A bayesian network-based early prediction method for network hotspot events

Abstract

The application discloses a network hotspot event early prediction method based on a Bayesian network, and comprises the following steps: sampling and selecting a source data set, and performing label classification on each piece of data to obtain event set sample data; the event set sample data is divided into a training set and a test set; for each sample event in the event set sample data, early information data associated with various types of information is extracted as input features, and a Bayesian network model is constructed based on the feature data to obtain a joint probability distribution representation; based on the joint probability distribution representation and the training set, Bayesian network model parameter estimation is performed to obtain an optimized hotspot event early prediction model; sample events in the test set are input into the optimized hotspot event early prediction model to predict whether the sample events will become hotspot events, and a prediction result probability is obtained. The application solves the problems of high feature extraction complexity and poor early prediction effect of existing hotspot event prediction methods.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and event prediction, and particularly relates to a network hotspot event early prediction method based on a Bayesian network. BACKGROUND

[0002] Nowadays, online social networks such as microblog, WeChat, Twitter and the like have gradually penetrated into various aspects of people's daily life, such as society, economy and politics. Due to the strong openness, timely and rapid information dissemination and low cost of social networks, social networks are a real event mapping network and an important platform for the first publication and rapid dissemination. For example, official media accounts publish relevant events on microblog, various users comment, like and forward, and can form rapid dissemination in a short time, leading to explosive growth of the heat of a certain event, thus becoming a hotspot event. Therefore, to some extent, the hotspot event reflects the focus of the public at that time in life, economy or politics. Therefore, the prediction research on the hotspot event has very important practical significance, whether it is applied in public opinion monitoring, emergency management or recommendation system and the like.

[0003] Most of the existing hotspot event prediction methods are based on the prediction of the change of the hotspot in time series, and the process of event evolution is captured by observing the change of time series data information within a certain time to predict the hotspot event. However, the existing hotspot event prediction method has high complexity in network data feature extraction and poor early prediction effect. SUMMARY

[0004] In view of the above problems in the prior art, the present application provides a network hotspot event early prediction method based on a Bayesian network, which solves the problem of high complexity in network data feature extraction and poor early prediction effect of the existing hotspot event prediction method.

[0005] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows: a network hotspot event early prediction method based on a Bayesian network, comprising the following steps:

[0006] S1: sampling and selecting a source data set, and classifying each data to obtain event set sample data, and dividing the event set sample data into a training set and a test set;

[0007] S2: for each sample event in the event set sample data, extracting associated early information data as input features, and constructing a Bayesian network model based on the feature data to obtain a joint probability distribution representation;

[0008] S3: based on the joint probability distribution representation and the training set, performing Bayesian network model parameter estimation to obtain an optimized hotspot event early prediction model;

[0009] S4: Input the sample events in the test set into the optimized hot event early prediction model to predict whether the sample events will become hot events and obtain the prediction result probability.

[0010] The beneficial effect of the above scheme is: the present invention obtains an event data set, extracts information features from the data set, constructs a Bayesian network model and estimates model parameters to predict whether an event can become a hot event, thereby achieving early prediction of hot events while ensuring efficiency, and solving the problem that the existing hot event prediction methods have high complexity in extracting network data features and poor early prediction effect.

[0011] Furthermore, S1 includes the following sub-steps:

[0012] S1-1: Randomly sample the source data set according to the set ratio;

[0013] S1-2: Merge tweets with the same event or topic name tags in the randomly sampled data set to form a complete network event, including tweet information discussed by users, and label the complete network event with categories;

[0014] S1-3: Based on the category label, the tweet information is divided into T continuous time windows, and the size of each window is set to t hours;

[0015] S1-4: Divide the sample data into training set and test set according to the set ratio.

[0016] The beneficial effect of the above further solution is: through the above technical solution, the acquired source data is processed through sampling, merging and category labeling to obtain a data set for model training and testing.

