JP2016006583A - Method and system for classifying noise post in social network service - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a system for classifying an unknown SNS client and extracting an SNS client which outputs unnecessary contributed data in a social network service.SOLUTION: Based on contributed data in an SNS, a classification model for classifying an SNS client, which outputs the contributed data, by purpose is generated, and an unknown SNS client is classified. By learning the classification model, it is adaptable to a new type of SNS client appearing in the future.

Description

  The present invention relates to a noise posting classification method and system in a social network service. Specifically, with respect to posting data in a social network service (hereinafter referred to as SNS) such as Twitter (registered trademark), client application software for posting to the service or viewing other people's posts ( For example, a learning (classification) model for classifying Twitter for Android (registered trademark) by purpose is generated. The present invention relates to a method and system for classifying an unknown SNS client by a classification model and, in particular, extracting an SNS client that outputs unnecessary post data (noise).

  2. Description of the Related Art Conventionally, post data posted in SNS is collected and analyzed to be used for marketing and the like. For example, Patent Document 1 discloses an apparatus that analyzes a degree of conformity when posting posted data electronically as an evaluation value, and ranks the posted data based on the evaluation value. The evaluation in Patent Document 1 is a method of decomposing a sentence in posted data into a plurality of words, comparing each word with a predefined score setting table, calculating a score in units of words, and then calculating a total score of the entire sentence It is. For example, the score in units of words is a so-called weighting value, and by setting a large score for inappropriate words, the total score of post data including many inappropriate words becomes larger. Thereby, it is possible to rank the posting data and analyze the degree of conformity in electronic posting.

JP 2006-268303 A

  However, the analysis in Patent Document 1 requires a certain number of words in the posted data and is not suitable for analyzing SNS posted data with a small number of words.

  In recent years, the use of SNS has become complicated, and the types of post data are various, such as automatic posting of standard sentences and advertisements / campaigns. In addition, the definition of noise varies depending on the purpose of post data analysis. In addition, an API for generating a posting application is open to the public, and anyone can generate an SNS client. In other words, new post data and noise always occur, and in addition, an unknown SNS client also newly occurs, so that it is assumed that there is something that cannot be determined by existing static data reference. In such a situation, in order to perform analysis for each post data and identify noise, a short individual sentence such as post data by an SNS client is not suitable as input data for analysis.

  Also, the user has a tendency to use (categorize) SNS clients properly according to the usage (for example, category 1: posting of the user himself is a posting application such as Twitter, category 2: high score of game application) Posts associated with user actions such as acquisition and reading of a predetermined electronic book are action recording applications, classification 3: campaign posting is a dedicated application for each company, classification 4: automatic posting is bot application, and the like). The classification of the SNS client can be considered that classification 1 has the strongest user's intention, weakens as the number increases, and classification 4 has the weakest. The category 2 can read the user's thoughts based on the content and the posting timing, and the category 3 can determine the user's degree of preference. In Category 4, it can be said that there is no user intention due to automatic posting by the robot. That is, it can be said that the stronger the user's intention in the posted data, the higher the importance for the user, and the higher the usefulness for a company or the like used for marketing.

  From the above, there is a need for a method and system for performing analysis in units of SNS clients, not in units of posted data by SNS clients, and extracting SNS clients that output noise.

Therefore, the present invention focuses on the classification of SNS clients, and in order to solve the above-described problem, the present invention can be executed by a computer for executing a method for classifying SNS clients that output the posted data based on the posted data in the SNS. A computer program having instructions, the method comprising:
Obtaining the post data;
Extracting at least an SNS client identifier, an account identifier, and text content from the acquired post data, and calculating a feature amount for each SNS client, wherein the feature amount includes an average compression rate and a total compression rate; The average compression rate includes at least the average compression rate for each SNS client identifier when the body content is combined and compressed for each of the SNS client identifier and the account identifier, and the total compression rate is , The compression ratio when the body content is combined and compressed for each SNS client identifier, and
Classifying SNS clients that output the acquired post data using the feature quantity as input data of a learning model, wherein the learning model extracts the feature quantities of a plurality of SNS clients from post data in the SNS. An explanatory variable and a step of generating the classification of the SNS client as an objective variable are provided.

  In the invention described in the preceding paragraph, the feature amount further includes an average number of posts, and the average number of posts is the number of posts calculated for each of the SNS client identifier and the account identifier for each SNS client identifier. The average number of posts.

