KR102126911B1 - Key player detection method in social media using KeyplayerRank - Google Patents

Abstract

The present invention relates to a method of detecting a key player by topic on social media by using keyplayerrank. The method of detecting a key player by topic on social media by using keyplayerrank includes the steps of: (a) collecting, by a news data collection module, news articles related to a specific category, and preprocessing, by the news data preprocessing module, the collected news articles; (b) training, by an article text classification module, an SVM through training data and classifying a text through the trained SVM model; (c) extracting, by a subject extraction module, a subject and keywords for a subject by performing LDA on data found to be related to a specific category; (d) collecting, by a Twitter data collection module, Twitter data, and preprocessing, by a Twitter data preprocessing module, Twitter data including subject keywords among the collected Twitter data; (e) training, by the Twitter text classification module, the SVM through the training data, and classifying the text through the SVM model trained by using the Twitter data including the subject keyword as test data; (f) classifying, by a user classification module, users on data found to be related to the subject; (g) calculating, by an influence index and subject similarity calculation module, an influence index and subject similarity; and (h) selecting, by a key player detection module, a key player for each subject by using the influence index and subject similarity. The most influential key players for each subject can be extracted by simultaneously considering the influence index and subject similarity.

Description

Key player detection method in social media using KeyplayerRank}

The present invention relates to a method for detecting a key player for each subject on social media using KeyplayerRank, and more specifically, to classify users by subject based on text written by the user on social media and to consider the impact index and subject similarity at the same time. It relates to a key player detection method for each topic on social media using KeyplayerRank, which can extract a potent key player.

As the number of social media users increases steadily every year, more and more people are using social media to acquire data or information, and communicating with other users on social media, and the words of the most influential users on social media when acquiring data or information. Trust and follow are happening frequently.

Accordingly, as social media has developed, attempts to identify social influence in social media have been made in various ways in recent years. In particular, among social media, research to find the most influential person on Twitter is the most active.

Similar to the present invention, there are prior studies using Google's PageRank algorithm, and Tunkelang (2009) determines the influence on Twitter, and calculates the probability that followers will read all tweets including retweets from key players. TunkRank was applied. However, TunkRank is an algorithm that focuses on retweets without considering the content of Twitter.

Weng et al. (2010) proposed TwitterRank, which calculates the in-degree (number of entries from a graph to one node in a graph, when an existing social network analysis method finds a key player, where one user follows another) Simple impact measurement using the number of followers) does not accurately represent the concept of influence, and PageRank improves the in-degree considering the connection structure of the entire network, but shows the degree of interest among users affecting Twitter users. He pointed out that he could not be considered. Therefore, TwitterRank was applied to compensate for the disadvantages of in-degree and page rank by considering the connection structure and topic similarity between Twitter users. In contrast to TunkRank, the TwitterRank algorithm developed a ranking method focusing on topic similarity.

If you want to find a key player that only considers topic similarity, you can use TwitterRank to find it. On the other hand, if you want to find an influential user without limiting to a specific topic such as celebrities, politicians, etc., you can use TunkRank's algorithm.

On the other hand, if you obtain and add two themes of similarity and influence separately, the following contradiction may occur. For example, if an influential user does not talk about a topic and mentions several topics little by little, there is no expert knowledge on the topic. Although it cannot be said to be a key player, previous studies refer to this user as a key player. Such a key player can be said to be a key player biased only to influence.

In other words, if a key player is derived by simply combining the influence index and the subject similarity, a result highly dependent on the influence of the subject itself can be derived. That is, for a case in which the subject to detect a key player is not a subject that is attracted by many people, the key player of the subject has a relatively low impact index compared to the key players of other subjects, although the subject similarity is high in this case. Nevertheless, there is a problem in that a key player for a desired topic cannot be derived due to a low influence index. On the contrary, even if a user has a very high influence index, there is a disadvantage that the user may not be a key player of a topic of interest.

