A short-term prediction method for topic popularity considering competition and cooperation

Abstract

The present invention discloses a method for short-term prediction of topic popularity that considers competition and cooperation relationships. The method comprises the following steps: obtaining a document collection, extracting topics, and determining target topics and non-target topics; selecting non-target topics whose similarity to the target topic is less than a given similarity threshold as potential influence topics; dividing the document collection into time periods with time intervals t as time intervals, calculating the number of occurrences of the target topic and each potential influence topic within each time period, and forming a popularity vector for the topic; calculating a popularity lag vector for each potential influence topic based on the popularity vector of each potential influence topic; aggregating the popularity lag vectors of all potential influence topics to form a potential influence topic time lag matrix; and predicting the popularity of the target topic within the next time interval t based on the target topic popularity vector and the potential influence topic time lag matrix. The present invention has the beneficial effect of utilizing competition and cooperation relationships to improve the interpretability and accuracy of short-term popularity prediction.

Description

Technical Field

[0001] The present invention relates to the field of social network big data analysis and mining. More specifically, the present invention relates to a short-term prediction method for topic popularity that takes into account competition and cooperation relationships. Background Art

[0002] With everyone able to express their opinions on social media, the sheer number of topics emerging on these platforms far exceeds users' information processing capabilities. These topics compete for social media users' attention like animals competing for limited food. This means that topics often spread independently, influencing each other. To better recommend valuable and popular topics to customers, accurately predicting topic popularity is both necessary and challenging.

[0003] Existing topic popularity prediction methods often only consider the content and user attributes of a single topic, ignoring the interactive influence of multiple topics. Therefore, it is urgent to address the shortcomings of existing topic popularity prediction methods in ignoring the interactive influence of multiple topics, automatically identify competitors and partners from massive amounts of information, and leverage these relationships to improve the interpretability and accuracy of popularity predictions. Summary of the Invention

[0004] An object of the present invention is to solve at least the above problems and to provide at least the advantages which will be described hereinafter.

[0005] Another object of the present invention is to provide a method for short-term prediction of topic popularity that takes into account competition and cooperation relationships, which utilizes competition and cooperation relationships to improve the interpretability and accuracy of popularity prediction.

[0006] In order to achieve these purposes and other advantages according to the present invention, a method for short-term prediction of topic popularity considering competition and cooperation relationships is provided, comprising the following steps:

[0007] Obtain document collection, extract topics, and determine target topics and non-target topics;

[0008] Take non-target topics whose similarity to the target topic is less than a given similarity threshold as potential influencing topics;

[0009] For the document collection, the time period is divided into time intervals with time t as the time interval. The number of times the target topic and each potentially influential topic appear in each time period is calculated to form the popularity vector of the topic.

[0010] Based on the popularity vector of each potentially influential topic, a popularity lag vector of each potentially influential topic is calculated;

[0011] The popularity lag vectors of all potentially influential topics are aggregated to form a potential influential topic time lag matrix;

[0012] The popularity of the target topic in the next time interval t is predicted based on the target topic popularity vector and the potential influence topic time lag matrix.

[0013] Preferably, topics are extracted from the document collection using regular expressions.

[0014] Preferably, the method for calculating the similarity between the non-target topic and the target topic comprises the following steps:

[0015] For each extracted topic, traverse the document collection to obtain the document collection containing the topic;

[0016] For each topic, the texts in the document collection are connected end to end to form the topic text;

[0017] Segment each topic text and remove stop words to obtain the topic word set for each topic;

[0018] Take all topics and topic word sets as input to train the Doc2vec model;

[0019] Based on the trained Doc2vec model, we obtain the topic vectors of the target topic and each non-target topic.

[0020] For each non-target topic, the Euclidean distance between its topic vector and the target topic vector is calculated as the similarity between the non-target topic and the target topic.

[0021] Preferably, the document set is divided into time periods by taking the time point with the target topic as the starting point and the deadline of the document set as the end point, wherein, in chronological order, the time interval of the first time period is ≤ t.

[0022] Preferably, the document set is divided into a time period, and the popularity vector of the target topic is recorded as y tar,q =(y tar,1 ,…,y tar,a ) T , where y tar,a represents the number of times the target topic appears in the ath time period, and T represents the vector transpose.

[0023] Preferably, the popularity of the target topic in the next time interval t is expressed as: Among them, μ tar,a+1 is the popularity of the target topic at time a+1 estimated based on the target topic popularity vector; X i,a+1-m β i is the popularity of the target topic at time a+1 estimated based on the time-lag matrix of potential influencing topics, and the regression coefficient βi =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T .