  Furthermore, in the invention described in the preceding paragraph, the extracting includes further extracting a posting date and time from the acquired posting data, and the feature amount further includes a posting interval average standard deviation, and the posting interval average The standard deviation is an average value for each of the SNS client identifiers calculated from the respective standard deviations by calculating a posting interval for each of the SNS client identifier and the account identifier from the posting date and time.

  As described above, according to the present invention, it is possible to generate a classification model for classifying SNS clients that output post data based on the purpose based on post data in the SNS and classify unknown SNS clients. Thereby, in particular, an SNS client that outputs noise can be extracted. In addition, when analyzing post data for a certain purpose, it is possible to exclude SNS clients that output post data that causes noise from the batch, and post more efficiently than before. Data analysis can be performed. Furthermore, by learning the classification model, it is possible to cope with a SNS client of a type newly generated in the future.

It is a figure which shows the system configuration | structure which concerns on one Embodiment of this invention. It is a figure which shows the data stored in the SNS client data storage part which concerns on one Embodiment of this invention. It is a figure which shows the data stored in the teacher data storage part which concerns on one Embodiment of this invention. It is a figure which shows the classification model which concerns on one Embodiment of this invention. It is a figure which shows the data stored in the classification model data storage part which concerns on one Embodiment of this invention. It is a flowchart which shows the classification model production | generation process which concerns on one Embodiment of this invention. It is a flowchart which shows the SNS client classification | category process which concerns on one Embodiment of this invention. It is a flowchart which shows the judgment method of the learning pattern performed in one Embodiment of this invention.

  Hereinafter, referring to the attached drawings, based on post data in SNS, a SNS client that classifies SNS clients that output the post data according to purpose can be generated, and an SNS client can be classified. The client classification system and method will be described in detail.

  First, the outline of this system will be described. FIG. 1 is a diagram showing a system configuration according to an embodiment of the present invention. In FIG. 1, an SNS post data storage server 100 installed in a data center or the like and managed by each SNS provider company or the like is connected to user terminals 102a, 102b,..., And 102n (hereinafter collectively) via the Internet 104. .., And 103n (hereinafter collectively referred to as “mobile terminal 103”) and the mobile terminals 103a, 103b,.

  A user who uses the SNS uses the SNS provided by each SNS post data storage server 100 using the user terminal 102 or the mobile terminal 103. Post data posted by each user is transmitted to the SNS post data storage server 100 and collected in the database. Although the SNS post data storage server 100 is shown as a single server in FIG. 1, it can also be configured as a distributed system including a plurality of servers. In addition, since the SNS post data storage server 100 exists for each SNS client, in reality, many SNS post data storage servers 100 are connected to the Internet 104.

  An SNS client classification server 101 is connected to the Internet 104. Post data aggregated in each SNS post data storage server 100 is transmitted to the SNS client classification server 101, and is classified on a SNS client basis on the SNS client classification server 101 using a classification model.

  Next, the configuration of the SNS client classification server 101 will be described in detail. In FIG. 1, only a necessary functional configuration is shown assuming a single server computer.

  The SNS client classification server 101 includes a CPU 110 and a storage device 116 including a RAM 111, an input device 112, an output device 113, a communication control device 114, and a nonvolatile storage medium (ROM, HDD, etc.) via a system bus 115. It has a connected configuration. The storage device 116 includes a program storage area for storing a software program for performing each function of the SNS client classification system, and a data storage area for storing data handled by the software program. Each means of the program storage area described below is actually an independent software program, its routines and components, etc., which are called from the storage device 116 by the CPU 110, are expanded in the work area of the RAM 111, and a database is appropriately stored. Each function is performed by being executed sequentially with reference to the reference.

  The data storage area in the storage device 116 includes an SNS client data storage unit 130, a teacher data storage unit 131, a classification model data storage unit 132, and a post data storage unit 133 when only those related to the present invention are listed. Both are fixed storage areas secured in the storage device 116.

