JP5848199B2 - Impact prediction device, impact prediction method, and program - Google Patents

Description

  The present invention relates to an apparatus, a method, and a program for evaluating how many people a content transmitted by a person on social media reaches in the future. The user of this apparatus can obtain the result of ranking the content based on the predicted influence of the content calculated by this apparatus.

  Social media is represented by Twitter (registered trademark) services, blogs, etc., and users constitute a social network based on links and follow-up relationships. A user is described in information (written in natural language). When a document, image, video, URL, etc.) is transmitted, the information can be browsed by neighbor users who have a relationship with the user or general users who have visited the user's site.

  As a first conventional technique in the above technical field, there is a technique using the concept of burst (Non-Patent Document 1) proposed by Kleinberg et al. (Non-Patent Document 4). This is to predict how many people the information will be quoted in the future, based on the fact that the interval at which the transmitted information is quoted by other people is shortened. This method is mainly used for extracting keywords that are becoming popular topics in the world. Web documents with time information including the keywords are treated as document streams, and the number of documents in the document stream increases rapidly. It is determined whether or not the keyword is in a burst state by extracting the portion that is present.

  As a conventional second technique, there is a topic extraction method using information written on an electronic bulletin board of Matsumura et al. (Non-patent Document 2). A keyword common to writing on an electronic bulletin board, a reply to it, and a comment is extracted, and a topic that is the center of the topic is extracted based on how many times the keyword is used.

  As a conventional third technique, there is a technique for predicting the future influence of a user based on past behavior on social media (Twitter (registered trademark)) (Non-Patent Document 3). The influence of a user is defined by the number of times information transmitted by the user is cited by another person. Based on how many people follow the user on Twitter (registered trademark), how many people the user is following, and the cumulative number of times the information sent by the user has been cited in the past, For example, the number of times (influence) cited by other users in the next month is predicted.

Jon Kleinberg.Bursty and hierarchical structure in streams.In Proc.the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2002. Masahiro Matsumura, Yukio Osawa, Mitsuru Ishizuka. Dissemination model of influence in text communication. Transactions of the Japanese Society for Artificial Intelligence Vol.17 No.3, pp.259-267,2002. Eytan Bakshy, Jake M. Hofman, Winter A. Mason and Duncan J. Watts. Everyone's an influencer: quantifying influence on twitter. In Proc. The 4th ACM International Conference on Web Search and Data Mining, pp. 65-74, 2011. Toshiaki Fujiki, Yasuyuki Minamino, Yasuhiro Suzuki, Manabu Okumura. Discovery of burst in document stream. IPSJ Research Report, 2003- NL-160, pp. 85-92, 2004. Quinlan, J. R. Induction of decision trees.Machine Learning, 1: 81-106, 1986. Quinlan, J. R. C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers, 1993.

  However, since the first conventional technique is a method that relies heavily on citation information (quoting intervals), most people or no one knows (unquoted) information soon after it was published. The value could not be evaluated. The first conventional technology can extract information that has already been talked about (already cited by many people), but information that is not quoted at this time but is likely to be talked about in the future. Could not be extracted.

  The conventional second technique is also a method for extracting information that has already been discussed, as in the conventional first technique.

  In addition, the third conventional technique predicts the future influence on the user basis, and is not a technique for predicting the value of the information unit transmitted by the user. Simply, a method of propagating the predicted influence estimated for each user to the transmission information can be considered. However, it is rare that a certain user always transmits valuable information, and there is a problem that information such as simple monologues and greetings is unfairly evaluated.

  The present invention has been made in view of the above points, and in the technology for calculating an evaluation value indicating the future influence of content transmitted on social media, even if the results cited by other people have been obtained. It is an object of the present invention to provide a technique that can calculate an evaluation value indicating a future influence with high accuracy even with little or no information.

In order to solve the above-described problem, the present invention is an influence prediction apparatus that calculates an evaluation value indicating a future influence of a content transmitted on social media at the present time,
The content is transmitted from each of a user information storage device storing user information including information related to a link between users on social media, and a citation information storage device storing citation information indicating mutual citation status between contents. The influence which calculates | requires the influence evaluation value which shows the future influence of the said transmission user at the present time by acquiring the feature-value about the transmission user who is a user and using a prediction model based on the said feature-value Prediction means;
Topicity evaluation for obtaining the average number of citations of the content per unit time from the generation of the content transmitted by the transmitting user to the present time as the topicality evaluation value indicating the topicality of the content at the current time Means,
A weight at the present time calculated so that the elapsed time from the time when the content is transmitted to the present time becomes smaller is added to the influence evaluation value, and the weight is calculated so as to increase as the elapsed time becomes larger. And adding a weight to the topic evaluation value, and calculating a comprehensive evaluation value indicating the future influence of the content from the weighted influence evaluation value and the topic evaluation value. Time-weighted combining means to output
It is comprised as an influence prediction apparatus characterized by having.

The influence prediction means includes topic extraction means for extracting a topic of the content transmitted by the transmission user, and the influence prediction means is extracted from the citation information storage device by the topic extraction means. The feature value of the calling user in the topic is acquired, and based on the feature value, the prediction value is used to obtain an evaluation value indicating the future influence of the calling user in the topic at the present time. It may be calculated as a value.

  The influence prediction device includes: a content feature amount extraction unit that extracts a feature amount of the content transmitted by the transmitting user; and a prediction model based on the feature amount of the content extracted by the content feature amount extraction unit. Content influence prediction means for calculating a content influence evaluation value indicating the future influence of the content, and the time weighted combination means from the time when the content is transmitted. The weight at the present time calculated so as to decrease as the elapsed time up to the current time increases, is added to the influence evaluation value and the content influence evaluation value, and is calculated so as to increase as the elapsed time increases. Current weights are added to the topical evaluation values, and these weighted influence evaluation values are added. May be output by calculating an overall evaluation value indicating the future impact of the content from said topicality evaluation value and the content influence evaluation value.

  Moreover, according to this invention, the influence prediction method which the said influence prediction apparatus performs is provided. Furthermore, according to the present invention, there is provided a program for causing a computer to function as each means of the influence prediction apparatus.

  According to the present invention, even if there is little or no track record cited by other people, the future influence estimated from the information sender and the future influence estimated from the characteristics of the contents In addition, the future value can be predicted with high accuracy.

  In particular, according to the present invention, a weight that decreases as the elapsed time from when the content is transmitted to the present time is added to the influence evaluation value, and a weight that increases as the elapsed time increases Since the comprehensive evaluation value is calculated from the weighted influence evaluation value and topicality evaluation value, it is possible to calculate the influence evaluation value with high accuracy.

