Cross-domain information pushing method and device and corresponding model training method and device

Abstract

According to the embodiment of the invention, a cross-domain information pushing technology concept is provided, after user representations of a source domain and a target domain are fused, filtering is carried out based on an information bottleneck theory, and data with relatively small relevance with the target domain is screened out, so that compression representations corresponding to the users are obtained. And matching the user compression representation with the information representation of the candidate information to obtain a prediction result (such as representation through the matching degree with each piece of candidate information). In a model training process, by minimizing model loss, undetermined parameters in a model can be adjusted at least towards the following targets: maximizing similarity between compression representation and representation corresponding to a user in a target domain; and the difference between the prediction result and the sample label is minimized. In this way, the cross-domain information pushing model based on the target adjustment is used for pushing the information of the target domain, and the accuracy of cross-domain information pushing can be improved.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of computer technology, and in particular, to cross-domain information push and corresponding model methods and devices. Background Art

[0002] Information push usually pushes information that meets the needs of users as much as possible based on data such as user behavior. Cross-domain recommendation (CDR) can use user behavior (e.g., global behavior accumulated in online service platforms) in a data-rich source domain (also called reference information domain) to learn transferable knowledge to help improve the recommendation of a data-sparse target domain (also called target information domain, such as a small program or micro application in an online service platform). By supplementing the target domain with relatively weak information, the recommendation effect of the target domain can be improved.

[0003] However, although the source domain information is relatively rich, not all source domain information is beneficial to the target domain tasks. For example, some accidental clicks and random clicks that users may have in the source domain are not an indication of user interest. If this type of information is also migrated to the target domain, it may have a negative impact on the target domain information push service.

[0004] Therefore, how to filter the source domain information and migrate valid data to the target domain as much as possible in the cross-domain information push business is a technical issue worthy of study. Summary of the invention

[0005] One or more embodiments of this specification describe a cross-domain information push and corresponding model method and device to solve one or more problems mentioned in the background technology.

[0006] According to a first aspect, a training method for a cross-domain information push model is provided, comprising multiple update cycles. In a single update cycle, the method comprises: obtaining a first representation of a current sample user in a source domain, a second representation in a target domain, and an information representation of candidate information by processing a fused knowledge graph, wherein the fused knowledge graph comprises a basic graph and a bipartite graph of user and candidate information; fusing the first representation and the second representation into a fused representation; performing information compression on the fused representation to determine a compressed representation of the current sample user, wherein the compressed representation is a weighted fusion result of the fused representation and a first noise, the first noise is determined based on a distribution sampling satisfied by each compressed representation of the current batch of sample users, and a weighted weight of the first noise is determined by a Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter; determining a model loss to update pending parameters in the cross-domain information push model, wherein the model loss comprises a representation loss and a prediction loss, the representation loss is determined by comparing the compressed representation with the first representation and the second representation, and the prediction loss is determined by a prediction result obtained by predicting the compressed representation in combination with the information representation.

[0007] In one embodiment, the method of obtaining a first representation of the current sample user in the source domain, a second representation in the target domain, and information representation of the candidate information by processing the fused knowledge graph includes: using a graph neural network to perform feature fusion on a fused knowledge graph consisting of a basic graph and a bipartite graph in the source domain, and obtaining information representations corresponding to the first representation and each piece of candidate information in the source domain; using a graph neural network to perform feature fusion on a fused knowledge graph consisting of a basic graph and a bipartite graph in the target domain, and obtaining information representations corresponding to the second representation and each piece of candidate information in the target domain.

[0008] In one embodiment, fusing the first representation and the second representation into a fused representation includes: processing the first representation and the second representation in one of a weighted, averaged, and summed manner to obtain the fused representation.

[0009] In one embodiment, the distribution satisfied by each compressed representation is a normal distribution determined by using the corresponding element in the mean matrix of each compressed representation as the mean and the corresponding element in the variance matrix as the variance.

[0010] In one embodiment, the prediction result is the matching scores corresponding to the current sample user on each candidate information, and the prediction loss is described by one of the following cosine similarity, Euclidean distance, cross entropy, and KL divergence between two vectors to express the degree of deviation between the two vectors: the matching vector composed of each matching score and the selected vector determined according to the actual selection of the candidate information by the current sample user.

[0011] In a further embodiment, in the target domain prediction result, the matching score corresponding to a single candidate information is determined by fusing the compressed representation with the second representation, and determining the similarity between the fusion result and the information representation of the single candidate information as the matching score.

[0012] In one embodiment, the model loss also includes filtering loss, and the model loss is a weighted sum of the compression loss, the prediction loss, and the filtering loss. The filtering loss is the mutual information between the better subgraph from the source domain to the target domain and the source domain subgraph under the condition of the target domain subgraph.

[0013] In one embodiment, the sum of the weighted weight of the fused representation and the weighted weight of the first noise is 1.

[0014] According to the second aspect, a method for information push is provided, which is used to use user information in a source domain to assist in information push in a target domain, the method comprising: obtaining a third representation of the current user in the source domain, a fourth representation in the target domain, and each information representation corresponding to each candidate information in the target domain through processing of a fused knowledge graph, the fused knowledge graph comprising a basic graph and a bipartite graph of users and candidate information; fusing the third representation and the fourth representation into a first fused representation; performing information compression on the first fused representation to determine a first compressed representation of the current user, wherein the first compressed representation is a weighted fusion result of the first fused representation and second noise, the second noise is determined based on distribution sampling satisfied by the compressed representation of training samples, and the weighted weight of the second noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter; based on matching the first compressed representation with each information representation, determining the target information to be pushed from each candidate information in the target domain.

[0015] In one embodiment, the process of processing the fused knowledge graph to obtain the third representation of the current user in the source domain, the fourth representation in the target domain, and the information representations corresponding to the candidate information in the target domain includes: using a graph neural network to perform feature fusion on the fused knowledge graph composed of the basic graph and the source domain bipartite graph to obtain the third representation; using a graph neural network to perform feature fusion on the fused knowledge graph composed of the basic graph and the target domain bipartite graph to obtain the fourth representation and the information representations corresponding to the candidate information in the target domain.

