A time-domain psychological topic propagation method

Abstract

The application belongs to the technical field of topic propagation, and particularly relates to a topic propagation method based on time domain psychology, which comprises the following steps: obtaining topic propagation data, the data comprising user information, marketing topic information, user relationship network and community information; extracting relevant attributes of the topic propagation data; inputting the relevant attributes into a topic propagation model to obtain a topic propagation prediction result; and performing topic propagation according to the topic propagation prediction result; the application provides a measurement standard for user state conversion by analyzing the uncertainty of the psychological state of users in different time domains and the relationship between adjacent nodes; and the influence of cross-community marketing topic propagation and its conversion problem are analyzed, the influence of marketing topics in the cross-community space and the information propagation law are fully analyzed, and the cross-community marketing topic propagation law is analyzed in depth and comprehensively.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of topic propagation, and particularly relates to social platform marketing topic data analysis, and more particularly to an e-commerce platform marketing topic cross-community propagation model. BACKGROUND

[0002] With the popularity and development of the Internet, people can conveniently and quickly obtain information from various Internet social platforms. These social platforms have also become important positions for the spread of various information on the Internet. With the popularity of social networks and the rapid growth of the number of users, more and more enterprises realize the importance of marketing on social networks. For example, Facebook, Instagram, WeChat and other social network platforms. These social platforms have become an important channel for enterprises to interact with consumers, promote products and establish brand image. By using social networks for marketing, enterprises can interact with users in real time, accurately target the target audience, and expand brand influence through sharing and word-of-mouth propagation.

[0003] Currently, the research on cross-community marketing topic propagation mainly has two directions: one is to establish an information propagation model based on the classic popular disease transmission model SIR, and to predict the propagation trend of cross-community marketing topics on the Internet by improving the SIR model. The second is to summarize the characteristics of the actual propagation trend of topics in the network, and to predict the forwarding behavior of users based on deep learning algorithms combined with propagation characteristics.

[0004] For the cognitive driving measurement problem of users, Pu et al. designed an adaptive knowledge accumulation framework according to the cognitive process in the human brain. The framework can continuously learn in multiple fields and even generalize to new and unseen fields. Yang et al. proposed an extensible hierarchical heterogeneous multi-core architecture applied to learning to analyze various human cognitive activities, including behavior selection, context-dependent learning, etc. From the above research, it can be found that scholars mostly analyze and model the relationship between cognition and behavior, and there is relatively less research on extracting and analyzing the multi-dimensional and complex causes of cognitive driving. For the problem of uncertainty of user's time domain psychological state, Hao et al. proposed a multi-scale sentence embedding method CAMSE, which encodes sentences into embedding tensors based on context self-attention and multi-scale technology. Huang et al. integrated emotional intelligence and attention mechanism to improve the LSTM network, and used the emotional semantic information hidden in the text theme and context to realize the representation and classification of emotion. Obviously, the above literature tends to be more inclined to emotion classification and grading when studying psychological state, and the behavioral impact of user psychological changes in continuous time still needs further research. For the multi-level nature of cross-community marketing topic propagation, Yu et al. proposed to identify the process of key nodes spreading from the center to the outside as a regression problem. Wen et al. used Jensen-Shannon divergence and log-sigmoid conversion function to construct community networks to show the relationship between communities. In summary, scholars have conducted in-depth research on the multi-level coupling and mutual influence of cross-community topic propagation, but further research is still needed on the mutual connection and transformation between communities. SUMMARY

[0005] In summary, the current topic propagation research methods have the following problems: 1. How to effectively measure the multi-dimensional and complex causes of user's own cognitive driving has become a challenge. 2. In the process of cross-community marketing topic propagation, in addition to the factors of the topic itself, the time domain psychological state of the user plays an important role in the attention of the topic and the propagation in the related community; however, analyzing the specific influence of this psychological state in topic propagation has become a challenge. 3. In the process of different community marketing topic propagation, the situation and characteristics of topic propagation in different communities are different, but the communities need to be associated and converted. How to combine the differences between communities and the state conversion between community users has become a challenge.

