Derived topic propagation prediction method based on emotion migration and evolutionary game

Abstract

The invention belongs to the technical field of internet application, and particularly relates to a derivative topic propagation prediction method based on emotion migration and evolutionary game, which comprises the following steps: acquiring user topic propagation data, and preprocessing the topic propagation data; related attributes are extracted from the preprocessed topic propagation data; dividing the extracted related attributes into a training set and a test set; performing sentiment analysis and topic propagation analysis on the data in the training set to obtain a topic-user sentiment matrix; inputting the topic-user emotion matrix into a derivative topic propagation prediction model for training, and testing the trained derivative topic propagation prediction model by adopting a test set to obtain an optimal derivative topic propagation prediction model; the optimal derivative topic propagation prediction model is adopted to process data to be propagated, and a derivative topic propagation prediction result is obtained; according to the invention, the performance and accuracy of the model are optimized.

Description

Technical Field

[0001] The present invention belongs to the field of Internet application technology, and specifically relates to a derivative topic propagation prediction method based on sentiment migration and evolutionary game. Background Art

[0002] In today's digital age, social networks have become a key channel for information dissemination and public opinion formation. User interactions on these platforms, such as posting, commenting, liking and forwarding, not only reflect personal emotional tendencies, but also promote the spread and evolution of topics. For example, hot topics on Weibo often trigger heated discussions among the public and become the focus of public attention. With the popularity of social networks and the increase in the number of users, massive amounts of data are generated and stored on the platforms. These data contain valuable information that can provide insights into social trends and predict the spread of topics.

[0003] In recent years, researchers have conducted a lot of research on early topic propagation prediction, mainly using collaborative filtering algorithms, traditional machine learning algorithms, and deep learning algorithms. These methods are committed to understanding user interests and behavioral patterns, topic characteristics, and propagation trends. With the continuous expansion of e-commerce platform data and the gradual maturity of deep learning technology, nonlinear models such as neural networks and deep learning have gradually become the mainstream choice for marketing data analysis and prediction models. Alireza et al. proposed an integrated transfer learning multimodal sentiment analysis model based on weighted convolutional neural networks, which utilized the output fusion of text and image classifiers, and finally used the extended Dempster–Shafer theory to determine sentiment polarity at the decision level.

[0004] At present, there are still some problems in the prediction of early derivative topic propagation: 1. The sparsity of early topics. There are fewer participants in early topics, and the user behavior data such as forwarding and commenting on the same topic are not enough to support the prediction of derivative topic propagation, which will inevitably affect the prediction accuracy of the entire propagation model. 2. The complexity of users' emotional tendencies towards topics. Different users have different emotional tendencies for different topics. The complexity of topic emotional tendencies will lead to the problem of dimensional disaster in the model training and reasoning process. How to effectively low-rank vectorize complex features will be a challenge. 3. There is an adversarial and cooperative relationship between native emotions and derived emotions in the process of derivative topic propagation. When native emotions and derived emotions confront each other, conflicts and contradictions will arise, which may have an adverse effect on the propagation of derivative topics. When native emotions and derived emotions cooperate with each other, it may promote the propagation of derivative topics. The adversarial and cooperative relationship between native emotions and derived emotions will have different effects on the propagation of derivative topics. What strategies to adopt to promote or inhibit the propagation of derivative topics is an urgent problem to be solved. Summary of the invention

[0005] In order to solve the above problems in the prior art, the present invention proposes a derivative topic propagation prediction method based on sentiment migration and evolutionary game, which includes:

[0006] Obtain user topic propagation data and pre-process the topic propagation data; the topic propagation data includes original topic data, derived topic data and user data;

[0007] The preprocessed original topic data and derived topic data are fused by using transfer learning method, and the fused topic data are processed by generative adversarial network to generate specific types of topic features; the user-topic relationship matrix is ​​constructed based on the topic features and user data;

[0008] The ME2vce algorithm is used to perform sentiment analysis on the user-topic relationship matrix to obtain the user-topic sentiment knowledge graph;

[0009] Obtain relevant attributes of user data, merge relevant attributes with user-topic sentiment knowledge graph, and obtain user feature matrix;

[0010] The user feature matrix is ​​input into the pre-trained derivative topic propagation prediction model to obtain the derivative topic propagation prediction result.