[0016] In one embodiment, the distribution satisfied by each compressed representation of the training samples is: a normal distribution determined by using each element in the mean matrix of each compressed representation of the training samples as the mean and the corresponding element in the variance matrix as the variance.

[0017] In one embodiment, fusing the third representation and the fourth representation into the first fused representation includes: processing the third representation and the fourth representation in one of a weighted, averaged, and summed manner to obtain the first fused representation.

[0018] In one embodiment, determining the target information to be pushed from the candidate information in the target domain based on the matching of the first compression representation with the various information representations includes: determining the matching scores respectively corresponding to the first compression representation and the various information representations; and selecting a number of candidate information as the target information to be pushed to the current user according to the matching scores.

[0019] According to a third aspect, a training device for a cross-domain information push model is provided, which is used to train the cross-domain information push model through multiple update cycles. The device includes an acquisition unit, a fusion unit, a compression unit, and an adjustment unit. In a single update cycle:

[0020] The acquisition unit is configured to acquire a first representation of the current sample user in the source domain, a second representation in the target domain, and information representation of the candidate information by processing the fused knowledge graph, wherein the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information;

[0021] The fusion unit is configured to fuse the first representation and the second representation into a fused representation;

[0022] The compression unit is configured to perform information compression on the fused representation to determine a compressed representation of the current sample user, wherein the compressed representation is a weighted fusion result of the fused representation and a first noise, the first noise is determined based on the distribution sampling satisfied by each compressed representation of the current batch of sample users, and the weighted weight of the first noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter;

[0023] The adjustment unit is configured to determine a model loss to update pending parameters in the cross-domain information push model, the model loss including a representation loss and a prediction loss, the representation loss being determined by comparing the compressed representation with the second representation, and the prediction loss being determined by a prediction result obtained by predicting the compressed representation in combination with the information representation.

[0024] According to a fourth aspect, there is provided an information push device, which is used to assist in information push in a target domain by using user information in a source domain, and the device includes:

[0025] an acquisition unit configured to acquire, by processing a fused knowledge graph, a third representation of the current user in a source domain, a fourth representation in a target domain, and respective information representations corresponding to respective candidate information in the target domain, wherein the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information;

[0026] a fusion unit, configured to fuse the third representation and the fourth representation into a first fused representation;

[0027] a compression unit, configured to perform information compression on the first fused representation to determine a first compressed representation of the current user, wherein the first compressed representation is a weighted fusion result of the first fused representation and a second noise, the second noise is determined based on a distribution sampling satisfied by the compressed representation of the training sample, and the weighted weight of the second noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter;

[0028] The matching unit is configured to determine the target information to be pushed from each candidate information in the target domain based on the matching between the first compression representation and each information representation.

[0029] According to a fifth aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method described in the first aspect or the second aspect.

[0030] According to a sixth aspect, a computing device is provided, comprising a memory and a processor, wherein an executable code is stored in the memory, and when the processor executes the executable code, the method described in the first aspect or the second aspect is implemented.

[0031] Through the method and device provided by the embodiments of this specification, in the cross-domain information push scenario, the source domain bipartite graph and the target domain bipartite graph are respectively fused with the basic knowledge graph to form a corresponding fused knowledge graph. Based on the processing of the fused knowledge graph, the user representation of the user in the source domain and the target domain, as well as the information representation of the candidate information, are determined respectively, so that the user representation of the source domain and the target domain are fused and then the information bottleneck theory is used to filter out the data with less correlation with the target domain to obtain the compressed representation corresponding to the user. Then the user compressed representation is matched with the information representation of the candidate information to obtain the prediction result (such as determining the degree of matching with each candidate information). During the model training process, the model loss of each update cycle may include prediction loss and compression loss. By minimizing the model loss, the undetermined parameters in the model can be adjusted toward the following goals: maximize the similarity between the compressed representation and the second representation; minimize the gap between the prediction result and the sample label. During the information push process, the prediction result obtained by the cross-domain information push model adjusted according to the above goals is used to determine the target information to be pushed to the corresponding user.

[0032] In this way, the information push process under the technical conception of this specification can use the user data in the source domain to push information to the target domain where the data is sparse, and use the information bottleneck theory to filter out the information in the source domain that provides less gain to the target domain in the process of determining the target information, reduce the impact of invalid information, and push information to the target domain more accurately, thereby improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0034] Figure 1 A schematic diagram of a specific scenario of this specification is shown;

[0035] Figure 2 A schematic diagram of a graph neural network training architecture under conventional techniques is shown;

[0036] Figure 3 A schematic diagram showing entities and association relationships in a knowledge graph;

[0037] Figure 4 A schematic diagram of a cross-domain information push model and its training architecture according to the technical concept of this specification is shown;

[0038] Figure 5 A schematic diagram of the training process of a cross-domain information push model according to an embodiment of this specification is shown;

[0039] Figure 6 A schematic diagram of a cross-domain information push process according to an embodiment of the present specification is shown;

[0040] Figure 7 A schematic block diagram showing a training device for a cross-domain information push model according to an embodiment of the present specification;

[0041] Figure 8 A schematic block diagram of a cross-domain information push device according to an embodiment of the present specification is shown. DETAILED DESCRIPTION

[0042] The technical solution provided in this specification is described below in conjunction with the accompanying drawings.

[0043] Figure 1 A schematic diagram of a specific application scenario of cross-domain product recommendation in this specification is shown. Figure 1As shown, the application scenario is an application scenario of an online service platform, which can be a social platform, a payment platform, etc., which can support the embedding of multiple micro-applications or applets, such as Figure 1 Micro Application 1, Mini Program 1, Mini Program 2, etc. The online service platform can connect the user IDs of various micro applications or mini programs. For example, the user ID of user 1 is a mobile phone number or identity information code. Figure 1 As shown, it is assumed that Mini Program 1 and Mini Program 2 are both information push applications, such as online shopping applications, food delivery applications, rental applications, news information applications, etc. A user can access multiple micro applications or mini programs through a terminal device.