[0024] Preferably, according to y tar,q , construct a local linear trend model, as follows:

[0025]

[0026]

[0027] Among them, μ tar,j represents the horizontal trend part of the target topic at the jth time period;

[0028] v tar,j-1 represents the slope of the target topic level trend at j-1 time periods;

[0029] ε tar,j-1 , White noise representing a horizontal trend;

[0030] ∈ tar,j-2 , White noise representing the slope of the horizontal trend;

[0031] When j≤a, y tar,j As μ tar,j The local linear trend model is input, and the maximum likelihood estimation and Kalman filter algorithm are used to iteratively obtain v tar,a-1 ,∈ tar,a-1 , ε tar,a The estimated value of further iterations is used to obtain the horizontal trend part μ of the target topic in the a+1 time period tar,a+1 .

[0032] Preferably, there are d potentially influential topics in total, and the popularity vector of the i-th potentially influential topic is recorded as y i,q =(y i,1 ,…,y i,a ) T ,i∈{1,…,d}, where y i,a represents the number of times the i-th potentially influential topic appears in the a-th time period;

[0033] Given the lag order l to be considered, l<a, the popularity lag vector of the i-th potentially influential topic is recorded as x i,a-m =(x i,m+1,a-m ,…,x i,a,a-m ) T,i∈{1,…,d},m∈{1,…,l-1}, where x i,m+1,a-m represents the m-th order lag term of the i-th potentially influential topic in the m+1-th time period.

[0034] Preferably, use Indicates the potential impact of topics on the target topic h tar The influence of popularity, i.e.

[0035] but The likelihood function can be expressed as follows:

[0036] Where i∈{1,…,d};

[0037] means The j-th element in , j∈{3,…,a}, is the influence part at the j-th moment.

[0038] x i,j-1 It's X i,a-m The vector consisting of the elements in the j-1th row;

[0039] is the j-th row and j-th column element in the diagonal matrix R;

[0040] Assuming the regression coefficient β i Obeying the nail-board distribution, we get: P(β i ,γ,R β )=P(β i |γ,R β )P(R β |γ)P(γ), where

[0041] Setting β i The mean is 0 and the covariance matrix is ​​R β Normal distribution; K is the number of non-target topics and their lag terms to be introduced into the impact part; π k β i The probability that the kth item is introduced into the model, and when β i The kth term β k ≠0, γ k =1; when β k = 0, γ k =0;

[0042] By using the MCMC method to obtain β i , based on the estimated β i And the potential impact topic delay matrix X at time a+1 i,a+1-m , using the following formula: Predict the impact part X of the a+1 period i,a+1-m β i .

[0043] Preferably, when β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T The coefficient β in i,a-c When ,i∈{1,…,d},c∈{1,…,m} is positive and significantly different from 0, the potential influencing topic is a cooperative topic of the target topic;

[0044] When β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T The coefficient β in i,a-c When ,i∈{1,…,d},c∈{1,…,m} is negative and significantly different from 0, the potential influencing topic is a competitive topic of the target topic.

[0045] The present invention has at least the following beneficial effects:

[0046] The method makes up for the shortcomings of existing short-term prediction methods of topic popularity, such as ignoring the interactive influence of multiple topics and being unable to automatically identify competitive topics and cooperative topics from massive information, and uses the relationship between competition and cooperation to improve the interpretability and accuracy of popularity prediction. Compared with the AR method and dynamic linear method commonly used in traditional time series prediction in one-step forward prediction tasks, it is found that the method proposed in this patent has improved the Cumulative Absolute Error (CAE) indicator by at least 15.46% and the Root Mean Squared Error (RMSE) indicator by at least 16.97% compared with the comparison method.

[0047] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flowchart of a method for short-term prediction of topic popularity that takes into account competition and cooperation relationships, as described in one of the technical solutions of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below with reference to the embodiments so that those skilled in the art can implement the invention with reference to the description.

[0050] like Figure 1 As shown in FIG, a short-term prediction method for topic popularity considering competition and cooperation relationships includes the following steps:

[0051] Step 1: Define topic popularity, target topics, and non-target topics

[0052] On social media platforms such as Weibo and Twitter, users often use hashtags (for example, in the form of “#keyword#” on Weibo) to mark topics and organize discussions on the topics.