  The SNS client data storage unit 130 stores data related to an SNS client (for example, an application such as Twitter) that outputs post data. FIG. 2 is a diagram illustrating data stored in the SNS client data storage unit 130 according to an embodiment of the present invention. The SNS client data in FIG. 2 includes a “client name” indicating the name of the SNS client, a “category ID” for classifying the SNS clients by purpose, and a “category name” indicating the content thereof. “Category ID” is, for example, “1” in the case where the SNS client is a posting application that outputs posted data by the user such as Twitter, and the action of outputting posted data accompanying the user's action such as high score acquisition of the game application. "2" for recording applications, "3" for each company's dedicated application that outputs campaign post data, and "bot" for automatic output of fixed phrases registered in the dictionary at regular intervals. Stores “4”. In this case, it can be considered that the smaller the numerical value of the category ID is, the more the SNS client outputs post data with a strong user intention. Therefore, for example, post data with a category ID “4” having a weak user intention is considered to be noise data, and an embodiment in which the post data is not displayed by filtering is assumed. The SNS client data is generated by extracting the “client name” included in the posted data. However, since “category ID” and “category name” cannot be determined at the time of extraction, both items are data records. It is empty data at the time of generation. After record generation, an administrator or the like manually registers a certain amount of data for “category ID” and “category name” (they need not be all but may be the minimum data necessary for learning), and will be described later. By performing the learning process, the data after learning can be registered again.

  The teacher data storage unit 131 stores teacher data for the classification model. FIG. 3 is a diagram illustrating data stored in the teacher data storage unit 131 according to an embodiment of the present invention. The teacher data in FIG. 3 includes “client name”, each feature amount, and “category ID”. The feature amounts are “average compression rate”, “total compression rate”, “post interval average standard deviation”, and “average number of posts”, and these are explanatory variables for generating a classification model. The calculation of each feature amount will be described later. Further, “category ID” is an objective variable for the explanatory variable of each feature amount. Based on these, the classification model data shown in the next section is generated. As for the category ID in this data, the learned data can be registered again by performing a learning process described later, like the category ID of the SNS client data in the previous section.

  The classification model data storage unit 132 stores learning (classification) model data generated based on teacher data. The classification model in the present invention is a binary classifier that takes an explanatory variable (feature value) as an input and quantitatively analyzes whether the objective variable (category ID) can be explained by the explanatory variable (regression analysis). The binary classifier can be implemented using various conventional techniques (linear regression, decision tree, logistic regression, neural network, support vector machine (SVM), perceptron, etc.).

  FIG. 5 is a diagram showing data stored in the classification model data storage unit 132 according to an embodiment of the present invention. FIG. 4 is a diagram showing a classification model according to an embodiment of the present invention. A decision tree is used in the classification model of FIG. The data corresponding to the classification model of FIG. 4 is the classification model data of FIG. The classification model data in FIG. 5 includes a “node ID” that uniquely identifies each tree node, a “parent node ID” that indicates a parent node ID of each node ID, and a “sibling” that indicates the order between nodes having the same parent node. “Inter-order”, “content” indicating the contents of each node, and “category ID”. In the present embodiment, “order between siblings” is “1” for a node hanging on the Yes edge and “2” for a node hanging on the No edge for the same parent node. If there are more types of edges for the same parent node, numbers 1, 2, 3,... Can be assigned from the left node, for example.

  When the classification model data of FIG. 4 is described using the classification model of FIG. 5, since the “node ID” is “0” for the root node [average compression rate <0.78] and there is no parent node, “parent” “Node ID” and “order between siblings” are empty data.

  Next, “node IDs” of “average number of posts <5” and “average number of posts <2.7” which are child nodes of the root node are “1” and “6”, respectively. The “parent node ID” of both nodes stores “0” which is the root node. In the “order between siblings”, “1” is stored for the node “average number of posts <5” hanging on the edge of Yes, and “2” is stored for the node “average number of posts <2.7” hanging on the edge of No. Is done.

  Similarly, “automatic posting” which is a child node of the node “average number of posts <5” and “node ID” of “total compression ratio <= 0.55” are “2” and “3”, respectively. Here, the node “automatic posting” is a terminal node (indicated as a double line block in FIG. 4), and the SNS client to be learned is classified into the bot application indicated by the category ID “4”. Indicates that Therefore, “4” is stored in “Category ID”. Other nodes are similarly configured. Each of the data in FIGS. 2, 3, and 5 is an embodiment, and does not prevent addition or deletion of data items.

  The posted data storage unit 133 stores posted data provided from each SNS posted data storage server 100. The post data will be described later.