[0017] Furthermore, the tweet information in S1-2 includes the content of the forwarded tweet, the forwarding time, the forwarding user ID and the original tweet ID.

[0018] The beneficial effect of the above further solution is: through the above technical solution, tweets with the same event or topic are merged according to the tweet content, forwarding time, forwarding user ID and original tweet ID.

[0019] Furthermore, the category marking in S1-2 is to set a threshold for marking as a hot event by analyzing the number of tweets forwarding the event. If the number of tweets forwarding the event is greater than the threshold, it is marked as 1, indicating a hot event. If the number of tweets forwarding the event is less than the threshold, it is marked as 0, indicating a non-hot event.

[0020] The beneficial effect of the above further solution is: through the above technical solution, the threshold of hot events is determined according to the number of tweet forwardings. When the threshold is exceeded, it is a hot event, otherwise it is a non-hot event, thereby completing the marking of network events.

[0021] Further, S2 includes the following sub-steps:

[0022] S2-1: design and extract information features of multiple dimensions as inputs of the event early prediction model, including original tweet information, time series forwarding information, user attributes and network structure;

[0023] S2-2: construct a Bayesian network model according to the inputs of the event early prediction model, the Bayesian network model including emotion score E, forwarding frequency and S, change rate V of S, change acceleration A of S, user number U, total number of fans F, maximum length L of forwarding chain, community number C and event category Y;

[0024] S2-3: based on the node attributes and structure of the Bayesian network model, obtain the joint probability distribution function p(Y, E, S, V, A, U, F, L, C), the formula is

[0025] p(Y, E, S, V, A, U, F, L, C) = p(Y) p(E|Y) p(S|Y) p(V|Y, S) p(A|Y, V) p(U|Y) p(F|Y) p(L|Y) p(C|Y)

[0026]

[0027] Wherein, p(Y) is the prior probability distribution of event category being a hot spot or a non-hot spot, p(E|Y) is the conditional probability distribution of E corresponding to Y, p(S|Y) is the conditional probability distribution of S corresponding to Y, p(V|Y, S) is the conditional probability distribution of V corresponding to Y and S, p(A|Y, V) is the conditional probability distribution of A corresponding to Y and V, p(U|Y) is the conditional probability distribution of U corresponding to Y, p(F|Y) is the conditional probability distribution of F corresponding to Y, p(L|Y) is the conditional probability distribution of L corresponding to Y, and p(C|Y) is the conditional probability distribution of C corresponding to Y.

[0028] The beneficial effects of the above further scheme are: through the above technical scheme, feature design and extraction are performed as inputs of the model, and then a Bayesian network model is constructed, and based on the node attributes of the existing model and the structure of the Bayesian network model, a joint probability distribution function is obtained.

[0029] Further, the original tweet information in S2-1 includes the initial semantic information of the event, and since emotion is related to the popularity of the tweet, for the i-th event, the event original text emotion feature E i as the original tweet information, the event original text emotion feature E i The formula is

[0030] E i ​SnowNLP(sentiments(text i ))

[0031] wherein, SnowNLP represents a Chinese text processing library, sentiments represents a sentiment feature extraction algorithm package in the text processing library, text i is the original tweet of the ith event.

[0032] The above further scheme has the beneficial effect that the feature extraction of the sentiment in the early prediction model is realized through the above technical scheme.

[0033] Further, the time sequence forwarding information in S2-1 includes the original tweet and the forwarded tweet, and the selected tweet forwarding frequency feature includes the text forwarding frequency within the observation time and the average change rate V i of the sum of the text forwarding frequency and i and the acceleration A i of the sum of the text forwarding frequency:

[0034] The sum of the text forwarding frequency S i is:

[0035]

[0036] wherein, f i (j) is the forwarding frequency of the ith event in the jth time period;

[0037] The average change rate V i of the sum of the text forwarding frequency is:

[0038]

[0039] wherein, V i (1) is the rate of the window of the former T / 2, V i (2) is the rate of the window of the latter T / 2;

[0040] The acceleration A i of the sum of the text forwarding frequency is:

[0041]

[0042] The above further scheme has the beneficial effect that the feature extraction of the tweet forwarding frequency in the early prediction model is realized through the above technical scheme.

