CN111967271A - Analysis result generation method, device, equipment and readable storage medium - Google Patents

Abstract

The present application discloses a method, apparatus, device and readable storage medium for generating analysis results, and relates to the field of machine learning. The method includes: acquiring a heterogeneous graph structure of a target heterogeneous graph, where the heterogeneous graph structure includes node data and edge data; determining initial node features corresponding to the node data and semantic aspects corresponding to the edge data; taking the heterogeneous graph structure as The random variable is a hidden variable in terms of semantics. The initial node features are embedded to obtain the heterogeneous graph features of the target heterogeneous graph. The heterogeneous graph features include updated semantic features and updated node features. Taking the semantic aspect as the hidden variable, the heterogeneous graph is embedded, and the implicit semantics is obtained from the heterogeneous graph, so as to update the node feature vector and the semantic feature vector, and avoid updating the node feature by setting the meta path. The update efficiency of node features is improved, and the execution accuracy is higher in the generation of downstream analysis results.

Description

Method, apparatus, device and readable storage medium for generating analysis results

technical field

The embodiments of the present application relate to the field of machine learning, and in particular, to a method, apparatus, device, and readable storage medium for generating an analysis result.

Background technique

Graph embedding is an algorithm that converts a property graph into a vector or set of vectors. A heterogeneous graph is a graph data structure that contains multiple types of nodes and multiple types of edges. That is, a heterogeneous graph embedding is Refers to an algorithm that expresses the structure and semantic information of heterogeneous graphs as node vectors, and finally obtains updated node features in the embedding space. Nodes that are semantically related are close to each other, and nodes that are semantically unrelated are far away from each other.

In the related art, the heterogeneous graph is embedded by setting the meta-path, wherein the meta-path is used to represent the sequence of node types with specific semantic meaning, which is predefined by the developer according to the domain knowledge, and the meta-path is used to guide the heterogeneous graph. The random walk of , obtains positive samples that conform to the meta-path and negative samples that do not conform to the meta-path, and embeds the positive samples and negative samples into the model to update the node features.

However, due to the limited number of artificially defined meta-paths, it cannot cover some potential and complex node semantic dependencies, resulting in the lack of node feature expression and the poor update accuracy of node features.

SUMMARY OF THE INVENTION

Embodiments of the present application provide a method, apparatus, device, and readable storage medium for generating an analysis result, which can improve the update accuracy of node features. The technical solution is as follows:

In one aspect, a method for generating an analysis result is provided, the method comprising:

obtaining a heterogeneous graph structure of the target heterogeneous graph, where the heterogeneous graph structure includes node data and edge data;

determining the initial node feature corresponding to the node data and the semantic aspect corresponding to the edge data;

Taking the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, the initial node feature is embedded to obtain the heterogeneous graph feature of the target heterogeneous graph, and the heterogeneous graph feature includes an update Semantic features and update node features;

The updated semantic feature and the updated node feature are analyzed to generate an analysis result corresponding to the node data.

In another aspect, a device for generating analysis results is provided, the device comprising:

an acquisition module, configured to acquire a heterogeneous graph structure of a target heterogeneous graph, wherein the heterogeneous graph structure includes node data and edge data, the node data corresponds to initial node features, and the edge data corresponds to semantic aspects;

a determination module, configured to determine the initial node feature corresponding to the node data and the semantic aspect corresponding to the edge data;

The embedding module is used to embed the initial node feature with the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable to obtain the heterogeneous graph feature of the target heterogeneous graph. The graph features include update semantic features and update node features;

An analysis module, configured to analyze the updated semantic feature and the updated node feature, and generate an analysis result corresponding to the node data.

In another aspect, a computer device is provided, the computer device includes a processor and a memory, the memory stores at least one instruction, at least a section of a program, a code set or an instruction set, the at least one instruction, the at least one A piece of program, the code set or the instruction set is loaded and executed by the processor to implement the method for generating the analysis result as described in any of the foregoing embodiments of the present application.

In another aspect, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, at least one piece of program, code set or instruction set, the at least one instruction, the at least one piece of program, the code The set or instruction set is loaded and executed by the processor to implement the method for generating the analysis result as described in any of the foregoing embodiments of the present application.

In another aspect, a computer program product or computer program is provided, the computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method for generating an analysis result described in any of the foregoing embodiments.

The beneficial effects brought by the technical solutions provided in the embodiments of the present application include at least:

Taking the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, the heterogeneous graph is embedded, and the implicit semantics are obtained from the heterogeneous graph, so as to update the node feature vector and the semantic feature vector. Update, obtain more accurate node feature vectors and semantic feature vectors, and avoid updating node features by setting meta-paths, which improves the update efficiency of node features, and has higher task execution accuracy in downstream prediction tasks.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present application more clearly, the following briefly introduces the drawings that are used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative effort.

1 is a schematic diagram of extracting node features of a heterogeneous graph by setting a meta-path in the related art provided by an exemplary embodiment of the present application;

2 is a schematic diagram of a heterogeneous graph attention network extracting node features of a heterogeneous graph in the related art provided by an exemplary embodiment of the present application;

FIG. 3 is a schematic framework diagram of an overall solution provided by an exemplary embodiment of the present application;

4 is a flowchart of a method for generating an analysis result provided by an exemplary embodiment of the present application;

5 is a flowchart of a method for generating an analysis result provided by another exemplary embodiment of the present application;

6 is a flowchart of a method for generating an analysis result provided by another exemplary embodiment of the present application;

7 is a structural block diagram of an apparatus for generating an analysis result provided by an exemplary embodiment of the present application;

8 is a structural block diagram of an apparatus for generating an analysis result provided by another exemplary embodiment of the present application;

FIG. 9 is a structural block diagram of a server provided by an exemplary embodiment of the present application.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

First, briefly introduce the terms involved in the embodiments of the present application:

Artificial Intelligence (AI): It is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, including both hardware-level technology and software-level technology. The basic technologies of artificial intelligence generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

Machine Learning (ML): It is a multi-field interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. It specializes in how computers simulate or realize human learning behaviors to acquire new knowledge or skills, and to reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications are in all fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, teaching learning and other techniques.

Heterogeneous Graph: Also known as Heterogeneous Information Network, it is a graph data structure that includes multiple types of nodes and multiple types of edges. Illustratively, taking the application scenario of the recommendation system as an example, the entities in the recommendation system include user accounts, news published by users, and recommended content uploaded by advertisers; associations between entities include friend relationships between user accounts, user accounts Interaction with the dynamic (such as: likes, reposts, accounts mentioned in the dynamic, etc.). For the above recommendation system, the entities are represented by nodes in the heterogeneous graph, and the associations between entities are represented by the edges between the nodes in the heterogeneous graph.

