A method, apparatus, electronic device and storage medium for entity classification

Abstract

This disclosure provides an entity classification method, apparatus, electronic device, and storage medium. The method includes: acquiring multiple entity relationship networks corresponding to multiple entity nodes; the same entity node corresponding to different entity relationships in different entity relationship networks; performing multi-label prediction on the multiple entity relationship networks based on multiple trained base learners to obtain the multi-label prediction result of each entity node in each entity relationship network; and determining the final multi-label prediction result corresponding to each entity node based on the multi-label prediction result of each entity node in each entity relationship network. This disclosure, from the perspective of multi-view learning, performs multi-label prediction based on multiple entity relationship networks, which can more fully explore the relationships between entities, making the multi-label prediction results for entity nodes more accurate.

Description

Technical Field

[0001] This disclosure relates to the field of computer application technology, and more specifically, to an entity classification method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the development of information technology, people have entered the era of "big data". Every day, massive amounts of diverse data emerge. On the one hand, this greatly enriches the semantic information of entities such as users, articles, and products. On the other hand, it makes it difficult for people to discover the required pattern features from the huge amount of data.

[0003] To systematically, standardizedly, and precisely describe and analyze the characteristics of entities, a standardized label set is typically established for labeling entities. However, given the massive scale of data, manually labeling the data is clearly an impossible task. Therefore, label classification and prediction technology has emerged and has been widely applied in various fields such as computer vision, natural language processing, bioinformatics, and information retrieval.

[0004] Label classification prediction technology is a technique that predicts the label of new data by summarizing the patterns of known entity labels. Traditional classification problems assume that an instance is associated with only one label, i.e., single-view classification prediction. However, in practice, a sample usually has multiple labels, making traditional classification techniques no longer applicable. Summary of the Invention

[0005] This disclosure provides at least one entity classification method, apparatus, electronic device, and storage medium to achieve multi-view classification prediction with high accuracy.

[0006] In a first aspect, embodiments of this disclosure provide an entity classification method, including:

[0007] Retrieve multiple entity relationship networks corresponding to multiple entity nodes; the same entity node corresponds to different entity relationships in different entity relationship networks;

[0008] Multi-label prediction is performed on the multiple entity relationship networks based on multiple pre-trained base learners to obtain the multi-label prediction result of each entity node in each entity relationship network.

[0009] Based on the multi-label prediction results of each entity node in each entity relationship network, the final multi-label prediction result corresponding to each entity node is determined.

[0010] In one possible implementation, when the multi-label prediction result includes prediction scores for multiple candidate labels; determining the final multi-label prediction result for each entity node based on the multi-label prediction result for each entity node in each entity relationship network includes:

[0011] For the target candidate label among the multiple candidate labels, based on the multi-label prediction results of each entity node in each entity relationship network, determine the label feature vector of each entity node for the target candidate label;

[0012] Based on the trained meta-learner, the label feature vectors of multiple entity nodes for the target candidate label are sorted and learned to determine the final multi-label prediction result for each entity node.

[0013] In one possible implementation, determining the label feature vector for each entity node relative to the target candidate label based on the multi-label prediction result of each entity node in each entity relationship network includes:

[0014] For the target entity relationship network in the plurality of entity relationship networks, the target prediction score that matches the target candidate label is selected from the predicted scores of the entity node for the plurality of candidate labels in the target entity relationship network;

[0015] The target prediction scores selected from the multiple entity relationship networks are combined to obtain the label feature vector of the entity node for the target candidate label.

[0016] In one possible implementation, the step of ranking and learning the label feature vectors of multiple entity nodes for the target candidate label based on the trained meta-learner to determine the final multi-label prediction result for each entity node includes:

[0017] For the target entity node among the multiple entity nodes, the label feature vector of the target entity node for the target candidate label is input into the trained meta-learner to determine the predicted score after the ranking of the multiple candidate labels corresponding to the target entity node.

[0018] Based on the sorted predicted scores, the multi-label prediction results corresponding to the target entity node are determined.

[0019] In one possible implementation, determining the multi-label prediction result corresponding to the target entity node based on the sorted predicted score includes:

[0020] If the predicted score after sorting is greater than a preset threshold, the candidate label corresponding to the predicted score after sorting is determined as the multi-label prediction result corresponding to the target entity node.

[0021] In one possible implementation, the meta-learner is trained according to the following steps:

[0022] Obtain a sample training set that includes multiple sample feature vectors, wherein each dimension value of the sample feature vector points to an entity node label pair;

[0023] Iterate through each sample feature vector in the training set and determine the feature vector of each sample. -gradient;

[0024] Based on the multiple sample feature vectors and each sample feature vector - Gradient construction of regression trees;

[0025] Based on the correspondence between each leaf node of the regression tree and each entity node, the ranking score of each entity node in the meta-learner to be trained is updated to obtain the trained meta-learner.

[0026] In one possible implementation, the multi-label prediction of the multiple entity relationship networks based on multiple trained base learners, to obtain the multi-label prediction result of each entity node in each entity relationship network, includes:

[0027] For the target entity relationship network in the multiple entity relationship networks, a target base learner corresponding to the target entity relationship network is used to perform multi-label prediction on the target entity relationship network to obtain the multi-label prediction result of each entity node in the target entity relationship network.

[0028] In one possible implementation, when the target base learner includes a fully connected input layer, a representation learning module, and a fully connected output layer, the step of using the target base learner corresponding to the target entity relationship network to perform multi-label prediction on the target entity relationship network, and obtaining the multi-label prediction result of each entity node in the target entity relationship network, includes:

[0029] The original feature vector of each entity node in the target entity relationship network is input into the fully connected input layer of the target base learner to determine the dimensionality-reduced feature vector output by the fully connected input layer for each entity node; and,

[0030] The reduced-dimensionality feature vector is input into the representation learning module of the target base learner to determine a hidden feature vector containing low-order signals and similar node information; and,

[0031] The hidden feature vector is input into the fully connected output layer of the target base learner to determine the multi-label prediction result of each entity node in the target entity relationship network.

[0032] In one possible implementation, a target base learner comprising a fully connected input layer, a representation learning module, and a fully connected output layer is trained according to the following steps:

[0033] Obtain a sample entity relationship network, wherein some entity nodes in the sample entity relationship network have multi-label annotation results;

[0034] The fully connected input layer is used to perform dimensionality reduction representation on each entity node in the sample entity relationship network, and the node dimensionality reduction implicit representation of each entity node after dimensionality reduction transformation is determined.

[0035] The graph convolutional layer included in the representation learning module is used to perform low-order signal attention learning on the dimensionality-reduced latent representation of the nodes, resulting in a node attention latent representation for each entity node after attention learning; and the conditional random field layer included in the representation learning module is used to perform node similarity learning on the node attention latent representation, resulting in a node similarity latent representation for each entity node after node similarity learning.

[0036] The fully connected output layer is used to perform multi-label prediction on the node similarity implicit representation of each entity node to obtain the prediction result;

[0037] The target base learner is adjusted based on the prediction results and the multi-label annotation results to obtain a trained target base learner.

[0038] In one possible implementation, when multiple graph convolutional layers are included, the step of using the graph convolutional layers included in the representation learning module to perform low-order signal attention learning on the node dimensionality-reduced latent representation to obtain a node attention latent representation for each entity node after attention learning includes:

[0039] The dimensionality-reduced implicit representation of the nodes is used as the initial representation for each graph convolutional layer;

[0040] For the current graph convolutional layer other than the first graph convolutional layer, perform the following steps:

[0041] The sample entity relationship network and the node attention implicit representation output by the previous graph convolutional layer before the current graph convolutional layer are input into the graph attention layer in other graph convolutional layers to determine the node attention implicit representation output by the graph attention layer.

[0042] Based on the initial representation, the node attention implicit representation output by the graph attention layer, and the training parameters of the current graph convolutional layer, the node attention implicit representation output by the current graph convolutional layer is determined.

[0043] In one possible implementation, determining the node attention latent representation output by the current graph convolutional layer based on the initial representation, the node attention latent representation output by the graph attention layer, and the training parameters of the current graph convolutional layer includes:

[0044] A first graph convolution operator is determined based on a first weighted summation operation between the initial representation and the node attention implicit representation output by the graph attention layer; and a second graph convolution operator is determined based on a second weighted summation operation between the training parameters of the current graph convolution layer and the identity mapping matrix corresponding to the current graph convolution layer.

[0045] Based on the first graph convolution operator and the second graph convolution operator, the node attention implicit representation output by the current graph convolutional layer is determined.

[0046] In one possible implementation, the step of using the conditional random field layer included in the representation learning module to perform node similarity learning on the node attention latent representation, to obtain a node similarity latent representation for each entity node after node similarity learning, includes:

[0047] For each entity node, a maximum conditional probability function is constructed for that entity node; the maximum conditional probability function is determined by a first difference between the node similarity implicit representation of the entity node and the node attention implicit representation of the entity node, and a second difference between the node similarity implicit representation of the entity node and the node similarity implicit representations of other entity nodes among the plurality of entity nodes excluding the entity node.

[0048] Given that the conditional probability function has reached its maximum value, the node similarity implicit representation of each entity node is determined.