[0044] exist Figure 1 In the example, user 1 can have multiple access records (which may include browsing, clicking links, purchasing products, etc.) through mini program 1, such as records of browsing products 1, product 2, etc. When the user visits mini program 2, mini program 2 can push corresponding candidate information to user 1. The candidate information in mini program 2 may include product a, product b (it may also be document d, news f, picture e), etc. In practice, the candidate information in mini program 2 and the candidate information in mini program 1 may have completely different content systems, description styles, etc. For example, the candidate information of mini program 1 is products, and the candidate information in mini program 2 is news, etc. In the scenario where products are pushed to user 1 through mini program 2 with sparse historical data, mini program 1 can be used as the source domain and mini program 2 as the target domain. By utilizing the data of user 1 in mini program 1, information can be pushed to user 1 in mini program 2, which is cross-domain recommendation.

[0045] Conventional cross-domain recommendation models usually use rich source domain data to assist in modeling the data-sparse target domain, obtain representations of user and candidate information through graph models, and then perform training and prediction for downstream tasks.

[0046] Figure 2 FIG. 1 shows a schematic diagram of an implementation architecture of a cross-domain recommendation in conventional technology. Figure 1 As shown in the implementation framework, the graph neural network can be used to process the knowledge graph to obtain the user representation vector h u and the product representation vector h i , then, by characterizing the user vector h u and the product representation vector h iPreference prediction is performed, and based on the obtained prediction score, it is determined whether the user is interested in the corresponding product, and the product is pushed based on the determination result. Here, the entities in the knowledge graph may include user U, product I, and product-related entities E (hereinafter referred to as product associations in this specification). Among them, the product association can be an associated entity determined based on the product description information or attribute information. Usually, products and product associations are associated with users through user clicks and purchase behaviors, and product associations can be connected to products through various types of associations. For example, Figure 3 In the specific example shown, the user's click and purchase of brand A computer appliances makes it associated with the product "brand A computer appliances", which contains three entities "brand A", "computer", and "appliance". Then, through the connection relationships such as associated products, required components, and subcategories, these entities can be associated with more product associations, such as "phone", "screen", "chip", "mobile device", etc. Among them, the knowledge graph can be used as follows Figure 1 The online service platform, various mini-programs, etc. shown in the figure are constructed by crawling data.

[0047] In conventional techniques, graph neural networks often rely on learning from source domains (such as Figure 1 The sample data obtained from the applet 1 in the example is trained and applied to the source domain or target domain (such as Figure 2 Mini Program 1, Mini Program 2, etc. in the knowledge graph). It is easy to understand that the target domain and the source domain can be the same information domain or different information domains. A single training sample can include a data pair consisting of a user, the candidate information pushed (such as a product), and a sample label of whether the user selects (such as clicks, purchases, uses, etc.) the corresponding information. Each entity in the knowledge graph can correspond to a corresponding representation vector. The initial representation vector can be extracted from the entity attribute information. For example, for a user entity, the initial representation vector can be constructed by its occupation, consumption frequency, consumption amount, etc. For example, in the feature dimension representing the consumption amount, the consumption amount "less than 100" corresponds to the feature value "1".

[0048] like Figure 2 As shown in the figure, in the conventional cross-domain information push model training process, the knowledge graph data can be processed by the graph neural network to obtain the user representation h u and product characterization i , then based on the user representation h u and product characterization i The prediction score of each user-item data pair in the training sample is obtained and compared with the label data (such as whether the user accepts the recommendation information) to obtain the model loss, and then the pending parameters in the graph neural network are adjusted according to the model loss to update the graph neural network.

[0049] Among them, processing knowledge graph data through graph neural networks is usually a process of aggregating representation vectors (i.e., representations) of neighbor nodes (or neighbor entities) in the graph neural network. A single-layer graph neural network can usually aggregate representation vectors of neighbor entities of a predetermined order (such as 2nd order). For a single training sample, a preference score is predicted for the user u and candidate information (such as product) i. The current model loss can be determined by comparing the preference score with the corresponding label data. Furthermore, the pending parameters in the graph neural network can be adjusted according to the model loss. Among them, the pending parameters in the graph neural network may include aggregation parameters, importance coefficients of each layer, and the like.

[0050] The trained graph neural network processes the knowledge graph to obtain representation vectors for each user and each product. Using these representation vectors, the prediction score for the candidate information can be determined through the prediction network, and the information of the target domain can be pushed based on the prediction score.

[0051] Although the above transferable learning method through knowledge graph can achieve good results in some migrations, the impact of sample bias caused by user misoperation on the results is rarely considered. In view of this, under the technical concept of this specification, in order to make full use of source domain data to push information to the target domain, the data of the source domain and the target domain are associated, and the data is filtered in combination with the information bottleneck theory, and invalid information is shielded as much as possible for information recommendation in the target domain, so as to improve the accuracy of information push and thus improve the user experience.

[0052] Figure 4 The following is a schematic diagram showing a cross-domain information push model architecture and its training diagram under the technical concept of this specification. Figure 4 As shown in FIG. 1 , the model architecture includes at least: a representation layer, a fusion layer, an information compression layer, and a prediction layer. The representation layer is used to process the knowledge graph through a graph neural network to obtain user representations of the source domain and the target domain. The fusion layer is used to fuse the user representations of the source domain and the target domain to obtain a fused representation. The information compression layer can filter the fused representation to obtain more effective data, such as a compressed representation. The prediction layer can match the compressed representation with the information representation of the candidate information to obtain a prediction result to obtain the target information of the information push service.

[0053] Combine the following Figure 5 , Figure 6 The embodiments shown in the drawings further describe the technical concept of the present specification in detail.

[0054] Figure 5This is a training flow chart of a cross-domain information push model in one embodiment of this specification. The execution subject of this process can be any computer, device, or server with certain computing capabilities. It can be understood that the training process of the cross-domain information push model includes multiple update cycles. In a single update cycle, the current batch of training samples can be used to update the pending parameters in the cross-domain information push model. Figure 5 As shown, a single cycle in the training process of the cross-domain information push model includes the following steps: step 501, through the processing of the fused knowledge graph, the first representation of the current sample user in the source domain, the second representation in the target domain, and the information representation of the candidate information are obtained, and the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information; step 502, the first representation and the second representation are fused to obtain a fused representation; step 503, information compression is performed on the fused representation to determine the compressed representation of the current sample user, wherein the compressed representation is a weighted fusion result of the fused representation and the first noise, the first noise is determined based on the distribution sampling satisfied by each compressed representation of the current batch of sample users, and the weighted weight is determined by the Bernoulli sampling of the discrete mapping under the first predetermined parameter; step 504, the model loss is determined to update the pending parameters in the cross-domain information push model, the model loss includes the representation loss and the prediction loss, the representation loss is determined by comparing the compressed representation with the second representation, and the prediction loss is determined by the prediction result obtained by predicting the compressed representation in combination with the information representation.