It is a block diagram of the information propagation prediction apparatus 100 in 1st embodiment of this invention. It is a flowchart of the process which the time weight combination means 108 in 1st embodiment of this invention performs. It is an example of the information stored in the user information storage device 102 in the first embodiment of the present invention. It is an example of the information stored in the content information storage device 103 in the first embodiment of the present invention. It is an example of the information stored in the quotation information storage device 103 in the first embodiment of the present invention. It is an example of the data used for the learning of the user influence prediction means 106 in 1st embodiment of this invention. It is an example of the data explaining the score and elapsed time which the time weight combination means 108 in 1st embodiment of this invention calculates. It is an example of a display by the search part 101 and the output part 109 in 1st embodiment of this invention. It is a block diagram of the information propagation prediction apparatus 200 in 2nd embodiment of this invention. It is an example of the data stored in the content information storage device 104 in 2nd embodiment of this invention. It is a block diagram of the information propagation prediction apparatus 300 in 3rd embodiment of this invention.

Embodiments of the present invention will be described below with reference to the drawings. The embodiment described below is only an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.
[First embodiment]
Hereinafter, a first embodiment of the present apparatus will be described with reference to the drawings.

  FIG. 1 shows a block diagram of an information propagation prediction apparatus 100 in the first embodiment of the present invention. The information propagation prediction device 100 shown in the figure includes a search unit 101, a user information storage device 102, a citation information storage device 103, a content information storage device 104, an evaluation target content extraction unit 105, a user influence prediction unit 106, and topicality evaluation. It comprises means 107, time weighted combining means 108, and output section 109. Among them, the evaluation target content extraction unit 105 is connected to the content information storage device 104 and the citation information storage device 103, and the user influence prediction unit 106 is the content information storage device 104, the citation information storage device 103, and the user information storage device 102. The topicality evaluation unit 107 is connected to the content information storage device 104 and the citation information storage device 103. The information propagation prediction device may be referred to as an influence prediction device.

  In the example illustrated in FIG. 1, the information propagation prediction device 100 includes a search unit 101, a user information storage device 102, a citation information storage device 103, a content information storage device 104, an evaluation target content extraction unit 105, and a user influence prediction unit. 106, topicality evaluation means 107, time weighted combining means 108, and output section 109, but the configuration of the information propagation prediction apparatus 100 is not limited to this. For example, the information propagation prediction apparatus 100 includes a search unit 101, an evaluation target content extraction unit 105, a user influence prediction unit 106, a topicality evaluation unit 107, a time weighted combination unit 108, and an output unit 109. The storage device 102, the citation information storage device 103, and the content information storage device 104 may be provided outside the information propagation prediction device 100 and connected to the information propagation prediction device 100 via a network. The configuration in which the user information storage device 102, the citation information storage device 103, and the content information storage device 104 are provided outside the information propagation prediction device can be similarly adopted in other embodiments.

  The information propagation prediction device 100 in a form connected to the user information storage device 102, the citation information storage device 103, and the content information storage device 104 via a network includes, for example, a computer having a CPU, a memory, and the like. This can be realized by executing a program corresponding to the processing operation of the means. Further, for the information propagation prediction apparatus 100 having the user information storage device 102, the citation information storage device 103, and the content information storage device 104 therein, for example, the user information storage device 102, the citation information storage device 103, and the content This can be realized by causing a computer having a database system corresponding to the information storage device 104 to execute a program corresponding to the processing operation of each processing means of the embodiment. The above program may be stored and distributed in a storage medium such as a portable memory, installed in a computer, or downloaded from a server on a network and installed in the computer. The point which can implement | achieve an information propagation prediction apparatus by making a computer run a program is the same also in other embodiment. Hereinafter, each component will be described in detail.

<Each storage device>
The user information storage device 102 stores user information that can be analyzed by the information propagation prediction device 100, reads user information according to a request from a predetermined functional unit of the information propagation prediction device 100, and requests the information. Send to the original.

  Each typical user information stored in the user information storage device 102 includes a user ID for uniquely identifying a user, attribute information such as the number of links, the number of linked links, the number of outgoing information, a self-introduction sentence, and a time stamp. It consists of. A link is data indicating interest from one user to another user. For example, when Twitter (registered trademark) is targeted, the number of links is the number of followers to other users, the number of linked links is the number of followers from other users, and the number of outgoing information is the total number of tweets in the past. The time stamp is an attribute value representing the time when each piece of user information is acquired. That is, it is possible to follow the transition of user information related to a certain user (number of links, number of links, number of transmitted information, etc.) in time series. The user information storage device 102 is configured as, for example, a Web server that holds Web pages, a database server that includes a database, a storage device that includes a database, and the like.

  The content information storage device 104 stores content information that can be analyzed by the information propagation prediction device 100. The content information storage device 104 reads the content information in response to a request from a predetermined functional unit of the information propagation prediction device 100, and requests the information. Send to the original.

  The content may be any information such as text, still image, video, music, and audio that can be viewed on a medium such as a website, CD-ROM, or DVD. Typical content information stored includes a content ID for uniquely identifying the content, a user ID that generated the content, created date information (may be interpreted as date information that the content was transmitted), text, Consists of still images, moving images, music, audio information and the like. The content information storage device 104 is configured, for example, as a Web server that holds Web pages, a database server that includes a database, a storage device that includes a database, and the like.

  The citation information storage device 103 stores citation information that can be analyzed by the information propagation prediction device 100, reads citation information according to a request from a predetermined functional means of the information propagation prediction device 100, and requests the information. Send to the original.

  Cited information is data indicating that transmission or propagation of certain information from person to person has occurred. Specifically, a certain content, or part of it, is transmitted by another user (content sent by). It is information indicating that it is mentioned or cited. For example, if the URL of a web page A is included in a web page B created by another user, B is treated as quoting A. When Twitter (registered trademark) is targeted, if a tweet A is retweeted by another user's tweet B, B is treated as quoting A. In addition, when a part of the content, for example, a URL or a sentence included in the content is quoted, it may be treated as quoted information. The quotation information is a concept including an operation of adding a certain content to a bookmark. The citation information storage device 103 is configured as, for example, a Web server that holds a Web page, a database server that includes a database, a storage device that includes a database, and the like.