[0049] In one possible implementation, constructing the maximized conditional probability function corresponding to the entity node includes:

[0050] For each entity node, obtain the first difference between the node similarity implicit representation of the entity node and the node attention implicit representation of the entity node, the second difference between the node similarity implicit representation of the entity node and the node similarity implicit representation of other entity nodes among the plurality of entity nodes, and the node similarity between the entity node and the other entity nodes.

[0051] The second difference is multiplied by the node similarity to determine the product result.

[0052] The summation of the products between the entity node and each of the other entity nodes is performed to obtain the second difference sum.

[0053] The maximized conditional probability function is determined based on a third weighted summation operation between the first difference and the second difference.

[0054] In one possible implementation, adjusting the target base learner based on the prediction result and the multi-label annotation result to obtain a trained target base learner includes:

[0055] Obtain the first weight parameter used to adjust the number of positive and negative samples, and the second weight parameter used to adjust the contribution of hard-to-distinguish samples;

[0056] Based on the first weight parameter, the second weight parameter, and the difference between the prediction result and the multi-label annotation result, the target loss function value of the target base learner is determined;

[0057] The training parameters of the target base learner are adjusted at least once based on the target loss function value to obtain a trained target base learner.

[0058] Secondly, embodiments of this disclosure also provide an entity classification device, comprising:

[0059] The acquisition module is used to acquire multiple entity relationship networks corresponding to multiple entity nodes; the same entity node corresponds to different entity relationships in different entity relationship networks;

[0060] The prediction module is used to perform multi-label prediction on the multiple entity relationship networks based on multiple trained base learners, and obtain the multi-label prediction result of each entity node in each entity relationship network;

[0061] The classification module is used to determine the final multi-label prediction result for each entity node based on the multi-label prediction result of each entity node in each entity relationship network.

[0062] Thirdly, embodiments of this disclosure also provide an electronic device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, they perform the entity classification method as described in the first aspect and any of its various embodiments.

[0063] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the entity classification method as described in the first aspect and its various embodiments.

[0064] By employing the aforementioned entity classification method, apparatus, electronic device, and storage medium, when multiple entity relationship networks corresponding to multiple entity nodes are obtained, multi-label prediction can be performed on the multiple entity relationship networks based on multiple trained base learners. This yields the multi-label prediction result for each entity node in each entity relationship network. Then, based on the multi-label prediction results for each entity node in each entity relationship network, the final multi-label prediction result for each entity node is determined. In other words, this disclosure, from the perspective of multi-view learning, performs multi-label prediction based on multiple entity relationship networks, which can more fully explore the relationships between entities, making the multi-label prediction results for entity nodes more accurate.

[0065] Other advantages of this disclosure will be explained in more detail below with reference to the accompanying drawings.

[0066] It should be understood that the above description is merely an overview of the technical solution of this disclosure, so as to enable a clearer understanding of the technical means of this disclosure, and thus enable its implementation in accordance with the contents of the specification. To make the above and other objects, features, and advantages of this disclosure more apparent and understandable, specific embodiments of this disclosure are illustrated below. Attached Figure Description

[0067] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort. Furthermore, the same reference numerals denote the same components throughout the drawings. In the drawings:

[0068] Figure 1 A flowchart of an entity classification method provided by an embodiment of this disclosure is shown;

[0069] Figure 2 The flowchart illustrates a specific method for training a base learner in the entity classification method provided in this embodiment of the present disclosure.

[0070] Figure 3The flowchart illustrates a specific method for constructing implicit representations in the entity classification method provided in this embodiment of the present disclosure.

[0071] Figure 4 This diagram illustrates the application of the entity classification method provided in the embodiments of this disclosure.

[0072] Figure 5 A schematic diagram of an entity classification device provided in an embodiment of this disclosure is shown;

[0073] Figure 6 A schematic diagram of an electronic device provided in an embodiment of the present disclosure is shown. Detailed Implementation

[0074] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0075] In the description of embodiments disclosed herein, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, numbers, steps, behaviors, components, portions or combinations thereof disclosed herein, and are not intended to exclude the possibility of the presence of one or more other features, numbers, steps, behaviors, components, portions or combinations thereof.

[0076] Unless otherwise stated, " / " means "or". For example, A / B can mean A or B. In this article, "and / or" is merely a way of describing the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone.

[0077] The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more.

[0078] Research has found that multi-label classification prediction problems generally have the following characteristics: (1) Entities in the real world are not independent of each other but have mutual relationships. For example, there are social relationships among users on e-commerce platforms, and similar social relationships generally mean similar purchasing preferences. In the process of multi-label learning, adding the correlation between entities as a priori knowledge into the model can help to capture the hidden distribution patterns more accurately. (2) In actual tasks, the same thing can be described from multiple ways or angles, which constitutes multiple views of the thing. For example, a video clip contains information in multiple modalities such as images, audio and video, and subtitles, and a news article has information in multiple parts such as title, abstract, and body. Compared with single-view representation, multi-view contains more complementary information and can describe things more comprehensively and diversely, thereby helping the learner to obtain better prediction performance. (3) In real tasks, the distribution of label labeling frequency is often very unbalanced and generally follows a class-scale-free distribution. That is to say, a small number of labels label a large number of samples, while the remaining most labels are only used to label a few samples, and the ratio of positive and negative samples of the labels is seriously unbalanced.

[0079] A series of studies have been conducted on multi-label learning in related technologies. For example, comparative document 1 provides a lazy multi-label learning algorithm based on the traditional K-Nearest Neighbors (KNN) algorithm and the maximum a posteriori probability rule. Its main idea is that the set of labels for a sample can be determined by the sample's nearest neighbors.

[0080] The ML-KNN algorithm can be divided into five steps:

[0081] a) Step 1: Given an existing labeled training sample set ,in , For tag sets. Each sample Calculate its nearest neighbors .

[0082] b) Step 2: For the tag set Each tag in Calculate the samples to be classified Does the tag set contain? Prior probability:

[0083]

[0084] in, Indicates label This incident occurred, The total number of samples in the training set. For smoothing terms, it is usually taken as... Then calculate the frequency array. and :

[0085]

[0086] here, Indicates that the sample in the sample set is the first Tag The labeled and the sample In the neighborhood The number of labeled samples is exactly The total number of samples. Among them,

[0087]

[0088] Indicates sample of Among the neighbors, the first Tag Total number of occurrences.

[0089] c) Step 3: Calculate the samples to be classified of Neighbor .

[0090] d) Step 4: For each tag in the tag set Statistical analysis of samples to be classified of Neighbor The Middle Tag Frequency of occurrence:

[0091]

[0092] e) Step 5: Predict the sample to be classified Annotation results:

[0093]

[0094] According to Bayes' theorem, we have

[0095]

[0096] and

[0097]

[0098] here, Indicates the sample to be classified of Labeled in the neighborhood The number of labeled samples is exactly Under these conditions, the sample Also tagged The probabilities are marked. This indicates that the sample is to be classified. Tag In the case of the annotation, its Labeled in the neighborhood The number of labeled samples is exactly The possibility.

[0099] However, this algorithm has the following technical drawbacks:

[0100] 1) This algorithm does not explicitly utilize real entity relationship network information. In fact, it implicitly constructs an entity relationship network based on the nearest neighbor concept: each node has k neighboring nodes, and the edge weights are calculated from the Euclidean distance of the feature vectors between nodes. However, it is obvious that the graph constructed in this way cannot accurately describe the relationships between nodes. Some nodes may only be similar to a few nodes, while some nodes may be related to many nodes. Simply defining a threshold for the degree of a node is unreasonable, and the constructed k-nearest neighbor graph usually deviates from the existing entity relationship network in the actual problem.

[0101] 2) The algorithm performs poorly in predicting rare categories when the samples are imbalanced. This is because the algorithm calculates the proportion of samples in their k-nearest neighbors that are labeled with that label. If a label appears very infrequently in the training set, there will be almost no samples in the k-nearest neighbors labeled with that label, leading to significant distortion in the calculated posterior probability.

[0102] 3) This algorithm is based on a single-view design, which simply combines all sample features into a single feature vector, and cannot be directly applied to multi-view data.

[0103] Furthermore, prior art document 2 in the related art also provides a collaborative multi-label propagation algorithm, CMLP. Compared with the classic single-label propagation algorithm, CMLP is based on the "collaborative assumption," which takes the correlation between labels into account in the model, assuming that the prediction result of a single label comes not only from the contribution of the label itself, but also from the contributions of other labels.

[0104] Let the characteristic matrix of the instance be... ,in For the number of instances, Let be the feature dimension. This is the label matrix of the labeled instances, where This represents the number of labeled instances. The number of labels. A nonnegative matrix. Let be the adjacency weight matrix of the entity relationship network. Let for The normalized propagation matrix, in which It is a diagonal matrix. .

[0105] The CMLP algorithm consists of two main steps:

[0106] a) Step 1: Estimate the label association matrix.

[0107] Based on the "co-operation hypothesis," the authors introduce a label association matrix. ,in Reflects the first The first label is for the first The contribution size of each label, and Therefore, the author believes that the final prediction result should be the original prediction result. And introduced the correlation matrix The trade-off between the prediction results, i.e.