[0043] Further, the features of the user attributes in S2-1 are the total number of participating users U i and the total number of fans F i in the continuous time period, and the formula is

[0044]

[0045]

[0046] wherein, is the number of active users in the jth time period, F i k is the number of fans owned by the kth user.

[0047] The beneficial effect of the above further scheme is that the feature extraction of the user attributes in the early prediction model is realized through the above technical scheme.

[0048] Further, the features of the network structure in S2-1 are the maximum length L of the forwarding chain and the number C of communities.

[0049] The beneficial effect of the above further scheme is that the feature extraction of the network structure in the early prediction model is realized through the above technical scheme.

[0050] Further, S3 includes the following steps:

[0051] S3-1: Obtain the prior probability distribution of each independent node variable through parameter estimation, the prior probability distribution of the hot event Y1 includes p(E|Y1), p(S|Y1), p(V1|Y1), p(V2|Y1), p(A|Y1), p(U|Y1), p(F|Y1), p(L|Y1) and p(C|Y1), and the prior probability distribution of the non-hot event Y0 includes p(E|Y0), p(S|Y0), p(V1|Y0), p(V2|Y0), p(A|Y0), p(U|Y0), p(F|Y0), p(L|Y0) and p(C|Y0);

[0052] S3-2: Calculate the conditional probability distribution function of the feature node based on the prior probability distribution;

[0053] S3-3: Based on the conditional probability distribution function, jointly obtain the joint distribution function p(X i ,Y1) of the input features of the hot event Y1, and jointly obtain the joint distribution function p(X i ,Y0) of the input features of the non-hot event Y0;

[0054] S3-4: Perform parameter estimation and optimization on the joint distribution function p(X i ,Y1) of Y1 and p(X i ,Y0) of Y0 through the coordinate descent algorithm, to obtain an optimized early prediction model of the hot event.

[0055] The above further scheme has the beneficial effect that: through the above technical scheme, the parameter estimation of the Bayesian network model is realized, and the objective function is maximized on the basis of the structure of the Bayesian network model. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A flowchart of a network hotspot event early prediction method based on a Bayesian network.

[0057] Figure 2 A Bayesian network model structure diagram.

[0058] Figure 3 A principle diagram of a network hotspot event early prediction method based on a Bayesian network. DETAILED DESCRIPTION

[0059] The application will be further described below in combination with the drawings and specific embodiments.

[0060] As Figure 1 shown, a network hotspot event early prediction method based on a Bayesian network comprises the following steps:

[0061] S1: sample a source data set, and label classify each data to obtain event set sample data, and divide the event set sample data into a training set and a test set;

[0062] S2: for each sample event in the event set sample data, extract associated early information data as input features, and construct a Bayesian network model based on the feature data to obtain a joint probability distribution representation;

[0063] S3: based on the joint probability distribution representation and the training set, perform Bayesian network model parameter estimation to obtain an optimized hotspot event early prediction model;

[0064] S4: input the sample events in the test set into the optimized hotspot event early prediction model to predict whether the sample events will become hotspot events, and obtain a prediction result probability.

[0065] S1 comprises the following steps:

[0066] S1-1: randomly sample the source data set according to a set proportion;

[0067] S1-2: merge the tweets with the same event or topic name label in the randomly sampled data set to form complete network events, including the tweet information discussed by users, and label the complete network events by categories;

[0068] S1-3: based on the category label, divide the tweet information into T continuous time window data, and set the size of each window to t hours;

[0069] S1-4: Divide the sample data into training set and test set according to the set ratio.