[0055] First, in step 501, by processing the fused knowledge graph, a first representation of the current sample user in the source domain, a second representation in the target domain, and information representation of the candidate information are obtained.

[0056] Knowledge graph is a kind of graph data, which is used to describe the relationship between entities. Entities can be anything in nature, and here can be various business entities involved in actual business, such as users (which can be user IDs), products (product names), documents, words, attributes, etc. Specifically, such as "computer" (item), "record a travel experience" (document), "office" (attribute), etc. The fused knowledge graph can be a knowledge graph obtained by fusing multiple graphs.

[0057] Here, in the information push scenario, the fused knowledge graph may include a basic graph, as well as a bipartite graph of users and candidate information. The basic knowledge graph may be a large knowledge graph constructed by collecting entity data in various business scenarios, such as a general knowledge graph constructed based on business data collected by various micro-applications or applets on a payment platform (such as life payment, asset credit accounts, online shopping, leasing, etc.). The basic knowledge graph can cover a wider range of businesses and is used for the expansion of business knowledge. A bipartite graph, also called a bipartite graph, is a special model in graph theory. Let G = (V, E) be an undirected graph. If the vertex V can be divided into two non-intersecting subsets (A, B), and any edge E in the graph k The two associated vertices V i and V j Belong to these two different vertex sets (such as V i ∈A,V j ∈B), then the graph G is called a bipartite graph. In the bipartite graph in the information push scenario, the two types of nodes can be users (such as represented by user) and candidate information (such as represented by item). The bipartite graph here can include the bipartite graph of the source domain and the bipartite graph of the target domain, and the corresponding candidate information is the candidate information of the source domain. Candidate information can be various information used to push to users, such as products, documents, etc.

[0058] The basic graph and the target domain bipartite graph and the source domain bipartite graph can be associated with each other to form a fused knowledge graph. The two types of vertices in the bipartite graph can correspond to two types of entities in the basic graph, such as users and products. The association method is, for example, the association of the user's unique identifier (ID, such as a mobile phone number, etc.) and the association of the candidate information's information identifier (such as a product name or number, etc.).

[0059] As an example, use Indicates that there is an edge between user u and candidate information item i in the source domain, Y i→e Indicates that there is an edge between candidate information item i and other entity entity e. Similarly, Indicates that there is an edge between candidate information item i and user u in the source domain, Y e→i Indicates that there is an edge between other entity entity e and candidate information item i, Y e→e Indicates that there is an edge between two other entities. Other entities are entities other than users and candidate information in the basic graph. The adjacency matrix of the fused knowledge graph can be expressed as:

[0060]

[0061] The neural network used to process the knowledge graph is usually a graph neural network (denoted as a GNN encoder, such as LightGCN, etc.). By processing the knowledge graph through a graph neural network, the information of neighbor entities can be aggregated for each entity, thereby obtaining the representation of each entity, for example, a user representation can be obtained for a user entity, and a product representation can be obtained for a product entity. Among them, the aggregation process of the neighbor entity information by a single layer of the graph neural network is: the representation of the current entity and its neighbor nodes in the previous layer is fused, and the fusion process is, for example, weighted average, weighted sum, median, maximum value, etc., which will not be repeated here. According to one embodiment, for each entity in the knowledge graph, a single layer of the graph neural network can be used to fuse the representation vectors of its neighbor entities of a predetermined order (such as 2nd order). According to another embodiment, for a single user or product entity, a single layer of the graph neural network can be used to fuse the representation vectors of its product association neighbors of a predetermined order (such as 2nd order). In other embodiments, a single layer of the graph neural network can be used to select neighbor entities in other ways to fuse their representation vectors, which will not be repeated here.

[0062] In addition, the entity representation obtained by processing the knowledge graph through the graph neural network can be the processing result of the last layer of the graph neural network, or it can be the weighted sum of the processing results of each layer, which is not limited here.

[0063] As a specific example, at layer l, the representation of a single node is in, is the incentive matrix, is the adjacency matrix, E S(l-1) is the representation of the node at the l-1 layer. Assuming the number of layers of the graph neural network is L, the user representation of the source domain is finally obtained and information representation Recorded as a set

[0064] Similarly, by processing the bipartite graph of the target domain with the basic knowledge graph through the graph neural network, the user representation of the target domain can be obtained. and information representation Recorded as a set

[0065] In the embodiment of this specification, the above graph neural network can correspond to the representation layer of the cross-domain information push model, and the user representation of the source domain It can be recorded as the first representation, the user representation of the target domain It can be recorded as the second representation. The current sample user can be a user in a single training sample in the current batch of samples.

[0066] Next, through step 502, the first representation and the second representation are fused to obtain a fused representation.

[0067] The fusion of the first representation and the second representation can be achieved by summing, averaging, weighting, etc., which are not limited here. The fused representation fuses the data of the current sample user in the source domain and the target domain. This step 502 can correspond to the fusion layer of the cross-domain information push model.

[0068] Then, in step 503, information compression is performed on the fused representation to determine a compressed representation of the current sample user.

[0069] In order to filter out more effective information from the source domain, the fused representation is compressed based on the information bottleneck theory. Under the information bottleneck theory, the bottleneck part of the system can limit the flow and processing of information transmission. Under the technical concept of this specification, noise can be used to replace the information in the fused representation that is irrelevant to the target domain task. To this end, the compressed representation can be determined by the weighted fusion result of the fused representation and the first noise at the information compression layer. For example, for user v, the compressed representation is recorded as Then we have: ∈ represents the first noise, λ v is the weight parameter. Information that has little effect on the result after adding noise is unimportant information, while information that has a greater impact on the result under noise interference is more important information, so that important information can be filtered out.