[0108]

[0109] To solve the incidence matrix Starting from the above equation, the author derives a ridge regression optimization problem:

[0110]

[0111] in, and They represent and The List, For the degree of synergy, This is the regularization parameter.

[0112] b) Step 2: Obtain the prediction results.

[0113] Let the output of the model be... ,in The objective function of the CMLP algorithm is defined as:

[0114]

[0115] in, For the introduction of intermediate variables, , and These are the parameters for the balance term. Representation matrix The OK. It represents the Frobenius norm. The paper employs an alternating iterative method for optimization, with initial values ​​for the variables... and .

[0116] First, fix renew ,have The recursive steps for gradient descent are:

[0117]

[0118] in, For learning rate, .

[0119] Then, fix renew ,have Closed-form solution:

[0120]

[0121] in, .

[0122] Repeat the above two steps until convergence, then convert the output into the final prediction result. .

[0123] However, this algorithm has the following technical drawbacks:

[0124] 1) The entity relationship network in this algorithm is actually a k-nearest neighbor graph constructed from entity feature vectors. Therefore, the algorithm only utilizes the entity's features, not the network. In some practical scenarios, entity feature vectors and entity relationship networks often coexist, but this algorithm does not take this situation into account.

[0125] 2) Although the algorithm explicitly utilizes the entity relationship network through the label propagation algorithm, its receptive field is limited to the first-order neighborhood of a node, and it cannot acquire broader connectivity information. Furthermore, the label propagation algorithm can only capture linear graph topologies and cannot capture complex nonlinear structures.

[0126] 3) This algorithm is designed for situations where an entity has only one type of relationship network, and cannot be applied to scenarios with multiple entity relationship networks. If it is to be applied, the multi-view network must be compressed into a single-view network beforehand through weighted averaging, consistency ensemble, or other means. This will inevitably lead to potential information loss and noise introduction, affecting the final prediction performance of the model.

[0127] Although a large amount of research has focused on entity multi-label classification algorithms, the applicant found after reviewing a large amount of existing technical materials that no one has yet approached multi-label prediction from the perspective of multi-view learning, based on multiple entity relationship networks.

[0128] To fill this gap, the applicant proposed using a Graph Convolutional Network (GCN) to extract information from entity relationship networks and integrating information from multiple views through a ranking learning algorithm to achieve better prediction results.

[0129] To facilitate understanding of this embodiment, a detailed description of the entity classification method disclosed in this disclosure is provided first. The entity classification method provided in this disclosure is generally executed by an electronic device with a certain computing capability. This electronic device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a computing device, etc. In some possible implementations, this entity classification method can be implemented by a processor calling computer-readable instructions stored in memory.

[0130] See Figure 1 The diagram shows a flowchart of an entity classification method provided in an embodiment of this disclosure. The method includes steps S101 to S103, wherein:

[0131] S101: Obtain multiple entity relationship networks corresponding to multiple entity nodes; the same entity node corresponds to different entity relationships in different entity relationship networks;

[0132] S102: Perform multi-label prediction on multiple entity relationship networks based on multiple trained base learners to obtain the multi-label prediction result of each entity node in each entity relationship network;

[0133] S103: Based on the multi-label prediction results of each entity node in each entity relationship network, determine the final multi-label prediction result corresponding to each entity node.

[0134] To facilitate understanding of the entity classification method provided in this disclosure, the application scenarios of this method will be described in detail below. The entity classification method in this disclosure can be applied to any application domain that requires multi-label prediction, such as social media applications that classify users using multiple labels. The final multi-label prediction result corresponding to each entity node can be a multi-label prediction of user loyalty, activity level, topic preferences, consumption preferences, etc.

[0135] The entity classification method provided in this embodiment can perform multi-label prediction on multiple entity relationship networks (corresponding to multiple views) based on multiple trained base learners when multiple entity relationship networks are obtained, so as to better explore the entity relationships between multiple entity nodes under multiple views, thereby making the determined multi-label prediction results more accurate.

[0136] Here, before using the base learner for multi-label prediction, the entity relationship network can first be modeled using the input entity set. This modeled entity relationship network more closely reflects the complex relationships between samples in the real world, rather than their independent nature. Through the input entity relationship network, embodiments of this disclosure can more easily uncover implicit similarities between samples and utilize this similarity to improve the accuracy of multi-label classification.

[0137] The embodiments disclosed herein can be specifically designed for modeling situations where there are multiple relationship networks between entities. They effectively utilize the complementary and consistent characteristics that multiple views usually possess, making up for the problem that the classification information contained in a single view is weak, and helping to improve the final prediction performance of the model by leveraging the advantages of different views.

[0138] Furthermore, multi-view learning avoids the overlearning problem that arises from piling up multiple data sources into high-dimensional single-view data, where the model becomes too complex and incorporates data noise into the classification rules. In addition, in this embodiment, each entity relationship network can have a corresponding base learner for multi-label prediction. That is, the multi-view learning framework allows each view to train its own classifier, thus preserving the characteristics of its respective data source.

[0139] Regardless of the base learner, the overall framework structure can be the same. For example, it can include a fully connected input layer for feature dimensionality reduction, a representation learning module for graph convolution operations, and a fully connected output layer suitable for prediction and classification.

[0140] In practical applications, for a target entity relationship network in a multi-entity relationship network, a target base learner corresponding to the target entity relationship network is used to perform multi-label prediction on the target entity relationship network, so as to obtain the multi-label prediction result of each entity node in the target entity relationship network.

[0141] The target entity relationship network here can be any one of multiple entity relationship networks, each of multiple entity relationship networks, or all of multiple entity relationship networks. This disclosure does not impose specific limitations in this regard. Considering that different entity relationship networks will correspond to node features that can represent different views, the target entity relationship network here can be each network. Thus, for each network, multi-label prediction can be performed using the corresponding trained base learner, thereby obtaining the multi-label prediction result for each entity node in each entity relationship network.

[0142] In the actual process of multi-label prediction for target entity relationship networks, the following steps can be taken:

[0143] Step 1: Input the original feature vector of each entity node in the target entity network into the fully connected input layer of the target base learner, and determine the dimensionality-reduced feature vector output by the fully connected input layer for each entity node;

[0144] Step 2: Input the dimensionality-reduced feature vector into the representation learning module of the target base learner to determine the hidden feature vector containing low-order signals and similar node information;

[0145] Step 3: Input the hidden feature vector into the fully connected output layer of the target base learner to determine the multi-label prediction result of each entity node in the target entity relationship network.

[0146] Here, the fully connected input layer included in the base learner can be used to extract the dimensionality-reduced feature vector for each node, so as to better extract the hidden feature vector through the representation learning module. Then, multi-label prediction can be achieved through the fully connected output layer.

[0147] The representation learning module in the multi-label prediction scheme provided in this disclosure addresses the problem that the receptive field of existing algorithms (such as Comparison Technique 1 and Comparison Document 2) is limited to the first-order neighborhood of nodes. By stacking multiple graph convolutional layers, low-order and high-order topological information in the network is fully integrated, thereby uncovering potential node similarity characteristics. Through the construction of a deeper network structure, a broader receptive field can be obtained, combining more shallow features to represent more complex abstract features. Benefiting from the deep architecture, even when facing large-scale entity relationship networks, this disclosure embodiment can achieve a relatively global understanding of the entire network structure.

[0148] It should be noted that the nodes and entity nodes here can be used as the same term for reference.

[0149] This is primarily because, generally speaking, close interactions may exist between two indirectly adjacent nodes in a social network. For example, users A and B frequently share posts, as do users A and C. Although users B and C do not know each other, it can be inferred that they have similar preferences for posts. In this example, users B and C are not directly adjacent (1-hop neighborhood), but rather connected through user A's secondary adjacent nodes (2-hop neighborhood). Therefore, considering higher-order network topology helps uncover potential entity similarity characteristics, making the representation learning module here more practically significant.

[0150] Considering the crucial role of training the base learner in achieving multi-label prediction, the following example illustrates the training process using the target base learner, specifically including the following steps:

[0151] Step 1: Obtain the sample entity relationship network. Some entity nodes in the sample entity relationship network have multi-label annotation results.

[0152] Step 2: Use a fully connected input layer to perform dimensionality reduction on each entity node in the sample entity relationship network, and determine the implicit node dimensionality reduction representation of each entity node after dimensionality reduction transformation;

[0153] Step 3: Use the graph convolutional layer included in the representation learning module to perform low-order signal attention learning on the dimensionality reduction latent representation of nodes, and obtain the node attention latent representation of each entity node after attention learning; and use the conditional random field layer included in the representation learning module to perform node similarity learning on the node attention latent representation, and obtain the node similarity latent representation of each entity node after node similarity learning.

[0154] Step 4: Use the fully connected output layer to perform multi-label prediction on the node similarity latent representation of each entity node to obtain the prediction result;

[0155] Step 5: Adjust the target base learner based on the prediction results and multi-label annotation results to obtain the trained target base learner.

[0156] Simply stacking a large number of graph convolutional modules can lead to the attenuation of low-order topological information by the time it reaches deep graph convolutional layers, resulting in weak low-order topological signals in the final node embedding representation. In practical applications, while some second-order adjacent nodes may have relatively close interactions in user social networks, the likelihood of a close relationship between a pair of high-order nodes is generally low. Therefore, low-order topological information contributes more to the final prediction result than high-order topological information.