[0070] Random sampling: Considering that the sample size of the source data set is too large, this embodiment sets a 5% ratio to randomly sample the data as the data support for model prediction; data set division: select a 7:3 ratio to divide the data set into training set and test set, where the ratio of positive samples to negative samples in both the training set and the test set is 3:7.

[0071] The tweet information in S1-2 includes the content of the forwarded tweet, the forwarding time, the forwarding user ID and the original tweet ID.

[0072] The category marking in S1-2 is to set a threshold for marking as a hot event by analyzing the number of tweets forwarding the event. If the number of tweets forwarding the event is greater than the threshold, it is marked as 1, indicating a hot event. If the number of tweets forwarding the event is less than the threshold, it is marked as 0, indicating a non-hot event.

[0073] S2 includes the following steps:

[0074] S2-1: Design and extract information features from multiple dimensions as input to the early event prediction model, including original tweet information, time-series forwarding information, user attributes, and network structure;

[0075] S2-2: Construct a Bayesian network model based on the input of the early event prediction model. The Bayesian network model includes the sentiment score E, forwarding frequency S, the change rate of S V, the change acceleration of S A, the number of users U, the total number of fans F, the maximum length of the forwarding chain L, the number of communities C and the event category Y, such as Figure 2 shown.

[0076] There are correlations between different node attributes. Based on feature selection methods and expert analysis, it is clear that the rate of change (V) of forwarding frequency is related to the forwarding frequency and S over a given period. The acceleration (A) of S also depends on the rate of change (V) of S. The relationship between acceleration (A) and rate (V) is considered because even with the same acceleration, different speeds can lead to different probabilities. For example, two events may have the same forwarding acceleration, but the one with the faster speed may have a greater chance of becoming a hot topic in the future. Therefore, a directed edge from S to V is connected with a directed edge from V to A.

[0077] S2-3: Based on the node attributes and structure of the Bayesian network model, the joint probability distribution function p(Y, E, S, V, A, U, F, L, C) is obtained, and the formula is

[0078] p(Y, E, S, V, A, U, F, L, C) = p(Y)p(E|Y)p(S|Y)p(V|Y, S)p(A|Y, V)p(U|Y)p(F|Y)p(L|Y)p(C|Y)

[0079] p(A|Y, V)p(U|Y)p(F|Y)p(L|Y)p(C|Y)

[0080] Where p(Y) is the prior probability distribution of event category being hot or non-hot, p(E|Y) is the conditional probability distribution of E corresponding to Y, p(S|Y) is the conditional probability distribution of S corresponding to Y, p(V|Y, S) is the conditional probability distribution of V corresponding to Y and S, p(A|Y, V) is the conditional probability distribution of A corresponding to Y and V, p(U|Y) is the conditional probability distribution of U corresponding to Y, p(F|Y) is the conditional probability distribution of F corresponding to Y, p(L|Y) is the conditional probability distribution of L corresponding to Y, and p(C|Y) is the conditional probability distribution of C corresponding to Y.

[0081] The original tweet information in S2-1 includes the semantic information of the initial event. Since the sentiment is related to the popularity of the tweet, for the i-th event, the event original text sentiment feature E i is selected as the original tweet information. i The formula is

[0082] E i = SnowNLP(sentiments(text i ))

[0083] Where SnowNLP represents a Chinese text processing library, sentiments represents a sentiment feature extraction algorithm package in the text processing library, text i is the original tweet of the i-th event.

[0084] The Chinese text processing library SnowNLP is used to extract the sentiment feature and calculate the original tweet sentiment score, which represents the probability of semantic positivity. The score closer to 0 indicates that the sentiment is more negative, and vice versa. The score closer to 1 indicates that the sentiment information is more positive.

