A data push method, apparatus, electronic device and storage medium

Abstract

This invention provides a data push method, apparatus, electronic device, and storage medium, applied in the field of data push technology. It involves determining a user identifier and constructing a user set based on that identifier; determining an item identifier and constructing an item set based on that identifier; using the user set and item set to determine a node set and a first edge set; generating an interaction graph based on the node set and the first edge set to express the interaction behavior between the user and the item; inputting the interaction graph into a pre-trained model to output a predicted score expressing the relationship between the user and the item; and pushing data based on the predicted score. This achieves score prediction for the relationship between unknown users and items, and by using the predicted score, it can automatically recommend data about target items to users while ensuring the accuracy of the recommendations.

Description

Technical Field

[0001] This invention relates to the field of data push technology, and in particular to a data push method, a data push device, an electronic device, and a computer-readable storage medium. Background Technology

[0002] With the advancement of technology, data push systems are widely used in various industries. Among them, user-item recommendation systems are used to predict users' "ratings" or "preferences" for items or services. Accurate data push can help improve the user experience. Therefore, how to improve the efficiency of data push has become a technical problem that needs to be overcome by those skilled in the art. Summary of the Invention

[0003] The present invention provides a data push method, apparatus, electronic device, and computer-readable storage medium to solve the problem of how to improve the efficiency of data push.

[0004] This invention discloses a data push method, which may include:

[0005] Determine the user identifier, and construct a user set for the user based on the user identifier;

[0006] Identify item identifiers and construct an item set for each item based on those identifiers;

[0007] The node set and the first edge set are determined using the user set and the item set;

[0008] An interaction graph is generated based on the node set and the first edge set to represent the interaction behavior between the user and the item;

[0009] The interaction graph is input into a pre-trained model to output a predicted score that expresses the relationship between the user and the item, and data is pushed based on the predicted score.

[0010] Optionally, the pre-trained model is configured with a prediction function and further includes:

[0011] Determine the important graph structures for the interaction graph;

[0012] Determine the difference measurement function for the interaction graph;

[0013] The difference in classification performance between the important graph structure and the interaction graph is reduced by using the difference measurement function and prediction function through a preset algorithm.

[0014] Optionally, it also includes:

[0015] Determine the complement graph for the interaction graph, and the complement set for the complement graph;

[0016] The dependency of the complement on the important graph structure is minimized by a preset algorithm.

[0017] Optionally, the step of determining the important graph structures for the interaction graph includes:

[0018] In the interaction graph, edges are determined to represent the interaction relationship between the user and the item;

[0019] The edges are determined as the scaling factors of the important graph structure;

[0020] Determine a first relation coefficient used to express the interaction relationship between the user and the item in the important graph structure;

[0021] Determine a second relationship coefficient to represent the interaction relationship between the user and the item in the supplementary graph;

[0022] The generation probability for important graph structures is calculated using the scaling factor, the first relation factor, and the second relation factor.

[0023] The important graph structures for the interaction graph are determined based on the generation probability.

[0024] Optionally, the step of determining the important graph structure for the interaction graph based on the generation probability includes:

[0025] Generate a graph autoencoder based on a graph converter architecture;

[0026] The graph autoencoder is used to determine the important graph structures for the interaction graph based on the generation probability.

[0027] Optionally, it also includes:

[0028] Nodes representing the user or item are identified from the interaction graph;

[0029] Determine the anchor nodes from the nodes, and calculate the distances between all the anchor nodes and the target node;

[0030] The association weight between the target node and the anchor node is determined by the distance.

[0031] Optionally, it also includes:

[0032] Normalization is performed on the associated weights.

[0033] Optionally, the pre-trained model includes a propagation layer and further includes:

[0034] The target node is embedded in the propagation layer based on the association weight.

[0035] Optionally, it also includes:

[0036] A global topology embedding matrix for the target node is determined by a linear transformation matrix; the global topology embedding matrix is ​​used to generate higher-order topology information.

[0037] Optionally, it also includes:

[0038] The target node is embedded into a multi-head self-attention mechanism model, and a probability score for the edge is calculated.

[0039] Optionally, it also includes:

[0040] The probability score is used to sample the first edge set, and a sampling subset is constructed based on the sampling results.

[0041] Optionally, it also includes:

[0042] Determine the second edge set for the aforementioned important graph structure;

[0043] The normalized importance between the target node and the anchor node is determined using the probability score.

[0044] The interaction probability between the user and the item is calculated based on the normalized importance and the second edge set.

[0045] Optionally, the step of inputting the interaction graph into a pre-trained model to output a predicted score expressing the relationship between the user and the item, and pushing data based on the predicted score, includes:

[0046] The interaction graph is input into a pre-trained model to output a predicted score based on the interaction probability, which expresses the relationship between the user and the item, and data is pushed based on the predicted score.

[0047] Optionally, the pre-trained model is generated by training in the following manner:

[0048] Generate a mask image autoencoder;

[0049] Training is performed to ensure the independence between the important graph structure and the complementary graph;

[0050] Generate an interaction model to output an expression of the interaction relationship between the user and the item.

[0051] Optionally, the step of generating the graph autoencoder includes:

[0052] Generate a mask image for the interaction graph;

[0053] The edges of the mask image are sampled, and the sampling edge probability is determined;

[0054] Determine the selection probability of the target node and the target edge corresponding to the target node in the mask graph autoencoder;

[0055] The embedding position of the target node in the mask graph autoencoder is determined using the sampling edge probability and the selection probability.

[0056] Optionally, the step of training on the independence between the important graph structure and the complementary graph includes:

[0057] The supplementary image is generated using different sampling rates;

[0058] Based on the sampling rate, determine the target sampling rate loss;

[0059] And based on the target sampling rate loss, the similarity between the important graph structure and the complement graph is minimized.