  Next, among software programs stored in the program storage area in the storage device 116, only those related to the present invention are listed. Posted data acquisition means 120, SNS client data generation means 121, feature quantity calculation means 122, teacher Data generation means 123, classification model data generation means 124, and SNS client classification means 125 are provided. These means are executed by the CPU 110.

  Post data acquisition means 120 acquires post data provided from each SNS post data storage server 100 and stored in post data storage unit 133.

  The SNS client data generation unit 121 extracts the SNS client name from the acquired post data, generates SNS client data based on the SNS client name, and stores it in the SNS client data storage unit 130.

  The feature amount calculation unit 122 extracts an SNS client identifier, an account identifier, a text content, a posting date and the like from the acquired posting data, and extracts each feature amount (average compression rate, total compression rate, posting interval average standard deviation, average). Number of posts).

  The teacher data generation unit 123 generates teacher data from the generated SNS client data and each calculated feature amount, and stores the teacher data in the teacher data generation unit 123.

  The classification model data generation unit 124 generates classification (learning) model data using each feature amount of the teacher data as an explanatory variable and the category ID as an objective variable, and stores the classification (learning) model data in the classification model data storage unit 132.

  The SNS client classification unit 125 classifies the SNS clients by using each feature amount related to the classification target SNS client as an input of the classification model. In addition, the SNS client classification unit 125 can reflect the classification result in the SNS client data storage unit 130 and the teacher data generation unit 123.

  Next, the classification model generation processing of the present invention will be described along the flow. FIG. 6 is a flowchart showing classification model generation processing according to an embodiment of the present invention. First, in step 101, post data acquisition means 120 acquires post data provided from each SNS post data storage server 100 and stored in post data storage unit 133. Here, in one embodiment, post data is acquired by using an SNS post data storage server 100 as an SNS client name, account name (post user name), text content, post date and time, etc. embedded in the source code. To receive from. Note that the SNS client name and the account name only need to be able to identify the SNS client and the account, and are arbitrary identifiers including an ID and the like. In addition, acquisition of post data can narrow down acquisition conditions such as data posted in a specific period. In another embodiment, the post data itself is data that has been narrowed down in advance.

  When the posted data is acquired, the SNS client data generating unit 121 extracts the SNS client name from the acquired posted data, and generates SNS client data (FIG. 2) based on the SNS client name (step 102). The generated SNS client data is stored in the SNS client data storage unit 130. Here, the “category ID” and “category name” in the SNS client data are empty data because they are not yet classified at the time of data generation. However, for subsequent classification (learning) processing, data must be manually registered for the “category ID” and “category name” of all SNS client data or some SNS clients whose location is clearly known. Can do. Subsequent classification (learning) is performed using the manually registered category data as an initial value.

  Next, when the posting data is acquired, the feature amount calculation unit 122 extracts the SNS client name, the account name, the body content, the posting date and time from the acquired posting data, and each feature amount (average compression rate, total compression rate) , Posting interval average standard deviation, average number of postings) is calculated (step 103). Each feature amount will be described in detail.

  The average compression rate is an average compression rate for each SNS client when post data (text content) is compressed for each SNS client and account. When the same account makes the same or similar posting using a predetermined SNS client, the compression rate becomes high (the file size after compression becomes small), and when there are many combinations of such SNS clients and accounts, The average compression ratio becomes higher. For example, when the SNS client is a bot application that automatically outputs a fixed sentence registered in the dictionary at regular intervals, the posted data for each account tends to be the same or similar, and the compression rate also increases. Note that the bot application can post set contents not only at regular intervals but also at timings set in advance by the user, such as designated date and time. In addition, although there is a fixed time, strictly speaking, a random time of about several seconds is given for each posting timing, and then automatic posting is performed. In one embodiment, compression is performed using a general compression algorithm by combining body contents (character strings) in post data. Therefore, the average compression rate is calculated by combining the contents of the text in the posted data for the same SNS client and the same account, and compressing the average compression rate for each SNS client for each compression rate ( This is the average compression rate for each SNS client).

In one embodiment, the average compression rate V _1C (cεC) for each SNS client C and for each account included in the SNS client set C can be calculated by the following equation.