[0006] To solve the problems existing in the prior art, the present application provides a topic propagation method based on time domain psychology, which comprises the following steps:

[0007] Step A: Obtain user historical topic propagation data, which includes user information, user relationship network, marketing topic information and social platform community network information; and pre-process the user historical topic propagation data;

[0008] Step B: extracting relevant attributes from the preprocessed user topic propagation data;

[0009] Step C: constructing a topic propagation model according to the relevant attributes, the topic propagation model comprising: a user-driven cognitive metric mechanism, a time-domain psychological state description mechanism, and a cross-community marketing information propagation model; the user-driven cognitive metric mechanism is used to obtain a proportionality coefficient of a user cognitive description mechanism parameter; the time-domain psychological state description mechanism is used to obtain a high-dimensional vector representation of a user; and the cross-community marketing information propagation model predicts a topic propagation result according to the high-dimensional vector representation of the user and the proportionality coefficient of the user cognitive description mechanism parameter;

[0010] Step D: predicting a propagation trend of a topic to be propagated according to the topic propagation model.

[0011] Advantages of the present application:

[0012] The present application quantifies the multi-dimensional and complex causes of cognitive driving by using a latent causal algorithm from the perspective of user cognitive driving causes; the present application uses word vector embedding and LSTM to depict the time-domain psychology of a user, and more accurately captures the changes in the psychological state of the user; finally, a cross-community marketing topic propagation model is proposed according to the correlation between communities, in combination with user cognitive driving and time-domain psychology, so as to deeply and comprehensively analyze the cross-community marketing topic propagation rules. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flowchart of the present application;

[0014] Figure 2 A schematic diagram of the latent causal algorithm of the present application for measuring user cognitive driving factors;

[0015] Figure 3 Word vectors and LSTM of the present application for measuring user time-domain psychology;

[0016] Figure 4 A state transition diagram of the CC-SIR model of the present application;

[0017] Figure 5 A schematic diagram of the CC-SIR model of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] A time domain psychological topic propagation method, the method comprising:

[0020] Step A: obtaining user historical topic propagation data, the data comprising user information, user relationship network, marketing topic information and social platform community network information; preprocessing the user historical topic propagation data;

[0021] Step B: extracting relevant attributes from the preprocessed user topic propagation data;

[0022] Step C: constructing a topic propagation model according to the relevant attributes, the topic propagation model comprising: a user-driven cognitive metric mechanism, a time domain psychological state description mechanism and a cross-community marketing information propagation model; the user-driven cognitive metric mechanism is used to obtain the proportion coefficient of the user cognitive description mechanism parameter; the time domain psychological state description mechanism is used to obtain the high-dimensional vector representation of the user; the cross-community marketing information propagation model predicts the topic propagation result according to the high-dimensional vector representation of the user and the proportion coefficient of the user cognitive description mechanism parameter;

[0023] Step D: predicting the propagation trend of the topic to be propagated according to the topic propagation model.

[0024] As Figure 1 The overall flow of the present application is shown, which shows that the input of the present application is the basic information of the user under the social platform, the historical behavior data of the user and the basic information of the marketing topic, and the output after the prediction model is the prediction result of whether the user will order the marketing topic commodity.

[0025] In this embodiment, the specific process of the present application comprises the following steps:

[0026] S1: obtaining data. The way to obtain data is to download open source data set on the network. Here, what needs to be obtained is the interaction of the marketing topic in its life cycle and the historical behavior data of the user who has interaction with the marketing topic. The marketing topic participation needs to obtain the time of the marketing topic interaction, the basic information of the interactive user and the basic information of the marketing topic; the historical behavior of the marketing topic interactor includes the marketing topic information interacted by the user in the history. And the data is preprocessed.

[0027] S2: extracting relevant attributes. According to the platform active user basic information and historical behavior under the life cycle of a certain marketing topic, the basic information and interaction information of the marketing topic are extracted.

