representation aggregates more neighbor node information, including deeper user characteristics and/or push object characteristics, which can make scoring more accurate and push more reasonable.

In other embodiments, the node representation trained by the scoring model may also be directly used to participate in the scoring prediction.

In this specification, the "first" in words such as the first sampling representation and the first node, and the corresponding "second" (if any) in the text are only for the convenience of distinction and description, and do not have any limiting meaning.

While the foregoing describes certain embodiments of the specification, other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. Multitasking and parallel processing are also possible, or may be advantageous, in certain embodiments.

Fig. 4 is a schematic block diagram of a scoring model training device provided in an embodiment. The scoring model is used to predict the ratings of several users on several push objects to be pushed; the scoring model is trained based on a relationship network comprising a plurality of nodes, and the multiple nodes include nodes representing users and nodes representing push objects, and edges representing connection relationships between nodes; the parameters to be learned of the scoring model include node representations. The device can be deployed in a service platform, and the service platform can be realized by any device, device, platform, device cluster, etc. that have computing and processing capabilities. This device embodiment corresponds to the method embodiment shown in FIG. 2 . The device 400 includes:

The aggregation module 410 is configured to determine an aggregated representation of a node based on the node representations of multiple nodes and the original adjacency matrix of the relationship network through a graph neural network; wherein the aggregated representation aggregates neighbor user features and/or Neighbors push object characteristics;

The construction module 420 is configured to construct a continuous distribution of node representations based on the aggregated representations and random parameters of nodes, and sample the continuous distribution of node representations by determining the value of the random parameters to obtain the first comparison target A sampling representation and a second sampling representation; wherein, the node representation continuous distribution includes: a node representation continuous distribution containing user characteristics and a node representation continuous distribution containing push object characteristics;

a loss module 430 configured to be based on a difference between the first sampled representation and the second sampled representation of the same node, and a difference between the first sampled representation and the second sampled representation of different nodes, determining a first loss, determining a predicted loss based on the first loss;

The update module 440 is configured to update the node representations of multiple nodes in the direction of reducing the prediction loss.

In one embodiment, the graph neural network is implemented using a graph convolutional network; the graph convolutional network includes several convolutional layers; the aggregation module 410 includes a first determination sub-module and a second determination sub-module; (Fig. not shown in

The first determination sub-module is configured such that any convolutional layer outputs the intermediate representation of any node in the following manner: based on the intermediate representation of the node’s neighbors output by the previous convolutional layer, determine the node’s position in the convolutional layer intermediate representation;

The second determination sub-module is configured to, for any node, determine the aggregate representation of the node based on the intermediate representations of the node output by several convolutional layers.

In one embodiment, the node characterizes a continuous distribution conforming to a Gaussian distribution; the construction module 420 includes a third determination submodule and a first construction submodule; (not shown in the figure)

The third determining submodule is configured to use the aggregation representation as a mean value, and determine the corresponding variance based on the mean value;

The first construction submodule is configured to construct a continuous distribution of node representation conforming to Gaussian distribution based on the mean value, the variance and the random parameter.

In one embodiment, the first construction submodule is specifically configured as:

Using the variance as the coefficient of the random parameter, constructing the first term of the node representing the continuous distribution;

Based on the sum of the first term and the mean value, the node representation continuous distribution is constructed.

In one embodiment, the random parameter is random Gaussian noise with a mean of 0 and a variance of 1.

In one embodiment, the loss module 430, when determining the predicted loss based on the first loss, includes:

determining a second loss based on the difference between the mean and the variance and the mean and variance of a preset Gaussian distribution, respectively;

A predicted loss is determined based on the first loss and the second loss.

In one embodiment, the device 400 further includes:

A reconstruction module (not shown in the figure), configured to predict the probability of interconnection between multiple nodes based on the node representation continuous distribution after constructing the node representation continuous distribution; based on the probability, determine the relationship network The reconstructed adjacency matrix of ;

The loss module 430, when determining the predicted loss based on the first loss, includes:

determining a third loss based on the difference between the reconstructed adjacency matrix and the original adjacency matrix;

A predicted loss is determined based on the first loss and the third loss.

In one embodiment, the reconstruction module, when predicting the probability of interconnection between multiple nodes, includes:

For any first node and second node, the similarity between the node representation continuous distribution of the first node and the node representation continuous distribution of the second node is used as the input parameter of the activation function, and the obtained output The value is used as the probability of interconnection between the first node and the second node.