Graph Embedding: It is an algorithm that converts attribute graphs into vectors or vector sets. Heterogeneous Graph Embedding (HGE) refers to an algorithm that expresses the structure and semantic information of heterogeneous graphs as node vectors. The updated node features satisfy semantically related nodes in the embedding space are close to each other, and semantically irrelevant nodes are far away from each other.

In the related art, the heterogeneous graph is embedded by setting a meta-path, wherein the meta-path refers to a sequence of node types with specific semantic meaning. For example, a meta-path "A-P-C" is set in the heterogeneous graph network of paper citations, and the meta-path indicates that the paper (Paper) written by the author (Author) is published in an academic conference (Conference). In the related art, meta-path-based heterogeneous graph embedding methods include the following two categories:

First, use the meta-path to randomly walk on the heterogeneous graph, use the predefined meta-path as a conditional filter, use the meta-path as a positive sample, and use the meta-path as a negative sample, and embed the positive and negative samples as input features. In the model, the output gets updated nodes;

Illustratively, please refer to FIG. 1, taking the paper citation heterogeneous graph 100 as an example, in which an institutional node 110, an author node 120, a paper node 130 and an academic conference node 140 are set. A random walk is performed on 100, multiple positive sample sequences are sampled from the input heterogeneous graph 100, and then encoded and input into the Skip-Gram model 160. The Skip-Gram model 160 predicts the probability of occurrence of the node's context node, that is, the probability of "co-occurrence" of the selected node and its neighbor nodes, so that the probability of co-occurrence of nodes in the positive sample sequence is as large as possible. Similarly, the model will randomly sample multiple negative samples from the heterogeneous graph 100 and input them into the Skip-Gram model 160, so that the probability of the nodes in the negative samples co-occurring is as small as possible. Through this constraint, the updated node features in the embedding space satisfy the condition that semantically related nodes are close to each other, and semantically unrelated nodes are far away from each other.

Second, the meta-path is used as a priori information to guide the graph neural network to learn the weight of the neighbor nodes corresponding to the central node, and weight the node features according to the weight during aggregation to obtain the updated node.

Schematically, please refer to FIG. 2, taking Heterogeneous Graph Attention Network (HAN) as an example, learning is performed at two levels of node level 210 and semantic level 220, wherein the attention mechanism of node level 210 is mainly Learning the weights between nodes and their neighbors, the attention mechanism of semantic level 220 mainly learns the weights based on different meta-paths.

However, due to the limited number of artificially defined meta-paths, it cannot cover some potential and complex node semantic dependencies, resulting in the lack of node feature expression and the poor update accuracy of node features.

The application scenarios involved in the embodiments of the present application are described with examples in conjunction with the above-mentioned nomenclature introductions:

Taking the application scenario of the recommender system as an example, the entities in the recommender system include user accounts, news published by users, and recommended content uploaded by advertisers; the associations between entities include the friend relationship between user accounts, and the interaction between user accounts and news. Circumstances (such as: likes, retweets, dynamic mentions of accounts, etc.). For the above recommendation system, the entities are represented by nodes in the heterogeneous graph, and the associations between entities are represented by the edges between the nodes in the heterogeneous graph. Through the generation method of the analysis result provided by the embodiment of the present application, the update node feature and update semantic feature corresponding to the heterogeneous graph of the recommendation system are obtained, and the updated node feature and the updated semantic feature imply the information contained in the heterogeneous graph of the recommendation system. semantic and structural information.

After determining the heterogeneous graph features, the heterogeneous graph features are applied in content recommendation. Illustratively, perform association prediction on the first node corresponding to the user account and the second node corresponding to the candidate recommended content according to the heterogeneous graph feature, and obtain an association prediction result, where the association prediction result is used to indicate the user account's effect on the candidate recommended content. Predict programs of interest. Optionally, when performing association prediction, the heterogeneous graph feature is used as an input parameter, and a content recommendation model is input. Relevance prediction, obtain the association prediction result, and output the target recommended content recommended to the target user account. The content recommendation model is a pre-trained neural network model.

Optionally, the heterogeneous graph features include updated node features and semantic features between nodes, and the content recommendation model is used to determine the degree of association between the first node and the second node according to the updated node features and semantic features. , and according to the degree of relevancy, the candidate recommended content associated with the target user account with a relatively high degree of relevancy is determined as the target recommended content.

In the above example, the application in a recommender system is taken as an example for illustration. The generation method of the analysis result provided in this application can also be applied to other heterogeneous graphs to perform embedding processing to obtain updated node features and semantic features, and update them. The latter feature is applied to the scenario of the downstream task, which is not limited in this embodiment of the present application.

Illustratively, please refer to FIG. 3 , which shows a schematic frame diagram of an overall solution provided by an exemplary embodiment of the present application. As shown in FIG. 3 , taking two semantic aspect types as an example, for the heterogeneous graph 300 , take the structure of the heterogeneous graph 300 as a random variable, take the semantic aspect 310 as a hidden variable, embed the initial node features of the heterogeneous graph 300, and take the semantic aspect 320 as a hidden variable, and use the semantic aspect 320 as a hidden variable. Embedding is performed to obtain updated node features and semantic features.

In conjunction with the above-mentioned noun introduction and application scenario, the method for generating the analysis result provided by this application is described. Taking the method applied to the server as an example, as shown in FIG. 4 , the method includes:

In step 401, a heterogeneous graph structure of the target heterogeneous graph is obtained, and the heterogeneous graph structure includes node data and edge data.

其中，ν表示异质图的节点集合，也即异质图的节点数据，ε表示异质图的边集合，也即边数据，每个节点的映射由

ν→T确定，T表示节点的类型，每条边的类型由映射ψ：ε→R确定，R表示边的类型。对于异质图而言，|T|+|R|>2。In one example, the heterogeneous graph structure is expressed as

Among them, ν represents the node set of the heterogeneous graph, that is, the node data of the heterogeneous graph, ε represents the edge set of the heterogeneous graph, that is, the edge data, and the mapping of each node is given by

ν→T is determined, T represents the type of node, and the type of each edge is determined by the mapping ψ:ε→R, where R represents the type of edge. For heterogeneous graphs, |T|+|R|>2.

Step 402: Determine the initial node feature corresponding to the node data and the semantic aspect corresponding to the edge data.

The initial node feature corresponds to attribute data of each node, and the attribute data includes the node type, node name, etc. corresponding to the node. Illustratively, the target heterogeneous graph is a recommender system network, which includes node A, the node type corresponding to node A is user account, and the node name corresponding to node A is "00123", which is used to represent the account identifier of the user account.

The semantic aspect corresponds to the relationship between the nodes in the heterogeneous graph. Illustratively, taking a movie recommendation system as an example, the node types include: 1. User account; 2. Movie name; 3. Movie category; 4. Director; 5. Leading role. The semantic aspects include, for example, movies watched by the user account, movie categories preferred by the user account, directors preferred by the user account, actors preferred by the user account, and the like.