[0028] S3: Model establishment. First, the latent causal algorithm is used to quantify the cognitive-driven multi-dimensional and complex causes. Then, in order to more accurately capture the changes of the user's psychological state, the word vector embedding and LSTM are used to depict the user's time-domain psychology. Finally, according to the correlation between communities, the cross-community marketing topic propagation model is proposed based on the user's cognitive drive and time-domain psychology.

[0029] In the embodiment, the data source acquisition includes:

[0030] S11: Obtain original data. The original data can be obtained through the public API of a social network or by directly downloading an existing data source.

[0031] S12: Simple data cleaning. Generally, the original data obtained is unstructured and cannot be directly used for data analysis. Through simple data cleaning, most of the unstructured data can be structured. For example, deleting duplicate data, cleaning invalid information, etc.

[0032] In the embodiment, the specific process of extracting the relevant attributes includes that in the social platform, the judgment of a user on a cross-community marketing topic is influenced by many factors, such as the activity degree of the user, the influence of adjacent users on the user, the heat of the topic when the user contacts the topic, the topic diffusion and the correlation between communities, etc.

[0033] S21, user driving force Dri(ui);

[0034] After the user publishes the topic information, the user's fans browse, comment and forward the information, so that the topic gradually spreads in the social network. Num[read(ui)], Num[comt(ui)] and Num[ret(ui)] respectively represent the average browsing quantity, the average comment quantity and the average forwarding quantity of all microblogs published by the user ui in a period of time before the topic is initiated. Since the browsing quantity is usually much larger than the other two and the browsing quantity has little influence on the topic propagation, a weakening coefficient μ is added before the average browsing quantity. The driving force of the user in the information propagation process can be defined as:

[0035]

[0036] S22, user activity Act(u i );

[0037] The user activity represents the activity degree of the user u i in the recent topic propagation. It is generally believed that the higher the activity of the user, the more likely the user is to participate in the topic propagation, and the greater the driving force of the user on the topic propagation.

[0038] The definition of the user activity is:

[0039] Act(u i )=η×Num[orig(u i )]+Num[retw(u i )]

[0040] where Num[orig(u i )] and Num[retw(u i )] represent the number of original microblog and the number of retweet microblog of user u i published at a certain time before the topic spread. Generally, the number of original microblog is much less than the number of retweet microblog, so a y[0, 1] is defined as a weakening coefficient to reduce the influence of the number of original microblog on the calculation of user activity.

[0041] S23, topic popularity P(t);

[0042] Topic popularity refers to the traffic size of a topic, that is, the degree of popularity of a topic. Each relatively independent topic will experience a process of gradually increasing heat and then gradually declining, and the traffic of the topic will slowly decline and slowly reach a low level of final value, representing that the topic has left the public view. And this decay process is similar to the half-life property of physical elements. This paper introduces a half-life function (1 / 2) to describe the decay process of topic traffic, and the definition of message heat is as follows:

[0043]

[0044] where t and t0 represent the current time and the starting time of the topic spread heat decay process, and T is the average propagation period of the topic.

[0045] S24, topic diffusion Diff(topic);

[0046] Topic diffusion refers to the degree of propagation of a topic in a social network, that is, the popularity and propagation of a topic. The high and low of topic diffusion reflects the popularity and propagation of a topic in a social network, and is an important indicator to measure the popularity and propagation of a topic in a social network. High topic diffusion means that the popularity and propagation of a topic in a social network are large, and the content and activities of the topic are easy to attract more people's attention and participation. Low topic diffusion means that the popularity and propagation of a topic in a social network are small, and the content and activities of the topic are not easy to attract more people's attention and participation.

[0047] Diff(topic)=Num[brow(topic)]+Num[comm(topic)+Num[like(topic)]+Num[C(topic)

[0048] S25, community correlation I(A, B);

[0049] The community correlation degree represents the correlation degree between communities. The correlation degree can be represented by community similarity. The higher the similarity, the more similar the propagation trend of the community topic, and vice versa. Measuring the community correlation degree can help to study the propagation law of similar community marketing topics.