In addition, the prior art Republic of Korea Patent Publication No. 10-2014-0062635 (2014.05.26. published) step of receiving a keyword from the user; Transmitting a search request message including the received keyword and receiving a search response message corresponding to the transmitted search request message; Upon receiving the search response message, extracting a search list listed according to a predetermined influence ranking by an influential user related to the keyword from the received search response message; Sequentially displaying influential users related to keywords received from the user based on the extracted search list according to a predetermined influence ranking; And storing the extracted search list. A method for searching an influential user in social media is disclosed. However, there is a limit to extracting influencers that reflect similarities and influences of each topic at the same time because the user is displayed according to a predetermined influence ranking after finding the influencers related to the keyword.

Republic of Korea Patent Publication No. 10-2014-0062635 (2014.05.26. Publication, the name of the invention: a device, system and method for searching for influential users in social media)

The present invention has been devised to solve the above problems, and the object of the present invention is to consider the impact index and the subject similarity by analyzing messages generated on social media at the same time, and key players on social media that are not biased to one indicator ( By detecting the influencers), it is to provide a method for detecting a key player for each topic on social media using KeyplayerRank, which can obtain reliable information about the topic through the most influential users by topic and filter out unnecessary information.

In order to achieve the above object, the present invention is (a) news data collection module to collect news articles related to a specific category specified from the news data DB through crawling, the news data pre-processing module is collected ' A pre-processing step of extracting nouns by removing idioms from the TEXT' field and analyzing morphemes; (b) The article text classification module selects features to be used for analysis for classifying news article texts, indexes texts expressing features by feature weights, and then, a certain percentage of article texts extracted from nouns in step (a) Dividing into training data and test data to train the SVM through the training data, and classifying the text through the trained SVM model; (c) extracting a subject and subject-specific keywords by subjecting the subject extraction module to LDA on data found to be related to a specific category by text classification through the SVM model in step (b); (d) Twitter data that the Twitter data collection module collects Twitter data in the same period as the news article collection period from the Twitter data DB, and Twitter data including subject-specific keywords extracted through the LDA among the Twitter data collected by the Twitter data pre-processing module. A pre-processing step of extracting nouns by removing idioms from the'text' field of their own and morphological analysis; (e) The Twitter text classification module trains SVM through training data in which Twitter data is added to a certain number of article data, and SVM trained using Twitter data including subject-specific keywords in step (d) as test data. Classifying text through a model; (f) the user classification module classifying the data on the data found to be related to the subject by text classification through the SVM model in step (e); (g) an influence index and subject similarity calculation module calculate an impact index using attribute information by the user classification, and calculate a subject similarity using a'Jaccard Index', and (h) a key player detection module It characterized in that it consists of selecting the user having the highest KpRank value among the users as the key player for each topic by using the influence index and the subject similarity.

In addition, in the present invention, the step (b), (b1) is characterized by selecting the words to be used to classify the text by performing the labeling according to whether or not the article text extracted from the noun is related to the category Selection step; (b2) generating an article text DTM applying TF-IDF as a feature weight using the selected features, and (b3) training the training data through the training data by dividing the training data and the test data in a constant ratio using the DTM. It consists of classifying the article text by applying it to the test data with the trained SVM model, and comparing the result of classifying the article text by applying the test data with the trained SVM model to the actual labeled result, and extracting the correct classification. Is done.

In addition, in the present invention, the step (c), the result of classifying the article text through the labeling result and the SVM model, generates a LDA model by generating a data text that is related to a specific category as a DTM, and an LDA model. It is characterized by extracting keywords for each subject by determining'num_topics','passes', and'number of keywords', which are necessary parameters to perform the task, and determining the subject from the extracted keywords.

In addition, in the present invention, the step (e) is characterized by extracting the correct classification by comparing the results of classifying the Twitter text and the actual labeled results by applying to the test data with a trained SVM model.

In addition, in the present invention, the step (f) is classified into a retweet user and a retweet user to retrieve attribute information of Twitter users, but if the retweet user is not stored in the storage device, use the Twitter REST API. It is characterized in that the attribute information of the user is retrieved from the Twitter data DB.