The initial feature vector is expressed as _X∈R ^|ν|×din , where |ν| represents the number of nodes and din represents the dimension of the input initial feature vector.

The semantic aspect is expressed as A(aspect), and the semantic aspect also corresponds to a semantic quantity. The semantic quantity is used to indicate the quantity of the semantic aspect type indicated by the edge data, that is, the semantic quantity is specified according to a priori condition.

Step 403 , using the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, embed the initial node feature to obtain the heterogeneous graph feature of the target heterogeneous graph.

In an example, the heterogeneous graph structure is input into the probability graph generation model, the heterogeneous graph structure is used as a random variable, and the semantic aspect is a hidden variable, and initial node features are embedded through the probability graph generation model.

Inputting the heterogeneous graph structure into the probability graph generation model includes inputting the initial node features and the above-mentioned semantic quantities into the probability graph generation model, so that the probability graph generation model updates the initial node features in combination with the input content.

Probabilistic Graphical Models (PGMs) refers to a mathematical tool that represents probability distribution through graph structure, in which the nodes of the probability graph represent random variables, and the edges of the graph represent the dependencies between random variables.

Heterogeneous graph features include update semantic features and update node features.

Among them, the probability map generation model also corresponds to the generation model parameters. In response to the generation model parameters being trained parameters, the initial node features are directly embedded through the probability map generation model to obtain updated semantic features and updated node features; response Since the generated model parameters are untrained parameters, it is necessary to iteratively adjust the model parameters according to the heterogeneous graph structure.

The process of generating updated node features and semantic features by the probability graph generation model is expressed as the following formula 1:

Formula 1: p _φ (G, A, X) = p _φ (G│A, X)p _φ (A|X)p(X)

Among them, φ represents the generative model parameters of the probability graph generative model, and the distribution of the initial node feature X is non-parametric. In the formalization of the above generation process, it is assumed that the heterogeneous graph is generated by the input initial node features and semantic hidden variables, that is, the joint probability distribution of G, A, and X is expanded by the chain rule, and the initial node feature X Together with the unobserved semantic aspect latent variable A, the topology of the heterogeneous graph is generated.

The generative model parameters φ of the generative model can be solved by means of maximum likelihood estimation, that is, maximizing logp _φ (G|X). Due to the existence of hidden variables in the model, this maximum optimization problem cannot be solved directly, so the variational Expectation Maximization (vEM) algorithm is used to solve it.

Optionally, determine the maximum likelihood estimation function logp _φ (G|X) of the model parameters corresponding to the heterogeneous graph structure and initial node features, adjust the model parameters through the maximum likelihood estimation function, and use the adjusted model parameters to pass The probabilistic graph generation model embeds the initial node features.

Among them, the maximum likelihood estimation function is transformed into a variational distribution q _θ (A|X) and a posterior distribution p _φ (A|G, X) expression, the variational distribution corresponds to the variational model parameter θ, and the posterior distribution corresponds to The model parameter φ is generated, the variational model parameters are adjusted according to the divergence requirements between the variational distribution and the posterior distribution, and the generated model parameters are adjusted according to the adjusted variational model parameters. Illustratively, the maximum likelihood estimation function is shown in the following formula 2:

Formula two:

表示似然函数的一个下界，称为证据下界(Evidence Lower Bound，ELBO)，通过对该证据下界的最大化处理，实现对最大似然估计函数的最大化求解。Among them, KL represents the KL divergence (Kullback-Leibler Divergence), which is used to measure the difference between two probability distributions,

Represents a lower bound of the likelihood function, called the Evidence Lower Bound (ELBO). By maximizing the evidence lower bound, the maximum likelihood estimation function can be maximized.

The variational model parameters and generative model parameters are adjusted by the vEM algorithm. The vEM algorithm is divided into E steps and M steps. E step represents the inference process, and M step represents the learning process. In the E step, according to the divergence requirements between the variational distribution and the posterior distribution, first of all The variational model parameter θ is adjusted, and the generated model parameter φ is adjusted in combination with the adjusted variational model parameter θ in step M. The variational model parameters θ and generative model parameters φ are iteratively updated by the vEM algorithm. The E step and the M step are described separately.

Step E represents the inference process. The probability map generation model needs to estimate the posterior distribution p _φ (A|G, X). In the estimation process, due to the unknown and complex relationship between the latent variable A and the node feature X, so , the estimation result cannot be obtained directly. Therefore, according to the vEM algorithm, the posterior distribution p _φ (A|G, X) is fixed at step E, and the true posterior distribution is approximated by updating the variational distribution q _θ (A|X).

Optionally, a graph neural network is used to instantiate the variational distribution q _θ (A|X) according to amortized inference, denoted as GNN _θ , and the variational distribution is resized according to mean-field theory. It is written as the following formula three:

Formula three:

where K represents the number of facet latent variables in the heterogeneous graph. a _k represents the k-th aspect variable.

The optimization objective of GNN _θ is defined as the following formula 4:

Formula four:

According to the above optimization objective formula, the optimization of the variational distribution q _θ (A|X) needs to satisfy the following formula 5:

Formula five:

Among them, C is const, is a constant. Illustratively, taking one aspect variable a ₀ as an example, to illustrate the derivation process of the above formula 5, please refer to the following formula 6:

The logF(a ₀ ) in the above derivation formula is expanded as shown in the following formula 7:

Formula seven:

在KL散度为零时取到，此时KL散度内部的两个分布相等。即参考如下公式八：According to the above derivation, the optimal

Taken when the KL divergence is zero, at which point the two distributions inside the KL divergence are equal. That is, refer to the following formula eight:

Formula eight:

The M step represents the learning process. In this process, the variational distribution q _θ (A|X) is fixed, the posterior distribution p _φ (A|G, X) is updated, and the optimization process of the parameter φ is as follows:

Formula nine:

where logp _φ (G|A,X) and logp _φ (A|X) are instantiated as a graph neural network GNN _φ .

That is, the probability graph generation model includes the above-mentioned GNN _θ and GNN _φ , the initial node features are embedded through GNN _θ and GNN _φ , and the updated semantic features and updated node features are outputted.

Step 404 , analyze the updated semantic feature and the updated node feature, and generate an analysis result corresponding to the node data.

Optionally, the association prediction is performed on the node data in the heterogeneous graph structure according to the heterogeneous graph feature, and the association prediction result is obtained as the analysis result.

The association prediction result is used to indicate the predicted association relationship between node data.

According to the association prediction result, follow-up application of the heterogeneous graph structure is performed.

First, the node data includes a first node and a second node, wherein the first node corresponds to the account data, and the second node corresponds to the candidate recommended content; the first node and the second node are correlated according to the characteristics of the heterogeneous graph, and get The associated prediction result, where the associated prediction result is used to indicate that the account data is interested in the prediction of the candidate recommended content.