[0050]

[0051] Wherein, A i , B j represent the keyword set extracted by different communities, and I represents mutual information.

[0052] In this embodiment, the model includes: constructing a user-driven cognitive measurement mechanism, constructing a time-domain psychological description mechanism, and constructing a cross-community marketing topic propagation model. In the first stage, the algorithm design of the user-driven cognitive measurement mechanism is constructed. Starting from the multi-dimensionality and complexity of user cognitive driving factors, considering the advantages of potential causal algorithm in analyzing multi-factor and potential factor model, a marketing topic user cognitive driving measurement method based on potential causal algorithm is proposed. First, the personal information, social relationship and community information of the user are vectorized, the users of different communities participating in the topic T are extracted, the inclination score model is established according to the user information and user topic acceptance, and the causal effect is calculated according to the control group and experimental group model results. The normalized proportion coefficient of the user cognitive description mechanism parameter is obtained. In the second stage, starting from the uncertainty of the user's psychological state in different time domains, a user time-domain psychological state description mechanism based on word vector embedding and LSTM is constructed, which is effective in representing high-dimensional space and time sequence space information. First, the user's published information is one-hot coded as the input of the word vector model, then the model is trained, and finally the high-dimensional vector representation of the user is obtained according to the model parameter matrix. In the third stage, a cross-community marketing information propagation model CC-SIR is proposed. First, mutual information is used to measure the correlation between communities, then the user cognitive driving and time-domain psychological state are combined to define the cross-community model state transition equation, and finally the cross-community marketing topic propagation model CC-SIR is obtained.

[0053] The specific steps are as follows:

[0054] S31: Construct a user-driven cognitive measurement mechanism.

[0055] First, the inclination score model is established, the personal information, social relationship and community information of the user are vectorized, the inclination score model is used to obtain the causal effect of the user, and the cognitive driving force coefficient is obtained according to the average causal effect, such as Figure 2As shown. First, the information vector serves as the model's input, X. During the data screening process, the data is divided into a treatment group (with the factor applied) and a control group (without the factor applied) based on whether a specific factor is applied. The marketing topic T that all users browsed together is determined, and the target variable Y is set based on whether the user accepted the content of the topic: 1 for acceptance and 0 for non-acceptance.

[0056] In order to evaluate the user's acceptance of the topic, generally speaking, if the user's subsequent browsing of related topics increases significantly after browsing the topic, the marketing topic is considered successful and the target variable is set to 1. The calculation method of acceptance Acc is as follows:

[0057]

[0058] Among them, num before Indicates the number of views on related topics before browsing topic T, num after It represents the number of views on related topics after browsing topic T, and Acc represents the final acceptance, which is the target variable Y.

[0059] First, normalize the input data. Assume that Xi represents the i-th eigenvalue in the feature matrix. The normalized value is Xnorm, which is defined as follows:

[0060] Xnorm=(Xi-Xmin) / (Xmax-Xmin)

[0061] Where Xmin and Xmax represent the minimum and maximum values ​​of the feature across all users, respectively. Each user’s feature vector is a row of data, i.e., a 1×n vector, where n is the number of features.

[0062] Use the Sigmoid function to convert the linear combination into a probability value and obtain the final predicted value:

[0063]

[0064] Among them, the Sigmoid function is defined as:

[0065] Sigmoid(x)=(1+e -x ) -1

[0066] Use mean square error as the loss function Loss of the propensity score model:

[0067]

[0068] After the propensity score model is completed and the prediction is finished, each sample will get a propensity score. Then, sample matching is performed, and for each intervention sample, a control sample with the highest score is matched.