<Search unit 101>
The search unit 101 receives a search condition for narrowing down a content set to be evaluated from the user. The search condition is a condition related to attribute values held by the content information storage device 104. For example, when date information is used as a condition, it is possible to specify a condition such as “within several hours after creation”. In addition, when each content has a text attribute, it is possible to perform a narrowing operation using a search condition including a certain keyword. It is also possible to narrow down by a combination of a plurality of attributes, for example, date information and keywords. Inquiries using information stored in the quote information storage device 103 can also be made. For example, a condition such as “5 or more cited contents” can be designated. The search unit 101 includes, for example, a keyboard, OCR, pen input, touch panel input, voice recognition device, means for reading a text file placed on the network, and the like. The search unit can be realized by a device driver such as a mouse, menu screen control software, or the like.

<Evaluation Target Content Extraction Unit 105>
The evaluation target content extraction unit 105 makes an inquiry to the content information search device 104 and the citation information storage device 103 based on the search condition specified by the search unit 101, and acquires a set of content IDs that match the search condition. .

<User influence predicting means 106>
The user influence predicting means 106 acquires the user set that has transmitted these contents based on the content set that is the output of the evaluation target content extracting means 105, and determines the future influence of each user included in the user set. It is a means of prediction. First, the user influence prediction means 106 makes an inquiry to the user information storage device 102 based on the content ID set, and acquires the user set that has transmitted these contents. The following operations are repeated for each acquired user.

  First, the feature amount of the user is extracted. The feature amount is the latest number of links, the number of links, the number of transmitted information, etc. stored in the user information storage device 102. In addition, the citation information storage device 103 is also inquired, and the total number of times the content generated by each user has been cited by other users in the past, the content generated by each user in the past month is sent to other users. A statistic indicating past performance such as the total number of times quoted and the average number of times cited by other users when each user makes one post is calculated and used as a feature value for each user. Finally, based on these past feature amounts, the number of times each user is cited in a future fixed period is predicted. A model that predicts the number of times quoted in a future fixed period using the past feature quantity as an input and outputs the model can be configured using an existing prediction model in the field of machine learning such as a regression tree. The regression tree is a model that predicts the real value of the objective variable based on the value of the explanatory variable. That is, in this apparatus, the number of times a user is cited (objective variable) in a future fixed period is predicted based on the user feature value (explanatory variable) known so far. The fixed period is, for example, one day, one week, one month, etc., but in which range the number of times cited is predicted depends on the learning method of the prediction model. The predicted user influence value (in this example, the number of times the user is cited in a certain period in the future) is sent to the time weighted combining means 108. Moreover, it is good also as the estimated frequency | count of quotation per content (user and) instead of per user. For example, a value obtained by calculating the number of citations per user expected in one week and dividing the value by the number of contents that the user transmits on average in one week may be adopted.

  Hereinafter, a regression tree that is one of the prediction models will be described.

  The regression tree is a decision tree when the objective attribute is a numerical type. In the regression tree, a division test for dividing data into each intermediate node of the tree is shown. Each intermediate node also has two downward branches for the two possible values that result from applying the split test. On the other hand, each terminal node has “the number of times a user is cited in a certain period in the future”, which is noticed here. Each user in the data used for learning is classified into one of the terminal nodes using this regression tree. The vertex nodes of the regression tree correspond to the entire data set. A regression tree can easily express knowledge and laws obtained from data as IF-THEN rules for split tests from the top node to the end of the tree. The label of the terminal node corresponds to the IF-THEN rule conclusion part. For example, a user who has “the number of links is 100 or more” and “the total number of times quoted in the past is 100 times” can make a prediction that “it will be quoted 10 times within a week in the future”.

  Hereinafter, creation of learning data for learning a regression tree, which is a prediction model in this apparatus, and a learning method will be described. In addition, this Embodiment is an example of a structure of the apparatus for demonstrating a prediction phase, and the following learning phases including learning data preparation are performed as a preliminary process.

  The regression tree data obtained by this pre-processing is stored, for example, in storage means such as a memory in the user influence prediction means 106 and used for user influence prediction. Further, this pre-processing itself may be performed by the user influence prediction unit 106 or may be performed by another unit.

  First, a past time t for dividing the data is determined. Next, a user set stored in the user information storage device 102 is acquired. Next, the user information storage device 102 and the citation information storage device 103 are inquired, and for each user in the user set, the feature amount of each user is calculated based on information before time t, To do. That is, the number of links, the number of links, the number of outgoing information, and the total number of times content generated by each user before time t is cited by other users. At the same time, based on the information after time t, the total number of times each user has been cited by another person is calculated and used as an objective variable. Data obtained by converting each user in the user set into information including a combination of explanatory variables and objective variables is used as learning data. The above is the creation method of learning data.

  Next, a prediction model and a regression tree are created based on the learning data. In the regression tree, the regression tree is learned by performing a division test that minimizes the mean square error or maximizes the evaluation function based on the interclass variance. A simple method is to divide the data from the top of the tree recursively by an optimal division test based on the mutual information. ID3 uses an optimal division test in terms of mutual information (Non-Patent Document 5), and C4.5 uses a division test that optimizes the gain ratio (Non-Patent Document 6). A regression tree may be created using an algorithm. In addition, the REPTree function in the free software WEKA is an algorithm that performs division from the viewpoint of information gain and distribution, and has the advantage of faster calculation than C4.5. It may be used.

  The regression tree created by the above learning method is a model that predicts the number of times a user is cited in a certain period in the future based on the user's characteristics and influence until a certain point in time. The time range in which the regression tree predicts the number of citations is determined depending on how much time range after time t is taken into consideration. For example, when the number of citations in one week after time t is considered, the regression tree is a model for predicting the number of citations expected in one week. Alternatively, the number of predicted citations per content, not per user, may be used. That is, the number of citations expected in one week may be divided by the number of contents that the user transmits on average in one week.

<Topicity evaluation means 107>
The topicality evaluation means 107 is based on the citation information storage device 103 and the content information storage device 104 and how much each content included in the content set that is the output of the evaluation target content extraction means 105 becomes a topic at the current time. It is a means to evaluate whether or not. The topicality evaluation unit 107 makes an inquiry to the content information storage device 104 and acquires the date on which each content was generated. Next, the citation information storage device 103 is inquired, and the topicality of the content is evaluated based on the citation information observed between the date when each content was generated and the current time. As an example of the topicality evaluation method, a method in which the average number of citations per unit time from the date when the content is generated to the current time is used as the topicality evaluation value can be considered. The current time is t, the content to be evaluated is c, Δt is the time elapsed since the content was generated until the current time, and n (c, t) is the current time t. Assuming the total number, the topicality evaluation value h (c, t) is as follows.
h (c, t) = n (c, t) / Δt
In addition to this, any existing method for evaluating topicality may be used, such as a method using a burst of Non-Patent Document 1 in which a difference from a steady state is observed. The scale of the topicality evaluation value is aligned with the influence predicted by the user influence prediction means 106. For example, when the user influence prediction means 106 predicts and outputs the expected number of citations per week, the topicality evaluation value is also output as the average number of citations with a unit time of one week. The topicality evaluation value evaluated for each content is sent to the time weighted combining means 108.