Optionally, according to the association prediction result, send candidate recommended content with a higher predicted interest degree to the target account; or, according to the association prediction result, predict the number of target accounts with higher interest degree corresponding to the candidate recommended content, so as to obtain the candidate recommended content. Predict the effect of promotion.

Optionally, in the process of association prediction, the heterogeneous graph features are input into a recommendation model, the recommendation model is a machine learning model obtained by pre-training, and the association prediction is performed on the first node and the second node through the recommendation model, and the association is obtained. forecast result.

Second, the node data includes a third node and a fourth node, wherein the third node corresponds to the abnormal description content, and the fourth node corresponds to the abnormal state type, wherein the abnormal description content is the content describing the abnormal situation of the program, and the abnormal state The type is the diagnosis result of the program running state corresponding to the abnormal description content. The association prediction is performed on the third node and the fourth node according to the feature of the heterogeneous graph, and the association prediction result is obtained, and the association prediction result is used to indicate the abnormal state type prediction corresponding to each abnormal description content.

Optionally, an abnormal state type with a high degree of correlation corresponding to the target abnormality description content is determined according to the correlation prediction result, so that the program abnormality is processed with a solution corresponding to the abnormal state type.

Optionally, in the process of association prediction, the heterogeneous graph features are input into an anomaly prediction model, the anomaly prediction model is a machine learning model obtained by pre-training, and the association prediction is performed on the third node and the fourth node through the anomaly prediction model. , get the correlation prediction result.

To sum up, the method for generating the analysis result provided by this embodiment uses the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, so as to perform embedding processing on the heterogeneous graph, and obtain implicit information from the heterogeneous graph. The semantics of the node feature vector is updated, and the semantic feature vector is updated, so as to obtain more accurate node feature vector and semantic feature vector, and avoid updating the node feature by setting the meta-path, which improves the update efficiency of the node feature. , the task execution accuracy in downstream tasks is higher.

In this embodiment, in the probability graph generation model, the maximum likelihood estimation function is assisted by variational distribution, thereby deriving a graph neural network for updating node features and semantic features, which improves the accuracy of node updating .

Illustratively, the algorithm of the overall solution framework provided by the embodiments of the present application is as follows:

initialnodefeatureX∈R^|ν|×din，aspectnumberKInput:

initialnodefeatureX∈R ^|ν|×din , aspectnumberK

Output: NodeembeddingH∈R ^|ν|×dout

1: whilenotconverged

2: E-Step: Inference Procedure

3: Updatenodeembeddingsandedgeweightsbyq _θ

4: InferenceaspectembeddingsAbyq _θ

5: Updateq _θ byEq(4)

6: M-Step: Learning Procedure

7: Updatenodeembeddingsandedgeweightsbyp _φ

8: Updatep _φ

9: endwhile

Among them, the first line Input of the algorithm indicates that the input parameters include heterogeneous graph structure G, initial node feature X and semantic quantity K. The semantic quantity is used to indicate the number of semantic aspect types indicated by the edge data, that is, the number of aspect hidden variables. The second line Output represents the output content, including updated node features and updated semantic features, where R is used to indicate the type of edge, |ν| represents the number of nodes, and d _out represents the dimension of the output updated feature vector.

In the algorithm process, step 1 represents the start of iterative update; step 2 represents the prediction process, which corresponds to the above E step; step 3 represents the update of node features and edge weights according to q _θ ; step 4 represents the prediction of aspect hidden vector A according to q _θ ; 5 represents updating q _θ according to the above-mentioned formula four; Step 6 represents starting the learning process, that is, corresponding to the above-mentioned M step; Step 7 represents updating node features and edge weights according to p _φ ; Step 8 represents updating p _φ ; Step 9 represents ending iteration.

In an optional implementation, when updating node features through the above GNN _θ and GNN _φ , it is achieved by decoupling edge weights and initial node features to obtain K aspect channels. FIG. 5 is another example of this application. The flow chart of the method for generating the analysis result provided by the embodiment, taking the method applied to the server as an example, as shown in FIG. 5 , the method includes:

In step 501, a heterogeneous graph structure of the target heterogeneous graph is obtained, and the heterogeneous graph structure includes node data and edge data.

Node data corresponds to initial node features, and edge data corresponds to semantic aspects.

初始特征向量表达为X∈R^|ν|×din，语义方面表达为A(aspect)。In one example, the heterogeneous graph structure is expressed as

The initial feature vector is expressed as X∈R ^|ν|×din and the semantic aspect is expressed as A(aspect).

Step 502 , input the heterogeneous graph structure of the target heterogeneous graph into the probability graph generation model.

Inputting the heterogeneous graph structure into the probability graph generation model includes inputting the initial node features and the above-mentioned semantic quantities into the probability graph generation model, so that the probability graph generation model updates the initial node features in combination with the input content.

Step 503, decoupling the edge weight and the initial node feature to obtain K aspect channels, where K is a positive integer.

The value of K corresponds to the semantic quantity, and the semantic quantity is used to indicate the quantity of semantic aspect types indicated by the edge data, wherein the value of K is predefined.

Optionally, the neural network structure adopted by the probability graph generation model is a graph neural network (Aspect-awareGNN, A ² GNN), that is, the above-mentioned GNN _θ and GNN _φ .

h_i表示神经网络层输出的特征，i表示第i个节点；更新语义特征表达为

a_k表示第k个方面隐变量，K表示语义方面的总数量。The content that the probability graph generation model needs to output includes the update semantic feature and the update node feature, where the update node feature is expressed as

h _i represents the feature output by the neural network layer, i represents the ith node; the update semantic feature is expressed as

a _k denotes the k-th aspect latent variable, and K denotes the total number of semantic aspects.

Step 504: Update the edge weights according to the node feature vector and the aspect feature vector of the kth aspect channel.

Since the hidden variables a _k of different aspects need to keep the conditions as independent as possible to describe the semantic aspects with different meanings in the heterogeneous graph, and each node can participate in the description of different semantic aspects, therefore, the node and its surrounding The co-occurrence probability between the neighbor nodes of the node needs to satisfy the aspect-aware condition, that is, the probability graph generation model needs to have the ability to model the co-occurrence probability between the node and its surrounding neighbor nodes.

Therefore, in this embodiment, edge data also corresponds to edge weights, and K aspect channels are obtained by decoupling edge weights and initial node features, where K is a positive integer, and the value of K corresponds to the semantic quantity. According to the kth aspect The node feature vector and aspect feature vector of the channel update the edge weights, and update the initial node features according to the updated edge weights. Among them, the node feature vector is obtained by mapping the initial node feature vector through the nonlinear mapping function, that is, the initial node feature is mapped through the nonlinear mapping function to obtain the node feature vector of the kth aspect channel; k aspect channels The aspect feature vector of is obtained by performing global graph pooling processing in the kth aspect channel, wherein, in this embodiment, the global graph pooling processing is implemented as global average pooling processing, and the global graph pooling processing can also It is implemented as other graph pooling technology, which is not limited in this embodiment. Illustratively, please refer to the following formula 10 for the global pooling process:

Formula ten:

W表示语义类型。Among them, GLOBALPOOL represents the global graph pooling process,

W represents the semantic type.