[0053] As an emerging information classification method on social media, hashtags are attracting more and more users with their features of increasing exposure and mobilizing user participation. Hashtags have also become an effective way to identify social media topics.

[0054] In view of this, the present invention defines the popularity of a topic as: the number of times a topic tagged with a hashtag is used by users in a specific time interval t (e.g., 1 hour, 1 day, etc.) in social media;

[0055] When it is necessary to predict the popularity of a target topic within a time interval t from the current moment, obtain the text data before the current moment to form a document set; take the time point with the target topic in the document set as the starting point, take the deadline of the document set as the end point, and divide the text in the document set into a text set under a time period according to a specific time interval t. Each time period is recorded in sequence as q=1, 2, 3, ..., a according to the time progress, where "q=1" represents the first time period, and the corresponding time period time interval is ≤t;

[0056] Define the topics of interest (topics to be predicted) in applications (social media such as Weibo and Twitter) as target topics, and define other topics in the document collection except the target topics as non-target topics;

[0057] Predict the popularity of the target topic in the next time period (a+1), which is the short-term popularity of the target topic.

[0058] Step 2: Obtain document collection, extract topics, and determine target topics and non-target topics

[0059] The fixed format of hashtags in social media makes it possible to extract topics using regular expressions. Taking Weibo as an example, hashtags use “#keyword#” to mark topics;

[0060] For each text in the document set W of Weibo, use formula (1) to match the topic tags in the text;

[0061]

[0062] Formula (1) represents: English characters (az, AZ) and traditional and simplified Chinese characters that start and end with a pound sign. model;

[0063] By matching the topic tags in each text in the document set W, we can get the topic set H = [h1,…,h n ], that is, there are n topics in total, and the topic of interest in the application is selected as the target topic h tar , the remaining topics are non-target topics h i ,i∈{1,…,n},i≠tar.

[0064] Step 3: Calculate the semantic similarity between target topics and non-target topics

[0065] First, for each topic h in the topic set H i ,i∈{1,…,n}, traverse the document set W and obtain the document set W containing the topic i ={w1,…,w m},i∈{1,…,n}, where w1 is a document, i.e., the m documents containing the topic constitute the document set corresponding to the topic;

[0066] Next, for each topic h i , the document set W i The text in the topic text is connected end to end i , that is, s i =w1+…+w m .

[0067] Preprocess the topic text, specifically including: for each topic text s i Segment the words, remove stop words, and get the topic word set w i ,i∈{1,…,n}.

[0068] Topic h i and topic word set w i As input, train the Doc2vec model;

[0069] Based on the trained Doc2vec model, we get the topic vector b of the target topic. tar , and the topic vector b of the non-target topic i ,i∈{1,…,n},i≠tar;

[0070] For each non-target topic, calculate the Euclidean distance between its topic vector and the target topic vector as the similarity dist between the non-target topic and the target topic i ,i∈{1,…,n},i≠tar.

[0071] Step 4: Take non-target topics whose similarity to the target topic is less than the given similarity threshold as potential influencing topics. For the document set, calculate the number of times the target topic and each potential influencing topic appear in each time period to form the popularity vector of the topic.

[0072] 4.1. Take non-target topics whose similarity to the target topic is less than a given similarity threshold as potential influencing topics

[0073] Given a similarity threshold δ, traverse the similarity dist between non-target topics and target topics i ,i∈{1,…,n},i≠tar, when dist i <δ, the corresponding non-target topic is the potential influence topic, and all potential influence topics constitute the potential competition-cooperation topic set H′={h1,…,h d}, that is, there are d potential influential topics, h d represents the dth potentially influential topic;

[0074] 4.2. For the document collection, calculate the number of times the target topic and each potentially influential topic appear in each time period to form the popularity vector of the topic

[0075] Based on the document set W, calculate each topic (target topic h tar and the d potential influence topics (h1,…,h d ))The number of times it appears in each time period forms the popularity vector y of the topic i,q =(y i,1 ,…,y i,a ) T ,i∈{1,…,d,tar}, where y i,a Indicates topic h i ,i∈{1,…,d,tar} is the number of times it appears in the ath time period, T represents the vector transpose;

[0076] When i∈{1,…,d}, it is the popularity vector y of d potential influential topics in the potential competition-cooperation topic set i,q =(y i,1 ,…,y i,a ) T ,i∈{1,…,d}, where y i,a represents the number of times the i-th potentially influential topic appears in the a-th time period (proceed to step 5);