(여기서,

는 사용자 i가 주제 내에서 쓴 각 트윗에 해당하는 각 리트윗된 횟수를 모두 더해 사용자 i가 주제 내에서 쓴 모든 트윗 수로 나눈 사용자 i의 평균 리트윗된 수이고,

는

와 같이 사용자 i의 평균 favorite 수를 나타내며,

와

는 주제 내의 전체 사용자들의 평균 리트윗된 수와 평균 favorite 수이고,

는

과

를 구한 것과 같이 사용자 i가 주제 내에서 쓴 각 트윗에 해당하는 각 팔로워 수를 모두 더해 사용자 i가 주제 내에서 쓴 모든 트윗 수로 나눈 평균 팔로워 수이고,

는 전체 사용자들의 평균 팔로워 수임)으로 산정하고, 상기 주제 유사도는 다음의 수학식,In addition, the present invention, in the step (g), the impact index is the following equation using'Follwers_count, RT_count, Favorite_count' among the attribute information of users,

(here,

Is the average number of retweets for user i divided by the number of all tweets written by user i within the topic by adding up each number of retweets corresponding to each tweet written by user i in the topic,

The

Like this, it represents the average number of favorite users i,

Wow

Is the average number of retweets and average favorites of all users in the subject,

The

and

The average number of followers divided by the total number of tweets that user i has written within the topic by adding up the number of each follower corresponding to each tweet that user i wrote within the topic, as obtained by

Is the average number of followers of all users), and the subject similarity is the following equation,

(여기서, A와 B는 각 사용자가 쓴 트위터 텍스트(명사들로만 이루어져 있음) 묶음임)으로 산정하는 것을 특징으로 한다.

(Here, A and B are characterized by calculating by Twitter text (consisting only of nouns) written by each user).

In addition, the present invention, in step (h), the KpRank value is the following equation,

(Where d is a probability that a Twitter user visits a Twitter page referring to a topic of interest within a specific topic).

As described above, the key player detection method for each topic on social media using the present invention, KeyplayerRank, is different from simply finding and adding the impact index and subject similarity, and simultaneously considering the impact index and subject similarity by analyzing messages generated on social media. By detecting key players (influencers) on social media that are not biased by one indicator, it is possible to obtain certain information about the subject through the most influential users for each subject and to filter out unnecessary information.

Furthermore, the present invention can utilize a marketing aspect to promote a company's product, and detect a key player for a specific subject to effectively promote it through the key player.

1 is a view showing the overall flow of a key player detection method for each topic on social media using KeyplayerRank according to the present invention.
Figure 2 is a view showing a part of the article text DTM applying TF-IDF as a feature weight in the present invention.
3 is a view showing a process of classifying users by subject in the present invention.
4 is a diagram showing a key player for each subject visualized on a map.
5 is a block diagram showing an embodiment of a system related to a key player detection method for each subject on social media using a KeyplayerRank according to the present invention.

If described in detail with reference to the accompanying drawings, preferred embodiments of the present invention configured as described above are as follows. It should be noted that the accompanying drawings and the description with reference to them are exemplified for easy understanding by those skilled in the art with respect to the present invention, and are not intended to limit the spirit and scope of the present invention. Will have to.

5 is a block diagram showing an embodiment of a system related to a method for detecting a key player for each subject according to the present invention, and the key player detection device for each subject 10 classifies and influences users by subject based on text written by the user on social media. Including the method of extracting the most influential key player for each subject by considering the index and subject similarity at the same time, news data collection module (11), news data pre-processing module (12), article text classification module (13), subject extraction module (14), Twitter data collection module (15), Twitter data pre-processing module (16), Twitter text classification module (17), user classification module (18), key player detection module (20) and influence index and subject similarity calculation module (19). The subject key player detection device 10 includes a server, a desktop, a laptop, or a portable terminal, and includes software for performing subject key player detection on social media using KeyplayerRank.

In addition, it is recommended that the data calculated or input and output by the subject key player detection device 10 is stored in the storage device 30. The subject key player detection device 10 may include a storage device 30.

The key player detection method for each topic on social media using the KeyplayerRank according to the present invention made as described above will be described with reference to the flowchart of FIG. 1.

News data preprocessing (S10)

First, it is a pre-processing of news data (for example, Chosun Ilbo data). In the present invention, the reason why the news article is used to extract the subject is because the text that can be written is limited to 140 characters or less due to the nature of Twitter, and is inappropriate for extracting the subject, and Twitter responds immediately to media news and receives public opinion. Because it is a social media that can be quickly identified, topics extracted from news articles can be applied to Twitter as described below.