[0060] Optionally, the step of generating an interaction model for outputting an expression of the interaction relationship between the user and the item includes:

[0061] Based on the embedding position of the target node in the mask map autoencoder, a loss function for the interaction model is determined;

[0062] An interaction model is generated based on the loss function to output an expression of the interaction relationship between the user and the item.

[0063] This invention also discloses a data push device, comprising:

[0064] The user identifier determination module is used to determine the user identifier and construct a user set for the user based on the user identifier;

[0065] An item identification determination module is used to determine item identifications and construct an item set for each item based on the item identifications.

[0066] The node set and first edge set determination module is used to determine the node set and the first edge set using the user set and the item set;

[0067] An interaction graph generation module is used to generate an interaction graph based on the node set and the first edge set to express the interaction behavior between the user and the item.

[0068] The data push module is used to input the interaction graph into a pre-trained model to output a predicted score that expresses the relationship between the user and the item, and push data based on the predicted score.

[0069] This invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0070] The memory is used to store computer programs;

[0071] When the processor executes a program stored in the memory, it implements the method described in the embodiments of the present invention.

[0072] This invention also discloses a computer-readable storage medium storing instructions that, when executed by one or more processors, cause the processors to perform the methods described in this invention.

[0073] The embodiments of the present invention have the following advantages:

[0074] This invention, through the following steps: determining a user identifier and constructing a user set based on the user identifier; determining an item identifier and constructing an item set based on the item identifier; using the user set and the item set to determine a node set and a first edge set; generating an interaction graph based on the node set and the first edge set to express the interaction behavior between the user and the item; inputting the interaction graph into a pre-trained model to output a predicted score expressing the association between the user and the item; and pushing data based on the predicted score, achieves the prediction of the association between unknown users and items. By using the predicted score, data about target items can be automatically recommended to users, ensuring the accuracy of the recommendation. Attached Figure Description

[0075] Figure 1 This is a flowchart of the steps of a data push method provided in an embodiment of the present invention;

[0076] Figure 2 This is a structural block diagram of a data push device provided in an embodiment of the present invention;

[0077] Figure 3 This is a hardware structure block diagram of an electronic device provided in an embodiment of the present invention;

[0078] Figure 4 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation

[0079] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0080] In practical applications, recommended items include: products, movies, music, news, books, academic papers, search queries, and other products. Improving the efficiency of data-driven recommendations has become a technical problem that those skilled in the art need to overcome.

[0081] In many vertical industries, recommendation systems are built for specific needs, specifically designed to perform recommendations from massive amounts of data in areas such as government policy recommendations, job recommendations, insurance recommendations, expert recommendations, restaurant recommendations, and other scenarios including food, financial services, online dating, and smart tourism. Taking e-commerce as a prime example, applications (APPs) provide customers with potential purchase options and suggestions, helping them decide what to buy and simulating a salesperson's assistance in the purchase process. The principle behind personalized recommendation systems is to recommend products and services that users are interested in based on their interests and historical purchasing behavior; the effectiveness of these recommendations directly impacts conversion rates.

[0082] Self-supervised learning (SSL) has become a popular solution to address the sparsity of interaction labels in recommender systems by generating auxiliary supervision signals from unlabeled data. In a recommender system, users and items are considered nodes, and user-item interactions are considered edges, such as purchasing, clicking, adding to cart, favoriting, and deleting, forming a graph (also known as an interaction graph). By integrating with graph neural network architectures for collaborative filtering, SSL-based graph augmentation techniques have proven effective in modeling user-item interactions using a limited number of training labels. Among related techniques, graph contrastive learning (GCL) is widely used in recommender systems. The key behind GCL-based recommender techniques is to obtain supervision signals from auxiliary learning tasks and then achieve the recommendation objective through SSL-enhanced collaborative training.

[0083] Graph comparison methods in related technologies aim to achieve training results by maximizing representational consistency among positive samples and minimizing similarity among negative samples, or by attempting to follow the principle of maximizing useful information, comparing multi-view learning of user-item interaction graphs with heuristic data augmentation to complete recommendations. For example, schemes may involve reconstructing a topological comparison view by randomly deleting user and item nodes and their connections; however, randomly disrupting the graph topology may lead to the loss of key relationships between users and items. Therefore, mining more available self-supervised signals to learn user preference information is crucial.

[0084] Specifically, related technologies based on collaborative filtering can provide users with accurate ranking results by leveraging their rating records of items or other behavioral patterns. These methods require utilizing historical information or features of users and items to calculate the similarity between entities, thereby providing recommendations based on the target user's preferences.

[0085] Content-based recommendation technology predicts items similar to those liked by the target user based on the content descriptions of the user and the items. Its basic principle is to match the target user's attributes with the item attributes in terms of content, and then return the user's interests and preferences to generate recommendation results.

[0086] Recommendation schemes based on deep learning can learn corresponding feature representations from massive amounts of data through deep nonlinear network structures, thereby predicting missing data. Examples include multilayer perceptrons, convolutional neural networks, recurrent autoencoders, and hierarchical self-attention mechanisms.

[0087] However, using related technologies for data push may present the following technical problems.

[0088] (1) The problem of sparse data features. Training traditional methods requires a large amount of explicitly labeled data, but in actual recommendation systems, most users only have a small amount of historical behavior data. The problem of sparse data features seriously restricts the improvement of recommendation accuracy.

[0089] (2) Manually constructed augmentation patterns are time-consuming and labor-intensive, and have weak generalization ability. Recommendation systems are limited by manually designed graph structure patterns, and their prediction performance is often unstable or even inefficient. In addition, the training model only considers user-item interaction signals, and the training process is prone to "overfitting".

[0090] (3) Weak generalization ability of the model. Neural network-based methods are prone to "overfitting", that is, they perform well on the training set but poorly on the test set. The reason for this is that the recommendation algorithm incorrectly learns the noisy features and does not strengthen the learning of important graph structures.