[0157] Based on this, the embodiments of this disclosure use two solutions: (1) Based on the idea of ​​"skip connection", an initial representation of the node is introduced in each convolutional layer, and an identity mapping is added to the transformation matrix so that the initial representation of the node and low-order topological information can be passed to deeper hidden layers; (2) A conditional random field layer is added after each graph convolutional module to ensure that the similarity information between adjacent nodes is preserved and to avoid the loss of similarity information after repeated graph convolution transformations. Through the above two solutions, the base learner in the embodiments of this disclosure can effectively alleviate the oversmoothing problem and reasonably balance the contribution of low-order and high-order topological information to the node embedding representation, thereby obtaining a better multi-label prediction effect than the original graph convolutional module.

[0158] Specifically, the process of determining the implicit representation of node attention mentioned above can be determined through the following steps:

[0159] Step 1: Use the dimensionality-reduced implicit representation of the nodes as the initial representation for each graph convolutional layer;

[0160] Step 2: For the current graph convolutional layer other than the first graph convolutional layer, perform the following steps:

[0161] (1) Input the node attention implicit representation of the sample entity relationship network and the output of the previous graph convolutional layer before the current graph convolutional layer into the graph attention layer in other graph convolutional layers to determine the node attention implicit representation output by the graph attention layer.

[0162] (2) Based on the initial representation, the node attention implicit representation output by the graph attention layer, and the training parameters of the current graph convolutional layer, determine the node attention implicit representation output by the current graph convolutional layer.

[0163] Furthermore, the process of determining the aforementioned implicit representation of node similarity can be achieved through the following steps:

[0164] Step 1: For each entity node, construct the maximum conditional probability function for that entity node. The maximum conditional probability function is determined by the first difference between the node similarity latent representation of the entity node and the node attention latent representation of the entity node, and the second difference between the node similarity latent representation of the entity node and the node similarity latent representations of other entity nodes among the multiple entity nodes.

[0165] Step 2: Given that the conditional probability function has reached its maximum value, determine the node similarity implicit representation for each node.

[0166] As can be seen, the base learner here takes entity label vectors as input and performs feature dimensionality reduction through a fully connected input layer. Then, the pre-processed feature representation is fed into the representation learning module in the entity relationship network. Then, graph convolution is performed to pass messages, incorporating the network's topological information, and conditional random fields are used to adjust the node representations to preserve local similarity features, thereby generating high-quality entity node embedding representations (i.e., node similarity latent representations).

[0167] These embedding representations are then refined through a multilayer perceptron in a fully connected output layer to obtain the final embedding representation, which provides a more accurate likelihood prediction of whether there is a relationship between nodes and labels.

[0168] To facilitate a further understanding of the training process of the base learner, the following section will combine... Figure 2 Provide an explanation of the overall process of the base learner.

[0169] Since the original feature vectors of entity nodes are often high-dimensional and sparse, directly feeding them into a graph convolutional network without processing would cause a sharp increase in computational pressure. Therefore, the first step of the base learner is to reduce the dimensionality of the original feature vectors of the entities through a fully connected input layer, mapping them into a low-dimensional dense vector.

[0170] Specifically, in this embodiment of the disclosure, the transformation function of the fully connected input layer is defined as follows:

[0171]

[0172] in, For the first individual entities The original feature vector, Indicates passing through the first The node implicit representation after dimensionality reduction transformation of the fully connected input layer of each base learner. The weight parameter matrix, This is the bias parameter. For the activation function, the LeakyReLU linear rectified function is selected in this embodiment, which is defined as follows:

[0173]

[0174] Where ρ∈(0,1) is a set constant, representing the slope of the third quadrant ray in the activation function.

[0175] like Figure 2 As shown, the representation learning module, as the core component of the base learner, can be divided into two main sub-modules: graph convolution and conditional random fields (CRF). First, to incorporate the topological information of the entity relationship network into the node embedding representation, this embodiment uses graph convolutional network technology. The graph convolutional network takes the entity node embedding representation obtained from the previous layer as input and performs message passing operations on the entity relationship network to generate a preliminary latent representation. Then, to preserve the similarity information within the direct neighborhood, the preliminary latent representation is further processed using a conditional random field to obtain a new node embedding representation. Such a representation learning module can be stacked in multiple layers (total number of layers denoted as L), using the output of the previous layer as the input of the next layer, thereby fully capturing both low-order and high-order topological structures in the network.

[0176] The graph convolution module in this embodiment is mainly divided into four parts: graph convolution, batch normalization (BN), activation function, and dropout.

[0177] The graph convolutional layer operates on the entity relationship network. It receives the embedded representations of entity nodes obtained in the previous step as input, introduces the topological structure information of the entity relationship network through a transformation function, and outputs new embedded representations of entity nodes.

[0178] The graph convolutional layer here comprises two steps: First, message passing is performed on the entity-relationship network through a graph attention layer (GAT); second, the results generated by the previous graph attention layer are further processed through initial representations and identity mappings to preserve low-order signals in the node representations. For a detailed implementation of the node attention implicit representation output by the current graph convolutional layer, please refer to [link to relevant documentation]. Figure 3 .

[0179] First, unlike the native graph convolution operation which assigns the same weight to all adjacent nodes and simply sums them, this embodiment employs a dynamic attention mechanism to adaptively learn the weights of adjacent nodes. This allows different neighbors to contribute varying degrees to the information of the central node, thereby mitigating the interference of noise on model learning and enhancing the model's robustness. (Based on entity nodes...) For example, the first The layer's transformation function is defined as:

[0180]

[0181] Among them, attention coefficient The calculation formula is:

[0182]

[0183] In the formula, For entities In the The embedding representations generated by the representation learning modules, and This is the output representation obtained by dimensionality reduction of the fully connected input layer. This indicates the relationship between nodes in the entity relationship network. The set of associated adjacent nodes (excluding) itself). For the first The learning weight parameter matrix of the attention layer in the layer diagram. This is the self-attention parameter vector to be learned, used for calculating the attention coefficients. (Symbols) This indicates a vector concatenation operation.

[0184] Because excessively deep graph convolutional layers can cause oversmoothing, compromising the final prediction accuracy, this embodiment introduces initial node representations and identity mapping techniques after the aforementioned graph attention layer to enhance the lower-level network signals and ensure their full preservation. Specifically, using entity nodes... For example, the first The layer's transformation function is defined as:

[0185]

[0186] in, The output representation obtained by dimensionality reduction of the fully connected input layer is the initial representation of the nodes. This is the identity mapping matrix. Hyperparameters Control the strength of the initial residual connection (corresponding to the first weighted summation operation). Control the strength of the identity mapping (corresponding to the second weighted summation operation). Generally, Using the following formula:

[0187]

[0188] To calculate, among which For hyperparameters, This indicates the current layer of the model. The formula shows that as... Keep increasing The value gradually decreases, and the role of identity mapping in convolution becomes increasingly prominent. Therefore, the deeper the layer, the less new high-order topological information is introduced, and more low-order topological information is preserved.

[0189] In deep neural networks, the distribution of inputs to each layer changes during training as the parameters of the previous layer vary; this phenomenon is called internal covariate shift. After various transformations within the layer, the output of the hidden layers differs from the distribution of the input signal, and this difference intensifies with increasing network depth. This leads to increasingly slower convergence of deep networks, making them difficult to train. In this embodiment, batch normalization is used to normalize the input signal in each round of stochastic gradient descent during training, ensuring that the mean of the output signal is 0 and the variance is 1 across all dimensions. This mitigates the internal covariate shift phenomenon and alleviates the gradient vanishing problem to some extent, resulting in faster and more stable training of deep networks.

[0190] The batch normalization process described above comprises four steps. Given a batch of input signals, i.e., the output of the aforementioned graph convolutional layer... Their empirical means and variances are as follows:

[0191]

[0192]

[0193] Then, the input signal is normalized by recentering and rescaling:

[0194]

[0195] in, Let be an arbitrarily small constant, the purpose of which is to ensure numerical stability. Finally, scale and shift the normalization result to obtain the final batch normalization transformation result:

[0196]

[0197] in, and All of these are parameters to be learned.

[0198] The third step of the graph convolution module is to apply a non-linear activation function to the mapping result of the neurons. In the disclosed embodiment, the LeakyReLU function can be used, resulting in the following output:

[0199]

[0200] In addition, random deactivation, as a commonly used regularization method, can effectively alleviate the overfitting problem of a model. Its core idea is to allow neurons to stop working with a given probability during training, so that they no longer participate in subsequent forward and backward propagation, thereby improving the model's generalization ability.

[0201] During the training phase of the base learner, the formula for calculating random deactivation is:

[0202]

[0203] in, Let represent a mask vector, where each element follows a probability of . A random variable with a Bernoulli distribution, i.e., a neuron has a probability... Reserved. Symbol This represents element-wise multiplication of vectors.

[0204] During the model testing phase, random deactivation was not used; all neurons remained active.