[0069] In order to better obtain the matching object, on the basis of extracting the important features of the user, the propensity score of the user is combined to perform full matching. Specifically, for each intervention group sample, a control sample is matched in the control group to obtain a sample matching tuple (X, X'). First, the propensity score difference is calculated: for the intervention group sample X and the control group sample X', the propensity scores are p(X) and p(X') respectively. The propensity score difference is defined as:

[0070] Δp = |p(X) - p(X')|

[0071] The user feature similarity is calculated. For each user feature vector X and X', the Euclidean distance is calculated:

[0072]

[0073] The comprehensive matching score is calculated. The propensity score difference and the feature similarity are combined to define the comprehensive matching score Score(X, X'):

[0074] Score(X, X') = a · Δp + β · Dist(X, X')

[0075] Wherein, a and β are weight parameters, used to balance the influence of the propensity score difference and the feature similarity on the matching score.

[0076] The optimal matching is selected. In the control group, for each intervention group sample X, the control sample X' with the minimum matching score is selected, that is:

[0077]

[0078] The lower the matching score Score(X, X') is, and G0 represents the control group, indicating that the X of the treatment group and the X' of the control group are closer to the matching tuple, and the result of estimating the causal effect is better.

[0079] According to the latent causal model, the individual causal effect of each individual is calculated. The individual causal effect represents the difference between the results of the individual under the intervention treatment and the non-intervention treatment. X" represents the user feature with the lowest matching score corresponding to X in the control group. The causal effect CE is calculated as follows:

[0080]

[0081] According to the calculation of the causal effect CE of each processing group user and the matched control group user, then according to whether the related factors are applied, the average causal effect of ATT in each case is calculated as follows:

[0082]

[0083] Wherein, n represents the number of people in the experimental group, and i represents the first person.

[0084] The average value of each influencing factor is calculated, and the proportional coefficient of the user cognitive driving force mechanism is obtained through normalization. The proportional coefficient β i As follows:

[0085]

[0086] In the social network, the factors that determine the influence of the message in the cross-community marketing topic propagation are both the message itself factors and the user factors and the community factors. The factors of the message itself such as the topic heat and the topic diffusion degree, the user factors such as the activity degree, the cognitive matching degree and the time domain psychology, and the community factors including the community acceptance and the correlation degree between communities. Comprehensive three kinds of factors, the cognitive driving force function of the user ui is constructed as:

[0087] Con(ui)=β0+β1*Soc(ui)+β2*Com(ui)+β3*Per(ui)

[0088] Wherein, the parameters β1, β2, β3 are the causal effect coefficients obtained by the causal inference from the potential causal model, β1, β2 and β3 respectively represent the weights of the social factors, the community factors and the personal factors, and describe the proportions of the three factors in the cognitive driving force.

[0089] S32: Constructing the time domain psychology description mechanism.

[0090] Extracting the information such as the comments published by the user within 24 hours before and after contacting the marketing topic,

[0091] Like, collect content, etc. With 30 minutes as a slice, describe the topic information browsed according to the user behavior,

[0092] Comments, likes, forwards and other information content depict the user's time domain psychological state, and the obtained information is high-dimensional vectorized through word vector embedding, to obtain the user's time domain psychological state feature set, such as Figure 3 As shown.

[0093] Using one-hot encoding of the user published information as the input data set D=(x1, y1), …, (xn, yn) of the word vector model, wherein x i ∈X is the input, and y iThe weight matrix W, W', W is the final required word vector matrix, W' is the vector mapping of the hidden layer output to a vector z consistent with the size of the vocabulary i The vector z i is converted to a probability distribution representing the predicted central word by the Softmax function At y i Cross loss:

[0094]

[0095] Where N is the total number of samples, is the cross-entropy loss value.

[0096] By iteratively minimizing the loss function, update W, W' using Adam backpropagation until convergence. Calculate the user's time domain vector using W and the text extracted from the keywords:

[0097] Vec(Text(ui,t))=X i *W

[0098] The resulting user word vector is used as input to the LSTM model to obtain the user's current time domain psychological state description.