In one embodiment, the device 400 further includes:

The clustering module is configured to cluster multiple nodes based on the continuous distribution of node representations after constructing the continuous distribution of node representations, to obtain clusters to which multiple nodes belong respectively;

The loss module 430, when determining the predicted loss based on the first loss, includes:

Determining a fourth loss based on a difference between the first sampled representation and the second sampled representation of nodes in a homogeneous cluster and a difference between the first sampled representation and the second sampled representation of nodes in a heterogeneous cluster ;

A predicted loss is determined based on the first loss and the fourth loss.

In an implementation manner, the loss module 430, when determining the predicted loss based on the first loss, includes:

Based on the first loss, the prediction loss is determined using a contrastive loss function.

Fig. 5 is a schematic block diagram of an apparatus for pushing users by using a scoring model provided by an embodiment. The scoring model is trained using the method provided in the embodiment shown in FIG. 2 . The apparatus 500 may be deployed on a service platform, and the service platform may be implemented by any apparatus, device, platform, device cluster, etc. that have computing and processing capabilities. This device embodiment corresponds to the method embodiment shown in FIG. 3 . Device 500 includes:

The determination module 510 is configured to determine the similarity between the corresponding user and the push object based on the continuous distribution of the node representation of the user node and the continuous distribution of the node representation of the push object node through the scoring model;

The scoring module 520 is configured to use the scoring model to input the similarity as an input parameter into an activation function, and use the obtained output value as the user's score on the pushed object;

The push module 530 is configured to select the several push objects based on the user's ratings on the several push objects, and push the user based on the selection result.

The foregoing device embodiments correspond to the method embodiments, and for specific descriptions, refer to the description of the method embodiments, and details are not repeated here. The device embodiment is obtained based on the corresponding method embodiment, and has the same technical effect as the corresponding method embodiment. For specific description, please refer to the corresponding method embodiment.

The embodiment of the present specification also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed in a computer, the computer is instructed to execute the method described in any one of FIG. 1 to FIG. 3 .

The embodiment of this specification also provides a computing device, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, the computer described in any one of Fig. 1 to Fig. 3 is implemented. described method.

Each embodiment in this specification is described in a progressive manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the storage medium and computing device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for relevant parts, please refer to part of the description of the method embodiments.

Those skilled in the art should be aware that, in the above one or more examples, the functions described in the embodiments of the present invention may be implemented by hardware, software, firmware or any combination thereof. When implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium.

The specific implementation manners described above further describe the purpose, technical solutions and beneficial effects of the embodiments of the present invention in detail. It should be understood that the above description is only the specific implementation of the embodiment of the present invention, and is not used to limit the protection scope of the present invention. Any modification and equivalent replacement made on the basis of the technical solution of the present invention , improvements, etc., should be included within the protection scope of the present invention.

Claims (15)