[0205]

[0206] After performing a series of operations in the graph convolution module, the conditional random field module can be used for feature learning of similar nodes. This is mainly because although graph convolution operations can encode connectivity information into the latent representation of nodes, as the nesting level increases, the similarity features between nodes in the first-order neighborhood will be diluted, resulting in the final latent representation of the nodes failing to accurately represent the actual contextual relationship of the nodes in the network. Therefore, in this embodiment, a conditional random field is used to preserve the similarity relationship in the latent representation of nodes.

[0207] Formalistically, given the implicit node representations produced by a graph convolution module Conditional random field models aim to maximize conditional probabilities.

[0208]

[0209] To predict node representations that include similarity information ,in This represents the partition function, which acts as a normalization factor. This represents the energy function. The energy function comprises two parts: a unary energy function and a pairwise energy function. In this embodiment, nodes are defined. The energy function is:

[0210]

[0211] in, and These are hyperparameters (corresponding to the third weighted summation operation) used to balance the two parts of the energy function; Used to represent the first difference, Used to represent the second difference, Represents a node and The similarity between nodes (i.e., node similarity). Obviously, this energy function ensures, on the one hand, that the difference between the node representation generated by the conditional random field processing and the original input node representation is limited, and on the other hand, it ensures that the node representation generated by the conditional random field processing retains the similarity properties of the graph.

[0212] In this embodiment of the disclosure, the conditional random field is solved using the mean-field approximation method. The solution process includes... Round iteration, where the first The formula for calculating the round of iteration is:

[0213]

[0214] Among them, nodes and Similarity between It can be calculated using the Gaussian function:

[0215]

[0216] In the formula, These are the parameters to be learned. After repeated iterations, the nodes are finally obtained. The implicit representation is:

[0217]

[0218] After experiencing After repeated iterations of the layer representation learning module, nodes that fully condense the network topology and similarity information were obtained. Implicit representation Using this representation, nodes can ultimately be generated through a fully connected output layer. Multi-label prediction results:

[0219]

[0220] in, It is the sigmoid function, that is

[0221]

[0222] is the base of the natural logarithm. This is the weight matrix of the parameters to be learned. The first in element Represents a node With tags The likelihood score predicted in the interval.

[0223] In real-world multi-label classification datasets, many labels are only used to label a small number of samples. Using the traditional cross-entropy loss function can lead to poor prediction performance of the final model on low-frequency labels. To alleviate the imbalance between positive and negative samples, in this embodiment, the training parameters can be adjusted by combining a first weight parameter for adjusting the number of positive and negative samples and a second weight parameter for adjusting the contribution of hard-to-distinguish samples.

[0224] In practical applications, the Focal loss function can be used. This loss function dynamically adjusts the contribution of easily distinguishable and difficult-to-distinguish samples to the loss function by balancing the weights between positive and negative samples, making the model training process focus more on those difficult-to-distinguish samples, thereby achieving better training results.

[0225] Specifically, the objective function is defined as follows:

[0226]

[0227] in, This represents the true label value on the training dataset (corresponding to the multi-label annotation result). This is the predicted value of the base learner (corresponding to the prediction result). Parameters (Corresponding to the first weight parameter) Adjusts the weights between positive and negative samples. (Corresponding to the second weight parameter) Adjust the weight of samples that are difficult to distinguish.

[0228] As can be seen, the present invention proposes to use the Focal loss function to train the base learner. By assigning a larger weight to the hard-to-distinguish samples, the influence of the samples on the loss function is expanded, causing the loss function to focus on the hard-to-distinguish samples. This can effectively alleviate the adverse effects of the high sparsity of positive samples and the extreme label imbalance on model training. It solves the problem in existing methods (such as comparison technique 1) where the influence of positive samples is too small, causing the decision interface to shift towards negative samples and resulting in poor model prediction accuracy. In particular, it can achieve better prediction performance on low-frequency labels.

[0229] In addition, the Adam optimizer can be used to optimize the objective function in this embodiment. The Adam optimizer, as a stochastic gradient descent algorithm based on adaptive first-order and second-order momentum estimation, is well-suited for large-scale neural network optimization problems due to its high computational speed and minimal memory consumption. By pre-setting parameters such as the learning rate and the exponential decay rate used for first-order and second-order momentum estimation, the objective function value is continuously reduced using the gradient descent algorithm, and backpropagation is used to optimize and update the model parameters until the objective function value converges, at which point training terminates.

[0230] Here, after obtaining the corresponding multi-label prediction results using each base learner, the final label prediction result can be determined by combining the ranking learning method. That is, in this embodiment of the disclosure, the ranking learning algorithm can be used as a meta-learner to integrate the preliminary prediction results obtained by the base learners on each view, thereby improving the generalization performance of the prediction. Specifically, this can be achieved through the following steps:

[0231] Step 1: For the target candidate label among multiple candidate labels, determine the label feature vector of each entity node for the target candidate label based on the multi-label prediction results of each entity node in each entity relationship network;

[0232] Step 2: Based on the trained meta-learner, sort and learn the label feature vectors of multiple entity nodes for the target candidate label to determine the final multi-label prediction result for each entity node.

[0233] Similar to the target entity relationship network, the target candidate label here can be one of multiple candidate labels, each candidate label, or a subset of candidate labels; in this case, it can be each candidate label. Thus, for each entity node, a label feature vector can be determined for each candidate label. That is, each entity node label pair can correspond to a label feature vector, and the vector value corresponds to the predicted score of the base learner.

[0234] By using the multiple label feature vectors determined by each entity node label pair as input to the meta-learner, the ranked prediction score of multiple candidate labels corresponding to each entity node can be determined. Then, by filtering through a preset threshold, the multi-label prediction result corresponding to the entity node can be determined.

[0235] The training process of the meta-learner can be explained in detail below, including the following steps:

[0236] Step 1: Obtain a sample training set containing multiple sample feature vectors, where each dimension value of the sample feature vector points to an entity node label pair.

[0237] Step 2: Traverse the feature vector of each sample in the training set and determine the feature vector of each sample. -gradient;

[0238] Step 3: Based on multiple sample feature vectors and each sample feature vector... - Gradient construction of regression trees;

[0239] Step 4: Based on the correspondence between each leaf node and each entity node in the regression tree, update the ranking score of each entity node in the meta-learner to be trained, and obtain the trained meta-learner.

[0240] Here, we can first construct the sample feature vectors of the meta-learner based on the prediction scores of the base learners. Let's denote the entity-relationship network... The above is a label The base learner trained is Then the entity it predicts Tag The likelihood score is .

[0241] For each label ,Will The predicted scores obtained by the seed learner are concatenated to form a score string, which serves as the input feature vector for the ranking learning model (i.e., the meta-learner). Specifically, for entities... With tags Yes, its eigenvectors are constructed as follows:

[0242]

[0243] Each score .

[0244] The final prediction result can then be given by the learning-to-rank (LTR) model. Learning-to-rank (LTR) is a machine learning technique widely used in search engines and recommendation systems. It is a supervised learning algorithm that ranks a set of candidate documents based on given query keywords, prioritizing relevant documents.

[0245] In practical applications, LambdaMART can be used for ranking learning. The core of this ranking learning method is to transform the ranking problem into a pairwise regression problem, that is, given the query entity... For tags and LambdaMART will predict Compare A more relevant likelihood score; the higher the value, the better compared to... , The more likely to be As marked.

[0246] The LambdaMART algorithm combines the advantages of Multiple Additive Regression Tree (MART) and LambdaRank, a pairwise ranking learning model based on neural networks.

[0247] Specifically, based on the idea of ​​MART, LambdaMART is essentially a gradient boosted decision tree (GBDT), whose final output is a linear combination of the outputs of a set of regression trees. Furthermore, LambdaMART replaces the gradient calculation part in the gradient boosted tree with the gradient in LambdaRank, thus making MART suitable for ranking tasks.

[0248] The training algorithm flow of LambdaMART is shown below. For simplicity, the entity is referred to as... With tags The eigenvectors corresponding to the constructed pairs are The corresponding tags Therefore, the size of the training dataset for the meta-learner .

[0249] In the initial training state, the predicted value for all samples is set to 0. After initialization, the ranking learning method here will iteratively generate... The system uses a tree to continuously fit the residuals, improving the final prediction performance. Each iteration consists of four parts.

[0250] The first part involves iterating through the training set and calculating the results for each sample. - Gradient and its derivative. Specifically, first for the sample... and ,definition

[0251]

[0252] in, and This indicates that the model provides the following for the sample at this point. and The prediction results. It is interchangeable and The change in the Normalized Discounted Cumulative Gain (NDCG) metric is evaluated after the ranking position. Assumptions and The sorting positions in the prediction results are respectively and ,but

[0253]

[0254] Wherein, IDCG (idea DCG) represents the DCG value under perfect sorting, or the DCG value calculated after arranging the samples in order of positive samples first and negative samples second.

[0255]

[0256] here, Indicating a perfect permutation The sorting position.

[0257] Next, for the samples , cumulatively with all other samples worth

[0258]

[0259] in, For all Index pairs ( A set consisting of ) For convenience, it can be abbreviated.

[0260]

[0261] Among them, record .

[0262] Therefore, right Differentiate, and we get

[0263]

[0264] The second part involves constructing a regression tree based on the feature vectors of the samples to fit the λ-gradient. The heuristic node splitting criterion used here is the minimum square error (MSE).