[0070] The first noise ∈ can be sampled and determined from a predetermined distribution, such as Gaussian noise, Poisson noise, salt and pepper noise, etc. Taking Gaussian noise as an example, a type of noise whose probability density function obeys Gaussian distribution (i.e., normal distribution). If the amplitude distribution of a Gaussian noise obeys Gaussian distribution, and its power spectrum density is uniformly distributed, it is called Gaussian white noise. Practice has proved that Gaussian white noise is a better solution for adding noise. In the case of adding Gaussian white noise, the predetermined distribution can be the corresponding mean matrix and variance matrix determined by the fusion representation of the current batch of training samples, such as: where μ H is the mean matrix of each fusion representation corresponding to the current batch of training samples, is the variance matrix of each fusion representation corresponding to the current batch of training samples (the single element in the variance matrix is ​​the variance of each element corresponding to each fusion representation). It can be understood that a single element in the mean matrix and a single element at the corresponding position in the variance matrix can determine a Gaussian distribution, and the element at the corresponding position of the first noise is obtained by sampling the corresponding Gaussian distribution. The sampled first noise can be applied to the current batch of training samples.

[0071] On the other hand, the added noise should be adjustable so that the parameters can be adjusted based on the training samples in each update cycle to filter out important information from the fused representation and obtain noise that can replace the task-irrelevant information in the target domain. To achieve this goal, the differentiable sampling weight parameter λ can be used. v In one embodiment, the weight parameter λ v It can be determined by Bernoulli sampling, denoted as λ v ~Bernoulli(Sigmoid(p v )(, p v The size is adjustable, usually a preset value, such as the first predetermined parameter. In order to make Bernoulli sampling differentiable, a discrete mapping (such as Gumbel Sigmoid) can be used to determine the weight parameter λ v .like: Among them, t is the temperature coefficient, which can be a preset hyperparameter, and m satisfies a uniform distribution between 0 and 1, as recorded as In other embodiments, other methods may be used to sample the weight parameter λ v , I will not go into details here.

[0072] In this way, the compressed representation can carry the important information in the fused representation and filter out the unimportant information.

[0073] Furthermore, in step 504, the model loss is determined to update the pending parameters in the cross-domain information push model.

[0074] Here, model loss can be used to describe the gap between the processing result obtained by processing the data through the model and the expected target. In this specification, on the one hand, it is hoped that the compressed representation contains as much information as possible related to the task of the target domain, so that the similarity between the compressed representation and the second representation can be maximized, and on the other hand, the gap between the predicted result and the sample label can be minimized.

[0075] Thus, the model loss may include compression loss and prediction loss. The compression loss is negatively correlated with the similarity between the compressed representation and the second representation, such as:

[0076]

[0077] That is, the higher the similarity between the compressed representation and the second representation, the smaller the compression loss. The similarity between the compressed representation and the second representation can be measured by cosine similarity, Euclidean distance, KL divergence, etc. between vectors.

[0078] The prediction loss is positively correlated with the gap between the prediction result and the sample label. It can be understood that the prediction result can be determined by processing the compressed representation through the prediction layer, for example, it can be the matching result of the current sample user and the candidate information of the source domain or the target domain, such as the vector composed of the matching degree on each candidate information. The sample label indicates the selection vector composed of one or more candidate information actually selected by the current sample user. For example, in the selection vector, the candidate information actually selected by the user is represented by a non-zero value (such as 1), and the candidate information not actually selected by the user is represented by 0.

[0079] The source domain prediction results are as follows: The target domain prediction results are as follows: in, represents matrix addition, represents the user representation of the target domain user i, represents the information representation of the candidate information j in the source domain, represents the information representation of the target domain candidate information k, and <> represents the inner product. Assume that the source domain sample label is recorded as Y S , the target domain sample label is recorded as Y T , the gap between the prediction result and the sample label can include at least one of the following: With Y S The gap With Y T The gap can be determined by the similarity (such as vector similarity, etc.) or distance (such as Euclidean distance, cross entropy, etc.) between the two. The higher the similarity, the smaller the gap, and the larger the distance, the larger the gap.

[0080] The model loss may include a weighted sum of compression loss and prediction loss. When determining the weighted sum, since the current batch of training samples may include at least one of the source domain training samples and the target domain training samples, the prediction loss may include the prediction loss of at least one of the source domain and the target domain. In the case where the prediction loss includes the source domain prediction loss and the target domain prediction loss, the two may correspond to different weights, such as the target domain prediction loss has a larger weight, so that the model loss reflects the data characteristics of the target domain more. In this way, in the process of minimizing the model loss, the compressed representation can contain as much information related to the task of the target domain as possible, and the prediction result is as close to the sample label as possible.

[0081] In addition, in order to filter out irrelevant information as much as possible and retain information that is relatively important to the prediction task, in some possible designs, the target domain subgraph G can also be minimized. T Under this condition, the better subgraph from the source domain to the target domain is the same as the source domain subgraph G S The mutual information between them. For example:

[0082]

[0083] in, is a better subgraph from the source domain to the target domain. According to relevant theories, the upper limit of the mutual information is Expected value B is the number of training samples in the current batch, and λ7 is the sampling weight parameter corresponding to the i-th training sample. Furthermore, the expected value can also be obtained through the target domain subgraph G T Under the condition, the better subgraph and the source domain subgraph G S The KL divergence between L KL It can be recorded as the filtering loss of the model.

[0084] At this time, according to a specific example, the model loss is the weighted sum of compression loss, prediction loss, and filtering loss, such as: Among them, α1, α2, and α3 are weight coefficients for weighting the corresponding loss terms, which can be set in advance. is the target domain prediction loss, is the source domain prediction loss, is the compression loss, In this way, while minimizing the model loss, the information gain of the information bottleneck can also be minimized, and irrelevant information can be filtered as much as possible.

[0085] By using the gradient back propagation of the model loss, the gradient data of each undetermined parameter in the cross-domain information push model can be obtained, and then the gradient data can be used to update each undetermined parameter in the model. The undetermined parameters may include at least one of the undetermined parameters in the graph neural network of the representation layer, the fusion layer, the information compression layer, and the prediction layer.

