Multi-view individual opinion prediction method and system based on graph representation learning

Abstract

The application provides a multi-view individual opinion prediction method based on graph representation learning, comprising the following steps: constructing a background knowledge graph, an interaction behavior network and an opinion network of an opinion holder respectively; constructing a target triple of an opinion prediction task by taking the opinion holder, a target opinion object and an opinion generation time as elements; obtaining a knowledge vector, an interaction behavior vector and an opinion vector of the opinion holder from the background knowledge graph, the interaction behavior network and the opinion network respectively; and outputting an opinion prediction result corresponding to the target triple by the classifier. The application also provides a multi-view individual opinion prediction system based on graph representation learning and a data processing device for multi-view individual opinion prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of knowledge graph, in particular to a method and system for individual user opinion prediction. BACKGROUND

[0002] With the rapid development of Web2.0 technology, the emergence of social platforms such as Facebook, Twitter, WeChat, microblog, Douban, and the advent of the big data era, the Internet has become the "fourth media" after newspapers, radio and television. The development of the Internet has promoted the generation of a large amount of user data. These massive data contain users' direct comments on various topics, and also contain users' browsing, liking, forwarding and other behaviors that do not directly express their opinion tendencies. These data directly or indirectly show users' opinion tendencies on different objects. Obtaining users' opinions from these data and mining valuable information from these data can not only help the government understand people's livelihood and formulate reasonable policies, but also help large companies monitor products and quickly correct problems, which has practical application value.

[0003] Existing individual opinion prediction methods can be broadly divided into two categories: individual opinion prediction methods based on opinion evolution rules and individual opinion prediction methods based on graph representation learning. Individual opinion prediction methods based on opinion evolution rules artificially formulate the interaction rules between opinions according to sociological theories, thereby predicting the evolution of opinions. Individual opinion prediction methods based on graph representation learning use graph representation learning models to model opinion holders and opinion objects to obtain vector representations of opinion holders and objects, and then use classifiers for prediction.

[0004] Individual opinion prediction methods based on opinion evolution rules artificially formulate the interaction and evolution rules between opinions according to sociological theories, and then simulate the evolution of opinions based on existing individual opinion data under a given opinion network structure, thereby predicting individual opinions. Sznajd et al. proposed the Sznajd model, the core idea of which is that two people are more effective in persuasion than one person, and the opinion of each individual is determined by its neighbors. This model defines simple interaction rules between individuals, believes that the interaction between individuals makes individuals easily influenced by others, so that individuals will completely abandon their own opinions and choose to completely adopt the opinions of other individuals. Ochrombel et al. improved the model by adding the concept of opinion leaders to the Sznajd model, which can completely change the opinions of other individuals and maintain their own opinions unchanged. Toscani et al. analyzed the factors of individual opinion change and the rules of global opinion assimilation using the Sznajd model.

[0005] Deffuant model designs the interaction of individual opinions as a stochastic process, that is, comparing the opinions of two randomly selected individuals, if the difference is less than a certain threshold, the two individuals interact, otherwise they do not interact. The HK model considers that the opinion of an individual is usually influenced by the opinions of multiple individuals, so after traversing all the neighbor nodes of an individual, the mean of the neighbor opinions that meet the threshold range is selected for interaction.

[0006] Bu et al. proposed the GK-means model, which uses a heuristic method to identify opinion leaders in the network, and then uses a dynamic game model to identify communities. Within the community, a dynamic opinion model based on the LHN index is used to simulate the evolution of the opinion matrix. He et al. proposed the DOMF model, which is based on the idea of maximizing opinions. First, use Q-learning theory to dynamically generate the opinions of nodes at each stage, then use a multi-stage heuristic model to select seed nodes, and predict the evolution of opinions based on opinion propagation.

[0007] The individual opinion prediction method based on graph representation learning regards users and opinion objects as nodes on a graph, and opinions as edges on a graph. The positive and negative polarity of the edge is the same as the polarity of the opinion. This method defines the model objective function based on sociological theories such as structural balance and social status, uses graph representation learning models on the network composed of user opinions to model users, and finally uses a classifier to predict opinions.