[0265] Specifically, for datasets , where the feature vector , Indicates the first 3D eigenvalues That is, sample of -gradient Now, we can enumerate all the features. All values ​​on Find the feature dimension that minimizes the following sum. and the values ​​on it :

[0266]

[0267] Among them, satisfying All samples fall into the left subtree middle, All samples fall into the right subtree middle, and Let represent the mean values ​​of all samples falling into the left and right subtrees, respectively. Thus, we obtain a simple binary regression tree with one root node and two leaf nodes. Then, we only need to repeatedly split the leaf nodes according to the optimal threshold, repeating the process. This will yield a tree containing A regression tree with leaf nodes.

[0268] For the The first tree One leaf node, This represents the set of all samples that fall into this leaf node.

[0269] Here, for the first The first tree The formula for calculating the multiplier value of a leaf node is as follows:

[0270]

[0271] The final part involves adding the regression tree learned in this round to the existing ranking learning model and updating the score for each sample:

[0272]

[0273] in, Indicator Sample Does it fall on a leaf node? If the condition is met, the value is 1; otherwise, it is 0. Parameter The learning rate set. Practice shows that compared to not setting a learning rate (i.e., ... Configure a smaller learning rate (generally) This can greatly improve the generalization performance of the model. This regularization method is called "shrinkage".

[0274] At this point, we have completed one iteration and built a regression tree.

[0275] In this embodiment of the disclosure, the final ranking learning model can be obtained by repeatedly performing the above four steps.

[0276]

[0277] in, The set number of iteration rounds.

[0278] Based on the above description, it can be understood that the problem to be solved by the embodiments of this disclosure is to perform multi-label classification prediction of unlabeled entities given multiple entity relationship networks and some entities with known labels. The prediction algorithm given in the embodiments of this disclosure consists of two stages. Based on the stacking strategy in ensemble learning, a base-learner with a deep graph neural network as the backbone is first used to fully mine and capture the low-order and high-order topological information in the entity relationship network. Then, a ranking learning algorithm is used as a meta-learner to integrate the preliminary prediction results obtained by the base learners on each view, thereby improving the generalization performance of the model. To facilitate a further understanding of the implementation logic of the above two stages, it can be combined with... Figure 4 Please provide a detailed explanation.

[0279] Formalistically, let For the tag set, where The total number of tags. (In the context of...) A set of entities middle, front The entity already has a label, recorded as the first... individual entities The label is ,in express Is it tagged? The value is 1 if annotated, and 0 otherwise; [Entity name] The feature vector is .

[0280] Furthermore, there are also... Different types of relationship networks It describes the relationships between entities from different perspectives. Among them, the first In the adjacency matrix of the interaction network on the nth view element Representing entities and The value is 1 if a correlation exists between the two elements, and 0 otherwise.

[0281] Furthermore, the entity relationship network involved in this embodiment is an undirected, unweighted graph, and satisfies... The problem this solution aims to solve is utilizing a known set of annotations. and multi-view entity relationship network Predict the remaining Entity labeling .

[0282] Given dataset ,in For the labeled entity set, This is the set of unlabeled entities. During model training, first... Divided into a certain proportion and Two parts, of which Used for training the base learner. Used for training meta-learners. This is the set of entities to be predicted and labeled.

[0283] Depend on Figure 4 As is known, unlike most existing algorithms (such as comparative techniques 1 and 2) which extract information from only a single relationship network, this embodiment of the disclosure adopts the "stacked" idea in ensemble learning technology, integrating the preliminary information captured from various entity relationship networks through a ranking learning model. By integrating the preliminary prediction results of each base learner through a meta-learner, the model can selectively filter valuable view information, achieving superior generalization performance compared to a single learner, and obtaining a final prediction result containing more comprehensive information and outperforming the base learners. This is mainly because in real-world scenarios such as social media, users do not exist independently but rather have certain social relationships, which can usually be modeled as social networks. Since users have multiple behaviors, the social networks between users are often not of a single type. For example, behaviors such as following, sharing, liking, commenting, and mentioning between users can all construct corresponding relationship networks. Each network reflects the closeness of the social relationships between users from one perspective, and modeling from only one relationship network will lead to a biased characterization.

[0284] Furthermore, the embodiments of this disclosure employ an ensemble learning framework to first train base classifiers for each of the multiple views separately, and then integrate them through a ranking learning algorithm, so as to achieve the goal that each view is "weak in individual learning and strong in ensemble learning", making the final prediction performance of the model better than any single-view base learner.

[0285] The entity classification method provided in this disclosure can be applied to various application fields. Taking social media applications as an example, operators typically build a tag system for users to describe them. For example, user tags may include loyalty, activity level, topic preferences, consumption preferences, etc.

[0286] This establishes an application scenario for multi-label user classification. Furthermore, since users exhibit diverse behaviors, their social networks are often not singular. For example, user behaviors such as following, sharing, liking, commenting, and mentioning can all construct corresponding relationship networks. Therefore, this further constructs a multi-view user social relationship network. Thus, the task of predicting user labels in social media can be modeled as the multi-view entity multi-label classification problem proposed in this embodiment.

[0287] First, feature vectors characterizing users are constructed based on their basic attributes, posting information, browsing history, and favorites. Then, a corresponding social relationship network is built based on user behaviors such as following, sharing, liking, commenting, and mentioning. Finally, a selected group of users is manually annotated with multiple labels to construct a standard dataset.

[0288] The second step involves inputting the constructed user social relationship network, user feature vectors, and manual annotations into the base learner based on graph convolutional networks proposed in this embodiment of the invention to obtain preliminary label prediction results for unlabeled users under each view.

[0289] Finally, the preliminary prediction scores generated by each base learner in the previous step are concatenated and input into the meta-learner based on the ranking learning algorithm proposed in this embodiment of the present disclosure to obtain the final user-labeled prediction results.

[0290] In the description of this specification, the references to terms such as "some possible implementations," "some implementations," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that implementation or example is included in at least one implementation or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same implementation or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more implementations or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different implementations or examples described in this specification, as well as the features of different implementations or examples.

[0291] Regarding the method flowcharts of embodiments of this disclosure, certain operations are described as different steps performed in a certain order. Such flowcharts are illustrative and not restrictive. Some steps described herein may be grouped together and performed in a single operation, some steps may be divided into multiple sub-steps, and some steps may be performed in an order different from that shown herein. The various steps shown in the flowcharts may be implemented in any way by any circuit structure and / or tangible mechanism (e.g., software running on a computer device, hardware (e.g., logic functions implemented by a processor or chip), and / or any combination thereof).

[0292] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0293] Based on the same inventive concept, this disclosure also provides an entity classification device corresponding to the entity classification method. Since the principle of the device in this disclosure for solving the problem is similar to that of the entity classification method described above, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0294] Reference Figure 5 The diagram shown is a schematic representation of an entity classification device provided in an embodiment of this disclosure. The device includes: an acquisition module 501, a prediction module 502, and a classification module 503; wherein,

[0295] The acquisition module 501 is used to acquire multiple entity relationship networks corresponding to multiple entity nodes; the same entity node corresponds to different entity relationships in different entity relationship networks;

[0296] Prediction module 502 is used to perform multi-label prediction on multiple entity relationship networks based on multiple trained base learners, and obtain the multi-label prediction result of each entity node in each entity relationship network;

[0297] The classification module 503 is used to determine the final multi-label prediction result for each entity node based on the multi-label prediction result of each entity node in each entity relationship network.

[0298] Using the aforementioned entity classification device, when multiple entity relationship networks corresponding to multiple entity nodes are obtained, multi-label prediction can be performed on multiple entity relationship networks based on multiple trained base learners. This yields the multi-label prediction result for each entity node in each entity relationship network. Then, based on the multi-label prediction results for each entity node in each entity relationship network, the final multi-label prediction result for each entity node is determined. In other words, this disclosure, from the perspective of multi-view learning, performs multi-label prediction based on multiple entity relationship networks, which can more fully explore the relationships between entities, making the multi-label prediction results for entity nodes more accurate.

[0299] In one possible implementation, when the multi-label prediction result includes predicted scores for multiple candidate labels, the classification module 503 is configured to determine the final multi-label prediction result for each entity node based on the multi-label prediction result for each entity node in each entity relationship network, according to the following steps:

[0300] For the target candidate label among multiple candidate labels, based on the multi-label prediction results of each entity node in each entity relationship network, determine the label feature vector of each entity node for the target candidate label;

[0301] Based on the trained meta-learner, the label feature vectors of multiple entity nodes for the target candidate label are sorted and learned to determine the final multi-label prediction result for each entity node.

[0302] In one possible implementation, the classification module 503 is configured to determine the label feature vector for each entity node relative to the target candidate label based on the multi-label prediction results of each entity node in each entity relationship network, according to the following steps:

[0303] For a target entity relationship network in a network of multiple entity relationships, select the target prediction score that matches the target candidate label from the predicted scores of entity nodes for multiple candidate labels in the target entity relationship network.

[0304] The target prediction scores selected from multiple entity relationship networks are combined to obtain the label feature vector of the entity node for the target candidate label.