[0011] Beneficial effects of the present invention:

[0012] The present invention adopts the ME2vec algorithm to construct the relationship between user emotions and topics, uses semantic relationships to infer the similarity and correlation of emotions between users, and introduces sparse representation to low-rank vectorize the derived topic feature space from the perspectives of time, space, and emotion. Finally, the native-derived topic propagation model of evolutionary game theory and adaptive mechanism is used to realize early prediction of derived topic propagation, simulate the strategy selection of nodes in topic propagation, and thus optimize the performance and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is the overall flow chart of the present invention;

[0014] Figure 2 Schematic diagram of enriching the native topic domain using the DATR-GAN algorithm of the present invention;

[0015] Figure 3 It is a schematic diagram of the ME2vec algorithm of the present invention;

[0016] Figure 4 Schematic diagram of the AEGT-GCN model of the present invention;

[0017] Figure 5 This is a flow chart of the derivative topic propagation prediction method based on sentiment transfer and evolutionary game of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] A derivative topic propagation prediction method based on sentiment transfer and evolutionary game, such as Figure 5 As shown, the method includes:

[0020] Obtain user topic propagation data and pre-process the topic propagation data; the topic propagation data includes original topic data, derived topic data and user data;

[0021] The preprocessed original topic data and derived topic data are fused by using transfer learning method, and the fused topic data are processed by generative adversarial network to generate specific types of topic features; the user-topic relationship matrix is ​​constructed based on the topic features and user data;

[0022] The ME2vce algorithm is used to perform sentiment analysis on the user-topic relationship matrix to obtain the user-topic sentiment knowledge graph;

[0023] Obtain relevant attributes of user data, merge relevant attributes with user-topic sentiment knowledge graph, and obtain user feature matrix;

[0024] The user feature matrix is ​​input into the pre-trained derivative topic propagation prediction model to obtain the derivative topic propagation prediction result.

[0025] like Figures 1 to 4 As shown, it indicates that the input of the present invention is the basic information of users on the social platform, native topic data and user relationship information, and the output after the prediction model is the prediction result of whether the user will participate in the topic.

[0026] The present invention comprises the following three steps:

[0027] S1: Get data online. The way to get data is to directly query the topic propagation data in the database provided by the enterprise. What needs to be obtained here is the interaction of topic propagation during its life cycle (such as forwarding, commenting, sharing, etc.), and the historical behavior data of users who have interacted with the topic propagation. The topic propagation participation situation needs to obtain the time when the user participated in the topic interaction, the basic information of the interactive user, and the basic information of the topic.

[0028] S2: Extract relevant attributes. Extract relevant attributes based on the basic information and historical behaviors of participating users in the life cycle of a topic activity, the content of the topic, and the interaction information.

[0029] S3: Build a model. First, the user relationship network can be obtained through the original social network, interaction frequency and other data; then the user-topic sentiment matrix is ​​constructed to map the user's emotional inclination towards different topics; finally, it is integrated with the original user basic data, and the graph convolutional neural network is used to build the derivative topic propagation prediction model AEGT-GCN, based on which the development trend of the user group towards a certain derivative topic is studied.

[0030] In this embodiment, obtaining the data source includes: obtaining raw data. The raw data can be obtained through the social network public API or by directly downloading the existing data source. Simple data cleaning. The raw data usually obtained is unstructured and cannot be directly used for data analysis. Most of the unstructured data can be structured through simple data cleaning. For example, deleting duplicate data, cleaning invalid information, etc.

[0031] In this embodiment, the attributes related to topic propagation data are extracted, including: user attributes, user relationship network, user historical behavior, topic information, social intimacy, user social influence, and message propagation heat.