[0086] After multiple update cycles of iterative updates, a cross-domain information push model that meets the training end condition can be trained. The training end condition can be at least one of the following: the update cycle reaches a predetermined number of cycles, the accuracy reaches a predetermined accuracy, the model loss approaches 0, the gradient approaches 0, the value of the undetermined parameter converges to a certain value, etc.

[0087] The trained cross-domain information push model can be used for information push in the target domain. Figure 5 The schematic diagram of the process of information push using the cross-domain information push model is shown. The execution subject of the information push process can be a computer, device, or server with certain computing capabilities. Figure 6As shown, according to an embodiment, the information push process may include: step 601, through the processing of the fused knowledge graph, obtaining the third representation of the current user in the source domain, the fourth representation in the target domain, and the information representations corresponding to the candidate information in the target domain, the fused knowledge graph includes the basic graph and the user and candidate information bipartite graph; step 602, fusing the third representation and the fourth representation into a first fused representation; step 603, performing information compression on the first fused representation to determine the first compressed representation of the current user, wherein the first compressed representation is a weighted fusion result of the first fused representation and the second noise, the second noise is determined based on the distribution sampling satisfied by the compressed representation of the training sample, and the weighted weight is determined by the Bernoulli sampling of the discrete mapping satisfying the first predetermined parameter; step 604, based on the matching of the first compressed representation with each information representation, determining the target information to be pushed from each candidate information.

[0088] First, in step 601, by processing the fused knowledge graph, the third representation of the current user in the source domain, the fourth representation in the target domain, and each information representation corresponding to each candidate information in the target domain are obtained.

[0089] Here, the fused knowledge graph may include a basic graph and a bipartite graph of users and candidate information. The basic knowledge graph may be a large knowledge graph constructed by collecting entity data in various business scenarios, such as a general knowledge graph constructed based on business data collected by various micro-applications or applets on a payment platform (such as life payment, asset credit account, online shopping, leasing, etc.). The basic knowledge graph can cover a wider range of businesses and is used for the expansion of business knowledge. The bipartite graph may include a bipartite graph of the source domain and a bipartite graph of the target domain, and the corresponding candidate information is the candidate information of the source domain. The candidate information may be various information used to push to users, such as commodities, documents, etc.

[0090] The basic graph and the target domain bipartite graph and the source domain bipartite graph can be associated with each other to form a fused knowledge graph. The two types of vertices in the bipartite graph can correspond to two types of entities in the basic graph, such as users and products. The association method is, for example, the association of the user's unique identifier (ID, such as a mobile phone number, etc.) and the association of the candidate information's information identifier (such as a product name or number, etc.).

[0091] By processing the fused knowledge graph with a graph neural network, entity representations corresponding to each entity including the user and the candidate information can be obtained. In the embodiment of this specification, the representation of the current user in the source domain can be recorded as the third representation, the representation in the target domain can be recorded as the fourth representation, and the representations corresponding to each piece of candidate information can be recorded as information representations.

[0092] Next, through step 602, the third representation and the fourth representation are fused into a first fused representation.

[0093] The fusion of the third representation and the fourth representation can be achieved by summing, averaging, weighting, etc., which are not limited here. The first fused representation fuses the data of the current user in the source domain and the target domain.

[0094] Then, in step 603, information compression is performed on the first fused representation to determine a first compressed representation of the current user.

[0095] In order to filter out more effective information from the source domain data, the first fused representation is compressed based on the information bottleneck theory. Under the technical concept of this specification, noise can be used to replace information in the fused representation that is irrelevant to the target domain task. Specifically, a second noise can be added to the first fused representation to obtain a first compressed representation after compression of the first fused representation. The way to add the second noise to the first fused representation can be to perform weighted fusion of the first fused representation and the second noise, and the weighted fusion result is used as the first compressed representation. Among them, the second noise can be determined by sampling based on the distribution satisfied by the compressed representations of multiple users. The compressed representations of multiple users can be compressed representations of multiple previous users, or compressed representations of part or all of the training samples. The distribution satisfied by the compressed representations of multiple users is determined by the statistical results of the corresponding multiple compressed representations. For example, the distribution satisfied by the compressed representations of multiple training samples is statistically analyzed, such as the normal distribution. Among them, μ H is the mean matrix of each fusion representation corresponding to multiple users, is the variance matrix of each fusion representation corresponding to the corresponding multiple users. The second noise is sampled according to the obtained distribution, such as ε, which obeys Specifically, a single element (such as the first element) in the second noise can be sampled from a normal distribution determined by the element at the corresponding position of the mean matrix as the mean and the element at the corresponding position of the variance matrix as the variance.

[0096] In order to make the noise amount controllable and offset the invalid information from the source domain as much as possible, the weight parameter λ of the first fusion representation and the second noise can be determined by Bernoulli sampling of the discrete mapping v Bernoulli sampling is determined by the corresponding probability parameter, denoted as λ v ~Bernoulli(Sigmoid(p v )(, p v The size is adjustable, usually a preset value, such as the first predetermined parameter. During the model training process and the data processing process of online information push, p v Can be consistent. The sampling process is recorded as: t is the temperature coefficient, which can be a preset hyperparameter, and m satisfies the uniform distribution between 0 and 1, such as m~Uniform(0,1).

[0097] The weight coefficient obtained by sampling can be used to replace the information irrelevant to the target domain in the first fusion representation with the second noise to obtain the first compressed representation. The first compressed representation is the coded representation of the current user after the user information of the target domain is supplemented by the source domain information.

[0098] Furthermore, in step 604, based on the matching between the first compression representation and each information representation, the target information to be pushed is determined from each candidate information.

[0099] It can be understood that in the process of processing the fusion graph through the graph neural network, it is equivalent to quantizing each entity in the graph in multiple predetermined dimensions. For example, the representation of a single entity is a 64-dimensional vector, which is equivalent to mapping the entity information to a quantitative representation in 64 dimensions.