[0305] In one possible implementation, the classification module 503 is configured to perform sorting and learning on the label feature vectors of multiple entity nodes for the target candidate label based on a trained meta-learner, and determine the final multi-label prediction result for each entity node according to the following steps:

[0306] For a target entity node among multiple entity nodes, the label feature vector of the target entity node for the target candidate label is input into the trained meta-learner, and the predicted score is determined after the ranking of the multiple candidate labels corresponding to the target entity node is determined.

[0307] Based on the predicted scores after sorting, the multi-label prediction results corresponding to the target entity nodes are determined.

[0308] In one possible implementation, the classification module 503 is used to determine the multi-label prediction result corresponding to the target entity node based on the sorted predicted score according to the following steps:

[0309] If the predicted score after sorting is greater than the preset threshold, the candidate label corresponding to the predicted score after sorting will be determined as the multi-label prediction result corresponding to the target entity node.

[0310] In one possible implementation, the classification module 503 is used to train the meta-learner according to the following steps:

[0311] Obtain a sample training set containing multiple sample feature vectors, where each dimension value of the sample feature vector points to an entity node label pair;

[0312] Iterate through the feature vectors of each sample in the training set and determine the feature vector of each sample. -gradient;

[0313] Based on multiple sample feature vectors and each sample feature vector - Gradient construction of regression trees;

[0314] Based on the correspondence between each leaf node and each entity node in the regression tree, the ranking score of each entity node in the meta-learner to be trained is updated to obtain the trained meta-learner.

[0315] In one possible implementation, the prediction module 502 is configured to perform multi-label prediction on multiple entity relationship networks based on multiple trained base learners according to the following steps, to obtain the multi-label prediction result of each entity node in each entity relationship network:

[0316] For a target entity relationship network in a multi-entity relationship network, a target base learner corresponding to the target entity relationship network is used to perform multi-label prediction on the target entity relationship network, so as to obtain the multi-label prediction result of each entity node in the target entity relationship network.

[0317] In one possible implementation, when the target base learner includes a fully connected input layer, a representation learning module, and a fully connected output layer, the prediction module 502 is configured to perform multi-label prediction on the target entity relationship network using the target base learner corresponding to the target entity relationship network, according to the following steps, to obtain the multi-label prediction result of each entity node in the target entity relationship network:

[0318] The original feature vector of each entity node in the target entity network is input into the fully connected input layer of the target base learner to determine the dimensionality-reduced feature vector output by the fully connected input layer for each entity node; and,

[0319] The dimensionality-reduced feature vectors are input into the representation learning module of the target base learner to determine hidden feature vectors containing low-order signals and information about similar nodes; and,

[0320] The hidden feature vectors are input into the fully connected output layer of the target base learner to determine the multi-label prediction result of each entity node in the target entity relationship network.

[0321] In one possible implementation, the prediction module 502 is used to train a target base learner comprising a fully connected input layer, a representation learning module, and a fully connected output layer according to the following steps:

[0322] Obtain the sample entity relationship network, where some entity nodes have multi-label annotation results;

[0323] The fully connected input layer is used to perform dimensionality reduction representation on each entity node in the sample entity relationship network, and the node dimensionality reduction implicit representation of each entity node after dimensionality reduction transformation is determined.

[0324] The graph convolutional layer included in the representation learning module is used to perform low-order signal attention learning on the dimensionality reduction latent representation of nodes, resulting in the node attention latent representation of each entity node after attention learning; and the conditional random field layer included in the representation learning module is used to perform node similarity learning on the node attention latent representation, resulting in the node similarity latent representation of each entity node after node similarity learning.

[0325] The fully connected output layer is used to perform multi-label prediction on the node similarity latent representation of each entity node to obtain the prediction result;

[0326] The target base learner is adjusted based on the prediction results and multi-label annotation results to obtain a trained target base learner.

[0327] In one possible implementation, when multiple graph convolutional layers are included, the prediction module 502 is configured to perform low-order signal attention learning on the node dimensionality-reduced latent representation using the graph convolutional layers included in the representation learning module, to obtain a node attention latent representation for each entity node after attention learning:

[0328] The dimensionality-reduced implicit representation of the nodes is used as the initial representation for each graph convolutional layer;

[0329] For the current graph convolutional layer other than the first graph convolutional layer, perform the following steps:

[0330] The sample entity relationship network and the node attention implicit representation output by the previous graph convolutional layer before the current graph convolutional layer are input into the graph attention layer in other graph convolutional layers to determine the node attention implicit representation output by the graph attention layer.

[0331] Based on the initial representation, the node attention implicit representation output by the graph attention layer, and the training parameters of the current graph convolutional layer, the node attention implicit representation output by the current graph convolutional layer is determined.

[0332] In one possible implementation, the prediction module 502 is configured to determine the node attention latent representation output by the current graph convolutional layer based on the initial representation, the node attention latent representation output by the graph attention layer, and the training parameters of the current graph convolutional layer, according to the following steps:

[0333] The first graph convolution operator is determined based on the first weighted summation operation between the initial representation and the node attention implicit representation output by the graph attention layer; and the second graph convolution operator is determined based on the second weighted summation operation between the training parameters of the current graph convolution layer and the identity mapping matrix corresponding to the current graph convolution layer.

[0334] Based on the first graph convolution operator and the second graph convolution operator, determine the node attention implicit representation of the output of the current graph convolutional layer.

[0335] In one possible implementation, the prediction module 502 is configured to perform node similarity learning on the node attention latent representation using the conditional random field layer included in the representation learning module, thereby obtaining a node similarity latent representation for each entity node after node similarity learning:

[0336] For each entity node, a maximum conditional probability function is constructed for that entity node. The maximum conditional probability function is determined by the first difference between the node similarity latent representation of the entity node and the node attention latent representation of the entity node, and the second difference between the node similarity latent representation of the entity node and the node similarity latent representations of other entity nodes among the multiple entity nodes.

[0337] Given that the conditional probability function is maximized to reach its maximum value, determine the node similarity implicit representation for each node.

[0338] In one possible implementation, the prediction module 502 is used to construct a maximized conditional probability function for the corresponding entity node according to the following steps:

[0339] For each entity node, obtain the first difference between the node similarity latent representation of the entity node and the node attention latent representation of the entity node, the second difference between the node similarity latent representation of the entity node and the node similarity latent representation of other entity nodes among multiple entity nodes, and the node similarity between the entity node and other entity nodes.

[0340] The second difference is multiplied by the node similarity to determine the product result.

[0341] The summation of the products between the entity node and each other entity node is performed to obtain the second difference sum.

[0342] The conditional probability function is determined by a third weighted summation operation between the first difference and the second difference.

[0343] In one possible implementation, the prediction module 502 is configured to adjust the target base learner based on the prediction results and multi-label annotation results according to the following steps to obtain a trained target base learner:

[0344] Obtain the first weight parameter used to adjust the number of positive and negative samples, and the second weight parameter used to adjust the contribution of hard-to-distinguish samples;

[0345] Based on the first weight parameter, the second weight parameter, and the difference between the prediction result and the multi-label annotation result, the target loss function value of the target base learner is determined;

[0346] The training parameters of the target base learner are adjusted at least once based on the target loss function value to obtain the trained target base learner.

[0347] It should be noted that the apparatus in the embodiments of this application can implement each process of the aforementioned method and achieve the same effect and function, which will not be elaborated here.

[0348] This disclosure also provides an electronic device, such as... Figure 6 The diagram shown is a schematic representation of an electronic device structure provided in this embodiment of the present disclosure, including: a processor 601, a memory 602, and a bus 603. The memory 602 stores machine-readable instructions executable by the processor 601 (e.g., ...). Figure 5The device contains the execution instructions corresponding to the acquisition module 501, prediction module 502, and classification module 503. When the electronic device is running, the processor 601 and the memory 602 communicate via the bus 603. When the machine-readable instructions are executed by the processor 601, the following processing is performed:

[0349] Retrieve multiple entity relationship networks corresponding to multiple entity nodes; the same entity node corresponds to different entity relationships in different entity relationship networks;

[0350] Multi-label prediction is performed on multiple entity relationship networks based on multiple pre-trained base learners to obtain the multi-label prediction result of each entity node in each entity relationship network.

[0351] Based on the multi-label prediction results of each entity node in each entity relationship network, the final multi-label prediction result corresponding to each entity node is determined.

[0352] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the entity classification method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0353] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the entity classification method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0354] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0355] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the descriptions of the apparatus, device, and computer-readable storage medium embodiments are simplified because they are substantially similar to the method embodiments; relevant details can be found in the descriptions of the method embodiments.

[0356] The apparatus, device, and computer-readable storage medium provided in this application correspond one-to-one with the method. Therefore, the apparatus, device, and computer-readable storage medium also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the apparatus, device, and computer-readable storage medium will not be repeated here.

[0357] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, apparatus (devices or systems), or computer-readable storage media. Therefore, this disclosure can take the form of a completely hardware implementation, a completely software implementation, or an implementation combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer-readable storage medium implemented on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-readable program code.