[0099] In the process of marketing topic topic propagation, due to the uncertainty of the psychological state of users at different times, the propagation of marketing topics is huge, different psychological states at different times will make users make completely different decisions, and the role of the user's adjacent node is also huge. Therefore, this paper introduces the LSTM network to analyze the user's time domain psychology and considers the influence of adjacent users. Based on the analysis in Chapter 2 and the output of the word vector embedding in the previous section, the output of the LSTM is obtained:

[0100] h t =σ(W o [h t-1 ,V t ]+b o )*tanh(σ(W f ·[h t-1 ,V t ]+b f ))*C t-1 +σ(W i ·[h t-1 ,V t ]+b i )*tanh(W C ·[h t-1 ,V t ]+b C )

[0101] Where, bo, bf, bi, bc are the bias of output gate, forget gate and input gate respectively.

[0102] O(t) = softmax(h t )

[0103] Considering the influence of adjacent nodes and topic influence, the time domain psychological coefficient δ(t) is:

[0104]

[0105] Wherein, δ(t) coefficient as a measure of the result of t moment user time domain psychological state.

[0106] S33, construct a cross-community marketing topic propagation model.

[0107] S331 cross-community state transition mechanism;

[0108] SIR propagation mechanism, research marketing topic propagation under the background of cross community. Due to the repeatability of cross community marketing topic life cycle, cross community marketing topic iteration propagation model CC-SIR is established, and cross model community topic interlayer state transition mechanism is constructed, as shown in Figure 4 For the convenience of model expression and simplification of theoretical expression, the present application takes three communities as an example for research, which are community 0, community 1 and community 2. The topic is first propagated in community 0, and after the forwarding of critical users, it is diffused to community 1. Similarly, the propagation path of community 2 is also through the forwarding of critical users in community 1.

[0109] Assume that for an independent topic, the infection rate is λ, that is, the probability from S state to I state is λ. When the topic is in cross community 1, 2, the infection rate and the community correlation degree are related to each other, and the topic propagation effect in two communities will tend to be the same.

[0110] The intercommunity topic propagation will produce interlayer conversion through the forwarding of critical users, and the conversion probability is directly related to the correlation degree of the two, and the correlation degree size threshold ε is specified, the cross community state conversion equation is given, and the cross model state conversion (i.e. a ->S b ) probability is as follows:

[0111]

[0112] S332: topic propagation mechanism;

[0113] Considering the repetition of the life cycle in the process of cross-community marketing topic, this paper multi-coupled SIR model according to different community layers, and constructed CC-SIR cross-community marketing topic propagation model. In this model, the users of each topic layer have three states: susceptible state (Susceptible), infected state (Infected), and immune state (Recovered). The construction of this model is based on the following assumptions (as shown in Figure 4

[0114] 1. Because the topic propagation has the characteristics of short time and outbreak, this paper does not consider the population dynamic factors such as birth, death, and flow in the research process. It is believed that the number of user nodes in the social network is equal at any time within the research period, so the sum of the state ratio of users in the same community layer model at any time:

[0115] S i +I i +R i =1

[0116] 2. Topic propagation is similar to the spread of infectious diseases. When a new user contacts a user who has already spread the topic, there will be a certain infection rate.

[0117] 3. In the process of cross-community marketing topic propagation, the critical users of different communities will directly inherit the susceptible state, infected state, and immune state of the topic.

[0118] According to the above assumptions, this paper defines the propagation rules of topics in social networks as follows: in the same community layer, when an infected node contacts a susceptible node, without considering the influence of other factors, it will convert the susceptible node to an infected node with a probability of λ. As the topic popularity decreases and time progresses, the infected node will transform into an immune node with a probability of μ.

[0119] Based on the CC-SIR model and the above propagation rules, the dynamic equations of k cross-community marketing topic communities are as follows:

[0120]

[0121] Because the information propagation in each marketing topic community layer is unidirectional, the state transition of users in the same layer is also unidirectional. That is, the user state can only change from susceptible state to infected state and then to immune state. Assuming that a user node ui has m neighbor nodes, and n of them forward the topic of this layer with a probability that follows a binomial distribution:

[0122]

[0123] The probability that any user ui forwards the topic of this layer at time t can be obtained as: ​

[0124]

[0125] In combination with the mean field theory, the infection rate at time t is obtained as Let α represent the community C i The correlation degree Rlv(ci-1, ci) with its predecessor community is defined, and the correlation degree threshold is defined as ε. The dynamic equation is obtained as follows:

[0126]

[0127] Through the combination of the mean field theory and the CC-SIR model, the final dynamic equation of the cross-community marketing topic is obtained, which accurately describes the propagation and evolution process of the topic in different communities, as shown in Equation (7). Figure 5 This model provides a theoretical basis for understanding and predicting the propagation trend of the topic in the complex social network.