[0085] The time series forwarding information in S2-1 includes the original tweet and the forwarded tweet. The tweet forwarding frequency feature includes the text forwarding frequency and the average change rate V i of the text forwarding frequency and S i and the acceleration A i of the text forwarding frequency and S i :

[0086] The text forwarding frequency and S i are as follows:

[0087]

[0088] where f i (j) is the retweet frequency of the ith event in the jth time period;

[0089] the average rate of change V i is:

[0090]

[0091] where V i (1) is the rate of the first T / 2 window, V i (2) is the rate of the second T / 2 window;

[0092] the acceleration A i of the sum of retweet frequencies is:

[0093]

[0094] The feature of the user attributes in S2-1 is the total number of participating users U i and the total number of fans F i in a continuous time period, and the formula is:

[0095]

[0096]

[0097] where, is the number of active users in the jth time period, F i k is the number of fans owned by the kth user.

[0098] The features of the network structure in S2-1 are the maximum length of the retweet chain L and the number of communities C.

[0099] The maximum length of the retweet chain L can approximately represent the depth of the current affected users, that is, the depth of the event information propagation. This feature is extracted through the retweet state. Based on the ID of each tweet under the event and the ID of the original tweet it retweets, a propagation tree can be constructed, with the event original tweet publishing user as the root node. L is equivalent to finding the deepest length in the tree from the root node u0, that is, the tree height.

[0100] The number of communities C contained in the active users reflects the propagation potential of the event in the social network to some extent. The more the number, the easier it is to affect more different user groups. The Louvain algorithm, which has a relatively low time complexity, is selected as the hierarchical clustering algorithm. Through a recursive manner, the communities are merged into a single node, and the modular clustering is performed on the compressed graph to maximize the modularity score of each community, realizing the detection of the number of communities C.

[0101] S3 comprises the following sub-steps:

[0102] S3-1: obtaining prior probability distribution of each independent node variable through parameter estimation, the prior probability distribution of the hotspot event Y1 comprises p(E|Y1), p(S|Y1), p(V1|Y1), p(V2|Y1), p(A|Y1), p(U|Y1), p(F|Y1), p(L|Y1) and p(C|Y1), and the prior probability distribution of the non-hotspot event Y0 comprises p(E|Y0), p(S|Y0), p(V1|Y0), p(V2|Y0), p(A|Y0), p(U|Y0), p(F|Y0), p(L|Y0) and p(C|Y0);

[0103] S3-2: based on the prior probability distribution, calculating the conditional probability distribution function of the feature node;

[0104] S3-3: based on the conditional probability distribution function, jointly obtaining the joint distribution function p(X i ,Y1) of the input features of the hotspot event Y1 and the joint distribution function p(X i ,Y0) of the input features of the non-hotspot event Y0;

[0105] S3-4: performing parameter estimation and optimization on the joint distribution function p(X i ,Y1) of Y1 and p(X i ,Y0) of Y0 through the coordinate descent algorithm, to obtain an optimized hotspot event early prediction model.

[0106] In an embodiment of the present application, in the process of Bayesian model training for model parameter learning and estimation, the objective function p(Y, E, S, V, A, U, F, L, C) is maximized on the basis of the Bayesian network model structure. For the input features X = [E, S, V, A, U, F, L, C], it is known from the calculation of probability density distribution and cumulative probability distribution that the prior probability of each node feature conforms to Beta distribution, Gamma distribution and Gaussian distribution.

[0107] Since the sentiment score E, forwarding frequency, and S are greater than 0 and less than 1, we assume that E and S follow a beta distribution, denoted as E~Beta and S~Beta. The acceleration of change A of S can contain both positive and negative values, so it is assumed to follow a Gaussian distribution, denoted as A~Gaussian. The rate of change V of S, the number of users U, the total number of followers F, the maximum length of the forwarding chain L, and the number of communities C are all positive and greater than 1, with a unimodal distribution, and therefore follow a gamma distribution, denoted as U~Gamma, F~Gamma, L~Gamma, and C~Gamma. Let Y1 represent a hot event and Y0 represent a non-hot event. Furthermore, Y1 and Y0 are considered discrete distributions, with p(Y1)=p(Y=1)=β∈[0,1], representing the proportion of hot events in the sample. Furthermore, it is obtained that the characteristics of events Y1 and Y0 obey the following prior probability distribution p(E|Y1), p(S|Y1), p(V1|Y1), p(V2|Y1), p(A|Y1), p(U|Y1), p(F|Y1), p(L|Y1), p(C|Y1), p(E|Y0), p(S|Y0), p(V1|Y0), p(V2|Y0), p(A|Y0), p(U|Y0), p(F|Y0), p(L|Y0), and p(C|Y0).