Here, ave a _{∈ A} () represents that an average value of arguments is calculated for each account a included in the account set A. Also, zip (str) represents that compression processing is performed on the character string str. Tw (C, a) is a post data set posted by the SNS client C using the account a. _{ΣtεTw (C, a)} str (t) represents concatenation of all character strings str of the posted data t included in the posted data set Tw (C, a).

  The total compression rate is a compression rate when the posted data is compressed for each SNS client. When similar postings are made between users who use the same SNS client, the compression rate increases. The calculation of the total compression rate is performed by combining the contents of the text in the posted data for each same SNS client and compressing, and each compression rate is the total compression rate for each SNS client.

In one embodiment, the total compression rate V _2C (cεC) for each SNS client C included in the SNS client set C can be calculated by the following equation.

The meaning of each notation is as shown in the explanation of the average compression rate V _1C (cεC).

  The posting interval average standard deviation is an average value for each SNS client calculated from the respective standard deviations by calculating the posting interval for each SNS client and account from the posting date and time in the posting data. When the SNS client is a bot application that automatically posts every fixed time, since the posting interval is constant, there is little variation in the posting interval, and the standard deviation and its average value are also reduced. Note that the standard deviation is used here, even if the current bot application posts at regular intervals, strictly speaking, after adding a random time of several seconds for each posting timing This is because some of them automatically post. Posting interval average standard deviation is calculated for each of the same SNS client and the same account by calculating the posting interval after arranging the posting date and time obtained from posting data in time series, and calculating the standard deviation of the posting interval. For each standard deviation, the average value is calculated for each identical SNS client (this is the posting interval average standard deviation for each SNS client).

In one embodiment, the posting interval average standard deviation V _3C (cεC) for each SNS client C and for each account included in the SNS client set C can be calculated by the following equation.

Here, ave a _{∈ A} () represents that an average value of arguments is calculated for each account a included in the account set A. Stddev _{tεTw (C, a)} () calculates the standard deviation of the argument for each posting data t included in the posting data set Tw (C, a) for each SNS client C and for each account a. Represents. minu t ediff _{(t i,} t _j) denotes calculating a difference between the post time of the post data _{t i} and _{t j} in minutes.

  The average number of posts is the average number of posts for each SNS client of the number of posts (posts) calculated for each SNS client and account. Applications that post for campaigns tend to have fewer posts. The average number of posts is calculated by counting the number of posts for the same SNS client and the same account and taking the average value for each same SNS client (this is the average number of posts for each SNS client). .

In one embodiment, the average number of posts V _4C (cεC) for each SNS client C and for each account included in the SNS client set C can be calculated by the following equation.

Here, ave a _{∈ A} () represents that an average value of arguments is calculated for each account a included in the account set A. Further, n (Tw (C, a)) is the number of elements of the posting data set Tw (C, a) for each SNS client C and for each account a.

  Returning to the processing flow of FIG. 6, when each feature amount is calculated in step 103, the teacher data generation unit 123 uses the SNS client data generated in step 102 and the teacher data (FIG. 3) is generated (step 104). The generated teacher data is stored in the teacher data storage unit 131. The “client name” and “category ID” in FIG. 3 can be stored by copying from the SNS client data (FIG. 2). In one embodiment, the “client name” and “category ID” can be linked by the person in charge of analysis. Each feature amount (average compression ratio, total compression ratio, posting interval average standard deviation, average posting) (Number) is the value itself calculated in step 103.

  Next, the classification model data generation unit 124 generates classification (learning) model data (FIG. 5) using each feature amount of the teacher data generated in step 104 as an explanatory variable and a category ID as an objective variable (step 5). 105). The generated classification model data is stored in the classification model data storage unit 132. The classification model (FIG. 4) that can be shown from the classification model data is a so-called decision tree, and is generated using a general decision tree generation algorithm. The detailed description of the classification model data and the relationship between the classification model data and the classification model are as described above. Further, it is not always necessary to use all the feature quantities used as explanatory variables when generating the classification model. For example, a learning model can be generated using only the average compression rate and the total compression rate as explanatory variables. After step 105, the process ends.

  Next, the SNS client classification process of the present invention will be described along the flow. FIG. 7 is a flowchart showing SNS client classification processing according to an embodiment of the present invention. First, in step 201, the posting data acquisition unit 120 acquires posting data related to the SNS client to be classified (hereinafter referred to as classification target posting data) from the posting data storage unit 133. The explanation about the posted data is the same as that in step 101.