[0008] Cartwright et al. extended the balance theory from ternary closure to arbitrary loops. As long as the number of negative edges in the loop is even, the loop is balanced, and if the number of negative edges is odd, the loop is unbalanced and the opinions in the loop will change. Cygan et al. generalized the balance theory, which states that users are more similar to objects with positive opinions and more distinct from objects with negative opinions. Social status theory is generally used in directed networks, where each person has a corresponding social status. Each person will generate a positive edge to a person with higher social status and a negative edge to a person with lower social status.

[0009] The SiNE model encodes the vector representation of nodes using a lookup table, and designs an objective function based on the extended balance theory. The objective function is similar to the negative sampling interval loss function, which replaces the positive samples with the similarity to the friend neighbors and replaces the negative samples with the similarity to the enemy neighbors. If there are only friend nodes around the node, the model will generate a fictitious node as the enemy neighbor of the target node. The SNE model learns the vector representation of each node by using the logarithmic bilinear model to predict the target node in the known path pre-node, which is inspired by the Skip-gram model. The model gives different weights to the positive and negative edges to capture the symbolic information of the edges on the path. The SIDE model introduces a random walk algorithm into the directed opinion network to generate a truncated random walk path, which converts the graph structure into a sequence structure, and learns the neighborhood information in the opinion network through the co-occurrence frequency of nodes in the random walk. At the same time, in the objective function, the model maximizes the likelihood probability of positive node pairs and minimizes the likelihood probability of negative node pairs.

[0010] The SGCN model is a signed GCN model. According to the structural balance theory, the SGCN model defines the balance set and the non-balance set, and uses two independent aggregators to aggregate the node information in the balance set and the non-balance set of the target node, respectively. In the objective function, the model requires the distance between positive node pairs to be smaller than the unconnected node pairs, and the distance between unconnected node pairs to be smaller than the negative node pairs. On this basis, the SiGAT model introduces the GAT model into the signed graph. According to the sign and direction of the edge, the model defines 38 kinds of triangular motifs (Motif), and uses independent aggregators to aggregate the information of neighbor nodes for different motifs, and finally uses MLP to aggregate the information of different motifs. SHINE introduces the background knowledge and interaction information of the person into the opinion prediction, and uses three independent autoencoders to extract the person vectors of the opinion network, the attention social network and the person attribute network, respectively, and then splices the three vectors into the embedding vector of the person, and uses the inner product of the vector for opinion prediction. The SNEA model introduces the self-attention mechanism into the signed graph, defines the balance set and the non-balance set of the node, and according to the structural balance theory, redefines the aggregation operation and the objective function in the signed graph, and calculates the weight of each node when aggregating. The SDGNN model considers the directionality of the opinion, defines four kinds of relationships according to the direction and polarity of the opinion, and divides the neighbors of each node into four sets, uses different aggregation functions for aggregation, and obtains the vector representation of the node, and designs a new loss function according to the social status theory.

[0011] However, the existing individual opinion prediction methods have two problems.

[0012] First, existing methods rely on the interaction between viewpoints to predict viewpoints. The prediction only uses explicit viewpoint data from users. However, since most users rarely express their own opinions directly, they generally only like, forward, or save articles published by other users. Therefore, there is relatively little explicit viewpoint data, and existing methods face the problem of data sparsity.

[0013] Secondly, existing methods do not consider the time of opinion publication. Existing methods either artificially construct virtual data to simulate the process of opinion evolution, or treat opinions published at different times as static data, without modeling the time information of opinion publication, ignoring users' historical opinions and changes in opinions. Summary of the Invention

[0014] To address the issues of sparse opinion data and lack of temporal modeling in existing technologies, this invention proposes a multi-view individual opinion prediction method based on graph representation learning. Specifically, it includes: constructing a background knowledge graph, an interaction behavior network, and an opinion network for opinion holders; constructing a target triplet for the opinion prediction task using the opinion holder, the target opinion object, and the opinion generation time as elements; obtaining the opinion holder's knowledge vector, interaction behavior vector, and opinion vector from the background knowledge graph, the interaction behavior network, and the opinion network, respectively; and passing the knowledge vector, the interaction behavior vector, and the opinion vector through a classifier to output the opinion prediction result corresponding to the target triplet.