The news data collection module 11 collects news articles from the news data DB through crawling. At this time, a specific category must be designated for subject extraction, which will be described later. In the present invention, in one embodiment, the category is defined as'artificial intelligence'. Fields imported from news articles related to the above categories are composed of'TITLE, CATEGORY, URL, DATE, TEXT' as shown in Table 1 below. For reference, the content of news articles related to the category'Artificial Intelligence' includes not only'Artificial Intelligence' but also news articles including'Artificial Intelligence' and'Intelligence'.

In the present invention, the news data pre-processing module 12 pre-processes only the'TEXT' field in the collected news articles. Since the article text is composed of several sentences, the process of first separating into sentences is called tokenization. Tokenization refers to the task of dividing a document or sentence into tokens (tokens, meaning strings with meanings, and morphemes and words). After the sentence is separated, it is tokenized once more in word units for morphological analysis within the sentence.The package used at this time is the'NLTK (Natural Language Toolkit)' package, which is a natural language processing and document analysis Python developed for education (Python ) It can be a package.

Removes stopwords (Chinese characters, numbers, English characters, and URL addresses) that are not meaningful in text analysis from tokenized text in sentence and word units, and uses only Korean nouns to extract topics from the text where the stopwords are removed. Therefore, in order to extract only nouns, morpheme analysis is performed in which words separated by word units are given part-of-speech (POS). The morpheme analyzer available here is a java-based Korean morpheme analyzer. Only nouns (general nouns (NNG) and proper nouns (NNP)) are extracted using KOMORAN of the KoNLPy package that gives 42 kinds of parts of speech. SVM is used for text classification in order to classify only text related to a category in the next process by storing the article text in which only nouns are extracted in the storage device 30.

SVM Article text classification through (S20)

Next, it is a text classification process through a support vector machine (SVM).

Here, SVM (Support Vector Machine) is a classification model that finds a decision surface that classifies positive examples from negative examples, and is a machine learning algorithm proposed by Vladimir Vapnik to solve the binary pattern recognition problem.

First, the reason why the SVM algorithm is suitable for text classification is that text is composed of a myriad of features (words), and therefore, a model that is overfitting to the data can be generated because the dimension is infinitely large. However, SVM is suitable for text classification because there are variables that can control these large dimensions. Second, irrelevant features can be distinguished in feature selection. SVM can classify documents considering all information by learning various features to distinguish irrelevant features. Finally, most texts can be classified linearly, and SVM is suitable for text classification because it basically finds a linear classification method. The linear SVM finds an optimal linear decision boundary that linearly separates data, where it is possible to find a decision boundary that best classifies different data.

In the present invention, document classification is performed through a linear SVM to extract a topic with relevant news article text from news data. The text classification process through the SVM can improve performance of a topic modeling result, which will be described later.

Here, the SVM process by the article text classification module 13 is performed as follows, after selecting a feature to be used for analysis for classifying a news article text, and after performing text indexing to express the feature by feature weight , Divide the data into training data and test data at a constant (e.g., 3:1) ratio. Then, SVM is trained through training data, and text is classified through a trained SVM model.

Looking specifically at this, in the pre-processing process, manual labeling is performed on the article texts extracted only from nouns as shown in Table 2 (Example: 5,000 article texts), for example, related to the category'Artificial Intelligence'. '1' is assigned to the class of the text, and'-1' is assigned to the text with unrelated content.

Document frequency is used as a feature selection process in order to form 5,000 article texts labeled in Table 2 into one corpus, and to select words useful for classifying texts. At this time, it is preferable to select the remaining words as features except for words whose document frequency (the number of times a certain word is repeated in one article text) is 2 or less. For reference, the texts in Table 2 are all selected words.

In the next process, text indexing, the weight of the term Term Frequency-Inverse Document Frequency (TF-IDF, used for information retrieval and text mining) is used as a feature weight to determine how important a word is within a specific document when there is a group of documents consisting of multiple documents. Statistical figures).