  Next, when the classification target post data is acquired, the feature amount calculation unit 122 extracts the SNS client name, the account name, the body content, the posting date and the like from the acquired classification target post data, and each feature amount (average compression rate) , The total compression rate, the posting interval average standard deviation, and the average number of postings) are calculated (step 202). The description of each feature amount is the same as that in step 103.

  When each feature amount of the classification target posted data is calculated, the SNS client classification unit 125 classifies the SNS client to be classified using each calculated feature amount as an input of the classification model generated in Step 105. The classification method considers each node (excluding the terminal node) in the classification model and an edge for each node as a conditional expression (for example, an if statement or a switch statement), and uses each feature amount of the SNS client to be classified as a route. This is done by proceeding from the node according to each condition and reaching one of the terminal nodes. The category related to the terminal node that has been reached is the category in which the SNS clients to be classified are classified.

  In step 203, when the SNS clients to be classified are classified, the SNS client classification unit 125 uses the classification result (category ID and category name) and each feature amount calculated in step 202 as the SNS client data storage unit. 130 and teacher data generation means 123 (steps 204 and 205). Thus, the classification model can be updated (learned) by executing the classification model generation process (FIG. 6) again. After step 205, the process ends. If the classification model is not updated with the initial value, steps 204 and 205 are not performed, and after step 203, the process ends.

  Whether or not to update (learn) the classification model and what kind of learning to perform can be determined according to a predetermined rule. FIG. 8 is a flowchart showing a learning pattern determination method to be executed in an embodiment of the present invention. First, in step 301, it is determined whether or not the posting period of post data to be analyzed is fixed within a certain range. When repeatedly analyzing data for various posting periods, the process proceeds to the No route. On the other hand, if it is determined that the posting period of the posting data to be analyzed is fixed within a certain range, the process proceeds to the Yes route and executes the classification process of FIG. 7 in accordance with the following execution pattern 1 (non-learning pattern) rule. can do.

Execution pattern 1 (non-learning pattern)
Once the classification model is generated, the SNS client is classified using the classification model, but the classification model is not learned. That is, the classification model generation process (FIG. 6) is only executed at the first time without reflecting the classification result and each feature amount (execution of steps 204 and 205).

  On the other hand, if it is determined in step 301 that the posting period of the posting data to be analyzed is not fixed within a certain range, the process proceeds to the No route, and it is determined whether or not a storage area for past data can be sufficiently secured (step 302). . When learning the classification model, more accurate classification and fine classification are possible by using past data. However, this requires a data area for storing past data. The storage amount of past data can be determined based on the accuracy of classification to be obtained, and when the past data exceeds the determined storage amount, the oldest data can be deleted in order.

  If it is determined in step 302 that a sufficient storage area for past data cannot be secured, the process proceeds to the No route, and the classification process of FIG. 7 is performed in accordance with the rule of the following execution pattern 2 (unclassified SNS client learning pattern). Can be executed.

Execution pattern 2 (uncategorized SNS client learning pattern)
After classifying the SNS clients using the classification model for the first period, the second SNS client is generated on the condition that a predetermined number or more of unknown SNS clients occur in a second period different from the first period. SNS clients are classified for the period. In the case of classification for the second period, if the classification target is a known SNS client, the classification model is updated using each feature amount already stored in the teacher data storage unit 131 in the previous classification. . On the other hand, in the case of an SNS client whose classification target is unknown, each feature amount is calculated from the posted data stored in the posted data storage unit 133. The unknown SNS client is classified using the calculated feature values of the unknown SNS client as the input of the updated classification model.

  On the other hand, if it is determined in step 302 that a sufficient storage area for past data can be secured, the process proceeds to the Yes route, and it is determined whether or not more accurate classification of the posted data is necessary (step 303). If it is determined in step 303 that more accurate classification is not required for the posted data, the process proceeds to the No route, and the classification process of FIG. 7 is executed in accordance with the following rules of execution pattern 3 (all learning patterns). be able to.

Execution pattern 3 (all learning patterns)
After the SNS client is classified using the classification model for the first period, the second period is set on condition that a predetermined number or more of post data is generated in a second period different from the first period. The SNS client is classified as a target. At the time of classification for the second period, each feature amount is calculated from the posted data stored in the posted data storage unit 133 regardless of whether the SNS client to be classified is known or unknown. The classification model is updated using each feature quantity already stored in the teacher data storage unit 131 in the previous classification, and the calculated feature quantities of the known and unknown SNS clients are used as the input of the updated classification model. To classify known and unknown SNS clients.