[0015] The multi-perspective individual viewpoint prediction method of this invention, wherein for the target triple (u, v, t), the knowledge vector Emb KG (u, v, t), the interaction behavior vector Emb IN (u, v, t) and the viewpoint vector Emb ON The vector (u, v, t) is uniformly pooled to obtain the final triplet representation vector Emb(u, v, t). Emb(u, v, t) is then input into the classifier for opinion prediction, and the obtained opinion polarity s is used as the opinion prediction result. Here, u represents the opinion holder, v represents the target opinion object, and t represents the opinion generation time. When s = 0, it means that u holds a negative opinion about v, and when s = 1, it means that u holds a positive opinion about v.

[0016] The multi-perspective individual viewpoint prediction method described in this invention uses an L-layer graph attention network as a knowledge graph encoder. KG Encode the background knowledge graph A knowledge graph triple (u, v, r) represents a relationship r between u and v; retrieve the Emb KG The steps for (u, v, t) include: through the Encoder KG ,from Learn about the background information of those who hold this viewpoint and obtain an Emb. KG (u, v, t); where, from Multiple knowledge graph triples are randomly selected as positive samples, and the same number of triples not present in the knowledge graph are randomly generated. Triples with the same structure as the knowledge graph triples are used as negative samples; the Encoder is trained using these positive and negative samples. KG And use cross-entropy as the objective function L KG .

[0017] A topological temporal coding model is used as the encoder for the interactive network. IN The network that encodes this interaction behavior A triple (u, v, t) with interactive behavior indicates that u and v interact at time t; obtain the Emb. IN The steps for (u, v, t) include: through the Encoder IN ,from Learn about the dynamic behavior of those who hold this viewpoint and obtain Emb. IN (u, v, t); where, from Multiple interaction behavior triples are randomly selected as positive samples, and the same number of triples that do not exist in the sample are randomly generated. Triples with the same structure as the interaction triples are used as negative samples; the Encoder is trained using these positive and negative samples. IN And use cross-entropy as the objective function L IN .

[0018] A topological temporal coding model is used as the view network encoder. ON Encoding this viewpoint network A view quadruple (u, v, s, t) indicates that u's view of v at time t is s;

[0019] Get Emb ON The steps for (u, v, t) include: through the Encoder ON ,from Learn about the viewpoints of those who hold this opinion and obtain an Emb. ON (u, v, t); where, from Multiple positive opinion quadruplets are randomly selected as positive samples, and an equal number of negative opinion quadruplets are selected as negative samples; the encoder is trained using these positive and negative samples. ON And use cross-entropy as the objective function L ON .

[0020] The multi-view individual opinion prediction method, wherein the loss function L of the opinion prediction task Total OP +λ1L KG +λ2L IN +λ3L ON ; L OP is a target function of the classifier, L KG is a target function of the background knowledge graph , L IN is a target function of the interaction behavior network , L ON is a target function of the opinion network , and λ1, λ2 and λ3 are hyperparameters.

[0021] The application further provides a multi-view individual opinion prediction system based on graph representation learning, comprising: an initialization module configured to construct a target triple for an opinion prediction task of an opinion holder by using three elements of an opinion holder, an opinion object and an opinion generation time; acquire a user information network of the opinion holder; and a prediction module configured to acquire a knowledge vector, an interaction behavior vector and an opinion vector of the opinion holder through the user information network, and output an opinion prediction result corresponding to the target triple through a classifier.

[0022] The multi-view individual opinion prediction system, wherein the task module comprises an opinion prediction module configured to acquire the opinion prediction result; for the target triple (u, v, t), the knowledge vector Emb KG (u, v, t), the interaction behavior vector Emb IN (u, v, t) and the opinion vector Emb ON (u, v, t) are uniformly pooled to obtain a final triple representation vector Emb(u, v, t), and Emb(u, v, t) is input into the classifier to perform opinion prediction, so that the obtained opinion polarity s is taken as the opinion prediction result; wherein u represents the opinion holder, v represents the opinion object, t represents the opinion generation time, s = 0 represents that the opinion holder u holds a negative opinion on the opinion object v, and s = 1 represents that the opinion holder u holds a positive opinion on the opinion object v.