[0032] S21: User attributes User (u i );

[0033] The user's own attributes include the user's age, the number of the user's fans and the number of the user's followers, which are related to the derivation of the topic. Therefore, the user's personal attributes are defined as:

[0034] S22: User relationship network R u ;

[0035] The user relationship network includes all participating users and the relationships between users, so the user relationship network is defined as:

[0036] R u = {U,E u}

[0037] S23: User historical behavior Action (u i );

[0038] User historical behavior includes the content and number of topics posted, the content and number of topics forwarded, and the content and number of topics commented on. These attributes reflect the user's behavioral preferences. Therefore, user historical behavior is expressed as:

[0039] Action(u i)=[public{n,c},forword{n,c},comment{n,c}]

[0040] Among them, public{n,c} represents the content c and the number n of topics published by users, forword{n,c} represents the content and the number of topics forwarded by users, and comment{n,c} represents the content and the number of topics commented by users.

[0041] S24: topic information topic(w);

[0042] Topic information includes information about the products in the topic, the start time, and the content. Therefore, the topic information is:

[0043] topic(w)=[product{w,p},time(w),content(w)]

[0044] Among them, product{w,p} represents the information related to product p in topic w, time(w) represents the start time of the topic, and content(w) represents the content of the topic.

[0045] S25: Social intimacy i ,u j};

[0046] Social intimacy refers to the behaviors of users forwarding, liking and commenting with other users. The higher the intimacy, the more likely users are to participate in each other's topics. Therefore, social intimacy is defined as:

[0047] intimacy i ,u j}=[retran{u i ,u j},like{u i ,u j},review{u i ,u j}]

[0048] Among them, retran{u i ,u j} indicates the number of topics forwarded by a user to another user, like{u i ,u j} represents the number of topics that a user likes another user, review{u i ,u j} represents the number of topics in which a user comments on another user.

[0049] S26: User social influence Inf(u i );

[0050] In social networks, the social influence of users is a key way to spread topics and can often determine the direction of things. The more influential a user is, the wider the scope of their dissemination will be. Therefore, the definition of user social influence is as follows:

[0051]

[0052] where fol(U i )、fol avg (net) respectively represent user U i The number of followers and the average number of followings for users in the network.

[0053] S27: message propagation heat MsgHeat (m);

[0054] MsgHeat(m) indicates the current heat of message m. Generally speaking, the higher the heat of a message, the easier it is to attract other users to take action on the message. The definition of message heat is as follows:

[0055]

[0056] Where Num[transpond(m)] represents the number of reposts of message m, and Num[fans(ui)] represents the original fan u of message m. i The number of fans, α and β are adjustable parameters, where α, β∈(0,1), which are used to neutralize the difference in the impact between the number of message forwarding and fans. And because the spread of messages will gradually decrease over time, this paper introduces a time decay function To simulate the change of message popularity over time, λ is the heat decay factor. The messages under different topics have different λ values, which are fitted by real data. t is the current time, t 0 The time when the message was originally created.

[0057] In this embodiment, the establishment of the model includes: the model is mainly divided into three stages: topic relationship construction, user-topic sentiment measurement and model algorithm design, such as Figure 1 In the first stage, attribute data that may affect user behavior are extracted from the original topic domain and the derived topic domain, and then the IT2vec representation learning algorithm is used to perform multi-domain iterative representation of this information to mine richer features hidden in the topic space.

[0058] At the same time, a generative adversarial network is introduced to enhance the implicit relationship between topics. In the second stage, the ME2vec algorithm is proposed to construct a derived sentiment knowledge graph based on the SentiWordNet dictionary, and semantic relationships are used to infer users' sentiment tendencies towards different topics. In the third stage, a native-derived topic propagation model AEGT-GCN based on evolutionary game theory and adaptive mechanism is established to achieve early derivative topic propagation prediction and simulate the strategy selection of nodes in topic propagation, thereby optimizing the performance and accuracy of the model.