<Time Weighted Combiner 108>
The time weighted combining unit 108 is a unit that combines the user influence prediction value calculated by the user influence prediction unit 106 and the topicality evaluation value calculated by the topicality evaluation unit 107 with time weighting. This means that when a certain content is created immediately, the user influence prediction value is emphasized as an evaluation value, and a topic importance evaluation value is emphasized as time passes after the creation of a certain content. Realized. For example, a method of combining by the following method is conceivable.

  Assume that it is desired to calculate the evaluation value f (c, u, t) of the current time t of the content c generated by the user u. Further, g (u, t) is a function that returns a user influence prediction value at the current time t calculated by the user influence prediction means 106, and h (c) is a function that returns a topicality evaluation value calculated by the topicality evaluation means 107. , t), the coupling equation is as follows.

上記の数式において、Δtはコンテンツが生成されてから現在時刻までに経過した時間、Tは時間減衰速度を決定するパラメータである。ｅ^{−Δｔ／Ｔ}は減衰関数であり時間の経過に従って0に近づく。また、（1−ｅ^{−Δｔ／Ｔ}）は時間の経過に従って１に近づく。ｅ^{−Δｔ／Ｔ}は減衰関数の一例であり、他の減衰関数を用いても良い。なお、時間の概念は、順序関係を表す時間インデクスをも含む概念であり、時間の経過を時間インデクスに基づく距離として時間重み付き結合手段１０８を実現しても良い。算出した値は、出力部１０９へと送られる。f(c,u,t)を計算するためのフローチャートを図２に示す。図２の手順に沿って、時間重み付き結合手段１０８が実行するf(c,u,t)を計算する処理を説明する。なお、以下の処理動作は一例である。また、他の実施の形態における時間重み付き結合手段の処理の流れも基本的に同様である。

In the above equation, Δt is a time elapsed from the generation of the content to the current time, and T is a parameter for determining a time decay rate. e ^{−Δt / T} is an attenuation function and approaches 0 as time passes. Further, (1−e− ^{Δt / T} ) approaches 1 as time passes. e− ^{Δt / T} is an example of an attenuation function, and other attenuation functions may be used. Note that the concept of time is a concept including a time index representing an order relationship, and the time-weighted combining unit 108 may be realized with the passage of time as a distance based on the time index. The calculated value is sent to the output unit 109. A flow chart for calculating f (c, u, t) is shown in FIG. A process of calculating f (c, u, t) executed by the time weighted combining unit 108 will be described along the procedure of FIG. The following processing operation is an example. The processing flow of the time-weighted combining means in the other embodiments is basically the same.

  First, the content set C, the attenuation parameter T, and the current time t are acquired and input (step 1). The content set C may be received from the topicality evaluation unit 107 or may be acquired from the evaluation target content extraction unit 105. The attenuation parameter T is set in advance and stored in a memory or the like. The current time t may be acquired from the clock of the device or may be acquired from the network. Then, the following processing is performed for each content c in the content set C (loop from S2 to S13).

  First, temporary variables x and y are initialized (step 3). Subsequently, the topicality evaluation value h (c, t) calculated by the topicality evaluation means 107 is substituted for x (step 4). Also, the user u who created the content c is acquired (step 5). Information about the user u who created the content c may be received from the user influence prediction unit 106 or may be acquired from the content information storage device 104.

  Next, it is checked whether or not the user u has been analyzed (step 6). If it has not been analyzed, the process proceeds to step 7, and if it has been analyzed, the process proceeds to step 9.

  In step 7 when the user u has not been analyzed, the user influence prediction means 106 calculates the user influence prediction value g (u, t), and the calculated user influence prediction value g (u, t) is represented by y. Assign to. Then, the combination of u and y is stored in a storage medium (such as a memory provided in the time weighted combining means 108) (step 8).

  In step 9 in the case where the user u has been analyzed, the influence predicted value of the user u in the storage medium is substituted for y.

  Subsequently, the time tc when the content c is created is acquired (step 10). The time tc when the content c is created may be received from the topicality evaluation unit 107 or may be acquired from the content information storage device 104.

  Next, t−tc is substituted for Δt (step 11), and the following calculation is performed to obtain f (c, u, t) (step 12).

＜出力部１０９＞
出力部１０９は、時間重み付き結合手段１０８の処理結果に基づき、検索部１０１で指定された条件を満たすコンテンツ集合を並び替えて表示出力するための機能部である。ここで、出力とは、ディスプレイへの表示、プリンタへの印字、音出力、外部装置への送信等を含む概念である。出力部１０９は、例えば、ディスプレイやスピーカ等の出力デバイスと出力デバイスのドライバソフトにより実現される。また、出力部１０９を、ディスプレイやスピーカ等の出力デバイスを含まない機能部としてもよい。この場合、例えば、装置外部に出力デバイスが存在し、出力部１０９は、出力デバイスのドライバソフトとして構成される。

<Output unit 109>
The output unit 109 is a functional unit for rearranging and displaying the content sets satisfying the conditions specified by the search unit 101 based on the processing result of the time weighted combining unit 108. Here, output is a concept including display on a display, printing on a printer, sound output, transmission to an external device, and the like. The output unit 109 is realized by, for example, an output device such as a display or a speaker and driver software for the output device. The output unit 109 may be a functional unit that does not include an output device such as a display or a speaker. In this case, for example, an output device exists outside the apparatus, and the output unit 109 is configured as output device driver software.

(Example)
Hereinafter, examples in the first embodiment will be described.

  3, 4, and 5 show examples of information stored in the user information storage device 102, the content information recording device 103, and the citation information storage device 104 in this embodiment.

  Regarding the information stored in the user information storage device 102 (FIG. 3), as described above, the transition of each user information can be followed in time series. The concept of link in FIG. 3 represents an interest from the user to the user, while the concept of quoting in FIG. 5 represents a concept of interest from the user (content created by the user) to the content. The content c400 of FIG. 4 is stored in the citation information of FIG. 5 because the information of the content c500 is cited in the text.