[0023] The application further provides a computer readable storage medium storing computer executable instructions, wherein when the computer executable instructions are executed, the multi-view individual opinion prediction method based on graph representation learning is realized.

[0024] ​The application further provides a data processing apparatus comprising the computer readable storage medium as described above, and when a processor of the data processing apparatus invokes and executes the computer executable instructions in the computer readable storage medium, the multi-view individual opinion prediction based on graph representation learning is realized. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 FIG. 1 is a schematic diagram of a multi-view individual opinion prediction system based on graph representation learning of the application.

[0026] Figure 2 FIG. 4 is a schematic diagram of an interaction behavior network coding of the application.

[0027] Figure 3 FIG. 6 is a schematic diagram of a data processing apparatus of the application. DETAILED DESCRIPTION

[0028] In order to make the objectives, technical solutions and advantages of the application clearer, the application will be further described in detail below with reference to the drawings. It should be understood that the specific implementation methods described herein are only used to explain the application and do not limit the application.

[0029] When the inventors conducted research on individual opinion prediction, they found that the prior art faced the problem of data sparsity. For most users with less opinion data, the prediction effect was poor, and even prediction was impossible. This is because the prior art only uses a single data of user opinion, and most users rarely express their opinions explicitly, so the opinion itself is relatively sparse, thereby causing the prior art to have the problem of data sparsity. At the same time, the inventors found that the prior art faced the problem of being unable to predict opinions at a specific time. This is because the prior art does not model the time information of the opinion.

[0030] Through research on the data sparsity problem, the inventors found that the data sparsity problem can be alleviated by introducing the background knowledge and interaction behavior information of the user, using a graph representation learning method to model user information from multiple perspectives. And through research on the dynamic graph representation learning method, the inventors found that the problem of predicting opinions at a specific time can be solved by combining the dynamic graph representation learning method with the unique sociological properties of opinion data and modeling the time information of the opinion.

[0031] Based on the above two points, the inventors propose a multi-view individual opinion prediction method based on graph representation learning, which uniformly represents the user background knowledge, interaction behavior information and opinion data, and combines the dynamic graph representation learning method with the properties of the opinion to model the time information, thereby realizing the prediction of the opinion of an individual user at a specific time.

[0032] The multi-view individual opinion prediction method of the application integrates user's opinion data, interaction behavior data and background knowledge data for comprehensive modeling, integrates the three types of user data using a multi-view user modeling framework; joint prediction is performed using multi-angle user information, the accuracy of individual opinion prediction is improved, and effective opinion prediction can also be performed for users lacking opinion data.

[0033] In addition, the dynamic graph representation learning method is combined with the sociological nature of the opinion to model the time information of the opinion; the prediction of the opinion of the user at a specific time is realized, and the accuracy of the individual opinion prediction is improved.

[0034] The multi-view individual opinion prediction method based on graph representation learning of the application will be described in detail below with reference to the accompanying drawings.

[0035] Firstly, the individual opinion prediction task is defined as: given the user background knowledge graph interaction behavior network and opinion network and the triplets (u, v, t) to be predicted, the opinion polarity s of the opinion holder (entity u) to the opinion object (entity v) at time t is predicted, s = 0 represents opposing or holding a negative opinion, and s = 1 represents supporting or holding a positive opinion. Wherein, the background knowledge graph is the knowledge graph triplets (u, v, r), indicating that there is a relationship r between entity u and entity v; the interaction behavior network is the interaction behavior triplets (u, v, t), indicating that entity u and entity v have an interaction behavior at time t; the opinion network is the opinion quadruplet (u, v, s, t), indicating that the opinion of entity u to entity v at time t is s, s = 0 represents, and s = 1 represents a positive opinion. As Figure 1 shown, the overall structure of the multi-view individual modeling framework of the application is divided into four main components, including knowledge graph encoding, opinion network encoding, interaction network encoding and opinion prediction. The knowledge graph encoding mainly encodes the background knowledge graph in the dynamic user information network to obtain the intrinsic representation and group label feature of the entity, and provides an initialization vector for the opinion network encoding and the interaction network encoding. The opinion network encoding and the interaction network encoding encode the opinion network and the interaction behavior network in the dynamic user information network respectively to obtain the opinion vector and the behavior vector of the triplets. The opinion prediction module adds the knowledge vector, the opinion vector and the behavior vector of the triplets, calculates the opinion value through a multi-layer neural network classifier, and outputs the prediction result of the individual opinion.