[0091] This invention explores how to improve the predictive performance of recommendation systems using self-supervised learning techniques. Unlike traditional contrastive learning, the Masked Autoencoders (MAE) of this invention use reconstructing target data as the data augmentation task, avoiding the limitations of manually designed data augmentation methods. Furthermore, this invention proposes an intelligent recommendation method based on Graph Transformer technology, which can automatically extract masked self-supervised signals with invariant cooperative relationships. Specifically, this invention fully utilizes the ability of Graph Transformer to explicitly encode node relationships to mine self-supervised signals beneficial to downstream recommendation tasks and enhances the robustness of the recommendation method. In addition, this invention develops a novel topology-aware Graph Transformer method and integrates it into the user-item interaction model to model cooperative principles. In this invention, the topological information of the user-item graph is regarded as the global context for graph position encoding. To adapt the proposed method to different recommendation environments, it can learn appropriate interaction patterns as self-supervised signals under the guidance of adaptive cooperative principles, thereby achieving the final recommendation results. Finally, extensive experimental results verify that the proposed method has better recommendation performance than state-of-the-art baseline models.

[0092] Reference Figure 1 The diagram illustrates a flowchart of a data push method provided in an embodiment of the present invention, which may specifically include the following steps:

[0093] Step 101: Determine the user identifier and construct a user set for the user based on the user identifier;

[0094] Step 102: Determine the item identifier and construct an item set for the item based on the item identifier;

[0095] Step 103: Use the user set and the item set to determine the node set and the first edge set;

[0096] Step 104: Generate an interaction graph based on the node set and the first edge set to express the interaction behavior between the user and the item;

[0097] Step 105: Input the interaction graph into the pre-trained model to output a predicted score that expresses the relationship between the user and the item, and push data based on the predicted score.

[0098] For example, in a data push system, there are I users U = {u1, u2, ..., uI} and J items P = {p1, p2, ..., pJ}. By labeling them 1-I and 1-J respectively, we can obtain the user set U and the item set P. The known user interaction behaviors are represented by a matrix A. For example, aij = 1 indicates that user ui and item pj have interaction behaviors, and aij = 0 indicates that user ui and item pj have no interaction behaviors.

[0099] To transform the user-item interaction matrix into an interaction graph, this invention defines a graph based on graph collaborative filtering technology. Where V = U∪P is the node set, and Let represent the set of edges corresponding to user-item interactions, i.e., the first edge set. Based on the above definition, the recommendation task in this embodiment of the invention can predict the rating of unknown user-item interactions, which is formalized as a prediction function. in, Θ represents the predicted rating of item pj by user ui, and Θ represents the model parameters.

[0100] This invention, through the following steps: determining a user identifier and constructing a user set based on the user identifier; determining an item identifier and constructing an item set based on the item identifier; using the user set and the item set to determine a node set and a first edge set; generating an interaction graph based on the node set and the first edge set to express the interaction behavior between the user and the item; inputting the interaction graph into a pre-trained model to output a predicted score expressing the association between the user and the item; and pushing data based on the predicted score, achieves the prediction of the association between unknown users and items. By using the predicted score, data about target items can be automatically recommended to users, ensuring the accuracy of the recommendation.

[0101] Based on the above embodiments, modified embodiments of the above embodiments are proposed. It should be noted that, in order to keep the description brief, only the differences from the above embodiments are described in the modified embodiments.

[0102] In an optional embodiment of the present invention, the pre-trained model is provided with a prediction function and further includes:

[0103] Determine the important graph structures for the interaction graph;

[0104] Determine the difference measurement function for the interaction graph;

[0105] The difference in classification performance between the important graph structure and the interaction graph is reduced by using the difference measurement function and prediction function through a preset algorithm.

[0106] Optionally, it also includes:

[0107] Determine the complement graph for the interaction graph, and the complement set for the complement graph;

[0108] The dependency of the complement on the important graph structure is minimized by a preset algorithm.

[0109] In practical applications, graph learning models of related technologies are prone to associating an event with unimportant environmental factors. For example, when users shop online, they may browse items randomly or make accidental clicks, or their accounts may contain connections with other users. These factors may be irrelevant to the final data push decision. However, general graph learning methods treat all of these equally, which may lead to the association with environmental factors and the learning of incorrect decision models.

[0110] This invention employs a graph invariance principle modeling method to identify key graph structures that can guide downstream data delivery tasks. Specifically, the objective function can be defined from two perspectives: minimizing information loss and minimizing the dependency of the complement, thereby minimizing information loss.

[0111] For example, it can be expressed by the following formula:

[0112] min D(f(R(G)), f(G))+I(R(G), C(R(G))) (1)

[0113] Where f(·) represents the prediction function, and R(·) and C(·) represent the important graph structure of G and the complement graph of G, respectively.

[0114] The first term aims to minimize the classification performance difference between the important structures R(G) and the original graph G. In the first term, D(·) can represent the difference measure function, thereby preserving the important structural information of the cooperative relationships in R(G).

[0115] The second term aims to minimize the dependency between the complement C(R(G)) and the important graph structure R(G).

[0116] Therefore, the graph structure information invariance principle modeling method proposed in this embodiment of the invention can reduce the noise impact of unknown interactions by utilizing the invariant relationship between users and items.

[0117] In an optional embodiment of the present invention, the step of determining the important graph structure for the interaction graph includes:

[0118] In the interaction graph, edges are determined to represent the interaction relationship between the user and the item;

[0119] The edges are determined as the scaling factors of the important graph structure;

[0120] Determine a first relation coefficient used to express the interaction relationship between the user and the item in the important graph structure;

[0121] Determine a second relationship coefficient to represent the interaction relationship between the user and the item in the supplementary graph;

[0122] The generation probability for important graph structures is calculated using the scaling factor, the first relation factor, and the second relation factor.