[0108] After obtaining the prior probability distribution of each independent event feature node, for the joint distribution function p(Y, E, S, V, A, U, F, L, C), it is also necessary to calculate the conditional probability distribution functions of the feature nodes V and A, including p(V|Y1, S), p(V|Y0, S), p(A|Y1, V), and p(A|Y0, V).

[0109] For the hot event Y1 input feature X i =[E i ,S i ,V i ,A i ,U i ,F i ,L i ,C i ], we can get the joint distribution function p(X i ,Y1), and similarly we can get the joint distribution function p(X i ,Y0).

[0110] By using the coordinate descent algorithm, for the joint distribution function p(X i ,Y1) and p(X i ,Y0) to perform parameter estimation and optimization. The specific steps are as follows:

[0111] The model objective is to obtain the optimal parameters ω1 and ω0 by estimating the target distribution function by maximum likelihood.

[0112] In addition, the setting event i corresponds to the category numerical representation m i For

[0113]

[0114] Wherein, Y i is the category of event i;

[0115] For a given training data set D, the number of k = |D|, the target likelihood function is

[0116]

[0117] Wherein, beta is the Y1 prior probability, indicating the proportion of hot events in the sample;

[0118] The parameter learning target problem is to maximize the likelihood function maxlogp (D|beta, omega0, omega1).

[0119] To ensure the simplicity and efficiency of model learning, the coordinate descent algorithm is used for parameter estimation learning and optimization. Unlike the gradient optimization algorithm which finds the minimum value of the function along the steepest descent direction of the gradient, the coordinate descent method maximizes the target function Q value along the coordinate axis direction.

[0120] As Figure 3 shown, the present application faces the available social network data, based on expert experience knowledge, designs multiple information features, including text information, timing information, user information, structure information, and extracts features through data preprocessing; secondly, considering that there may be a dependent relationship between different characteristic variables, a Bayesian network structure is constructed to build the dependent relationship between the modeling features; finally, the model parameter learning and optimization are completed through parameter estimation, and the early prediction of hot events is realized. The present application can fuse various event initial information data under low time complexity, associate event changes and information causality, and realize early and effective prediction of hot events.

[0121] Those skilled in the art will appreciate that the embodiments described herein are presented for the purpose of helping the reader to understand the principles of the present application, and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A method for early prediction of network hot events based on Bayesian networks, characterized in that: The following steps are involved: S1: Sample and select the source data set, classify the labels of each data, obtain the event set sample data, and divide the event set sample data into training set and test set; S2: For each sample event in the event set sample data, extract various related early information data as input features, and build a Bayesian network model based on the feature data to obtain a joint probability distribution representation; The S2 includes the following steps: S2-1: Design and extract information features from multiple dimensions as input to the early event prediction model, including original tweet information, time-series forwarding information, user attributes, and network structure; S2-2: Constructing a Bayesian network model based on the input of the early event prediction model, the Bayesian network model includes the sentiment score , forwarding frequency and 、 The rate of change 、 The change of acceleration , number of users , total number of fans , Maximum length of forwarding chain , number of communities and event categories ; S2-3: Based on the node attributes and structure of the Bayesian network model, the joint probability distribution function is obtained , the formula is in, is the prior probability distribution of the event category being hotspot or non-hotspot, To correspond Down The conditional probability distribution of To correspond Down The conditional probability distribution of To correspond and Down The conditional probability distribution of To correspond and Down The conditional probability distribution of To correspond Down The conditional probability distribution of To correspond Down The conditional probability distribution of To correspond Down The conditional probability distribution of To correspond Down The conditional probability distribution of ; S3: Based on the joint probability distribution representation and training set, the Bayesian network model parameters are estimated to obtain an optimized early prediction model for hot events; S4: Input the sample events in the test set into the optimized hot event early prediction model to predict whether the sample events will become hot events and obtain the prediction result probability.