[0036] I. Knowledge graph encoding

[0037] The knowledge graph encoder Encoder KGThe L-layer graph attention network is used for encoding the background knowledge graph to learn the background knowledge information of the user.

[0038] The multi-head attention module calculates the attention weight between the head entity u and the tail entity v according to the entity type and the relationship type. (u,r,v)

[0039]

[0040] wherein N(v) represents a neighbor node set of the entity v, and the type functions and are used for outputting the types of the entity and the relationship, and represent the vector of the head entity u and the tail entity v in the (l-1)-th layer model. and represent the i-th transition matrix controlled by the types of the entities u and v, and W ψ(r) represents the transition matrix controlled by the type of the relationship r, and d represents the dimension of the entity vector.

[0041] Next, the information propagation function calculates the information vector m (u,r,v) transmitted from the head entity u to the tail entity v according to the relationship r.

[0042]

[0043] Then, the aggregation function aggregates the message vectors transmitted by the neighbor nodes using the attention weight and updates the vector representation of the tail entity v.

[0044]

[0045] Finally, the output H KG of the last layer of the Encoder ( L) is the embedded vector V KG of the node in the background knowledge graph. KG The background knowledge vector Emb KG (u, v, t) of the viewpoint prediction triple (u, v, t) is the concatenation of the vectors of u and v.

[0046] The present application performs the link prediction subtask on the background knowledge graph, randomly extracts part of the triples in the background knowledge graph as positive samples, and randomly generates triples that do not exist in the background knowledge graph as negative samples, the number of positive samples and negative samples is the same, the task is to predict whether the triple exists in the background knowledge graph , so as to learn the individual background knowledge. The present application uses cross entropy as the objective function L KG.

[0047] II. Interacting Network Encoding

[0048] Interacting Network Encoder Encoder IN The interacting behavior network of the user is encoded using a topological temporal encoding model, and the dynamic behavior information of the user is learned, such as Figure 2 As shown in the figure, it comprises:

[0049] (1) Temporally constrained random walk strategy

[0050] Given a view prediction triple (u, v, t), the Encoder IN First, take u and v as the starting point, and do a temporally constrained random walk along the edge on the interacting behavior network, sample a length of w historical interaction path BW(u, t) = [I0,...I w-1 ]. The temporal constraint requires that the path timestamp be monotonically non-increasing during the sampling process. In general, the temporal relationship is a necessary condition for causality, so the monotonically non-increasing temporal constraint can help the model capture the causal relationship between interactions and enhance the model's learning of user interaction behavior. During the random walk process, the sampling probability function of the next hop node samples the interaction behavior that occurs at a similar time with a higher probability. The present invention samples N random walk paths of length w for the target u and v respectively, to form the node behavior context path set {W u , W v}.

[0051] (2) Dual-view anonymization

[0052] The present invention uses a dual-view anonymization strategy to anonymize the walk path set obtained by sampling to obtain a new anonymized path aw = [aw p , aw g ]. The dual-view anonymization strategy mainly includes path view anonymization and group view anonymization.

[0053] The path view anonymization module only considers a single path when anonymizing nodes, and the identification of the node is replaced with its first occurrence in the path. Given a walk path BW(u, t7) = [b, u, d, c, u], where the first occurrence of u in the path is 2, so the node u is anonymized as node 2, and the anonymized path is [1, 2, 3, 4, 2].