FIG. 2 shows a part of the DTM, and as shown therein, the article text DTM (Document Term Matrix (DTM) to which TF-IDF is applied as a feature weight is multiplied by the feature weight TF-IDF) and each row is a text (Table 2 In 5,000 article texts), each column represents the previously selected features (listing the words in Table 2), and each element in the matrix represents the TF-IDF weighted value.

Dividing the training data and test data to form an SVM model with the DTM in a ratio of 3:1 has 3,750 training data and 1,250 test data. The SVM model trained from the training data is applied to 1,250 test data to classify text, and accuracy and recall are obtained to verify whether the classification is correct.

Table 3 below is a trained SVM model applied to 1,250 test data, and the result of classifying the text (experimental results, Predicted) and the actual manually labeled result (Actual) comparing the error matrix (Confusion matrix), 1,250 1,083 of the data were correctly classified, showing an overall accuracy of about 86.64%, a high accuracy rate of about 72.46%, and a reproducibility of about 68.73%, which is the proportion of the text correctly predicted among actual related texts. to be.

Here, the precision is the ratio of texts that are actually related to the subject in the experimental results (Equation 1), and the recall is the ratio of texts that are actually classified as experimental results among the texts related to the subject (math Equation 2). For reference, the following Equation 3 is an expression representing accuracy, and the following Table 4 is a confusion matrix for explaining Equations 1 to 3.

LDA Topic modeling (S30)

Next, 200 data from the labeling result and the predicted experimental result through the SVM model are all related (class: 1) to the storage device 30, and the subject extraction module 14 is 200 LDA (Latent Dirichlet Allocation) is performed on dog data to extract topics and keywords for each topic. That is, as an embodiment of the present invention, only nouns are extracted from the article text, and LDA is performed with 200 data that are related to the category “Artificial Intelligence” in SVM.

Here, LDA is a probability model algorithm that extracts meaningful topics from a bundle of structured text data. Based on the distribution of the number of words for each topic, which is known in advance as a probability model for which topics exist in each document for a given document. It is one of the topic modeling methods that can predict which topics the document covers together by analyzing the distribution of the number of words found in the document.

In order to create an LDA model, 200 texts of nouns are first generated in DTM, and the number of times each word is mentioned in DTM is expressed in DTM, and then an LDA model is generated. The parameters used to perform the LDA model are num_topics and passes,'num_topics' is to determine the number of potential topics, and'passes' is to determine the number of times to repeat the model and the number of passes increases. Become sophisticated.

In the present invention, in one embodiment,'num_topics' is set to 10, and'passes' is set to 1,000 to perform LDA. The following Table 5 is stored in the storage device 30 as an LDA result for the category'Artificial Intelligence' through the number of model iterative learning times of 1,000 times, and the probability value, which is a number in Table 5, is the probability that keywords for each subject will be distributed to the subject of the corresponding number. And the number of keywords is 5.

In Table 5 above, the keywords'Isedol, Go, Alpha Go, AI, and the Great Power' appear as the subject number 1, and inferring from these keywords, the subject number 1 can be said to be'Alphago and Lee Sedol's Baduk Power'. . Since March 9-15, 2016, the Go match between Alpha Go and Lee Sedol became a hot issue and attracted attention due to the confrontation between artificial intelligence and human beings, so the collection period in one embodiment of the present invention was July 22, 2015 Since many news articles related to this have been published in the period from July 22, 2016, it can be assumed that the same topic as #1 appeared. That is, when keywords for each subject are extracted, the subject can be determined based on the corresponding keywords.

Twitter Data preprocessing (S40)

In the present invention, Twitter is described as an example using SNS.

The Twitter data is collected by the Twitter data collection module 15 through the REST-based open API from the Twitter data DB, but the same period as the news data collection period (for example, July 22, 2015 to July 22, 2016) Data), and Twitter data (eg, 3,001,589) after removing the duplicated data is used.

In the present invention, only the'text, user_id, user_s_name, followers_count, retweet_count, favorite_count' fields are used as shown in Table 6 below among various property information provided by the Twitter open API.'text' is used to analyze the contents of the tweet written by the user. Select,'use_id' and'user_s_name' are selected to know the user's information, and'followers_count','retweet_count', and'favorite_count' are used to obtain the influence index.