<Example of learning data for regression tree>
First, the learning data creation method of the regression tree of the user influence predicting means 106 will be described with reference to the examples of FIGS. Now, the date to divide the data is 2012/1/15 00:00:00. FIG. 6 shows learning data created for the user whose user ID is u1. The latest user information (number of links, number of links, number of outgoing information, etc.) viewed from the split date is acquired from the user information shown in FIG. 3, and at the same time, the content sent by the user u1 at the split date is sent to other users. The total number quoted is also calculated based on FIG. 4 (content information) and FIG. 5 (citation information). These are explanatory variables. Further, the total number of contents sent by the user u1 after the division date is quoted by other users is calculated and set as an objective variable. Since this is repeated for all users, the number of rows in the table of FIG. 6 is equal to the total number of users. When the scale of each variable is greatly different, the common logarithm and natural logarithm of the variable value may be calculated and used as a new variable value with the same scale.

  The date for dividing the data is set by combining the output of the topicality evaluation means 107 and the time scale. For example, when the topicality evaluation unit 107 calculates a topicality evaluation value per week, the regression tree also learns to output a predicted value of the number of times the user is cited per week. For example, data observed from one week ago to the present time may be used as an objective variable, and data observed during a period before that may be used as an explanatory variable. Thereby, the regression tree becomes a model for predicting the number of times the user is quoted from a certain point in time to one week later.

<Example of operation of topicality evaluation means 107>
Next, the operation of the topicality evaluation unit 107 will be described using an example. Assume that there are two contents A and B that were sent on January 1, 2012 at 10:00:00. As of January 1, 2012, 12:00:00, the number of times content A was quoted by another user was 100, and the number of times content B was quoted by another user was 10. In this case, when the number of citations per hour is calculated, the content A is 50 times per hour, the content B is 5 times per hour, and the topicality evaluation means 107 determines that the topicality of the content A is higher than that of the content B. .

<Operation Example of Time Weighted Combining Unit 108>
Next, the operation of the time weighted combining unit 108 will be described using an example. Now, assume that there is content A sent by user X on January 1, 2012 at 10:00:00 and content B sent by user Y at the same time. The attenuation parameter T is 6 hours, and the current time is January 1, 2012, 16:00:00. At the time of the current time, the topicality of the content A calculated by the topicality evaluation means 107 is 15 times per hour, and the topicality of the content B is also 100 times per hour. On the other hand, the predicted influence of the user X evaluated by the user influence prediction means 106 is 10 times per hour, and the predicted influence of the user Y is 100 times per hour. Here, it is assumed that the regression tree of the user influence prediction means 106 has been learned so as to predict the number of citations per user one hour after the current time. That is, the scales of the outputs of the topic evaluation means 107 and the user influence prediction means 106 are equal. The time-weighted combining means 108 performs the prediction based on the influence of the user at the time when the content is transmitted (January 1, 2012 10:00:00). Evaluate the value as 100. In this way, even information that is hardly quoted can be finally discovered quickly by making a prediction based on the influence of the caller. Times every hour).

  Moreover, the passage of time and the change in the predicted value will be described using actual examples. Now, assume that there is content A sent by user X on January 1, 2012 at 10:00:00 and content B sent by user Y at the same time. The attenuation parameter T is 6 hours. Further, after the content is transmitted, it is assumed that the topicality of the content A calculated by the topicality evaluation unit 107 is always 50 times and the topicality of the content B is always 5 times. On the other hand, the predicted influence of the user X evaluated at the time when the user influence prediction means 106 transmits the content was 5 times per hour, and the predicted influence of the user Y was 50 times per hour. Here, it is assumed that the regression tree of the user influence prediction means 106 has been learned so as to predict the number of citations per user one hour after the current time. The time-weighted combining means 108 performs the prediction based on the influence of the user at the time when the content is transmitted (January 1, 2012 10:00:00). Evaluate the value as 50. As time elapses, a value according to the result of topicality evaluation means 107 based on the number of citations is output. After 6 hours, the value of content A is evaluated as 33.4454, and the value of content B is evaluated as 21.5546. The transition of the elapsed time and score in this example is shown in FIG.

<Display Examples of Search Unit 101 and Output Unit 109>
An example of information displayed by the search unit 101 and information displayed by the output unit 109 in this embodiment is shown in FIG. The display format in the search unit 101 may be changed according to the format of the content information storage device 104. The output unit 109 can display the rankings according to the value of the time weighted combining unit 108 (for example, the higher the evaluation value, the higher the display). Further, it is possible to simultaneously display the related information of the content stored in the content information storage device 104.

[Second Embodiment]
Hereinafter, a second embodiment of the present apparatus will be described with reference to the drawings.

  FIG. 9 is a block diagram of an information propagation prediction apparatus 200 in the second embodiment of the present invention. The information propagation prediction apparatus 200 shown in the figure includes a search unit 101, a content information storage device 104, a citation information storage device 103, a user information storage device 102, an evaluation target content extraction unit 105, a topic evaluation unit 107, and a topic analysis unit 201. , The user influence predicting unit 202 with topic condition, and the combining unit 203 with time weight. Among these, the search unit 101, the content information storage device 104, the citation information storage device 103, the user information storage device 102, the evaluation target content extraction unit 105, and the topicality evaluation unit 107 are the same as in the first embodiment. Reference numerals are attached. In addition, it is good also as a structure provided with the topic analysis means 201 in the user influence prediction means 202 with a topic condition.

<Topic analysis means 201>
The topic analysis unit 201 determines what topic the content is transmitted from, for example, “politics”, “economics”, “sports”, etc., from each content of the content set that is the output of the evaluation target content extraction unit 105. Extract. Topic analysis can be realized using any existing topic analyzer such as SVM that classifies content into topics with supervising, PLSA or LDA that classifies content into topics without supervision. The topic information extracted from the content is transmitted to the topic conditional user influence predicting means 202.

<Topic Conditional User Impact Prediction Unit 202>
The user condition predicting means 202 with topic condition is a means for predicting the future influence on the topic to which the content belongs by the user who transmitted each content that is the output of the evaluation target content extracting means 105. The user influence prediction unit with topic condition 202 repeats the following operation for each content in the content set that is the output of the evaluation target content extraction unit 105.