[0123] The important graph structures for the interaction graph are determined based on the generation probability.

[0124] Optionally, the step of determining the important graph structure for the interaction graph based on the generation probability includes:

[0125] Generate a graph autoencoder based on a graph converter architecture;

[0126] The graph autoencoder is used to determine the important graph structures for the interaction graph based on the generation probability.

[0127] In its implementation, to achieve enhanced self-supervised learning, this invention can automatically extract important graph structures on graph G. Simultaneously, collaborative modeling can estimate the subgraph G. R The probability of an important graph structure of G.

[0128] For example, this can be achieved using the following formula:

[0129]

[0130] p(e|R(G))←GT(G,TE(H;Θ) E );Θ G (3)

[0131] In practical applications, the edges of an interaction graph can be used to represent the interaction relationships between users and items. The edges ρ can be determined first in the interaction graph G. R This represents the selection of edges from G as important graph structures. R The proportionality coefficient, where G C Not belonging to G RThe subgraph with edges. Additionally, the first relation coefficient e and the second relation coefficient e′ can be defined to represent the user-item relationship in the important graph structure and the complementary graph, respectively. To evaluate the above distribution probability, edge probabilities p(e|G) and p(e′|G) can be defined as the generation probability of the important graph structure part, and the important graph structure for the interaction graph can be determined based on the generation probability. Furthermore, a parameterized important graph structure generator can be defined using a node-embedded graph autoencoder, as shown in formula (3). The graph autoencoder GT(·) can be built on the Graph Transformer architecture, and the topology embedding module TE(·) is used to capture the cooperative signals in the entire graph, Θ E and Θ G These represent the learnable parameters, which can be determined during the global topology information injection stage. H can be defined to represent the global topology embedding matrix in the l-th layer. During the training process of this invention, the graph collaborative filtering task can be optimized based on the Bayesian Personalized Ranking (BPR) objective loss function, specifically implemented during the task adaptive modeling stage.

[0132] In an optional embodiment of the present invention, it further includes:

[0133] Nodes representing the user or item are identified from the interaction graph;

[0134] Determine the anchor nodes from the nodes, and calculate the distances between all the anchor nodes and the target node;

[0135] The association weight between the target node and the anchor node is determined by the distance.

[0136] Optionally, it also includes:

[0137] Normalization is performed on the associated weights.

[0138] Optionally, the pre-trained model includes a propagation layer and further includes:

[0139] The target node is embedded in the propagation layer based on the association weight.

[0140] Optionally, it also includes:

[0141] A global topology embedding matrix for the target node is determined by a linear transformation matrix; the global topology embedding matrix is ​​used to generate higher-order topology information.

[0142] During the global topology information injection phase, location-aware graph neural networks are highly capable of capturing global relationships. This embodiment of the invention enhances the modeling of collaborative principles by preserving higher-order user or item dependencies. A set of anchor nodes can be sampled first from the user-item interaction graph G = {V, E}. To represent the global topological embedding of users and items based on their connectivity with anchor nodes, embodiments of the present invention can calculate the target node v. k and each anchor node v a The distance d between k，a The target node can be any node other than the anchor node, and its distance is d. k，a Defined as a graph G from v k to v a The minimum number of edges that must be traversed is required, and then the association weight ω of each target node-anchor node pair can be calculated based on the obtained given distance. k，a .

[0143] For example, the association weight ω of each target node-anchor node pair is calculated. k，a This can be achieved using the following formula.

[0144]

[0145] Where q is a set distance threshold; this invention normalizes the node association weights; then, by considering the target node k and each anchor node v a ∈∨ A The association weights between nodes are used to optimize the embedding of the target node.

[0146]

[0147] in, Representing the target node v k In the embeddings of the l-th and (l-1)-th propagation layers, || represents the vector concatenation operation. To represent the global topological context information based on anchor nodes, this invention defines H to represent the global topological embedding matrix in the l-th layer; after propagation through the L-th layer graph information, the model embedding retains high-order topological information; finally, this information can be injected into the corresponding embedding to obtain the final topological embedding.

[0148]

[0149] Among them, W L This is a learnable linear transformation matrix. Through the above method, the parameterized importance graph structure generator proposed in this invention can capture global collaborative relationships and identify information patterns in the interactions between users and items.

[0150] In an optional embodiment of the present invention, it further includes:

[0151] The target node is embedded into a multi-head self-attention mechanism model, and a probability score for the edge is calculated.

[0152] Optionally, the probability score is used to sample the first edge set, and a sampling subset is constructed based on the sampling results.

[0153] In the user-item interaction modeling stage based on Graph Transformer, the important graph structure modeling in this embodiment of the invention can extract rating patterns of user-item interactions. These patterns can be used for self-supervised enhancement in a finitely supervised labeling environment. Therefore, this embodiment of the invention can employ a Transformer architecture and a self-attention mechanism to learn rating patterns.

[0154] In practical applications, learning user preference information that remains unchanged in the learning environment as a self-supervised signal can significantly reduce behavioral noise data. This noise may contain popularity bias and other confounding factors, which can negatively impact recommendation performance.

[0155] The parameterized important graph structure modeling process of this invention can encode the implicitly labeled node relationships on the Graph Transformer framework, that is, nodes with the same label are regarded as having such a relationship. In addition, in order to incorporate the location information of user and item nodes into the topology learning process, the globally topology-aware node embedding can also be processed using a multi-head (N-head) self-attention mechanism.

[0156] For example, node v k and v k′ The correlation between the m-th attention heads can be represented by the following formula.

[0157]

[0158] in, and This represents the transformation weights used to obtain the query and embedding. Since the attention scores encoded by the Graph Transformer capture the strength information of node dependencies, this invention aggregates multi-head attention scores to obtain the probability scores of relevant edges, selecting them as the basis for its principles. These principles correspond to a sampled subset of the core user-item interaction patterns, reflecting the user preference learning process with invariant representations, which can be exemplified by the following formula.