  Note that the learning pattern determination method to be executed shown in FIG. 8 is merely an embodiment, and the present invention is not limited to these determinations and learning patterns. In addition, by considering past data for analysis of posted data or performing analysis for each predetermined period, SNS clients can be further classified as follows.

Execution pattern 4 (Past data consideration learning pattern)
After classifying the SNS client using the classification model, if the SNS client classification result is the same as the previous classification result, the classification result and each feature amount are reflected, and the classification model generation process is executed again. Then, the classification model is learned. Since learning is performed only when the SNS client classification result is the same as the previous classification result, it is possible to cope with the blur of the classification result, and classification with higher accuracy is possible.

Execution pattern 5 (period-specific classification pattern)
The SNS client data (FIG. 2) and the teacher data (FIG. 3) have the concept of a period and are classified for each period. That is, even for the same SNS client, for each period, each feature amount is calculated from the posted data in the period and is classified into individual categories. For example, the SNS client “Patent” can be classified as the category “person” in the period “March 2014” but the category “automatic posting” in “April 2014”.

Claims (5)

A method of classifying SNS clients that output post data based on post data in a social network service (SNS), the method comprising:
Obtaining the post data;
Extracting at least an SNS client identifier, an account identifier, and text content from the acquired post data, and calculating a feature amount for each SNS client, wherein the feature amount includes an average compression rate and a total compression rate; The average compression rate includes at least the average compression rate for each SNS client identifier when the body content is combined and compressed for each of the SNS client identifier and the account identifier, and the total compression rate is , The compression ratio when the body content is combined and compressed for each SNS client identifier, and
Classifying SNS clients that output the acquired post data using the feature quantity as input data of a learning model, wherein the learning model extracts the feature quantities of a plurality of SNS clients from post data in the SNS. A method comprising: generating an explanatory variable and a classification of the SNS client as an objective variable.   The feature amount further includes an average number of posts, and the average number of posts is an average number of posts for each SNS client identifier of the number of posts calculated for each of the SNS client identifier and the account identifier. The method of claim 1.   The extracting includes further extracting a posting date and time from the acquired posting data, and the feature amount further includes a posting interval average standard deviation, and the posting interval average standard deviation is calculated from the posting date and time, The method according to claim 2, wherein a posting interval is calculated for each SNS client identifier and each account identifier, and is an average value for each SNS client identifier calculated from each standard deviation. A computer program having computer-executable instructions for causing a computer to execute a method of classifying SNS clients that output the posted data based on posted data in a social network service (SNS), wherein the computer program ,
Get the post data,
The SNS client identifier, the account identifier, and the text content are extracted at least from the acquired posted data, the feature amount for each SNS client is calculated, and the feature amount includes at least an average compression rate and a total compression rate, The average compression rate is an average compression rate for each SNS client identifier when the body content is combined and compressed for each of the SNS client identifier and the account identifier, and the total compression rate is the SNS client It is a compression ratio when the body content is combined and compressed for each identifier,
The SNS client that outputs the acquired post data is classified using the feature amount as input data of a learning model, and the learning model uses the feature amount of the plurality of SNS clients as explanatory variables from post data in the SNS, and A computer program generated using the classification of the SNS client as an objective variable. A server computer that classifies SNS clients that output post data based on post data in a social network service (SNS), the server computer comprising:
Means for obtaining the post data;
A means for extracting at least an SNS client identifier, an account identifier, and text content from the acquired post data, and calculating a feature amount for each SNS client, wherein the feature amount includes an average compression rate and a total compression rate. The average compression rate includes at least the average compression rate for each SNS client identifier when the body content is combined and compressed for each of the SNS client identifier and the account identifier, and the total compression rate is , Means that is a compression ratio when the body content is combined and compressed for each SNS client identifier;
The means for classifying SNS clients that output the acquired post data using the feature quantity as input data of a learning model, wherein the learning model uses the feature quantities of a plurality of SNS clients from post data in the SNS. A server computer comprising: means for generating an explanatory variable and a classification of the SNS client as an objective variable.

Images