[0358] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices or systems), and computer-readable storage media according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0359] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0360] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0361] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0362] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0363] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. Furthermore, although the operations of the methods of this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0364] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. An entity classification method, characterized in that, Applications to user tag prediction based on multi-view social relationship networks include: Obtain multiple entity relationship networks corresponding to multiple entity nodes; the same user entity node corresponds to entity relationships built based on different interaction behaviors in different social relationship networks; Multi-label prediction is performed on the multiple entity relationship networks based on multiple pre-trained base learners to obtain the multi-label prediction result of each entity node in each entity relationship network. Based on the multi-label prediction results of each entity node in each entity relationship network, determine the final multi-label prediction result corresponding to each entity node; When the multi-label prediction result includes prediction scores for multiple candidate labels, determining the final multi-label prediction result for each entity node includes: For the target candidate label among the multiple candidate labels, based on the multi-label prediction results of each entity node in each entity relationship network, determine the label feature vector of each entity node for the target candidate label; Based on the trained meta-learner, the label feature vectors of multiple entity nodes for the target candidate label are sorted and learned to determine the final multi-label prediction result for each entity node.

2. The method according to claim 1, characterized in that, The step of determining the label feature vector for each entity node for the target candidate label based on the multi-label prediction results of each entity node in each entity relationship network includes: For the target entity relationship network in the plurality of entity relationship networks, the target prediction score that matches the target candidate label is selected from the predicted scores of the entity node for the plurality of candidate labels in the target entity relationship network; The target prediction scores selected from the multiple entity relationship networks are combined to obtain the label feature vector of the entity node for the target candidate label.

3. The method according to claim 1, characterized in that, The process of ranking and learning the label feature vectors of multiple entity nodes for the target candidate label based on the trained meta-learner to determine the final multi-label prediction result for each entity node includes: For the target entity node among the multiple entity nodes, the label feature vector of the target entity node for the target candidate label is input into the trained meta-learner to determine the predicted score after the ranking of the multiple candidate labels corresponding to the target entity node. Based on the sorted predicted scores, the multi-label prediction results corresponding to the target entity node are determined.

4. The method according to claim 3, characterized in that, The step of determining the multi-label prediction result corresponding to the target entity node based on the sorted predicted score includes: If the predicted score after sorting is greater than a preset threshold, the candidate label corresponding to the predicted score after sorting is determined as the multi-label prediction result corresponding to the target entity node.

5. The method according to any one of claims 1 to 4, characterized in that, The meta-learner is trained according to the following steps: Obtain a sample training set that includes multiple sample feature vectors, wherein each dimension value of the sample feature vector points to an entity node label pair; Iterate through each sample feature vector in the training set and determine the feature vector of each sample. -gradient; Based on the multiple sample feature vectors and each sample feature vector - Gradient construction of regression trees; Based on the correspondence between each leaf node of the regression tree and each entity node, the ranking score of each entity node in the meta-learner to be trained is updated to obtain the trained meta-learner.

6. The method according to any one of claims 1 to 4, characterized in that, The multi-label prediction of the multiple entity relationship networks based on multiple trained base learners is obtained to obtain the multi-label prediction result of each entity node in each entity relationship network, including: For the target entity relationship network in the multiple entity relationship networks, a target base learner corresponding to the target entity relationship network is used to perform multi-label prediction on the target entity relationship network to obtain the multi-label prediction result of each entity node in the target entity relationship network.

7. The method according to claim 6, characterized in that, When the target base learner includes a fully connected input layer, a representation learning module, and a fully connected output layer, the step of using the target base learner corresponding to the target entity relationship network to perform multi-label prediction on the target entity relationship network, and obtaining the multi-label prediction result of each entity node in the target entity relationship network, includes: The original feature vector of each entity node in the target entity relationship network is input into the fully connected input layer of the target base learner to determine the dimensionality-reduced feature vector output by the fully connected input layer for each entity node; and, The reduced-dimensionality feature vector is input into the representation learning module of the target base learner to determine a hidden feature vector containing low-order signals and similar node information; and, The hidden feature vector is input into the fully connected output layer of the target base learner to determine the multi-label prediction result of each entity node in the target entity relationship network.

8. The method according to claim 7, characterized in that, Train the target base learner, which includes a fully connected input layer, a representation learning module, and a fully connected output layer, using the following steps: Obtain a sample entity relationship network, wherein some entity nodes in the sample entity relationship network have multi-label annotation results; The fully connected input layer is used to perform dimensionality reduction representation on each entity node in the sample entity relationship network, and the node dimensionality reduction implicit representation of each entity node after dimensionality reduction transformation is determined. The graph convolutional layer included in the representation learning module is used to perform low-order signal attention learning on the dimensionality-reduced latent representation of the nodes, resulting in a node attention latent representation for each entity node after attention learning; and the conditional random field layer included in the representation learning module is used to perform node similarity learning on the node attention latent representation, resulting in a node similarity latent representation for each entity node after node similarity learning. The fully connected output layer is used to perform multi-label prediction on the node similarity implicit representation of each entity node to obtain the prediction result; The target base learner is adjusted based on the prediction results and the multi-label annotation results to obtain a trained target base learner.

9. The method according to claim 8, characterized in that, In the case of multiple graph convolutional layers, the low-order signal attention learning is performed on the node dimensionality-reduced latent representation using the graph convolutional layers included in the representation learning module to obtain the node attention latent representation of each entity node after attention learning, including: The dimensionality-reduced implicit representation of the nodes is used as the initial representation for each graph convolutional layer; For the current graph convolutional layer other than the first graph convolutional layer, perform the following steps: The sample entity relationship network and the node attention implicit representation output by the previous graph convolutional layer before the current graph convolutional layer are input into the graph attention layer in other graph convolutional layers to determine the node attention implicit representation output by the graph attention layer. Based on the initial representation, the node attention implicit representation output by the graph attention layer, and the training parameters of the current graph convolutional layer, the node attention implicit representation output by the current graph convolutional layer is determined.

10. The method according to claim 9, characterized in that, The step of determining the node attention latent representation output by the current graph convolutional layer based on the initial representation, the node attention latent representation output by the graph attention layer, and the training parameters of the current graph convolutional layer includes: A first graph convolution operator is determined based on a first weighted summation operation between the initial representation and the node attention implicit representation output by the graph attention layer; and a second graph convolution operator is determined based on a second weighted summation operation between the training parameters of the current graph convolution layer and the identity mapping matrix corresponding to the current graph convolution layer. Based on the first graph convolution operator and the second graph convolution operator, the node attention implicit representation output by the current graph convolutional layer is determined.

11. The method according to claim 8, characterized in that, The step of using the conditional random field layer included in the representation learning module to perform node similarity learning on the node attention latent representation, to obtain the node similarity latent representation of each entity node after node similarity learning, includes: For each entity node, a maximum conditional probability function is constructed for that entity node; the maximum conditional probability function is determined by a first difference between the node similarity implicit representation of the entity node and the node attention implicit representation of the entity node, and a second difference between the node similarity implicit representation of the entity node and the node similarity implicit representations of other entity nodes among the plurality of entity nodes excluding the entity node. Given that the conditional probability function has reached its maximum value, the node similarity implicit representation of each entity node is determined.

12. The method according to claim 11, characterized in that, The construction of the maximum conditional probability function corresponding to the entity node includes: For each entity node, obtain the first difference between the node similarity implicit representation of the entity node and the node attention implicit representation of the entity node, the second difference between the node similarity implicit representation of the entity node and the node similarity implicit representation of other entity nodes among the plurality of entity nodes, and the node similarity between the entity node and the other entity nodes. The second difference is multiplied by the node similarity to determine the product result. The summation of the products between the entity node and each of the other entity nodes is performed to obtain the second difference sum. The maximized conditional probability function is determined based on a third weighted summation operation between the first difference and the second difference.

13. The method according to claim 8, characterized in that, The step of adjusting the target base learner based on the prediction result and the multi-label annotation result to obtain a trained target base learner includes: Obtain the first weight parameter used to adjust the number of positive and negative samples, and the second weight parameter used to adjust the contribution of hard-to-distinguish samples; Based on the first weight parameter, the second weight parameter, and the difference between the prediction result and the multi-label annotation result, the target loss function value of the target base learner is determined; The training parameters of the target base learner are adjusted at least once based on the target loss function value to obtain a trained target base learner.

14. A physical sorting device, characterized in that, Applications to user tag prediction based on multi-view social relationship networks include: The acquisition module is used to acquire multiple entity relationship networks corresponding to multiple entity nodes; the same user entity node corresponds to entity relationships built based on different interaction behaviors in different social entity relationship networks; The prediction module is used to perform multi-label prediction on the multiple entity relationship networks based on multiple trained base learners, and obtain the multi-label prediction result of each entity node in each entity relationship network; The classification module is used to determine the final multi-label prediction result for each entity node based on the multi-label prediction result of each entity node in each entity relationship network. When the multi-label prediction result includes prediction scores for multiple candidate labels, the classification module is specifically used to determine the final multi-label prediction result for each entity node according to the following steps: For the target candidate label among the multiple candidate labels, based on the multi-label prediction results of each entity node in each entity relationship network, determine the label feature vector of each entity node for the target candidate label; Based on the trained meta-learner, the label feature vectors of multiple entity nodes for the target candidate label are sorted and learned to determine the final multi-label prediction result for each entity node.

15. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and the machine-readable instructions, when executed by the processor, perform the entity classification method as described in any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the entity classification method as described in any one of claims 1 to 13.

Images