[0059] The process of processing the fused topic data using a generative adversarial network includes: randomly sampling the fused topic data and setting conditional information; inputting the randomly sampled data and conditional information into a topic data generator to generate specific topic data; using a topic discriminator to verify the probability that the generated topic data is real data; calculating the loss function of the generative adversarial network based on the calculated probability, and optimizing and updating the parameters in the topic discriminator based on the loss function; updating the parameters in the topic data generator based on the updated topic data discriminator parameters; the topic data generator and the topic discriminator repeatedly iterate the input data, and follow the rules of local optimization and global optimization until the topic data generated by the topic data generator is similar to the real topic data, thereby obtaining enhanced topic features.

[0060] The rules of local optimization and global optimization include: calculating the parameter gradients of the topic data generator and the topic discriminator according to the loss function; updating the parameters of the topic data generator according to the parameter gradients of the topic data generator to minimize the value of the loss function; updating the parameters of the topic discriminator according to the parameter gradients of the topic discriminator to maximize the value of the loss function.

[0061] The ME2vce algorithm is used to perform sentiment analysis on the user-topic relationship matrix, including: using the Dirichlet allocation model to randomly generate topic distribution for each document; using the Dirichlet distribution as the conjugate prior distribution of the polynomial to model the topic distribution and obtain the topic theme corresponding to the user; and constructing a user-topic sentiment knowledge graph based on the topic theme.

[0062] Specifically include:

[0063] S31: Early sparse topic compensation;

[0064] First, obtain a rich set of labeled original topic domain feature datasets D S , and the relatively sparse unlabeled derived topic domain dataset D t The original topic domain data is D s ={(x s1 ,y s1 ),...,(x sn,y sn )}, where x si ∈X is the input, y si ∈X is the corresponding output. Similarly, the derived topic domain data is D t ={x t1 ,...,x tn}, where input y ti ∈Y. Use P(X S ) and Q(X t ) represents the data distribution of the original topic domain and the derived topic domain. S ) and Q(X t ) based on the TCA domain adaptation method to find a P(X S ) and Q(X t ), which can retain the common configuration information of the data in both domains after conversion. That is, there exists a nonlinear mapping function φ:X→H, so that X′ S ={x′ si}={φ(x si )},X′ t ={x′ ti}={φ(x ti )}, and the expected final P(X s ′)=Q(X t ′). So we have:

[0065]

[0066] We only need to minimize the distance between the two distributions to obtain an optimal nonlinear mapping function φ, so that in P(X s )≠Q(X t ) in the case of P(Y s |X s ′)=P(Y t |X t ′). The subsequent TCA-based eigendecomposition algorithm learns the optimal solution of the nonlinear mapping function φ.

[0067] S32: Sentiment measurement between users and topics.

[0068] Considering that users have different emotional tendencies towards different topics, different emotional tendencies may directly lead to whether users directly participate in the topic, so effectively quantifying the emotional tendencies of users towards different topics has a positive impact on predicting the spread of derived topics. In order to further explore this phenomenon, we first regard the user's historical participation topic as a document d, and use Latent Dirichlet Allocation (LDA) to randomly generate a topic distribution θ for each document d d , for each word w in topic d, according to θd Randomly select a topic z, and then according to the word distribution φ of topic z z , generate word w. At the same time, through the conjugate prior distribution of Dirichlet distribution to multinomial distribution, the prior distribution of document topic distribution and topic word distribution:

[0069]

[0070] Among them, B(α,β) is the beta function, which is used to normalize the probability density so that the sum is 1. α and β are the parameters of the distribution and are random variables with values ​​in the range [0,1].

[0071] The probability mass function of the multinomial distribution is:

[0072]

[0073] Where n is the total number of trials, x i is the number of occurrences of the i-th result, p i is the probability of the occurrence of the i-th result, maximizing the lower bound of evidence, that is:

[0074] L(q)=Ε q [logp(x,z)]-E q [logq(z)]

[0075] Among them, q(z) is the variational distribution, p(x,z) is the joint distribution, and E represents the expectation.