When updating the edge weight, obtain the first node feature vector and the first aspect feature vector of the i-th node in the k-th aspect channel, and splicing to obtain the first splicing vector; obtain the j-th node in the k-th aspect channel The second node feature vector and the second aspect feature vector are spliced to obtain the second splicing vector; the third node feature vector and the third aspect feature vector of the mth node in the kth aspect channel are obtained, and the third splicing vector is obtained by splicing , i, j, m are all positive integers, and the mth node is the neighbor node of the ith node, determine the first semantic similarity of the first splicing vector and the second splicing vector, and determine the first The second semantic similarity between the splicing vector and the third splicing vector, and the edge weight is updated according to the first semantic similarity and the second semantic similarity.

Illustratively, please refer to the following formula 11 for the update process of edge weights:

Formula eleven:

为第l次迭代中的第一拼接向量，

为第l次迭代中的第二拼接向量，

为第l-1次迭代中的第i个节点和第j个节点之间的边权重，N(v_i)指示第i个节点的邻居节点，

为第l次迭代中的第三拼接向量，

为第l-1次迭代中的第i个节点和第m个节点之间的边权重。K表示衡量节点之间语义相似度的核函数。Among them, for the kth aspect channel,

is the first splice vector in the lth iteration,

is the second splice vector in the lth iteration,

is the edge weight between the ith node and the jth node in the l-1th iteration, N(v _i ) indicates the neighbor nodes of the ith node,

is the third splice vector in the lth iteration,

is the edge weight between the ith node and the mth node in the l-1th iteration. K represents a kernel function that measures the semantic similarity between nodes.

Step 505: Update the initial node feature according to the updated edge weight, and output the heterogeneous graph feature.

Heterogeneous graph features include update semantic features and update node features.

Illustratively, based on the update of the edge weight, the initial node features are decoupled according to the semantic aspect, please refer to the following formula 12:

Formula twelve:

为第l-1次迭代中的第一拼接向量。Among them, for the kth aspect channel,

is the first splice vector in the l-1th iteration.

The denominators in the above formula 10 and formula 11 are the normalization operations on the corresponding features in the neighborhood range. Through alternating updates of edge weights and node features, the model can capture different aspects of semantics in heterogeneous graphs. The final node feature is obtained by splicing the node features in different aspects of the channel, please refer to the formula 13 in the figure:

Formula Thirteen:

Among them, hi represents the updated node, W _out represents the feature of the neural network output, and z _{i ,k} represents the splicing vector obtained by splicing the node feature vector and the aspect feature vector in the k-th aspect channel after iteration.

Step 506 , analyze the updated semantic feature and the updated node feature, and generate an analysis result corresponding to the node data.

Optionally, the association prediction is performed on the node data in the heterogeneous graph structure according to the heterogeneous graph feature to obtain an association prediction result.

The association prediction result is used to indicate the predicted association relationship between node data.

To sum up, the method for generating the analysis result provided by this embodiment uses the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, so as to perform embedding processing on the heterogeneous graph, and obtain implicit information from the heterogeneous graph. The semantics of the node feature vector is updated, and the semantic feature vector is updated, so as to obtain more accurate node feature vector and semantic feature vector, and avoid updating the node feature by setting the meta-path, which improves the update efficiency of the node feature. , the task execution accuracy in downstream tasks is higher.

In this embodiment, for the A ² GNN neural network, K aspect channels are obtained through decoupling, and the hidden variables a _k of different aspects need to be kept as independent as possible, so as to describe the semantic aspects with different meanings in the heterogeneous graph, Make sure that the probability graph generation model has the ability to model the co-occurrence probability between a node and its neighboring nodes.

Illustratively, the algorithm of the A ² GNN neural network provided by the embodiment of the present application is shown in the following process:

initialnodefeatureX∈R^|ν|×din，aspectnumberK，layernumberLInput:

initialnodefeatureX∈R ^|ν|×din , aspectnumberK, layernumberL

Output: NodeembeddingH∈R ^|ν|×dout

1: Randomlyinitializea _k , 1≤k≤K

2: whilenotconvergedo

3: Calculate x i _,k =f _k (x _i ),

4: forlayerl=1,2,...,L,do

5: fork=1,2,...,K,do

6: for(i,j)∈εdo

7:

8: endfor

9: fori=1,2,...,|V|do

10:

11: endfor

12: ifl=Lthen

13:

14: _Updateak

15: endif

16: endfor

17: endfor

18: endwhile

Among them, the first line Input of the algorithm indicates that the input parameters include the heterogeneous graph structure G, the initial node feature X and the semantic quantity K. The semantic quantity is used to indicate the number of semantic aspect types indicated by the edge data, that is, the number of aspect hidden variables, and the number of neural network iterations L. The second line Output represents the output content, including updated node features and updated semantic features, where R is used to indicate the type of edge, |ν| represents the number of nodes, and d _out represents the dimension of the output updated feature vector.

步骤4表示判断迭代次数l是否达到L；步骤5表示判断当前针对的方面通道k是否达到K；步骤6表示判断当前边(i，j)是否属于异质图结构中的已知边；步骤7表示更新边权重；步骤8表示针对每一条边更新权重后，结束权重更新的for循环；步骤9表示判断当前已更新的节点i是否达到节点总数|V|；步骤10表示更新第k个通道中节点特征向量和语义特征向量的拼接向量；步骤11表示节点更新完毕后，结束节点特征更新的for循环；步骤12表示当迭代次数l达到设定的次数L时；步骤13表示通过拼接不同方面通道中的节点特征得到总的节点特征；步骤14表示更新语义特征；步骤15至步骤18表示当所有条件符合时，结束迭代过程。In the algorithm process, step 1 represents the random initialization aspect of the feature vector a _k ; step 2 represents the start of iterative update; step 3 represents the calculation of x _{i, k} and

Step 4 means judging whether the number of iterations l reaches L; Step 5 means judging whether the current aspect channel k reaches K; Step 6 means judging whether the current edge (i, j) belongs to the known edge in the heterogeneous graph structure; Step 7 Represents updating edge weights; Step 8 represents that after updating the weights for each edge, the for loop of weight update ends; Step 9 represents judging whether the currently updated node i reaches the total number of nodes |V|; Step 10 represents updating the kth channel The splicing vector of the node feature vector and the semantic feature vector; Step 11 indicates that after the node update is completed, the for loop of node feature update ends; Step 12 indicates that when the number of iterations l reaches the set number of times L; Step 13 indicates that by splicing different aspects of the channel The total node features are obtained from the node features in ; Step 14 represents updating the semantic features; Steps 15 to 18 represent that when all conditions are met, the iterative process is ended.