[0054] The group view anonymization module considers the entire walk path set when anonymizing nodes, and generates an anonymization vector as the group view identification of the node. The group view anonymization vector IGK a of node a can be regarded as the expected distribution of the node's occurrence position in the walk path set, and the value of the i-th dimension of the vector represents the frequency of the node appearing in the i-th dimension of the path. Finally, awg aw g = [IGK v , IGK a , IGK u , IGK b , IGK v ].

[0055] (3) Topological temporal encoding module

[0056] After obtaining the anonymized paths, the present application encodes the path set using the topological temporal encoding module. First, the present application uses a topological encoder to encode the path perspective anonymized paths. The present application first numbers the paths according to the lexicographic order of the anonymized paths as awf1, awf2,..., for example, when the path length w = 5, awf1 = (1, 2, 1, 2, 1), awf2 = (1, 2, 1, 2, 3), awf3 = (1, 2, 1, 3, 1). Then the present application counts the frequency of occurrence of awf k on the path set {AW u , AW v}. Each dimension of the topological vector TopoEmb(u, v, t) represents the frequency of occurrence of the corresponding anonymized path in the path set.

[0057]

[0058] Then the present application uses a temporal encoder to encode the group perspective anonymized paths. Considering that the path sampling satisfies the temporal constraint, the present application uses GRU to encode this kind of time series. Although GRU encodes the time sequence of interactions, it does not consider the time difference of interaction occurrence. Intuitively, the stronger the causality between behaviors that occur closer in time, the stronger the causality between behaviors that occur closer in time, so the present application proposes a temporal GRU module, which modifies the recurrent component of GRU, thereby modeling the influence of the time difference of interaction.

[0059] First, in order to be able to model the time difference, the present application uses Fourier transform to map the time difference value in the real number field to a d-dimensional vector t2v(Δt)

[0060] t2v(Δt) = [cos(w1Δt + φ1),..., cos(w d Δt + φ d )]

[0061] where w and φ are weight parameters of the Fourier transform function.

[0062] In each time step, the temporal GRU receives not only the hidden state vector from the previous moment and the input vector from the current moment, but also the time difference vector from the previous time step and the polarity of the viewpoint. This invention uses the time difference vector t2v(Δt) to encode the restart gate r. k This allows the time difference in interaction to control whether the previous state is stored, thus correcting the restart gate r. k The formula is as follows:

[0063] r k =σ(W r t2v(Δt k )+U r h k-1 )

[0064] Among them W r U r These are the weights and bias parameters of the temporal GRU, and σ is the sigmoid function. Δt k It is the time difference between the k-th interaction and the (k-1)-th interaction, initially set to 0. Path Temporal encoding vector The state vector h of the last time step of the temporal GRU w-1 The temporal encoding vector TempEmb(u,v,t) of the entire path set is a uniform pooling of the temporal vectors of all paths.

[0065] Ultimately, this invention concatenates the topological encoding vector TopoEmb(u, v, t) and the temporal encoding vector TempEmb(u, v, t), and then passes them through a single-layer neural network NN to obtain the interaction behavior vector Emb of the triple (u, v, t). IN (u, v, t)

[0066] Emb IN (u, v, t) = NN (Concat [TopoEmb (u, v, t), TempEmb (u, v, t)])

[0067] This invention performs an opinion prediction subtask on an interactive behavior network. A subset of edges (u, v, t) are randomly selected from the network as positive samples, and edges not present in the network are randomly generated as negative samples. The number of positive and negative samples is equal. The task is to predict whether a triple exists in the interactive behavior network ε. IN In this process, user behavior information is learned. This invention uses cross-entropy as the objective function L for the encoder subtask of the interactive network. IN .

[0068] III. Viewpoint Network Coding

[0069] The structure of the opinion network coding module is similar to that of the behavior network coding module, and is also coded using the topological time sequence coding model. Considering that the positive and negative polarities of the opinion have an influence on the transmission of path information, the path encoder of the module is modified in the application.