2. The method for early prediction of network hot events based on Bayesian network according to claim 1, characterized in that: The S1 includes the following steps: S1-1: Randomly sample the source data set according to the set ratio; S1-2: Merge tweets with the same event or topic name tags in the randomly sampled data set to form a complete network event, including tweet information discussed by users, and label the complete network event with categories; S1-3: Based on category tags, tweet information is divided into The data of continuous time windows is set, and the size of each window is set to Hour; S1-4: Divide the sample data into training set and test set according to the set ratio.

3. The method for early prediction of network hot events based on Bayesian network according to claim 2, characterized in that: The tweet information in S1-2 includes the forwarded tweet content, forwarding time, forwarding user ID and original tweet ID.

4. The method for early prediction of network hot events based on Bayesian network according to claim 3 is characterized in that: The category marking in S1-2 is to set a threshold for marking as a hot event by analyzing the number of tweets forwarding the event. If the number of tweets forwarding the event is greater than the threshold, it is marked as 1, indicating a hot event. If the number of tweets forwarding the event is less than the threshold, it is marked as 0, indicating a non-hot event.

5. The method for early prediction of network hot events based on Bayesian network according to claim 4 is characterized in that: The original tweet information in S2-1 includes the initial semantic information of the event. Since emotions are related to the popularity of tweets, events, select the emotional features of the original event As the original tweet information, the emotional characteristics of the original event The formula is in, Represents the Chinese text processing library, Represents the sentiment feature extraction algorithm package in the text processing library. For the The original tweet about the event.

6. The method for early prediction of network hot events based on Bayesian network according to claim 5, characterized in that: The time series forwarding information in S2-1 includes the original tweets and forwarded tweets, and the tweet forwarding frequency features are selected including the text forwarding frequency and , text forwarding frequency and average rate of change and the frequency and acceleration of text forwarding : Text forwarding frequency and for: in, For the The event in Frequency of forwarding within a time period; Frequency and average rate of change of text forwarding for: in, For the front The window rate, For the Queen The rate of the window; Frequency and acceleration of text forwarding for 。 7. The method for early prediction of network hot events based on Bayesian network according to claim 6, characterized in that: The characteristic of the user attribute in S2-1 is the total number of participating users in a continuous time period and total number of fans , the formula is in, For the The number of active users in a time period, For the The number of followers a user has.

8. The method for early prediction of network hot events based on Bayesian network according to claim 7, characterized in that: The network structure in S2-1 is characterized by the maximum length of the forwarding chain and number of communities .

9. The Bayesian network-based early prediction method for hot network events according to claim 8, characterized in that: The S3 includes the following sub-steps: S3-1: Obtain the prior probability distribution of each independent node variable through parameter estimation, hot events The prior probability distribution of 、 、 、 、 、 、 、 and , non-hotspot events The prior probability distribution of 、 、 、 、 、 、 、 and ; S3-2: Calculate the conditional probability distribution function of the feature node based on the prior probability distribution; S3-3: Based on the conditional probability distribution function, hot events The input features are combined to obtain The joint distribution function of , for non-hotspot events The input features are combined to obtain The joint distribution function of ; S3-4: Through the coordinate descent algorithm The joint distribution function of and of Perform parameter estimation and optimization to obtain an optimized early prediction model for hot events.

Images