  The user influence prediction means 202 with topic condition makes an inquiry to the user information storage device 102 and acquires the user who transmitted the content. Next, the user feature amount independent of the topic of the user is extracted. The topic-independent feature amounts are the latest number of links, the number of links, the number of transmitted information, etc. stored in the user information storage device 102. Next, a topic-dependent user feature quantity indicating the past performance of the user in the topic to which the content belongs is extracted. Specifically, the content information storage device 104 and the citation information storage device 103 are inquired, and the total number of times the content generated by each user regarding the topic is cited by other users in the past, the past month Statistics such as the total number of times the content generated by each user regarding the topic was quoted by other users, and the average number of times each user cited one post regarding the topic Is calculated as a topic-dependent feature quantity for each user.

  When each piece of content information in the content information storage device 104 already has a topic attribute (for example, in the case of data as shown in FIG. 10), a content set transmitted by each user regarding the topic is acquired, and the content set From the above, the above processing can be realized by counting the number of citations. If each piece of content information in the content information storage device 104 does not have a topic attribute, a step of adding topic information of each piece of content to each piece of content using the topic analysis unit 201 is added.

  Finally, based on these past feature quantities, the total number quoted for a certain topic for each user over a certain period in the future is predicted. Similar to the first embodiment, the past feature quantity is input, the total number quoted in a fixed period in the future is predicted, and the output model is an existing prediction model in the machine learning field such as a regression tree. Can be configured. The regression tree is a model that predicts the real value of the objective variable based on the value of the explanatory variable. That is, in this apparatus, the number of times a user is cited (objective variable) in a future fixed period is predicted based on the user feature value (explanatory variable) known so far. Similar to the user influence prediction unit 106 of the first embodiment, the number of citations in which period is predicted depends on the learning method of the regression tree, but needs to be set in accordance with the time scale of the topic extraction unit 107. There is. The topic condition-based user influence prediction value is sent to the time-weighted combining unit 203. Moreover, it is good also as the estimated frequency | count of quotation per content (user and) instead of per user. For example, a value obtained by calculating the number of citations per user expected in one week and dividing the value by the number of contents that the user transmits on average in one week may be adopted.

  Note that, as in the first embodiment, the learning phase of a prediction model such as a regression tree is performed as a preliminary process of the prediction phase. The generation of the prediction model may be performed using all data, or may be performed using data for each topic.

<Time Weighted Combiner 203>
The time weighted combining unit 203 in the second embodiment uses the topic condition user influence calculated by the topic condition user influence prediction unit 202 and the topicality evaluation value calculated by the topicality evaluation unit 107 as a time weight. It is a means of joining. As in the first embodiment, this means that, when a certain content is created, the user influence predicted value with a topic condition is regarded as important, and the topic becomes so long that a certain content has been created. This is realized by a function that places importance on the sex evaluation value. For example, a method of combining by the following method is conceivable. Assume that it is desired to calculate the evaluation value f (c, u, t) of the current time t of the content c generated by the user u. As a result of the topic analysis means 201, the topic to which the content c belongs is z _c . G (u, t, z) is a function that returns a topic conditional user influence prediction value at topic z at the current time t calculated by the topic conditional user influence prediction means 202, and the topicality calculated by the topicality evaluation means 107. Let h (c, t) be a function that returns an evaluation value. In this case, the coupling formula is as follows.

上記の数式におけるΔtはコンテンツが生成されてから現在時刻まで経過した時間、Tは時間減衰速度を決定するパラメータである。ｅ^{−Δｔ／Ｔ}は減衰関数であり時間の経過に従って0に近づく。なお、時間の概念は、順序関係を表す時間インデクスをも含む概念であり、時間の経過を時間インデクスに基づく距離として時間重み付き結合手段２０３を実現しても良い。算出した値は、出力部１０９へと送られる。

In the above equation, Δt is a time elapsed from the generation of the content to the current time, and T is a parameter for determining a time decay rate. e ^{−Δt / T} is an attenuation function and approaches 0 as time passes. Note that the concept of time is a concept that also includes a time index that represents an order relationship, and the time-weighted combining unit 203 may be realized with the passage of time as a distance based on the time index. The calculated value is sent to the output unit 109.

  The feature of the second embodiment compared to the first embodiment is that the user influence for each topic is taken into consideration. As a result, when an information sender sends content that is the same as or close to a topic that has been active in the past (politics if it is a politician), the content is judged to have a high value, and the topic has not been active in the past. When the same or similar content is transmitted, it is determined that the value of the content is low.

[Third embodiment]
Hereinafter, a third embodiment of the present apparatus will be described with reference to the drawings.

  FIG. 11 is a block diagram of an information propagation prediction apparatus 300 in the third embodiment of the present invention. The information propagation prediction apparatus 300 shown in the figure includes a search unit 101, a content information storage device 104, a citation information storage device 103, a user information storage device 102, an evaluation target content extraction unit 105, a topic evaluation unit 107, and a topic analysis unit 201. , A topic condition-based user influence prediction means 202, a content feature extraction means 301, a content feature influence influence prediction means 302, and a time weighted combination means 303. Among these, the search unit 101, the content information storage device 104, the citation information storage device 103, the user information storage device 102, the evaluation target content extraction unit 105, the topicality evaluation unit 107, the topic analysis unit 201, and topic influence user influence prediction The means 202 is the same as in the second embodiment.

<Content Feature Extraction Unit 301>
The content feature extraction unit 301 extracts the feature amount of each content for each content of the content set that is the output of the evaluation target content extraction unit 105. The characteristics of the content itself are the date of creation, if the content is text, statistics such as the document length obtained by analyzing the text, and if it is an image, image features such as color features and SIFT-BoW features Information such as the amount and whether the URL is included in the document. When using specific data such as Twitter (registered trademark) data, information specific to the service, such as whether or not it is a symbol indicating a reply from the user to the user, may be used. An inquiry is made to the content information storage device 104 based on the set of content IDs, information about these contents is acquired, and basic statistics are extracted. In addition, the topic analysis unit 201 may be used to extract information about what topic, such as “political economy” or “sports”, and may be used as a content feature. The extracted content feature is transmitted to the influence predicting means 302 based on the content feature amount.

<Influence Predicting Means 302 Based on Content Feature Amount>
The influence predicting means 302 based on the content feature amount is based on the content feature amount of each content itself obtained by the content feature extraction means 301 for each content of the content set that is the output of the evaluation target content extraction means 105. In addition, it is a means for predicting how much the content will be cited in the future. The model that predicts the total number quoted in a certain period in the future with the feature amount of the content as input, and uses the existing prediction model in the machine learning field such as regression tree as the user influence prediction means 106 and the like. Can be configured. The predicted value of the content influence is sent to the time weighted combining unit 303.