1. A training method of a scoring model, the scoring model is used to predict the scoring of several users to be pushed respectively to several push objects to be pushed; the scoring model is trained based on a relational network comprising a plurality of nodes, and a plurality of nodes Including a node representing a user and a node representing a push object, and an edge representing a connection relationship between nodes; the parameters to be learned of the scoring model include node representation; the method includes: Through a graph neural network, based on the node representations of multiple nodes and the original adjacency matrix of the relationship network, determine the aggregation representation of the node; wherein, the aggregation representation aggregates neighbor user features and/or neighbor push object features; Based on the aggregate representation and random parameters of nodes, construct a continuous distribution of node representations, and by determining the value of the random parameters, sample the continuous distribution of node representations to obtain the first sampling representation and the second sampling representation as the comparison target. Sampling representation; wherein, the node representation continuous distribution includes: a node representation continuous distribution containing user characteristics and a node representation continuous distribution containing push object characteristics; Determining a first loss based on the difference between the first sampled representation and the second sampled representation for the same node, and the difference between the first sampled representation and the second sampled representation for different nodes, based on said first loss determines a predicted loss; The node representations of a plurality of nodes are updated in the direction of reducing the prediction loss. 2. The method according to claim 1, the graph neural network is implemented using a graph convolutional network; the graph convolutional network includes several convolutional layers; the step of determining the aggregation representation of a node includes: Any convolutional layer outputs the intermediate representation of any node in the following way: based on the intermediate representation of the node's neighbor nodes output by the previous convolutional layer, determine the intermediate representation of the node in the convolutional layer; For any node, based on the intermediate representations of the node output by several convolutional layers, the aggregated representation of the node is determined. 3. The method according to claim 1, wherein the continuous distribution of the node representation conforms to a Gaussian distribution; The step of constructing a node to represent a continuous distribution includes: taking the aggregated characterization as the mean; determining a corresponding variance based on the mean; Based on the mean value, the variance and the random parameter, a continuous distribution of node representation conforming to the Gaussian distribution is constructed. 4. The method according to claim 3, the step of constructing a node characterizing a continuous distribution conforming to a Gaussian distribution, comprising: Using the variance as the coefficient of the random parameter, constructing the first term of the node representing the continuous distribution; Based on the sum of the first term and the mean value, the node representation continuous distribution is constructed. 5. The method according to claim 3, wherein the random parameter is random Gaussian noise with a mean value of 0 and a variance of 1. 6. The method of claim 3, said step of determining a predicted loss based on said first loss, comprising: determining a second loss based on the difference between the mean and the variance and the mean and variance of a preset Gaussian distribution, respectively; A predicted loss is determined based on the first loss and the second loss. 7. The method according to claim 1, after constructing the node representation continuous distribution, further comprising: Predicting the probability of interconnection between multiple nodes based on the node representation continuous distribution; determining a reconstructed adjacency matrix of the relational network based on the probability; The step of determining a predicted loss based on the first loss includes: determining a third loss based on the difference between the reconstructed adjacency matrix and the original adjacency matrix; A predicted loss is determined based on the first loss and the third loss. 8. The method according to claim 7, the step of predicting the probability of interconnection between multiple nodes, comprising: For any first node and second node, the similarity between the node representation continuous distribution of the first node and the node representation continuous distribution of the second node is used as the input parameter of the activation function, and the obtained output The value is used as the probability of interconnection between the first node and the second node. 9. The method according to claim 1, after constructing the node representation continuous distribution, further comprising: Based on the continuous distribution of the node representation, clustering the multiple nodes to obtain the clusters to which the multiple nodes belong respectively; The step of determining a predicted loss based on the first loss includes: Determining a fourth loss based on a difference between the first sampled representation and the second sampled representation of nodes in a homogeneous cluster and a difference between the first sampled representation and the second sampled representation of nodes in a heterogeneous cluster ; A predicted loss is determined based on the first loss and the fourth loss. 10. The method of claim 1, said step of determining a predicted loss based on said first loss, comprising: Based on the first loss, the prediction loss is determined using a contrastive loss function. 11. A method of pushing users by using a scoring model, wherein the scoring model is obtained by training according to the method of claim 1; the method comprises: Through the scoring model, based on the continuous distribution of the node representation of the user node and the continuous distribution of the node representation of the push object node, the similarity between the corresponding user and the push object is determined; Through the scoring model, the similarity is input into an activation function as an input parameter, and the obtained output value is used as the user's rating of the pushed object; Based on the user's ratings on several push objects, select the several push objects, and push the user based on the selection result. 12. A training device for a scoring model, the scoring model is used to predict the ratings of several users on several push objects to be pushed; the scoring model is trained based on a relationship network comprising a plurality of nodes, and the multiple nodes Including a node representing a user and a node representing a push object, and an edge representing a connection relationship between nodes; the parameters to be learned of the scoring model include node representation; the device includes: The aggregation module is configured to determine an aggregated representation of a node based on the node representations of multiple nodes and the original adjacency matrix of the relationship network through a graph neural network; wherein the aggregated representation aggregates neighbor user features and/or neighbors push object characteristics; A building module configured to construct a continuous distribution of node representations based on the aggregate representation and random parameters of nodes, and sample the continuous distribution of node representations by determining the value of the random parameters to obtain the first A sampling representation and a second sampling representation; wherein, the node representation continuous distribution includes: a node representation continuous distribution containing user characteristics and a node representation continuous distribution containing push object characteristics; a loss module configured to determine based on a difference between the first sampled representation and the second sampled representation of the same node, and a difference between the first sampled representation and the second sampled representation of a different node a first loss, determining a predicted loss based on the first loss; An update module configured to update the node representations of a plurality of nodes in the direction of reducing the prediction loss. 13. A device that utilizes a scoring model to push users, the scoring model is obtained by training according to the method of claim 1; the device comprises: The determination module is configured to determine the similarity between the corresponding user and the push object based on the continuous distribution of the node representation of the user node and the continuous distribution of the node representation of the push object node through the scoring model; The scoring module is configured to input the similarity as an input parameter into an activation function through the scoring model, and use the obtained output value as the user's score on the pushed object; The push module is configured to select the several push objects based on the user's ratings on the several push objects, and push the user based on the selection result. 14. A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed in a computer, it causes the computer to execute the method according to any one of claims 1-11. 15. A computing device, comprising a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, the method according to any one of claims 1-11 is implemented.