[0076] The sentiment analysis tool TextBlob is used to evaluate the user’s sentiment towards the topic and use it as the weight of the user-topic sentiment knowledge graph edge. At the same time, the user’s features such as age and gender are integrated into the feature matrix.

[0077] S33: Design of propagation prediction model.

[0078] In the field of social network derived topic research, few studies have considered both the relationship between users and the user-topic sentiment tendency. Based on the user relationship network and user topic sentiment graph mentioned in the previous two sections, this paper proposes a prediction model called AEGT-GCN by designing and fusing the time decay function. It uses a graph convolutional neural network as a deep learning model to capture the relationship and structural characteristics of nodes in graph data.

[0079] In this paper, the prediction of user group behavior is defined as a binary classification problem, that is, through the network-wide feature vector of the hot topic in time period t, predict whether the potential group users in time period t+1 will participate in the spread of the hot topic. The model input in this paper is as follows:

[0080] Feature matrix M t =N×F, where N represents the number of user nodes in the hot topic propagation network, and F represents the input feature dimension of each node, which includes the user's own attribute characteristics and external factor characteristics.

[0081] The user's adjacency matrix G at time t t = {U t ,E t}, represents the user full relationship network under the hot topic in time period t, where U t represents the set of user nodes in the propagation space at time t, E t Represents the set of edges between nodes in the propagation space.

[0082] The training of the derived topic propagation prediction model includes: constructing a training set, extracting features from the data in the training set to obtain a feature matrix; inputting the feature matrix into the graph convolutional neural network, adjusting the edge weights through the time decay function and the activity decay function to obtain a weight matrix; multiplying the feature matrix and the weight matrix and adding a bias, and then comparing it with the average matrix of the user-topic relationship network. Multiply; use Relu function to process the multiplication result and perform Dropout operation; use Sigmod activation function to represent the convolution output as probability values ​​of different classifications; calculate the loss function of the model according to the probability value, adjust the parameters of the model, and complete the model training when the loss function value converges.

[0083] This paper uses a two-layer graph convolutional neural network with Dropout as a hot topic behavior prediction model. First, the weights W and bias terms are randomly initialized. In the first layer of GCN, the feature matrix M is first t The edge weights are adjusted through the time decay function and the activity decay function.

[0084]

[0085] Among them, w i,j Represents user u i ,u j The weight coefficient between is the time decay function, is the activity decay function, σ is the time decay factor, τ is the activity decay factor, t i , t j The timestamps for input and output times.

[0086] The user feature matrix M t Multiply by the weight matrix W and add the bias, and Multiply. Then use the Relu function as the activation function of this layer, and perform the Dropout operation on the model training. Finally, use the Sigmod activation function to express the convolution output as the probability value of different classifications. The final specific formula is expressed as:

[0087]

[0088] Among them, W l is the weight matrix corresponding to the lth layer network in the graph convolutional network, Add self-loops I to the adjacency matrix G.

[0089]

[0090] Since the Sigmod function can compress the output value into the range of [0,1], an output value greater than 0.5 indicates that the user participates in the topic propagation in the current period, and a value less than 0.5 indicates that the user does not participate.

[0091] The loss function is defined as:

[0092]

[0093] Among them, N is the size of the experimental data set, Y i is the real sample label of the i-th user in the dataset, It is the output result of the user behavior prediction model.

[0094] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A derivative topic propagation prediction method based on sentiment transfer and evolutionary game, characterized in that: include: Obtain user topic propagation data and pre-process the topic propagation data; The topic propagation data includes original topic data, derived topic data and user data; The preprocessed original topic data and derived topic data are fused by using transfer learning method, and the fused topic data are processed by generative adversarial network to generate specific types of topic features; the user-topic relationship matrix is ​​constructed based on the topic features and user data; The ME2vce algorithm is used to perform sentiment analysis on the user-topic relationship matrix to obtain the user-topic sentiment knowledge graph; Obtain relevant attributes of user data, merge relevant attributes with user-topic sentiment knowledge graph, and obtain user feature matrix; The user feature matrix is ​​input into the pre-trained derivative topic propagation prediction model to obtain the derivative topic propagation prediction result.