[0077] When i=tar, it is the popularity vector y of the target topic tar,q =(y tar,1 ,…,y tar,a ) T , where y tar,a represents the number of times the target topic appears in the ath time period, and T represents the vector transpose (proceed to step 6.1);

[0078] Step 5: Based on the popularity vector of each potentially influential topic, calculate the popularity lag vector of each potentially influential topic, specifically:

[0079] Given the lag order l to be considered;

[0080] For each potential influence topic h in the potential competition-cooperation topic set i ,i∈{1,…,d}, the lag vector of the topic popularity is obtained by moving forward l order time;

[0081] x i,a-m =(x i,m+1,a-m ,…,x i,a,a-m ) T ,i∈{1,…,d},m∈{1,…,l-1}

[0082] Among them, x i,a-m Potentially impacted topics i The lag vector of the potential influence on the popularity of the topic is obtained by moving forward m times;

[0083] x i,m+1,a-m Indicates the potential impact topic h i The m-th order lag term in the m+1th time period, that is, topic h i The number of users used in the first time period;

[0084] Taking l=4 as an example, the potential impact topic h i The lag vectors that potentially affect the popularity of the topic obtained by moving forward by l order of time include:

[0085] x i,a-1 =(x i,2,a-1 ,…,x i,a,a-1 ) T ,i∈{1,…,d};

[0086] x i,a-2 =(x i,3,a-1 ,x i,3,a-2 ,…,x i,a,a-1 ,x i,a,a-2 ) T ,i∈{1,…,d};

[0087] x i,a-3 =(x i,4,a-1 ,x i,4,a-2 ,x i,4,a-3 ,…,x i,a,a-1 ,x i,a,a-2 ,x i,4,a-3 ) T ,i∈{1,…,d};

[0088] Aggregating the d lag vectors of the potential influence topic popularity forms a potential influence topic lag matrix, which is expressed as:

[0089] X i,a-m =(x 1,a-1 ,…,x 1,a-m ,…,x d,a-1 ,…,x d,a-m )∈R (a-1)×(d×m)

[0090] Where: R (a-1)×(d×m) Represents X i,a-m It is a matrix with (a-1) rows and (d×m) columns.

[0091] Step 6: Predict the popularity of the target topic in the next time interval t based on the target topic popularity vector and the potential influence topic time lag matrix. Specifically:

[0092] In modeling the target topic h tar When the popularity is high, its popularity vector y tar,q =(y tar,1 ,…,y tar,a ) T It is decomposed into two parts, namely the horizontal trend part μ of the popularity of the target topic tar,q And the influence of d potential influence topics in the potential competition-cooperation topic set on the popularity of the target topic X i,a-m β i ,(i∈{1,…,d}),is expressed as follows:

[0093] y tar,q =μ tar,q +X i,a-m β i +e i ,e i ~N(0,R),i∈{1,…,d}

[0094] Where: horizontal trend part μ tar,q =(μ tar,1 ,…,μ tar,a ) T , characterize the use of target topic h tar The popularity at time a is estimated based on the historical popularity of the target topic h.tar Estimation and prediction based on the time series information itself;

[0095] The d potential influencing topics in the potential competition-cooperation topic set have an impact on the target topic h tar The impact of popularity i,a-m β i ,i∈{1,…,d}, describes the time-lag matrix X of the other d potential influencing topics i,a-m For the target topic h tar The effect of popularity, the effect size is determined by the regression coefficient β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T portrayal;

[0096] e i ,e i ~N(0,R) is white noise with a mean of 0 and a covariance of R, and R is a diagonal matrix;

[0097] Based on the above analysis, the popularity of the target topic in the next time interval t is expressed as: Among them, μ tar,a+1 is the popularity vector of the target topic at time a+1, which is the horizontal trend part, estimated based on the popularity vector of the target topic; X i,a+1-m β i is the popularity vector of the target topic at time a+1 estimated based on the time-lag matrix of the potential influencing topic, i.e., the influence of the potential influencing topic on the popularity of the target topic;

[0098] The horizontal trend part μ is described in detail below. tar,a+1 and the impact part X i,a+1-m β i Calculation method:

[0099] (6.1) Horizontal trend part μ tar,a+1

[0100] Based on the popularity vector y of the target topic tar,q =(y tar,1 ,…,y tar,a ) T , construct a local linear trend model, as follows:

[0101]

[0102]

[0103] Where j = 3, ..., a, a + 1,

[0104] μ tar,j Represents the horizontal trend part μ tar,a+1 The jth element is the horizontal trend part of the target topic at the jth time period;