  Hereinafter, creation of learning data for regression tree learning in the present embodiment will be described. As described above, the learning phase including learning data creation is performed as pre-processing, and the regression tree data obtained by the pre-processing is stored in a storage unit such as a memory and referred to in the prediction phase.

  First, a content set stored in the content information storage device 104 is acquired.

  Next, using the content feature extraction unit 301, the features of the content itself regarding each content are extracted and used as explanatory variables. At the same time, an inquiry is made to the citation information storage device 103, and the number of times that the content is quoted within a certain period from the time when the content is transmitted is calculated as an objective variable. After each content in the content set is converted into information consisting of a combination of explanatory variables and objective variables, a regression tree is created based on all content information. The learning method of the regression tree is the same as that described in the user influence prediction unit 106 of the first embodiment. The regression tree created by the above learning method is a model that predicts the number of times the content is cited in a certain period in the future after a certain content is transmitted. When the objective variable is the number of citations one week after the content is transmitted, the created regression tree becomes a model for predicting the number of citations of the content one week after the content is transmitted. It should be noted that parameters such as “one week” are set in accordance with the outputs of the user influence prediction means 202 and the topicality evaluation means 107. For example, when the user influence predicting means 202 is a model that predicts the number of citations for “one month”, the influence predicting means 302 based on the content feature amount is also quoted “one month” after the content is transmitted. Learning is a model that predicts the number of times that is performed. That is, the objective variable is the number of citations one month after the content is transmitted.

<Time weighted combining means 303>
The time-weighted combining unit 303 according to the present embodiment includes the topic condition user influence calculated by the topic condition user influence prediction unit 202, the topicality evaluation value calculated by the topicality evaluation unit 107, and the content feature amount. This is a means for combining the predicted influence by the content feature amount calculated by the influence prediction means 302 with time weighting. As soon as a certain content is created, the means attaches importance to the predicted user influence value with topic condition and the predicted influence based on the content feature amount as an evaluation value. This is realized by a function that places importance on the topical evaluation value. For example, a method of combining by the following method is conceivable.

Assume that it is desired to calculate the evaluation value f (c, u, t) of the current time t of the content c generated by the user u. Results of the topic analysis unit 201, the topic of the content c was z _c. Also, g (u, t, z) is a function that returns a topic conditional user influence prediction value for topic z at the current time t calculated by the topic conditional user influence prediction means 202, and the topicality evaluation means 107 calculates. If the function that returns the topicality evaluation value is h (c, t) and the function that returns the predicted influence by the content feature amount calculated by the content feature amount influence prediction unit 302 is i (c), for example, It becomes as follows.

上記の数式において、Δtはコンテンツが生成されてから現在時刻まで経過した時間、Tは時間減衰速度を決定するパラメータである。ｅ^{−Δｔ／Ｔ}は減衰関数であり時間の経過に従って0に近づく。なお、関数g、関数i、関数hの出力のスケールは等しい。つまり、同じ時間区間での引用予測値を出力する。また、以下のような結合式で時間重み付き結合手段３０３を実現しても良い。

In the above equation, Δt is a time elapsed from the generation of content to the current time, and T is a parameter for determining a time decay rate. e ^{−Δt / T} is an attenuation function and approaches 0 as time passes. Note that the scales of the outputs of the function g, the function i, and the function h are equal. That is, the quote prediction value in the same time interval is output. Further, the time weighted combining means 303 may be realized by the following combining formula.

ただし、この式で求める場合、関数g、関数i、関数hが0から1の値を出力するように、正規化を施す必要がある。なお、時間の概念は、順序関係を表す時間インデクスをも含む概念であり、時間の経過を時間インデクスに基づく距離として時間重み付き結合手段３０３を実現しても良い。算出した値は、出力部１０９へと送られる。

However, when this formula is used, it is necessary to normalize so that the function g, function i, and function h output values from 0 to 1. Note that the concept of time is a concept that also includes a time index representing an order relationship, and the time-weighted combining unit 303 may be realized with the passage of time as a distance based on the time index. The calculated value is sent to the output unit 109.

  The feature of the third embodiment is that the feature of the content itself is taken into consideration. As a result, content such as “greeting” and “self-speaking” can be excluded, and tweets with deep contents such as opinions and reviews can be positively evaluated. In the third embodiment, content feature extraction means 301 and influence prediction means 302 based on content features are added to the second embodiment. The content feature extraction unit 301 and the content feature influence prediction unit 302 may be added, and evaluation may be performed in consideration of the feature of the content itself in the same manner as described above.

(Summary of embodiment)
As described above, according to the first embodiment of the present invention, there is an influence prediction device that predicts how many people a content sent by a person on social media will reach in the future. Provided. The influence prediction device includes a user influence prediction unit that predicts a future influence of the information sender based on a user feature amount indicating the past or current performance of the information sender, and transmitted content. Topicity evaluation means for determining the current topicality as a topicality evaluation value,
At a point in time immediately after the content is transmitted, the future influence of the content is predicted based on the future influence of the information sender who has transmitted the content, which is obtained by the user influence prediction means, A time-weighted combining unit that predicts the future influence of the content based on the topical evaluation value of the content obtained by the topical evaluation unit as time elapses from the transmission.

  In addition, according to the second embodiment of the present invention, there is provided an influence prediction device that predicts how many people a content sent by a person on social media will reach in the future. Topicity evaluation means for determining the current topicality of content as a topicality evaluation value, topic analysis means for analyzing what topic the content is transmitted about, and the topic extracted by the topic analysis means A topic-conditional user influence predicting means for predicting the future influence of the information sender on the topic based on a topic-dependent user feature amount indicating the past or present performance of the information sender, and content Information sender who sent the content obtained by the topic conditional user influence predicting means at a short time after the message was sent The future influence of the content is predicted based on the future influence on the topic, and based on the topicality evaluation value of the content obtained by the topicality evaluation means as time elapses after the content is transmitted. There is provided an impact prediction device having time-weighted combining means for predicting the future impact of content.