[0070] When using the time sequence GRU for path coding, the polarity s of the opinion is mapped into a d-dimensional vector o2v(s), where s belongs to {0, 1}, s = 0 represents opposition, and s = 1 represents support. The reset gate r of the opinion network time sequence GRU k The calculation formula is as follows:

[0071] r k =σ(W r t2v(Δt k )+W′ r o2v(s)+U r h k-1 )

[0072] The remaining formulas are the same as those of the interaction network encoder, and the opinion network coding vector Emb ON (u, v, t) is obtained. The application performs the opinion prediction subtask on the opinion network, extracts part of the edges (u, v, s, t) from the opinion network as training samples, and in the embodiment of the application, a plurality of positive opinion edges are randomly extracted from the opinion network as positive samples, and an equal number of negative opinion edges are randomly extracted from the opinion network as negative samples; the task is to predict the type of the opinion, so as to learn the opinion information of the user. The application uses cross entropy as the objective function of the opinion network coding subtask, and the objective function is L ON .

[0073] Four, opinion prediction

[0074] The individual opinion prediction task is to predict the opinion tendency s of the opinion holder u to the opinion object v at time t, given the triple (u, v, t). Through the previous three modules, the application obtains the background knowledge vector Emb KG (u, v, t), the interaction behavior vector Emb IN (u, v, t) and the opinion vector Emb oN (u, v, t) of the triple (u, v, t), uniformly pools the three vectors to obtain the final triple representation vector Emb(u, v, t), and inputs a multi-layer neural network to perform the final opinion prediction.

[0075] The application uses a multi-task learning framework, learns the relevant information of the user through subtasks on various networks, models the user from multiple perspectives, and the final loss function L Total is as follows. Wherein L OPis the objective function of using the final triple vector to predict the opinion, calculated using cross-entropy, and λ1, λ2 and λ3 are hyperparameters.

[0076] L Total = L OP + λ1L KG + λ2L IN + λ3L ON

[0077] Figure 3 is a schematic diagram of the data processing device of the present application. As shown in Figure 3 The computer readable storage medium of the present application stores computer executable instructions, and the computer executable instructions are executed by the processor of the data processing device to realize the multi-view individual opinion prediction based on graph representation learning described above. Those skilled in the art can understand that all or part of the steps in the above method can be completed by programs instructing related hardware (such as processor, FPGA, ASIC, etc.), and the programs can be stored in readable storage medium, such as read-only memory, disk or optical disk, etc. All or part of the steps of the above embodiments can also be realized by one or more integrated circuits. Accordingly, each module in the above embodiments can be realized in the form of hardware, such as by integrated circuit to realize its corresponding function, or in the form of software function module, such as by processor executing the program / instructions stored in the memory to realize its corresponding function. The embodiments of the present application are not limited to any specific form of combination of hardware and software.

[0078] Compared with the existing method, the present method can realize the prediction of individual opinion at a specific time, introduce user background knowledge data and interaction behavior data, model individual user information from multiple perspectives, and for the scene with sparse opinion data, the background knowledge data and interaction behavior data can be used for auxiliary prediction, the prediction accuracy is high, and the universality is wider; in addition, the dynamic graph representation learning model is combined with the sociological nature of the opinion, and the time information of the opinion is fully modeled, so that the opinion at a specific time can be predicted, which is closer to the real world situation.

[0079] The above embodiments are only used to illustrate the present application, and not to limit the present application. Those skilled in the related art can make various changes and modifications without departing from the spirit and scope of the present application, therefore all equivalent technical solutions also belong to the scope of the present application, and the patent protection scope of the present application should be defined by the claims.