[0105] v tar,j-1 It represents the slope of the horizontal trend of the target topic in the j-1 time period, which determines the angle between the trend line and the X-axis;

[0106] ε tar,j-1 , White noise representing a horizontal trend;

[0107] ∈ tar,j-2 , White noise representing the slope of the horizontal trend;

[0108] When j≤a, y tar,j As μ tar,j The local linear trend model is input, and the maximum likelihood estimation and Kalman filter algorithm are used to iteratively obtain v tar,a-1 , model parameters (∈ tar,a-1 , ε tar,a )

[0109] Based on the model parameters, v tar,a-1 , μ tar,a , use the following formulas (2') and (3') to realize the horizontal trend μ at time a+1 tar,a+1 predictions;

[0110]

[0111]

[0112] (6.2) Affecting part X i,a+1-m β i

[0113] use Indicates the potential impact of topics on the target topic h tar The influence of popularity, i.e. Among them, the regression coefficient β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-m ,…,β d,a-m ) T

[0114] but The likelihood function can be expressed as follows:

[0115]

[0116] Among them, i∈{1,…,d};

[0117] means The jth element in , j∈{3,…,a}, is h at the jth moment tar The remaining part of popularity after removing the horizontal trend (the impact part);

[0118] x i,j-1 It's X i,a-m The vector consisting of the elements in the j-1th row;

[0119] is the j-th row and j-th column element in the diagonal matrix R;

[0120] Assuming the regression coefficient β i Obey the nail plate distribution, that is:

[0121] P(β i ,γ,R β )=P(β i |γ,R β )P(R β |γ)P(γ) (5)

[0122]

[0123] Where: Assume β i To obey the mean of 0, the covariance matrix is ​​R β Normal distribution;

[0124] K is the number of non-target topics and their lag terms to be introduced into the impact part;

[0125] π k β i The probability that the kth item is introduced into the model, and when β i The kth term β k ≠0, γ k =1; when β k = 0, γ k =0;

[0126] By using the MCMC (Markov Chain Monte Carlo method), the γ in (4), (5), and (6) can be realized. k , β i 、R β estimates;

[0127] Based on the estimated parameters and the time lag matrix of potential impact topics at time a+1

[0128] X i,a+1-m =(x 1,a,…,x 1,a+1-m ,…,x d,a ,…,x d,a+1-m )∈R a×(d×m) ;

[0129] Using formula (4'),

[0130]

[0131] Predict the impact of time period a+1 on part X i,a+1-m β i ;

[0132] (6.3) The predicted value μ at time a+1 of the fusion horizontal trend part tar,a+1 And the potential influence topic on the target topic h tar The impact of popularity i,a+1-m β i , we can get the predicted popularity value of the target Hashtag at time a+1 Right now

[0133]

[0134] Step 7: Identify competing and cooperating topics of the target topic.

[0135] When β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T The coefficient β in i,a-c ,i∈{1,…,d},c∈{1,…,m} is positive and significantly different from 0, indicating that topic h i The j-order lag term of h tar There is a positive driving relationship between the popularity of , that is, the potential influencing topic is a cooperative topic of the target topic;

[0136] When β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T The coefficient β in i,a-c ,i∈{1,…,d},c∈{1,…,m} is negative and significantly different from 0, it means that the topic h i The j-order lag term of h tar There is a negative weakening relationship between the popularity of the topic and the target topic, that is, the potential influencing topic is a competing topic of the target topic.

[0137] The size of the coefficient represents the strength of the competition and cooperation relationship.

[0138] Experimental results:

[0139] Based on five topic sets extracted from five Weibo document collections, the method proposed in this patent is compared with the AR method and dynamic linear method commonly used in time series prediction. The prediction results show that the method proposed in this patent improves the Cumulative Absolute Error (CAE) index by at least 15.46% and the Root Mean Squared Error (RMSE) index by at least 16.97% compared with the comparison method, which shows the effectiveness of the method proposed in this patent in short-term prediction.

[0140]

[0141]

[0142] Where y tar,j is the actual value of the popularity of the target topic at time j, is the predicted value of the popularity of the target topic at time j using the method proposed in this patent or the comparative method.