  In addition, according to the third embodiment of the present invention, there is provided an influence predicting device that predicts how many people a content sent by a person on social media will reach in the future. Topicity evaluation means for determining the current topicality of content as a topicality evaluation value, topic analysis means for analyzing what topic the content is transmitted about, and the topic extracted by the topic analysis means A topic-conditional user influence predicting means for predicting the future influence of the information sender on the topic based on a topic-dependent user feature amount indicating the past or present performance of the information sender, and content Based on content feature extraction means for extracting its own feature quantity and the content feature quantity extracted by the content feature extraction means, The impact prediction means based on the content feature amount that predicts how many people the content will reach in the future, and the topic conditional user impact prediction means at the time immediately after the content is transmitted The future influence of the content is predicted based on the future influence on the topic of the information sender who sent the content, and the influence of the content itself obtained by the influence prediction means based on the content feature amount. A time-weighted combining unit that predicts the future influence of the content based on the topical evaluation value of the content obtained by the topical evaluation unit as time elapses from the transmission. Is provided.

  Each of the influence prediction apparatuses described above may include an evaluation target content extraction unit that specifies conditions for narrowing down the content for which influence is to be predicted. In addition, the user influence prediction means is calculated based on, for example, a statistic obtained based on how many people are interested in the current time, or on how many people the content transmitted in the past has reached. One of the statistics is used as a user feature for predicting the future influence of the information sender.

  In addition, the topicality evaluation means may be, for example, either a statistic obtained based on how many people the transmitted content reaches at the present time, or a statistic obtained based on the time interval to reach the people. Based on the above, the topicality evaluation value is obtained.

  In addition, the topic condition user influence prediction means, for example, the statistics obtained based on how many people are interested in the current time, or the content sent by the information sender in the past regarding the topic, Any of the statistics obtained based on how many people have been reached is used as a user feature for predicting the future influence of the information sender on the topic.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

DESCRIPTION OF SYMBOLS 100 Information propagation prediction apparatus 101 Search part 102 User information storage apparatus 103 Citation information storage apparatus 104 Content information storage apparatus 105 Evaluation object content extraction means 106 User influence prediction means 107 Topicity evaluation means 108 Time weighted combination means 109 Output section 201 Topic analysis means 202 Topic conditional user influence prediction means 203 Time weighted combination means 301 Content feature extraction means 302 Content feature influence prediction means 303 Time weighted combination means

Claims (7)

An impact prediction device that calculates an evaluation value that indicates the current impact of content sent on social media,
The content is transmitted from each of a user information storage device storing user information including information related to a link between users on social media, and a citation information storage device storing citation information indicating mutual citation status between contents. The influence which calculates | requires the influence evaluation value which shows the future influence of the said transmission user at the present time by acquiring the feature-value about the transmission user who is a user and using a prediction model based on the said feature-value Prediction means;
Topicity evaluation for obtaining the average number of citations of the content per unit time from the generation of the content transmitted by the transmitting user to the present time as the topicality evaluation value indicating the topicality of the content at the current time Means,
A weight at the present time calculated so that the elapsed time from the time when the content is transmitted to the present time becomes smaller is added to the influence evaluation value, and the weight is calculated so as to increase as the elapsed time becomes larger. And adding a weight to the topic evaluation value, and calculating a comprehensive evaluation value indicating the future influence of the content from the weighted influence evaluation value and the topic evaluation value. Time-weighted combining means to output
The influence prediction apparatus characterized by having. The influence prediction means includes topic extraction means for extracting a topic of the content transmitted by the transmission user, and the influence prediction means is extracted from the citation information storage device by the topic extraction means. The feature value of the calling user in the topic is acquired, and based on the feature value, the prediction value is used to obtain an evaluation value indicating the future influence of the calling user in the topic at the present time. It calculates as a value. The influence prediction apparatus of Claim 1 characterized by the above-mentioned. The influence prediction device
Content feature amount extraction means for extracting feature amounts of the content transmitted by the transmitting user;
Content impact prediction means for calculating a content impact evaluation value indicating a future impact of the content by using a prediction model based on the content feature amount extracted by the content feature amount extraction means; ,
Further comprising
The time weighted combining means is:
A weight at the present time calculated so as to decrease as the elapsed time from when the content is transmitted to the present time becomes larger is added to the influence evaluation value and the content influence evaluation value, and the elapsed time is increased. The weight at the present time calculated so as to become larger is added to the topical evaluation value, and the content evaluation is performed from the weighted influence evaluation value, the content influence evaluation value, and the topicality evaluation value. The influence evaluation apparatus according to claim 1, wherein a comprehensive evaluation value indicating a future influence of is calculated and output. An impact prediction method executed by an impact prediction device that calculates an evaluation value indicating a future impact of content transmitted on social media at the present time,
The content is transmitted from each of a user information storage device storing user information including information related to a link between users on social media, and a citation information storage device storing citation information indicating mutual citation status between contents. The influence which calculates | requires the influence evaluation value which shows the future influence of the said transmission user at the present time by acquiring the feature-value about the transmission user who is a user and using a prediction model based on the said feature-value A prediction step;
Topicity evaluation for obtaining the average number of citations of the content per unit time from the generation of the content transmitted by the transmitting user to the present time as the topicality evaluation value indicating the topicality of the content at the current time Steps,
A weight at the present time calculated so that the elapsed time from the time when the content is transmitted to the present time becomes smaller is added to the influence evaluation value, and the weight is calculated so as to increase as the elapsed time becomes larger. And adding a weight to the topic evaluation value, and calculating a comprehensive evaluation value indicating the future influence of the content from the weighted influence evaluation value and the topic evaluation value. A time-weighted combination step to output
An influence prediction method characterized by comprising: In the influence prediction step, the influence prediction apparatus extracts the topic of the content transmitted by the transmission user, acquires the feature amount of the transmission user in the topic from the citation information storage device , and the characteristic The evaluation value indicating the future influence of the calling user at the present time on the topic is calculated as the influence evaluation value by using a prediction model based on the amount. Impact prediction method. The influence prediction method is:
A content feature amount extracting step of extracting a feature amount of the content transmitted by the transmitting user;
A content influence prediction step of calculating a content influence evaluation value indicating a future influence of the content by using a prediction model based on the feature quantity of the content extracted in the content feature amount extraction step; ,
Further comprising
In the time-weighted combining step, the influence prediction device includes:
A weight at the present time calculated so as to decrease as the elapsed time from when the content is transmitted to the present time becomes larger is added to the influence evaluation value and the content influence evaluation value, and the elapsed time is increased. The weight at the present time calculated so as to become larger is added to the topical evaluation value, and the content evaluation is performed from the weighted influence evaluation value, the content influence evaluation value, and the topicality evaluation value. An impact prediction method according to claim 4 or 5, characterized in that a comprehensive evaluation value indicating future impact of is calculated and output.   The program for functioning a computer as each means of the influence prediction apparatus of any one of Claims 1 thru | or 3.

Images