[0159]

[0160] In this embodiment of the invention, sampling can be performed separately from the edge set based on the probability score of the edge. The sampling results can be used to construct a sampling subset and control the size of the sampling subset.

[0161] In an optional embodiment of the present invention, it further includes:

[0162] Determine the second edge set for the aforementioned important graph structure;

[0163] The normalized importance between the target node and the anchor node is determined using the probability score.

[0164] The interaction probability between the user and the item is calculated based on the normalized importance and the second edge set.

[0165] In practical applications, during the task-adaptive modeling stage, in order to perform task-level adaptation in the modeling of important graph structures, this invention provides specific tasks to assist in personalized recommendations.

[0166] Specifically, embodiments of the present invention can utilize important graph structures extracted from Graph Transformer to generate user preference predictions for items, which can be exemplified by the following formula.

[0167]

[0168] Among them, y i，j Indicates user u i Using item p j The probability, Indicates (v) k v k′ The predicted score of user-item interactions, ε R Important graph structure G representing sampling R The second side set, β k，k′ Represents anchor node v k and target node v k The importance of the normalization can be obtained through formula (8).

[0169] Optionally, the step of inputting the interaction graph into a pre-trained model to output a predicted score expressing the relationship between the user and the item, and pushing data based on the predicted score, includes:

[0170] The interaction graph is input into a pre-trained model to output a predicted score based on the interaction probability, which expresses the relationship between the user and the item, and data is pushed based on the predicted score.

[0171] This invention uses residual connections to embed the topology into h. k′ As input, then using each (ui y j The predicted probabilities of the interaction are scored, and the present invention uses BPR loss to guide the modeling process and optimize downstream task objectives.

[0172] For example, the predicted score based on the interaction probability output to express the association between the user and the item can be represented by the following formula.

[0173]

[0174] The paired training triples are formed by sampling users and items and satisfying the following conditions: To incorporate task-specific knowledge, this invention employs a topology-aware GraphTransformer architecture, which can provide task-aware parameter learning and collaborative principle modeling to assist in predictions for different recommendation scenarios.

[0175] Optionally, the pre-trained model is generated by training in the following manner:

[0176] Generate a mask image autoencoder;

[0177] Training is performed to ensure the independence between the important graph structure and the complementary graph;

[0178] Generate an interaction model to output an expression of the interaction relationship between the user and the item.

[0179] Optionally, the step of generating the graph autoencoder includes:

[0180] Generate a mask image for the interaction graph;

[0181] The edges of the mask image are sampled, and the sampling edge probability is determined;

[0182] Determine the selection probability of the target node and the target edge corresponding to the target node in the mask graph autoencoder;

[0183] The embedding position of the target node in the mask graph autoencoder is determined using the sampling edge probability and the selection probability.

[0184] Optionally, the step of training on the independence between the important graph structure and the complementary graph includes:

[0185] The supplementary image is generated using different sampling rates;

[0186] Based on the sampling rate, determine the target sampling rate loss;

[0187] And based on the target sampling rate loss, the similarity between the important graph structure and the complement graph is minimized.

[0188] Optionally, the step of generating an interaction model for outputting an expression of the interaction relationship between the user and the item includes:

[0189] Based on the embedding position of the target node in the mask map autoencoder, a loss function for the interaction model is determined;

[0190] An interaction model is generated based on the loss function to output an expression of the interaction relationship between the user and the item.

[0191] In a specific implementation, the pre-trained model can be trained in three stages: generating a masked graph autoencoder; training on the independence between the important graph structure and the complement graph; and generating an interaction model to output the interaction relationship between the user and the item.

[0192] In the stage of generating the masked graph autoencoder, user-item interaction patterns can be augmented using graph masked autoencoders. To achieve this, embodiments of the present invention can use an important graph structure-aware masked graph autoencoder to calculate scoring information. This masked graph autoencoder infers collaborative principles from the interaction graph to reconstruct the target data based on the autoencoder. Furthermore, to enhance the masked graph data G... M ={V, ε M Sampling is performed, where M represents the mask. In embodiments of the present invention, the reciprocal of the importance graph structure score can be used to mask irrelevant structures; exemplarily, this can be expressed by the following formula.

[0193]

[0194] Where pM(·) is the sampling edge probability of the sampling edge in the mask image. In the mask map autoencoder, the target node v is selected. k and target edge v k′ The probability is shown in Equation (8). Given that the mask image has a higher edge density than the important graph structure, this invention further constructs noise-resistant autoencoders; then, it uses an edge set... mask image G M It is used as the input to an autoencoder network and is defined as follows.

[0195]

[0196] Where S represents the final embedding of the autoencoder, and GT(·) and TE(·) represent the graph transformer and topology information encoder, respectively.

[0197] Furthermore, embodiments of the present invention can embed nodes of order L. To enhance the initial embedding, LightGCN encoding can be used, as it can achieve more complex encoding through a lightweight graph convolutional network; then the embedding S of the user-item interaction used to train the mask is reconstructed, which can be exemplified by the following formula.

[0198]

[0199] Among them, L MAE It is the training objective function for reconstructing the mask interaction relationships. This indicates that on graph G, edge (v) k v k′ The predicted score is calculated based on the cooperative principle modeling. Inspired by this cooperative principle modeling, the graph mask autoencoder proposed in this invention can be applied to various downstream recommendation tasks, and the principle-aware enhancement method proposed in this invention can reduce the impact of noisy data.