[0100] The information representation describes the information of the corresponding candidate information in each dimension, and the first compressed representation describes the effective information obtained by filtering out irrelevant information from the comprehensive information of the current user in the source domain and the target domain. The degree of correlation between the first compressed representation and the information representation represents the degree of interest of the current user in the corresponding candidate information. Therefore, the first compressed representation can be matched with each information representation, and the candidate information most suitable for the current user can be determined according to the degree of matching. Among them, the matching between the first compressed representation and the information representation can be described by cosine similarity, Euclidean distance, etc., and the final matching degree is measured by matching degree or matching score. The matching degree or matching score can be positively correlated with similarity indicators such as cosine similarity, or negatively correlated with distance indicators such as Euclidean distance.

[0101] The matching degree or matching score here can also be recorded as a prediction result. According to the matching degree or matching score, N candidate information in the target domain that is more in line with the current user needs can be determined as the target information. N is a predetermined value, or the number of information with a matching degree or matching score greater than a predetermined threshold.

[0102] It is worth noting that the words "first", "second", "third", "fourth", etc. that appear in the description of this specification are for the convenience of description and are used to distinguish related concepts in the technical solution. They do not indicate the order and do not constitute a substantial limitation on the related concepts.

[0103] Reviewing the above process, the cross-domain information push technical concept provided by the embodiment of this specification fuses the source domain bipartite graph and the target domain bipartite graph with the basic knowledge graph to form a corresponding fused knowledge graph. Based on the processing of the fused knowledge graph, the user representation of the user in the source domain and the target domain, as well as the information representation of the candidate information, are determined respectively, so that the user representation of the source domain and the target domain are fused and then the data with less correlation with the target domain is filtered out based on the information bottleneck theory, so as to obtain the compressed representation corresponding to the user. Then match the user compressed representation with the information representation of the candidate information to obtain the prediction result (such as represented by the degree of matching with each candidate information). During the model training process, the model loss of each update cycle may include prediction loss and compression loss. By minimizing the model loss, the undetermined parameters in the model can be adjusted toward the following goals: maximize the similarity between the compressed representation and the second representation; minimize the gap between the prediction result and the sample label. During the information push process, the prediction result obtained by the cross-domain information push model adjusted according to the above goals is used to determine the target information to be pushed to the corresponding user.

[0104] In this way, the information push process under the technical conception of this specification can use the user data in the source domain to push information to the target domain where the data is sparse, and use the information bottleneck theory to filter out the information in the source domain that provides less gain to the target domain in the process of determining the target information, reduce the impact of invalid information, and push information to the target domain more accurately, thereby improving user experience.

[0105] According to another embodiment, a training device for a cross-domain information push model is also provided. The training device for the cross-domain information push model can be set on various computers, devices, and servers with certain computing capabilities. Figure 7As shown, the training device 700 for the cross-domain information push model may include an acquisition unit 701, a fusion unit 702, a compression unit 703, and an adjustment unit 704. In a single update cycle: an acquisition unit 701 is configured to acquire a first representation of the current sample user in the source domain, a second representation in the target domain, and an information representation of the candidate information by processing the fused knowledge graph, wherein the fused knowledge graph includes a basic graph and a bipartite graph of user and candidate information; a fusion unit 702 is configured to fuse the first representation and the second representation into a fused representation; a compression unit 703 is configured to perform information compression on the fused representation to determine a compressed representation of the current sample user, wherein the compressed representation is a weighted fusion result of the fused representation and the first noise, the first noise is determined based on the distribution sampling satisfied by each compressed representation of the current batch of sample users, and the weighted weight of the first noise is determined by the Bernoulli sampling of the discrete mapping satisfying the first predetermined parameter; an adjustment unit 704 is configured to determine a model loss to update pending parameters in the cross-domain information push model, wherein the model loss includes a representation loss and a prediction loss, the representation loss is determined by comparing the compressed representation with the second representation, and the prediction loss is determined by a prediction result obtained by predicting the compressed representation in combination with the information representation.

[0106] According to another embodiment, a device for cross-domain information push is also provided. The device for cross-domain information push is used to assist in information push in a target domain by using user information in a source domain, and can be set in various computers, devices, and servers with certain computing capabilities. Figure 8 As shown, the device 800 for cross-domain information push may include:

[0107] The acquisition unit 801 is configured to acquire, by processing the fused knowledge graph, a third representation of the current user in the source domain, a fourth representation in the target domain, and respective information representations corresponding to respective candidate information in the target domain, wherein the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information;

[0108] A fusion unit 802 is configured to fuse the third representation and the fourth representation into a first fused representation;

[0109] A compression unit 803 is configured to perform information compression on the first fused representation to determine a first compressed representation of the current user, wherein the first compressed representation is a weighted fusion result of the first fused representation and a second noise, the second noise is determined based on a distribution sampling satisfied by the compressed representation of the training sample, and the weighted weight of the second noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter;

[0110] The matching unit 804 is configured to determine the target information to be pushed from each candidate information in the target domain based on the matching between the first compression representation and each information representation.

[0111] It is worth mentioning that Figure 7 , Figure 8 The devices 700 and 800 shown are respectively Figure 5 , Figure 6 The method flow described corresponds to Figure 5 , Figure 6 The corresponding description in the illustrated method embodiment is also applicable to the devices 700 and 800 and will not be repeated here.

[0112] According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute a combination of Figure 5 , Figure 6 The method described by et al.

[0113] According to another embodiment of the present invention, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the Figure 5 , Figure 6 The method described by et al.

[0114] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of this specification can be implemented by hardware, software, firmware, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0115] The specific implementation methods described above further explain in detail the purpose, technical solutions and beneficial effects of the technical concept of this specification. It should be understood that the above description is only the specific implementation method of the technical concept of this specification, and is not used to limit the protection scope of the technical concept of this specification. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solutions of the embodiments of this specification should be included in the protection scope of the technical concept of this specification.

Claims

1. A training method for a cross-domain information push model, including multiple update cycles. In a single update cycle, the method includes: By processing the fused knowledge graph, a first representation of the current sample user in the source domain, a second representation in the target domain, and information representation of the candidate information are obtained, wherein the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information; fusing the first representation and the second representation into a fused representation; Performing information compression on the fused representation to determine a compressed representation of the current sample user, wherein the compressed representation is a weighted fusion result of the fused representation and a first noise, the first noise is determined based on distribution sampling satisfied by each compressed representation of the current batch of sample users, and the weighted weight of the first noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter; Determine a model loss to update pending parameters in the cross-domain information push model, wherein the model loss includes a representation loss and a prediction loss, the representation loss is determined by comparing the compressed representation with the second representation, and the prediction loss is determined by a prediction result obtained by predicting the compressed representation in combination with the information representation.