Claims

1. A multi-view individual opinion prediction method based on graph representation learning, characterized in that, Comprise: respectively construct the background knowledge graph of the opinion holder, the interaction behavior network and the opinion network; With the viewpoint holder, the target viewpoint object, and the viewpoint generation time as elements, a target triple of a viewpoint prediction task is constructed ; From the background knowledge graph, the interaction behavior network and the opinion network, respectively acquire a knowledge vector , an interaction behavior vector and an opinion vector of the opinion holder; This knowledge vector The interaction behavior vector and the vector of this viewpoint Perform uniform pooling to obtain the final triplet representation vector. ,Will Input the data into a classifier to predict opinion polarity. As a result of opinion prediction; among which... This indicates that the holder of this viewpoint, This indicates the target viewpoint object. This indicates the moment when the viewpoint was generated. Time indicates right Holding a negative view, when Time indicates right Holding a positive view, which predicts the loss function of the task. ; Let the objective function of this classifier be... To encode this background knowledge graph The objective function, To encode this interactive behavior network The objective function, For encoding viewpoint networks The objective function, , and This is a hyperparameter.

2. The multi-view individual opinion prediction method of claim 1, wherein, Using a graph attention network of L layers as a knowledge graph encoder Encoder KG , encoding the background knowledge graph ; with a knowledge graph triple ( ), representing there is a relationship between ; Acquiring The step of acquiring KG includes: learning background knowledge information of the opinion holder by Encoder ; wherein, a plurality of knowledge graph triples are randomly selected from as positive samples, and a same number of triples that do not exist in and have a same structure as the knowledge graph triples are randomly generated as negative samples; Encoder KG is trained by using the positive samples and the negative samples, and cross entropy is used as an objective function .​ 3. The multi-view individual opinion prediction method of claim 1, wherein, Using a topological temporal encoding model as an interaction network encoder Encoder IN , encoding the interaction behavior network ; with an interaction behavior triple ( ), representing an interaction behavior between at a time instant ; Get The steps include: via Encoder IN ,from Learn about the dynamic behavior of those who hold this viewpoint and obtain... Among them, from Multiple interaction behavior triples are randomly selected as positive samples, and the same number of triples that do not exist in the sample are randomly generated. Triples with the same structure as the interaction triples are used as negative samples; the Encoder is trained using these positive and negative samples. IN And use cross-entropy as the objective function. .

4. The multi-view individual opinion prediction method of claim 1, wherein, A topological temporal coding model is used as the view network encoder. ON Encoding this viewpoint network ; Having a viewpoint quadruple ( ),express exist Always The viewpoint is ; Get The steps include: via Encoder ON ,from Learn about the viewpoints of those who hold this opinion, and obtain... Among them, from Multiple positive opinion quadruplets are randomly selected as positive samples, and an equal number of negative opinion quadruplets are selected as negative samples; the encoder is trained using these positive and negative samples. ON And use cross-entropy as the objective function. .

5. A multi-perspective individual opinion prediction system based on graph representation learning, employing the multi-perspective individual opinion prediction method based on graph representation learning according to any one of claims 1-4 to perform individual opinion prediction, characterized in that, Comprise: an initialization module, configured to construct a target triple for an opinion prediction task on the opinion holder with three elements of the opinion holder, the opinion object and the opinion generation time; obtain a user information network of the opinion holder; a prediction module, configured to respectively obtain a knowledge vector, an interaction behavior vector and an opinion vector of the opinion holder through the user information network; and output an opinion prediction result corresponding to the target triple through a classifier on the knowledge vector, the interaction behavior vector and the opinion vector.

6. The multi-view individual opinion prediction system of claim 5, wherein, The prediction module includes an opinion prediction module, used to obtain the opinion prediction result; for the target triplet , and the knowledge vector The interaction behavior vector and the vector of this viewpoint Perform uniform pooling to obtain the final triplet representation vector. ,Will The classifier is input to predict opinions, and the resulting opinion polarity s is used as the opinion prediction result; where, This indicates that the holder of this viewpoint, This indicates the object of the viewpoint. This indicates the moment when the viewpoint was generated. Time indicates right Holding a negative view, when Time indicates right Hold a positive view.

7. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer executable instructions are executed, the method for multi-perspective individual opinion prediction based on graph representation learning in any one of claims 1-4 is implemented. 8.A data processing apparatus comprising the computer readable storage medium in claim 7, when a processor of the data processing apparatus invokes and executes the computer executable instructions in the computer readable storage medium, the multi-perspective individual opinion prediction based on graph representation learning is implemented.

Images