[0200] In the training phase focusing on the independence between important graph structures and their complements, to achieve the independence of complements in important graph structure modeling, this invention discloses a learning component to extract the independence between cooperative signals and their corresponding complements, thereby reducing information redundancy. This invention can also minimize the important graph structure G through a contrastive regularization mechanism. R and sampling complement G C The interaction information between them; similar to the sampling method of the complement image, this invention uses different sampling rates to generate the complement image G. C For example, this can be represented by the following formula.

[0201] ε C ~p C (ε C |G)=p M (ε C |G) (14)

[0202] To ensure supplementary diagram G C To minimize the important graph structure G in higher-order representations, which may affect the non-noise edges that could affect independent regularization, this invention uses a low sampling rate loss. R And supplementary diagram G C The similarity between them.

[0203]

[0204] Among them, e k This indicates the embedding of an L-layer stacked LightGCN model, which can be used in the input graph (G R and G C ) recursively pass messages; τ is a hyperparameter of the temperature coefficient. Therefore, compared with the independent regularization loss LCIR, it can embed the principle of all nodes. and complement embedding By conducting comparative learning, the model's important graph structure co-modeling and the independence between complementary graphs are enhanced, thereby improving the noise suppression capability of this invention.

[0205] In the stage of generating an interaction model that expresses the interaction relationship between users and items, embodiments of the present invention can use user-item embeddings S for prediction to train the recommender. Therefore, the present invention defines the loss function as follows.

[0206]

[0207] In contrast, this invention maximizes the prediction of all correct user-item interactions and minimizes the prediction of all incorrect user-item interactions. Furthermore, this invention utilizes the masked graph G in the Graph Transformer. M Replace it with a known interaction graph G, and through To predict user u i and item p j The relationship between them. Finally, by combining multiple training objectives, the present invention achieves the final prediction result shown below.

[0208]

[0209] Here, λ1, λ2, and λ3 are regularization parameters used to balance the loss, and the last term is the Frobenius normal form parameter Θ.

[0210] Through the above methods, (1) this invention re-examines the self-supervised enhancement recommendation method by exploring a collaborative principle modeling scheme based on self-supervised learning technology. This invention not only realizes automated data augmentation in self-supervised learning, but also provides a collaborative principle awareness understanding mechanism to improve the robustness of the model. (2) Based on Graph Transformer, this invention first proposes a "collaborative relationship invariance principle" modeling method, then designs a task-aware adaptive method to alleviate the data bias problem, and finally uses Graph Autoencoder (GAE) technology to reconstruct the masked user-item interaction graph. (3) This invention verifies the effectiveness of the proposed method on multiple datasets, and compared with various advanced baseline models, the proposed method continuously achieves performance improvement under different experimental settings. (4) The user-item recommendation method constructed by this invention can make personalized recommendations based on user behavior records and automatically learn important graph structures to reduce the impact of noise. (5) The user-item intelligent recommendation method of this invention can be applied to various product recommendation, short video recommendation, smart government recommendation and other recommendation system processes, and has no limitations on application scenarios. (6) The proposed solution adopts a novel approach and a self-supervised learning method to solve the problem of sparse feature learning in recommendation systems, thereby improving the generalization performance of the model. (7) The proposed solution is loosely coupled, can quickly adapt to and access new business data, and can be trained based on historical user click logs without manual data labeling. (8) Compared with typical content-based recommendation systems, the proposed method solves the problem of modeling user historical behavior and optimizes the parameter learning process in the recommendation system, alleviating the limitations of the model itself. (9) Compared with most manually constructed recommendation strategies, the proposed method can automatically learn important graph structure features, improving the universality and stability of the recommendation. (10) When used, this invention supports general-purpose CPU and GPU solutions, without the need to adjust the server hardware in the data center, greatly reducing the cost of model training and the cycle of new business development and maintenance. (11) Through the recommendation method proposed in this proposal, optimization can also be achieved for scenarios such as smart government recommendation, smart insurance recommendation, and job recruitment, supporting loosely coupled use of downstream tasks.

[0211] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0212] Reference Figure 2 The diagram illustrates a structural block diagram of a data push device provided in an embodiment of the present invention, which may specifically include the following modules:

[0213] User identifier determination module 201 is used to determine user identifiers and construct a user set for a user based on the user identifiers;

[0214] The item identifier determination module 202 is used to determine the item identifier and construct an item set for the item based on the item identifier;

[0215] The node set and first edge set determination module 203 is used to determine the node set and the first edge set using the user set and the item set;

[0216] The interaction graph generation module 204 is used to generate an interaction graph based on the node set and the first edge set to express the interaction behavior between the user and the item.

[0217] The data push module 205 is used to input the interaction graph into a pre-trained model to output a predicted score that expresses the relationship between the user and the item, and push data based on the predicted score.

[0218] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0219] In addition, this invention also provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the various processes of the above-described data push method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0220] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described data push method embodiments and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0221] Figure 3 A schematic diagram of the hardware structure of an electronic device for implementing various embodiments of the present invention.

[0222] The electronic device 300 includes, but is not limited to, components such as: a radio frequency unit 301, a network module 302, an audio output unit 303, an input unit 304, a sensor 305, a display unit 306, a user input unit 307, an interface unit 308, a memory 309, a processor 310, and a power supply 311. Those skilled in the art will understand that... Figure 3 The electronic device structures shown are not intended to limit the electronic device. An electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements. In embodiments of the present invention, the electronic device includes, but is not limited to, mobile phones, tablet computers, laptops, PDAs, in-vehicle terminals, wearable devices, and pedometers.

[0223] It should be understood that, in this embodiment of the invention, the radio frequency unit 301 can be used for receiving and transmitting signals during information transmission or calls. Specifically, it receives downlink data from the base station and processes it with the processor 310; additionally, it transmits uplink data to the base station. Typically, the radio frequency unit 301 includes, but is not limited to, an antenna, at least one amplifier, a transceiver, a coupler, a low-noise amplifier, a duplexer, etc. Furthermore, the radio frequency unit 301 can also communicate with networks and other devices through a wireless communication system.