In an optional embodiment, the parameters of the generative model involved in the embodiment of the present application also need to be trained and adjusted through a loss function. FIG. 6 is a flowchart of a method for generating an analysis result provided by another exemplary embodiment of the present application. Taking the method applied to the server as an example, as shown in FIG. 6 , the method includes:

Step 601: Obtain a heterogeneous graph structure of the target heterogeneous graph, where the heterogeneous graph structure includes node data and edge data.

Node data corresponds to initial node features, and edge data corresponds to semantic aspects.

初始特征向量表达为X∈R^|ν|×din，语义方面表达为A(aspect)。In one example, the heterogeneous graph structure is expressed as

The initial feature vector is expressed as X∈R ^|ν|×din and the semantic aspect is expressed as A(aspect).

Step 602: Input the heterogeneous graph structure of the target heterogeneous graph into the probability graph generation model.

Inputting the heterogeneous graph structure into the probability graph generation model includes inputting the initial node features and the above-mentioned semantic quantities into the probability graph generation model, so that the probability graph generation model updates the initial node features in combination with the input content.

Step 603 , using the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, embed the initial node feature through a probability graph generation model, and output the heterogeneous graph feature of the target heterogeneous graph.

Heterogeneous graph features include update semantic features and update node features.

The probability graph generation model also corresponds to generation model parameters. In response to the generation model parameters being untrained parameters, the model parameters need to be iteratively adjusted first according to the heterogeneous graph structure.

Illustratively, as shown in the above formula 1, φ represents the generation model parameter of the probability map generation model.

The generative model parameters φ of the generative model can be solved by means of maximum likelihood estimation, that is, maximizing logp _φ (G|X). Due to the existence of hidden variables in the model, this maximum optimization problem cannot be solved directly, so the variational Expectation Maximization (vEM) algorithm is used to solve it.

Optionally, determine the maximum likelihood estimation function logp _φ (G|X) of the model parameters corresponding to the heterogeneous graph structure and initial node features, adjust the model parameters through the maximum likelihood estimation function, and use the adjusted model parameters to pass The probabilistic graph generation model embeds the initial node features.

Among them, the maximum likelihood estimation function is transformed into a variational distribution q _θ (A|X) and a posterior distribution p _φ (A|G, X) expression, the variational distribution corresponds to the variational model parameter θ, and the posterior distribution corresponds to The model parameter φ is generated, the variational model parameters are adjusted according to the divergence requirements between the variational distribution and the posterior distribution, and the generated model parameters are adjusted according to the adjusted variational model parameters.

Optionally, according to the updated semantic feature and the updated node feature, the parameters of the generative model can also be adjusted through the loss value.

Step 604: Input the updated semantic feature and the updated node feature into the preset loss function, and output the loss value corresponding to the update result.

The traditional undirected probabilistic graph model (also known as Markov network) follows the pairwise Markov property, that is, each node only has probabilistic dependencies between its first-order neighbors, and there is a conditional relationship between nodes that are not directly connected. independent. However, there are complex long-distance dependencies in heterogeneous graphs. This application uses a self-supervised training strategy to model the associations between non-directly adjacent nodes in heterogeneous graphs in the form of regularization. Illustratively, the nodes in the n-hop field around a given central node _vc are studied. First, design a specific rule to construct a mask subgraph G _mask , in which the edges contained in G _mask satisfy: (1) Keep the edges corresponding to the one-hop neighbors directly connected to the central node _vc in the original heterogeneous graph, and convert these edges as anchor points; (2) For edges other than anchor points, the edges in the original heterogeneous graph are masked, and the edges that are not directly connected to the central node _vc are supplemented. After constructing the G _mask , the self-supervised task aims to predict the type of the central node _vc by updating the node features. The prediction task is formalized as a classification task, and the cross-entropy loss function is used as the loss function. The form of the loss function is shown in the following formula 14:

Formula fourteen:

where I represents the indicator function.

Step 605: Adjust the probability map generation model according to the loss value.

Optionally, according to the loss value calculated by formula 14, adjust the parameters of the generation model, and after the adjustment, take the updated node feature as the initial node feature, and iteratively execute the method of embedding the initial node feature through the probability map generation model. step.

Optionally, when calculating the loss value, in addition to the cross-entropy loss function provided by the above formula 14, the total loss value can also be determined according to the objective function of the downstream task, that is, combined with various downstream tasks on heterogeneous graphs. Up, end-to-end training using the standard backpropagation algorithm.

Illustratively, the total loss function can be formalized as the following Equation 15:

Among them, L _d represents the loss function corresponding to the downstream task.

To sum up, the method for generating the analysis result provided by this embodiment uses the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, so as to perform embedding processing on the heterogeneous graph, and obtain implicit information from the heterogeneous graph. The semantics of the node feature vector is updated, and the semantic feature vector is updated, so as to obtain more accurate node feature vector and semantic feature vector, and avoid updating the node feature by setting the meta-path, which improves the update efficiency of the node feature. , the task execution accuracy in downstream tasks is higher.

Illustratively, in combination with the methods for generating analysis results provided by the above embodiments of the present application, the accuracy of node classification can be improved. Please refer to Table 1 below, where the values are expressed in percentages.

Table I

As shown in Table 1 above, the classification results of related technologies and this application are compared by using the node classification task. The databases used include the DataBase systems and Logic Programming (DBLP) database and the Association for Computing Machinery (ACM) database. And Internet Movie Database (Internet Movie DataBase, IMDB).

Among them, related technologies mainly include methods for processing homogeneous graphs: DeepWalk, Graph Convolutional Networks (GCN), and heterogeneous graph processing methods: metapath2vec, Heterogeneous Graph Attention Network (Heterogeneous Graph). Graph Attention Network, HAN), Graph Transformer Networks (GTN). It can be seen from the results in Table 1 that the method provided by this application is in the lead in the node classification accuracy.

FIG. 7 is a structural block diagram of an apparatus for generating an analysis result provided by an exemplary embodiment of the present application. As shown in FIG. 7 , the apparatus includes:

The obtaining module 710 is configured to obtain the heterogeneous graph structure of the target heterogeneous graph, wherein the heterogeneous graph structure includes node data and edge data, the node data corresponds to initial node features, and the edge data corresponds to semantic aspects;

A determination module 720, configured to determine the initial node feature corresponding to the node data and the semantic aspect corresponding to the edge data;

The embedding module 730 is configured to use the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, and embed the initial node feature to obtain the heterogeneous graph feature of the target heterogeneous graph. The qualitative graph features include update semantic features and update node features;

The analysis module 740 is configured to analyze the updated semantic feature and the updated node feature, and generate an analysis result corresponding to the node data.