[0143] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A short-term prediction method for topic popularity considering competition and cooperation, characterized by: The following steps are involved: Obtain document collection, extract topics, and determine target topics and non-target topics; Obtain non-target topics whose similarity to the target topic is less than a given similarity threshold as potential influencing topics; For the document collection, the time period is divided into time periods with time t as the time interval, and the number of times the target topic and each potential influencing topic appear in each time period is calculated to form the popularity vector of the topic; Based on the popularity vector of each potentially influential topic, a popularity lag vector of each potentially influential topic is calculated; The popularity lag vectors of all potentially influential topics are aggregated to form a potential influential topic time lag matrix; Predict the popularity of the target topic in the next time interval t based on the target topic popularity vector and the potential impact topic time lag matrix. For the document collection, it is divided into a time period, and the popularity vector of the target topic is recorded as y tar,q =(y tar,1 ,…,y tar,a ) T , where y tar,a represents the number of times the target topic appears in the ath time period, and T represents the vector transpose; The popularity of the target topic in the next time interval t is expressed as: Among them, μ tar,a+1 is the popularity of the target topic at time a+1 estimated based on the target topic popularity vector; X i,a+1-m β i is the popularity of the target topic at time a+1 estimated based on the time-lag matrix of potential influencing topics, and the regression coefficient β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T , Assuming there are d potentially influential topics, the popularity vector of the i-th potentially influential topic is recorded as y i,q =(y i,1 ,…,y i,a ) T ,i∈{1,…,d}, where y i,a represents the number of times the i-th potentially influential topic appears in the a-th time period; Given the lag order l to be considered, l<a, the popularity lag vector of the i-th potentially influential topic is recorded as x i,a-m =(x i,m+1,a-m ,…,x i,a,a-m ) T ,i∈{1,…,d},m∈{1,…,l-1}, where x i,m+1,a-m represents the m-th order lag term of the i-th potentially influential topic in the m+1-th time period.

2. The method for short-term prediction of topic popularity considering competition and cooperation as claimed in claim 1, characterized in that: Extract topics from a document collection using regular expressions.

3. The method for short-term prediction of topic popularity considering competition and cooperation as claimed in claim 1, characterized in that: The method for calculating the similarity between a non-target topic and a target topic includes the following steps: For each extracted topic, traverse the document collection to obtain the document collection containing the topic; For each topic, the texts in the document collection are connected end to end to form the topic text; Segment each topic text and remove stop words to obtain the topic word set for each topic; Take all topics and topic word sets as input to train the Doc2vec model; Based on the trained Doc2vec model, we obtain the topic vectors of the target topic and each non-target topic. For each non-target topic, the Euclidean distance between its topic vector and the target topic vector is calculated as the similarity between the non-target topic and the target topic.

4. The method for short-term prediction of topic popularity considering competition and cooperation as claimed in claim 1, characterized in that: For a document collection, the time period is divided backwards, starting from the time point with the target topic and ending at the deadline of the document collection. In the chronological order, the time interval of the first time period is ≤ t.

5. The method for short-term prediction of topic popularity considering competition and cooperation as claimed in claim 1, characterized in that: use Indicates the potential impact of topics on the target topic h tar The influence of popularity, i.e. but The likelihood function can be expressed as follows: Where i∈{1,…,d}; means The jth element in , j∈{3,…,a}, is the influence part at the jth moment; x i,j-1 It's X i,a-m The vector consisting of the elements in the j-1th row; is the j-th row and j-th column element in the diagonal matrix R; Assuming the regression coefficient β i Obeying the nail-board distribution, we get: P(β i ,γ,R β )=P(β i |γ,R β )P(R β |γ)P(γ), where Setting β i To obey the mean of 0, the covariance matrix is ​​R β Normal distribution; K is the number of non-target topics and their lag terms to be introduced into the impact part; π k β i The probability that the kth item is introduced into the model, and when β i The kth term β k ≠0, γ k =1; when β k = 0, γ k =0; By using the MCMC method to obtain β i , based on the estimated β i And the potential impact topic delay matrix X at time a+1 i,a+1-m , using the following formula: Predict the impact part X of the a+1 period i,a+1-m β i .

6. The method for short-term prediction of topic popularity considering competition and cooperation as claimed in claim 5, characterized in that: When β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T The coefficient β in i,a-c When ,i∈{1,…,d},c∈{1,…,m} is positive and significantly different from 0, the potential influencing topic is a cooperative topic of the target topic; When β i =(β 1,a-1 ,…,β 1,a-m ,…,β d,a-1 ,…,β d,a-m ) T The coefficient β in i,a-c When ,i∈{1,…,d},c∈{1,…,m} is negative and significantly different from 0, the potential influencing topic is a competitive topic of the target topic.

Images