[0224] The electronic device provides users with wireless broadband internet access through the network module 302, such as helping users send and receive emails, browse web pages, and access streaming media.

[0225] The audio output unit 303 can convert audio data received by the radio frequency unit 301 or the network module 302 or stored in the memory 309 into audio signals and output them as sound. Furthermore, the audio output unit 303 can also provide audio output related to specific functions performed by the electronic device 300 (e.g., call signal reception sound, message reception sound, etc.). The audio output unit 303 includes a speaker, a buzzer, and a receiver, etc.

[0226] Input unit 304 is used to receive audio or video signals. Input unit 304 may include a graphics processing unit (GPU) 3041 and a microphone 3042. The GPU 3041 processes image data of still images or videos acquired by an image capture device (such as a camera) in video capture mode or image capture mode. The processed image frames can be displayed on display unit 306. The image frames processed by GPU 3041 can be stored in memory 309 (or other storage media) or transmitted via radio frequency unit 301 or network module 302. Microphone 3042 can receive sound and process such sound into audio data. The processed audio data can be converted into a format that can be transmitted to a mobile communication base station via radio frequency unit 301 in telephone call mode.

[0227] The electronic device 300 also includes at least one sensor 305, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor includes an ambient light sensor and a proximity sensor. The ambient light sensor can adjust the brightness of the display panel 3061 according to the ambient light level, and the proximity sensor can turn off the display panel 3061 and / or backlight when the electronic device 300 is moved to the ear. As a type of motion sensor, an accelerometer sensor can detect the magnitude of acceleration in various directions (generally three axes). When stationary, it can detect the magnitude and direction of gravity and can be used to identify the posture of the electronic device (such as landscape / portrait switching, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, tapping), etc. The sensor 305 may also include a fingerprint sensor, pressure sensor, iris sensor, molecular sensor, gyroscope, barometer, hygrometer, thermometer, infrared sensor, etc., which will not be described in detail here.

[0228] The display unit 306 is used to display information input by the user or information provided to the user. The display unit 306 may include a display panel 3061, which may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), or the like.

[0229] User input unit 307 can be used to receive input numerical or character information, and generate key signal inputs related to user settings and function control of electronic devices. Specifically, user input unit 307 includes a touch panel 3071 and other input devices 3072. Touch panel 3071, also known as a touch screen, can collect touch operations performed by the user on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or near touch panel 3071). Touch panel 3071 may include two parts: a touch detection device and a touch controller. The touch detection device detects the user's touch position and the signal generated by the touch operation, and transmits the signal to the touch controller; the touch controller receives touch information from the touch detection device, converts it into touch point coordinates, and sends it to the processor 310, which receives and executes commands from the processor 310. In addition, touch panel 3071 can be implemented using various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to touch panel 3071, user input unit 307 may also include other input devices 3072. Specifically, other input devices 3072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, power buttons, etc.), trackballs, mice, joysticks, etc., which will not be described in detail here.

[0230] Furthermore, the touch panel 3071 can cover the display panel 3061. When the touch panel 3071 detects a touch operation on or near it, it transmits the information to the processor 310 to determine the type of touch event. Subsequently, the processor 310 provides corresponding visual output on the display panel 3061 based on the type of touch event. Although in Figure 3 In this embodiment, the touch panel 3071 and the display panel 3061 are two independent components to realize the input and output functions of the electronic device. However, in some embodiments, the touch panel 3071 and the display panel 3061 can be integrated to realize the input and output functions of the electronic device. The specific implementation is not limited here.

[0231] Interface unit 308 serves as an interface for connecting external devices to electronic device 300. For example, external devices may include a wired or wireless headphone port, an external power supply (or battery charger) port, a wired or wireless data port, a memory card port, a port for connecting a device with an identification module, an audio input / output (I / O) port, a video I / O port, a headphone port, and so on. Interface unit 308 can be used to receive input from external devices (e.g., data, power, etc.) and transmit the received input to one or more components within electronic device 300, or it can be used to transmit data between electronic device 300 and external devices.

[0232] The memory 309 can be used to store software programs and various data. The memory 309 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback, image playback, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory 309 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0233] The processor 310 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 309, and by calling data stored in the memory 309, it performs various functions and processes data, thereby providing overall monitoring of the electronic device. The processor 310 may include one or more processing units; preferably, the processor 310 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 310.

[0234] The electronic device 300 may also include a power supply 311 (such as a battery) for supplying power to various components. Preferably, the power supply 311 can be logically connected to the processor 310 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system.

[0235] In addition, the electronic device 300 includes some functional modules not shown, which will not be described in detail here.

[0236] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0237] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0238] like Figure 4 As shown, in another embodiment of the present invention, a computer-readable storage medium 401 is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the data push method described in the above embodiment.