In an optional embodiment, please refer to FIG. 8 , the embedding module 730 is further configured to input the heterogeneous graph structure into a probability graph generation model; taking the heterogeneous graph structure as a random variable, the semantic Aspects are hidden variables, and the initial node features are embedded through the probability graph generation model.

In an optional embodiment, the probability map generation model includes generating model parameters;

The embedded module 730 includes:

A determination unit 731, configured to determine the maximum likelihood estimation function of the model parameter corresponding to the heterogeneous graph structure and the initial node feature;

an adjustment unit 732, configured to adjust the model parameters through the maximum likelihood estimation function;

The embedding unit 733 is configured to use the adjusted model parameters to embed the initial node feature through the probability map generation model.

In an optional embodiment, the adjustment unit 732 is specifically configured to convert the maximum likelihood estimation function into a variational distribution and a posterior distribution, where the variational distribution corresponds to the variational model parameters, so The posterior distribution corresponds to the generation model parameters; the variational model parameters are adjusted according to the divergence requirements between the variational distribution and the posterior distribution; and the variational model parameters are adjusted according to the The generative model parameters are adjusted.

In an optional embodiment, the pre-embedding module 730 is further configured to input the update semantic feature and the update node feature into a preset loss function, and output the loss value corresponding to the update result;

The embedding module 730 is further configured to adjust the probability map generation model according to the loss value.

In an optional embodiment, the embedding module 730 is further configured to use the updated node feature as the initial node feature, and iteratively execute the embedding of the initial node feature by using the probability graph generation model A step of.

In an optional embodiment, the edge data also corresponds to an edge weight, and the semantic aspect corresponds to a semantic quantity, and the semantic quantity is used to indicate the quantity of the semantic aspect type indicated by the edge data;

The embedded module 730 includes:

The embedding unit 733 is configured to decouple the edge weight and the initial node feature to obtain K aspect channels, where K is a positive integer, and the value of K corresponds to the semantic quantity, wherein each aspect channel corresponds to a group node eigenvectors and aspect eigenvectors;

The adjustment unit 732 is configured to update the edge weight according to the node feature vector and the aspect feature vector of the kth aspect channel; update the initial node feature according to the updated edge weight.

In an optional embodiment, the embedding unit is further configured to map the initial node feature through a nonlinear mapping function to obtain the node feature vector of the kth aspect channel; A global graph pooling process is performed in each aspect channel to obtain the aspect feature vector of the kth aspect channel.

In an optional embodiment, the adjustment unit 732 is further configured to obtain the first node feature vector and the first aspect feature vector of the i-th node in the k-th aspect channel, and splicing to obtain the first splicing vector ; Obtain the second node feature vector and the second aspect feature vector of the jth node in the kth aspect channel, and splicing to obtain the second splicing vector; Obtain the mth node in the kth aspect channel. The third The node feature vector and the third aspect feature vector are spliced to obtain a third splicing vector, i, j, m are all positive integers, and the mth node is the neighbor node of the ith node;

The adjustment unit 732 is further configured to determine the first semantic similarity between the first splicing vector and the second splicing vector; determine the second semantic similarity between the first splicing vector and the third splicing vector ; Update the edge weights according to the first semantic similarity and the second semantic similarity.

In an optional embodiment, the node data includes a first node and a second node, the first node corresponds to account data, and the second node corresponds to candidate recommended content;

The analysis module 740 is further configured to perform association prediction on the first node and the second node according to the heterogeneous graph feature, and obtain an association prediction result as the analysis result, and the association prediction result is used to indicate the The predicted interest level of the account data in the candidate recommended content.

In an optional embodiment, the analysis module 740 is further configured to input the heterogeneous graph features into a recommendation model, where the recommendation model is a machine learning model obtained by pre-training; The node and the second node perform association prediction to obtain the association prediction result.

To sum up, the device for generating analysis results provided by this embodiment uses the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, so as to perform embedding processing on the heterogeneous graph, and obtain implicit information from the heterogeneous graph. The semantics of the node feature vector is updated, and the semantic feature vector is updated, so as to obtain more accurate node feature vector and semantic feature vector, and avoid updating the node feature by setting the meta-path, which improves the update efficiency of the node feature. , the task execution accuracy in downstream tasks is higher.

It should be noted that: the apparatus for generating analysis results provided in the above-mentioned embodiments is only illustrated by the division of the above-mentioned functional modules. The internal structure is divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus for generating an analysis result provided in the above embodiment and the embodiment of the method for generating an analysis result belong to the same concept, and the specific implementation process thereof is detailed in the method embodiment, which will not be repeated here.

FIG. 9 shows a schematic structural diagram of a server provided by an exemplary embodiment of the present application. Specifically:

The server 900 includes a central processing unit (CPU) 901 , a system memory 904 including a random access memory (RAM) 902 and a read only memory (ROM) 903 , and a connection system memory 904 And the system bus 905 of the central processing unit 901. Server 900 also includes mass storage device 906 for storing operating system 913 , application programs 914 and other program modules 915 .

Mass storage device 906 is connected to central processing unit 901 through a mass storage controller (not shown) connected to system bus 905 . Mass storage device 906 and its associated computer-readable media provide non-volatile storage for server 900 . That is, mass storage device 906 may include a computer-readable medium (not shown) such as a hard disk or a Compact Disc Read Only Memory (CD-ROM) drive.

Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory, or other solid-state storage Its technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, cassette, magnetic tape, magnetic disk storage or other magnetic storage device. Of course, those skilled in the art know that the computer storage medium is not limited to the above-mentioned types. The system memory 904 and mass storage device 906 described above may be collectively referred to as memory.

According to various embodiments of the present application, the server 900 may also run on a remote computer connected to a network through a network such as the Internet. That is, the server 900 can be connected to the network 912 through the network interface unit 911 connected to the system bus 905, or can also use the network interface unit 911 to connect to other types of networks or remote computer systems (not shown).

The above-mentioned memory also includes one or more programs, and the one or more programs are stored in the memory and configured to be executed by the CPU.

Embodiments of the present application also provide a computer device, the computer device includes a processor and a memory, and the memory stores at least one instruction, at least one piece of program, code set or instruction set, at least one instruction, at least one piece of program, code The set or instruction set is loaded and executed by the processor to implement the method for generating the analysis result provided by the above method embodiments.

Embodiments of the present application further provide a computer-readable storage medium, on which is stored at least one instruction, at least one piece of program, code set or instruction set, at least one instruction, at least one piece of program, code set or The instruction set is loaded and executed by the processor, so as to implement the method for generating the analysis result provided by the foregoing method embodiments.