[0128] The above examples further illustrate the purpose, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made to the present application within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A time-domain based psychological topic propagation method, characterized in that, The application relates to a method for constructing a topic propagation model based on user historical topic propagation data. The method comprises the following steps: Step A: obtaining user historical topic propagation data, which comprises user information, a user relationship network, marketing topic information and social platform community network information; Step B: extracting relevant attributes from the user historical topic propagation data, wherein the relevant attributes comprise user driving force, user activity, topic heat, topic diffusion degree and community correlation degree; The user driving force comprises the average number of views, the average number of comments and the average number of reposts within a period of time after the user publishes topic information; the average number of views is multiplied by a weakening coefficient, and then the sum of the average number of comments and the average number of reposts is calculated to obtain the user driving force; The user activity is the sum of the number of published topics and the number of reposted topics before the user propagates the topic; The topic heat is calculated by using a half-life function; The topic diffusion degree is the propagation degree of a topic in a social network, that is, the popularity and propagation force of the topic; The community correlation degree is calculated by extracting a keyword set of different communities; the similarity between two communities is calculated by using mutual information; a similarity threshold is set, and the calculated similarity is compared with the similarity threshold; if the calculated similarity is greater than the similarity threshold, the two communities are correlated; otherwise, the two communities are not correlated; Step C: constructing a topic propagation model according to the relevant attributes, wherein the topic propagation model comprises a user-driven cognitive measurement mechanism, a time-domain psychological state description mechanism and a cross-community marketing information propagation model; the user-driven cognitive measurement mechanism is used to obtain a proportion coefficient of a user cognitive description mechanism parameter; the time-domain psychological state description mechanism is used to obtain a high-dimensional vector representation of the user; and the cross-community marketing information propagation model is used to predict a topic propagation result according to the high-dimensional vector representation of the user and the proportion coefficient of the user cognitive description mechanism parameter; The user-driven cognitive measurement mechanism comprises the following steps: establishing a tendency score model; vectorizing user information, a user relationship network and community information; processing the vectorized data by using the tendency score model to obtain a causal effect of the user; and constructing a cognitive driving force function according to the causal effect, wherein the function is as follows: Con(ui) = beta0 + beta1 * Soc(ui) + beta2 * Com(ui) + beta3 * Per(ui) wherein beta0 is a hyperparameter, beta1, beta2 and beta3 are weight values of a social factor, a community factor and a personal factor, ui is a user, Soc(ui) is a user social information vector, Com(ui) is a user community information vector, and Per(ui) is a user personal information vector; The time-domain psychological state description mechanism comprises the following steps: obtaining comment data of the user on the topic; performing word vector embedding representation on the comment data to obtain a user time-domain psychological state feature set; and inputting the user time-domain psychological state feature set into an LSTM model to obtain a psychological state description of the user at a current time domain. The cross-community marketing information propagation model comprises: constructing a cross-community state transition mechanism and a topic propagation mechanism; constructing a cross-community state transformation equation according to the cross-community state transition mechanism, and calculating a cross-model state transition probability according to the cross-community state transformation equation; processing the topic according to the topic propagation mechanism, and obtaining the cross-community marketing information propagation model based on the cross-model state transition probability; Step D: predicting the propagation trend of the topic to be propagated according to the topic propagation model.

2. The time-domain based psycho-topic propagation method according to claim 1, wherein, The preprocessing of the user historical topic propagation data comprises: data cleaning of the historical topic propagation data to obtain structured data, and deleting repeated data and invalid data in the structured data.

Images