[0239] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

[0240] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0241] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0242] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0243] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0244] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0245] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0246] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data push method, characterized by, The method comprises the following steps: determining a user identifier and constructing a user set for a user based on the user identifier; determining an item identifier and constructing an item set for an item based on the item identifier; determining a node set and a first edge set using the user set and the item set; generating an interaction graph for expressing the interaction behavior between the user and the item based on the node set and the first edge set; inputting the interaction graph into a pre-trained model to output a prediction score for expressing the association relationship between the user and the item, and pushing data based on the prediction score; wherein the pre-trained model is provided with a prediction function, and the pre-trained model further comprises: determining an important graph structure for the interaction graph; determining a difference measurement function for the interaction graph; reducing the classification performance difference between the important graph structure and the interaction graph by a preset algorithm using the difference measurement function and the prediction function; determining a complement graph for the interaction graph and a complement set for the complement graph; minimizing the dependence of the complement set on the important graph structure by a preset algorithm; wherein the step of determining the important graph structure for the interaction graph comprises: determining an edge in the interaction graph for expressing the interaction relationship between the user and the item; determining the edge as a proportionality coefficient of the important graph structure; determining a first relationship coefficient for expressing the interaction relationship between the user and the item in the important graph structure; determining a second relationship coefficient for expressing the interaction relationship between the user and the item in the complement graph; calculating a generation probability for the important graph structure using the proportionality coefficient, the first relationship coefficient and the second relationship coefficient; determining the important graph structure for the interaction graph based on the generation probability; wherein the objective function is defined by minimizing the information loss and minimizing the dependence of the complement set, and the objective function is represented by the following formula: wherein, represents a prediction function, represents an important graph structure of G, represents a complement graph of G; The first term of the objective function aims to minimize the difference in classification performance between the important structure and the original graph G; the second term of the objective function aims to minimize the dependency of the complement and the important graph structure ​ 2. The method of claim 1, wherein, the step of determining the important graph structure for the interaction graph based on the generation probability comprises: generating a graph autoencoder based on a graph transformer architecture; determining the important graph structure for the interaction graph based on the generation probability using the graph autoencoder.

3. The method of claim 2, wherein, Further comprising: determining a node for expressing the user or the item from the interaction graph; determining an anchor node from the node and calculating the distance between all the anchor nodes and a target node; determining the association weight between the target node and the anchor node by the distance.

4. The method of claim 3, wherein, Further comprising: performing a normalization operation on the association weight.

5. The method of claim 4, wherein, The pre-trained model comprises a propagation layer, and further comprises: embedding the target node in the propagation layer based on the association weight.

6. The method of claim 5, wherein, Further comprising: determining a global topology embedding matrix for the target node by a linear transformation matrix; the global topology embedding matrix is used to generate high-order topology information.

7. The method of claim 6, wherein, Further comprising: embedding the target node into a multi-head self-attention mechanism model to calculate a probability score for the edge.

8. The method of claim 7, wherein, Further comprising: sampling in the first edge set using the probability score and constructing a sampled set based on the sampling result.

9. The method of claim 8, wherein, Further comprising: determining a second edge set for the important graph structure; determine a normalized importance between the target node and the anchor node based on the probability score; calculate an interaction probability between the user and the item based on the normalized importance and the second edge set.

10. The method of claim 9, wherein, The step of inputting the interaction graph into a pre-trained model to output a predicted score for expressing an association relationship between the user and the item, and pushing data based on the predicted score, comprises: inputting the interaction graph into a pre-trained model to output a predicted score for expressing an association relationship between the user and the item based on the interaction probability, and pushing data based on the predicted score.

11. The method of claim 10, wherein, The pre-trained model is trained by the following way: generating a mask graph autoencoder; training for independence between the important graph structure and the complementary graph; generating an interaction model for outputting an interaction relationship between the user and the item.

12. The method of claim 11, wherein, The step of generating a graph autoencoder comprises: generating a mask graph for the interaction graph; sampling edges of the mask graph and determining a sampling edge probability; determining a selection probability of the target node and a target edge corresponding to the target node in the mask graph autoencoder; determining an embedding position of the target node in the mask graph autoencoder using the sampling edge probability and the selection probability.

13. The method of claim 12, wherein, The step of training for independence between the important graph structure and the complementary graph comprises: generating the complementary graph using different sampling rates; determining a target sampling rate loss based on the sampling rate; and minimizing the similarity between the important graph structure and the complementary graph based on the target sampling rate loss.

14. The method of claim 13, wherein, The step of generating an interaction model for outputting an interaction relationship between the user and the item comprises: determining a loss function for the interaction model based on the embedding position of the target node in the mask graph autoencoder; generating an interaction model for outputting an interaction relationship between the user and the item based on the loss function.

15. A data push apparatus, characterized by comprising: comprise: a user identifier determination module configured to determine a user identifier and construct a user set for a user based on the user identifier; an item identifier determination module configured to determine an item identifier and construct an item set for an item based on the item identifier; a node set and first edge set determination module configured to determine a node set and a first edge set using the user set and the item set; an interaction graph generation module configured to generate an interaction graph for expressing an interaction behavior between the user and the item based on the node set and the first edge set; The data pushing module is configured to input the interaction graph into a pre-trained model to output a predicted score for expressing the association relationship between the user and the item, and push data based on the predicted score; wherein the pre-trained model is provided with a prediction function, and the pre-trained model further comprises: determining an important graph structure for the interaction graph; determining a difference measurement function for the interaction graph; reducing the difference in classification performance between the important graph structure and the interaction graph by using the difference measurement function and the prediction function through a preset algorithm; determining a complementary graph for the interaction graph, and a complement set for the complementary graph; and minimizing the dependence of the complement set on the important graph structure through a preset algorithm; The step of determining the important graph structure for the interaction graph comprises: determining an edge for expressing the interaction relationship between the user and the item in the interaction graph; determining the edge as a proportionality coefficient of the important graph structure; determining a first relationship coefficient for expressing the interaction relationship between the user and the item in the important graph structure; determining a second relationship coefficient for expressing the interaction relationship between the user and the item in the complementary graph; calculating a generation probability for the important graph structure by using the proportionality coefficient, the first relationship coefficient and the second relationship coefficient; determining the important graph structure for the interaction graph based on the generation probability; wherein the objective function is defined by minimizing the information loss and minimizing the dependence of the complement set, and the objective function is represented by the following formula: wherein, represents a prediction function, represents an important graph structure of G, represents a complement graph of G; The first term of the objective function aims to minimize the difference in classification performance between the important structure and the original graph G; the second term of the objective function aims to minimize the dependency of the complement and the important graph structure .

16. An electronic device, comprising: A computer readable storage medium having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the method of any one of claims 1-14.

17. A computer readable storage medium having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the method of any one of claims 1-14. ​ ​

Images