2. According to claim 1, a derivative topic propagation prediction method based on sentiment migration and evolutionary game is characterized in that: The fusion of the original topic data and the derived topic data includes: using the TCA domain adaptation method to map the topic data in the derived topic domain to the topic data in the native topic domain to obtain the fused topic data.

3. The derivative topic propagation prediction method based on sentiment transfer and evolutionary game according to claim 1 is characterized in that: The process of processing the fused topic data using a generative adversarial network includes: randomly sampling the fused topic data and setting conditional information; inputting the randomly sampled data and conditional information into a topic data generator to generate specific topic data; using a topic discriminator to verify the probability that the generated topic data is real data; calculating the loss function of the generative adversarial network based on the calculated probability, and optimizing and updating the parameters in the topic discriminator based on the loss function; updating the parameters in the topic data generator based on the updated topic data discriminator parameters; the topic data generator and the topic discriminator repeatedly iterate the input data, and follow the rules of local optimization and global optimization until the topic data generated by the topic data generator is similar to the real topic data, thereby obtaining enhanced topic features.

4. According to claim 3, a derivative topic propagation prediction method based on sentiment transfer and evolutionary game is characterized in that: The rules of local optimization and global optimization include: calculating the parameter gradients of the topic data generator and the topic discriminator according to the loss function; updating the parameters of the topic data generator according to the parameter gradients of the topic data generator to minimize the value of the loss function; updating the parameters of the topic discriminator according to the parameter gradients of the topic discriminator to maximize the value of the loss function.

5. The derivative topic propagation prediction method based on sentiment transfer and evolutionary game according to claim 1 is characterized in that: The ME2vce algorithm is used to perform sentiment analysis on the user-topic relationship matrix, including: using the Dirichlet allocation model to randomly generate topic distribution for each document; using the Dirichlet distribution as the conjugate prior distribution of the polynomial to model the topic distribution and obtain the topic theme corresponding to the user; and constructing a user-topic sentiment knowledge graph based on the topic theme.

6. The derivative topic propagation prediction method based on sentiment transfer and evolutionary game according to claim 1 is characterized in that: Obtaining the user feature matrix includes: analyzing the sentiment information contained in the document through TextBlob, and using the sentiment information as the weight of the user-sentiment knowledge graph edge; extracting the relevant attributes of the user data, and integrating the user's relevant attribute characteristics into the user-sentiment knowledge graph according to the weight to obtain the user feature matrix; wherein the relevant attributes include the user's age, gender, browsing preferences, and product review data.

7. The derivative topic propagation prediction method based on sentiment transfer and evolutionary game according to claim 1 is characterized in that: The derivative topic propagation prediction model adopts a two-layer graph convolutional neural network.

8. The derivative topic propagation prediction method based on sentiment transfer and evolutionary game according to claim 7 is characterized in that: The training of the derived topic propagation prediction model includes: constructing a training set, extracting features from the data in the training set to obtain a feature matrix; inputting the feature matrix into the graph convolutional neural network, adjusting the edge weights through the time decay function and the activity decay function to obtain a weight matrix; multiplying the feature matrix and the weight matrix and adding a bias, and then comparing it with the average matrix of the user-topic relationship network. Multiply; use Relu function to process the multiplication result and perform Dropout operation; use Sigmod activation function to represent the convolution output as probability values ​​of different classifications; calculate the loss function of the model according to the probability value, adjust the parameters of the model, and complete the model training when the loss function value converges.

9. The derivative topic propagation prediction method based on sentiment transfer and evolutionary game according to claim 8 is characterized in that: matrix The expression is: in, is a diagonal matrix, is the adjacency matrix of the user-topic relationship network.

10. The derivative topic propagation prediction method based on sentiment transfer and evolutionary game according to claim 1 is characterized in that: The loss function expression of the model is: Among them, N is the size of the experimental data set, Y i The real sample label of the i-th user marked in the dataset, It is the output result of the user behavior prediction model.

Images