2. The method of claim 1, wherein: The step of obtaining the first representation of the current sample user in the source domain, the second representation in the target domain, and the information representation of the candidate information by processing the fused knowledge graph includes: Using a graph neural network to perform feature fusion on a fused knowledge graph consisting of a basic graph and a source domain bipartite graph, to obtain information representations corresponding to the first representation and each piece of candidate information in the source domain; A graph neural network is used to perform feature fusion on the fused knowledge graph composed of the basic graph and the target domain bipartite graph to obtain the second representation and the information representations corresponding to each candidate information of the target domain.

3. The method of claim 1, wherein: The fusing the first representation and the second representation into a fused representation includes: The first representation and the second representation are processed in one of a weighted, averaged, and summed manner to obtain the fused representation.

4. The method of claim 1, wherein: The distribution satisfied by each compressed representation is: a normal distribution determined by using the corresponding elements in the mean matrix of each compressed representation as the mean and the corresponding elements in the variance matrix as the variance.

5. The method of claim 1, wherein: The prediction result is the matching score corresponding to each candidate information of the current sample user. The prediction loss is described by one of the following cosine similarity, Euclidean distance, cross entropy, and KL divergence between two vectors to express the degree of deviation between the two vectors: the matching vector composed of each matching score and the selected vector determined according to the actual selection of the candidate information by the current sample user.

6. The method of claim 5, wherein: In the target domain prediction results, the matching score corresponding to a single candidate information is determined in the following way: The compressed representation is fused with the second representation, and a similarity between the fusion result and the information representation of the single candidate information is determined as a matching score.

7. The method of claim 1, wherein: The model loss also includes filtering loss, which is a weighted sum of the compression loss, the prediction loss, and the filtering loss. The filtering loss is the mutual information between the better subgraph from the source domain to the target domain and the source domain subgraph under the condition of the target domain subgraph.

8. The method of claim 1, wherein: The sum of the weighted weight of the fused representation and the weighted weight of the first noise is 1.

9. A method for information push, for using user information of a source domain to assist in information push of a target domain, the method comprising: By processing the fused knowledge graph, a third representation of the current user in the source domain, a fourth representation in the target domain, and respective information representations corresponding to respective candidate information in the target domain are obtained, wherein the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information; fusing the third representation and the fourth representation into a first fused representation; Performing information compression on the first fused representation to determine a first compressed representation of the current user, wherein the first compressed representation is a weighted fusion result of the first fused representation and second noise, the second noise is determined based on a distribution sampling satisfied by the compressed representation of the training sample, and the weighted weight of the second noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter; Based on the matching between the first compression representation and each information representation, the target information to be pushed is determined from each candidate information in the target domain.

10. The method of claim 9, wherein: The process of obtaining the third representation of the current user in the source domain, the fourth representation in the target domain, and each information representation corresponding to each candidate information in the target domain by processing the fused knowledge graph includes: Using a graph neural network to perform feature fusion on a fused knowledge graph consisting of a base graph and a source domain bipartite graph, to obtain the third representation; A graph neural network is used to perform feature fusion on a fused knowledge graph consisting of a basic graph and a target domain bipartite graph to obtain information representations corresponding to the fourth representation and each candidate information of the target domain.

11. The method of claim 9, wherein: The distribution satisfied by each compressed representation of the training samples is: a normal distribution determined by taking each element in the mean matrix of each compressed representation of the training samples as the mean and the corresponding element in the variance matrix as the variance.

12. The method of claim 9, wherein: The fusing the third representation and the fourth representation into a first fused representation includes: The third representation and the fourth representation are processed in one of a weighted, averaged, and summed manner to obtain the first fused representation.

13. The method of claim 9, wherein: The determining the target information to be pushed from each candidate information of the target domain based on the matching of the first compression representation and each information representation includes: Determine respective matching scores corresponding to the first compression representation and respective information representations; According to the matching scores, several candidate information are selected as the target information to be pushed to the current user.

14. A training device for a cross-domain information push model, used for training the cross-domain information push model through multiple update cycles, the device comprising an acquisition unit, a fusion unit, a compression unit, and an adjustment unit, in a single update cycle: The acquisition unit is configured to acquire a first representation of the current sample user in the source domain, a second representation in the target domain, and information representation of the candidate information by processing the fused knowledge graph, wherein the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information; The fusion unit is configured to fuse the first representation and the second representation into a fused representation; The compression unit is configured to perform information compression on the fused representation to determine a compressed representation of the current sample user, wherein the compressed representation is a weighted fusion result of the fused representation and a first noise, the first noise is determined based on the distribution sampling satisfied by each compressed representation of the current batch of sample users, and the weighted weight of the first noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter; The adjustment unit is configured to determine a model loss to update pending parameters in the cross-domain information push model, the model loss including a representation loss and a prediction loss, the representation loss being determined by comparing the compressed representation with the second representation, and the prediction loss being determined by a prediction result obtained by predicting the compressed representation in combination with the information representation.

15. An information push device, used to use user information of a source domain to assist in information push of a target domain, the device comprising: an acquisition unit configured to acquire, by processing a fused knowledge graph, a third representation of the current user in a source domain, a fourth representation in a target domain, and respective information representations corresponding to respective candidate information in the target domain, wherein the fused knowledge graph includes a basic graph and a bipartite graph of users and candidate information; a fusion unit, configured to fuse the third representation and the fourth representation into a first fused representation; a compression unit, configured to perform information compression on the first fused representation to determine a first compressed representation of the current user, wherein the first compressed representation is a weighted fusion result of the first fused representation and a second noise, the second noise is determined based on a distribution sampling satisfied by the compressed representation of the training sample, and the weighted weight of the second noise is determined by Bernoulli sampling of a discrete mapping satisfying a first predetermined parameter; The matching unit is configured to determine the target information to be pushed from each candidate information in the target domain based on the matching between the first compression representation and each information representation.

16. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 13.

17. A computing device comprising a memory and a processor, characterized in that: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 13 is implemented.

Images