Embodiments of the present application also provide a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method for generating an analysis result described in any of the foregoing embodiments.

Optionally, the computer-readable storage medium may include: Read Only Memory (ROM, Read Only Memory), Random Access Memory (RAM, Random Access Memory), Solid State Drives (SSD, Solid State Drives), or an optical disc. The random access memory may include a resistive random access memory (ReRAM, Resistance Random Access Memory) and a dynamic random access memory (DRAM, Dynamic Random Access Memory). The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

Those of ordinary skill in the art can understand that all or part of the steps of implementing the above embodiments can be completed by hardware, or can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium. The storage medium mentioned may be a read-only memory, a magnetic disk or an optical disk, etc.

The above are only optional embodiments of the present application, and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection of the present application. within the range.

Claims (14)

1. a generation method of analysis result, is characterized in that, described method comprises: obtaining a heterogeneous graph structure of the target heterogeneous graph, where the heterogeneous graph structure includes node data and edge data; determining the initial node feature corresponding to the node data, and the semantic aspect corresponding to the edge data; Taking the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, the initial node feature is embedded to obtain the heterogeneous graph feature of the target heterogeneous graph, and the heterogeneous graph feature includes an update Semantic features and update node features; The updated semantic feature and the updated node feature are analyzed to generate an analysis result corresponding to the node data. 2. The method according to claim 1, characterized in that, using the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable, embedding the initial node features, comprising: inputting the heterogeneous graph structure into a probability graph generation model; Taking the heterogeneous graph structure as a random variable and the semantic aspect as a latent variable, the initial node feature is embedded through the probability graph generation model. 3. The method according to claim 2, wherein the probability map generating model comprises generating model parameters; Embedding the initial node feature through the probability map generation model includes: determining the maximum likelihood estimation function of the model parameters corresponding to the heterogeneous graph structure and the initial node feature; adjusting the model parameters through the maximum likelihood estimation function; The initial node feature is embedded through the probability map generation model with the adjusted model parameters. 4. The method according to claim 3, wherein the model parameters are adjusted by the maximum likelihood estimation function, comprising: Converting the maximum likelihood estimation function to be expressed by a variational distribution and a posterior distribution, the variational distribution corresponds to the variational model parameters, and the posterior distribution corresponds to the generation model parameters; adjusting the variational model parameters according to the divergence requirement between the variational distribution and the posterior distribution; The generative model parameters are adjusted according to the adjusted variational model parameters. 5. The method according to any one of claims 2 to 4, wherein after the output obtains the updated semantic feature and the updated node feature, it further comprises: Inputting the update semantic feature and the update node feature into a preset loss function, and outputting the loss value corresponding to the update result; The probability map generation model is adjusted according to the loss value. 6. The method according to claim 5, wherein after the outputting the heterogeneous graph feature of the target heterogeneous graph, the method further comprises: Taking the updated node feature as the initial node feature, iteratively executes the step of embedding the initial node feature by using the probability graph generation model. 7. The method according to any one of claims 1 to 4, wherein the edge data further corresponds to an edge weight, and the semantic aspect corresponds to a semantic quantity, and the semantic quantity is used to indicate that the edge data indicates the number of semantic aspect types; The embedding of the initial node features includes: Decoupling the edge weights and the initial node features to obtain K aspect channels, where K is a positive integer, and the value of K corresponds to the semantic quantity, wherein each aspect channel corresponds to a set of node feature vectors and aspects Feature vector; Update the edge weights according to the node feature vector and the aspect feature vector of the kth aspect channel; The initial node features are updated according to the updated edge weights. 8. The method according to claim 7, wherein, before updating the edge weights according to the node feature vector and the aspect feature vector of the kth aspect channel, the method comprises: Mapping the initial node feature through a nonlinear mapping function to obtain the node feature vector of the kth aspect channel; A global graph pooling process is performed in the kth aspect channel to obtain the aspect feature vector of the kth aspect channel. 9. The method according to claim 7, wherein, updating the edge weights according to the node feature vector and the aspect feature vector of the kth aspect channel, comprising: Obtain the first node feature vector and the first aspect feature vector of the i-th node in the k-th aspect channel, and splicing to obtain the first splicing vector; Obtain the second node feature vector and the second aspect feature vector of the jth node in the kth aspect channel, and splicing to obtain the second splicing vector; Obtain the third node feature vector and the third aspect feature vector of the mth node in the kth aspect channel, splicing to obtain a third splicing vector, i, j, m are all positive integers, and the mth node is the neighbor node of the i-th node; determining the first semantic similarity of the first splicing vector and the second splicing vector; determining the second semantic similarity of the first splicing vector and the third splicing vector; The edge weights are updated according to the first semantic similarity and the second semantic similarity. 10. The method according to any one of claims 1 to 4, wherein the node data includes a first node and a second node, the first node corresponds to account data, and the second node corresponds to candidate recommendation content; Analyzing the update semantic feature and the update node feature to obtain an analysis result corresponding to the node data, including: Perform association prediction on the first node and the second node according to the heterogeneous graph feature, and obtain an association prediction result as the analysis result, where the association prediction result is used to indicate that the account data recommends the candidate The predicted level of interest in the content. 11. The method according to claim 10, wherein the association prediction is performed on the first node and the second node according to the heterogeneous graph feature, and an association prediction result is obtained as the analysis result, include: Inputting the heterogeneous graph features into a recommendation model, where the recommendation model is a machine learning model obtained by pre-training; The association prediction is performed on the first node and the second node through the recommendation model to obtain the association prediction result. 12. A device for generating analysis results, wherein the device comprises: an acquisition module, configured to acquire a heterogeneous graph structure of a target heterogeneous graph, wherein the heterogeneous graph structure includes node data and edge data, the node data corresponds to initial node features, and the edge data corresponds to semantic aspects; a determination module, configured to determine the initial node feature corresponding to the node data and the semantic aspect corresponding to the edge data; The embedding module is used to embed the initial node feature with the heterogeneous graph structure as a random variable and the semantic aspect as a hidden variable to obtain the heterogeneous graph feature of the target heterogeneous graph. The graph features include update semantic features and update node features; An analysis module, configured to analyze the updated semantic feature and the updated node feature, and generate an analysis result corresponding to the node data. 13. A computer device, characterized in that the computer device comprises a processor and a memory, and the memory stores at least one instruction, at least a piece of program, a code set or an instruction set, the at least one instruction, the at least one A piece of program, the code set or the instruction set is loaded and executed by the processor to implement the method for generating an analysis result according to any one of claims 1 to 11. 14. A computer-readable storage medium, wherein the storage medium stores at least one instruction, at least one piece of program, code set or instruction set, the at least one instruction, the at least one piece of program, the code The set or instruction set is loaded and executed by the processor to implement the method for generating the analysis result according to any one of claims 1 to 11 .

Images