In order to analyze Twitter text among the Twitter fields, the Twitter data pre-processing module 16 pre-processes only the'text' field, and pre-processes only the Twitters that contain the keyword for each subject extracted through the LDA, and the pre-processing process preprocesses the article text. As described above, terminology removal and morpheme analysis are performed using the KoNLPy package. On Twitter, only Korean nouns are extracted and stored in the storage device 30 in order to calculate text classification through SVM and subject similarity for obtaining KeyplayerRank (hereinafter KpRank).

SVM through Twitter Text classification (S50)

The Twitter SVM process by the Twitter text classification module 17 is performed after pre-processing the Twitters including the keywords for each subject extracted through the above-mentioned LDA. In one embodiment, the subject 1 of the LDA results among the 3,001,589 Twitter data. SVM is performed with a total of 5,666 tweets extracted using the keywords'Lee Sedol, Go, Alpha Go, AI, and Korea'. The reason for performing SVM on Twitter is to extract users who wrote text related to the subject on Twitter as well as classifying the article text. In other words, when you pull out texts that correspond to the keywords for each topic on Twitter, you often get texts that are not related to the subject. For example, if you pull out texts that contain the word'event', you are collecting'travel photos.' The same text may be extracted. To reduce this error, we classify the text.

Here, the Twitter SVM process goes through the same process as SVM in the article text, and only the training data and test data are different. In one embodiment, the training data used in Twitter is composed of 5,000 article data and 400 randomly selected Twitter data, and it is not suitable for DTM formation to apply the SVM model trained only on article data to Twitter, so Twitter data is added to training data. It is an expression that determines 5,599 data as test data, excluding 67 data of which null value is found as a result of morpheme analysis among 5,666 tweeters.

The following Table 7 is a SVM model trained as shown in Table 3 above.It is applied to 5,599 test data, and the result is a Twitter text classification result (Predicted) compared to the actual manually labeled result (Actual). As a result, 5,277 out of 5,599 were correctly classified and had a high accuracy of about 94.25%, while the proportion of the predicted text that was related showed a low accuracy rate of about 48.78%. However, the correctly predicted proportion of the relevant texts is about 98.05%, showing a high reproducibility.

Both the labeled result and the experimental result predicted through the SVM model are stored in the storage device 30 for 301 data that are found to be related to the subject (class: 1), and user classification is performed by subject.

User classification by subject ( S60,S70 )

Next, it is a user classification process by topic using LDA results.

As an example, when proceeding with the first theme of'Alphago and Lee Sedol's Go game' among the LDA results in Table 5 above, as follows.

That is, as described above, as a result of extracting Twitters corresponding to the subject related to'Alphago and Lee Sedol's Go,' 6,266 Twitters were extracted from a total of 3,001,589 Twitter data, and 6,266 data were removed from duplicate data, resulting in a total of 5,666 I got two Twitters. In addition, the user classification module 18 extracts and classifies the user from the 301 data (Table 7) obtained by performing the SVM prior to selecting the Twitter related to the subject.

Looking at the extracted 301 Twitter texts, there were 201 texts that retweets other users' content, about twice as many as 100 texts written by the user, but due to the nature of Twitter, retweets that carry other users' texts frequently Because it is done, there are many retweets of text that contain the content of other users. The retweeted text classifies the retweets separately because RT_count is the number of retweets of the retweets.

In the above 301 data, the user classification classifies the one who did the retweet and the one who did not. In one embodiment, 34 users who did not retweet and 127 users who did not retweet out of 172 total users, and then checked the ID of the retweeted user in the text of the user who retweeted, and As a result, 11 retweets did not exist in the Twitter database collected because the number of fetched Twitters was limited when collecting Twitters using the REST API. Therefore, the screen name after'RT @'is extracted from the text of the retweeted user and stored in the storage device 30 separately. At this time, since the number of followers of the user does not exist in the DB and is unknown, the number of followers of each user is collected through the Twitter API, and property information of a total of 172 Twitter users including missing users is obtained.

As described above, users including keywords extracted through LDA topic modeling are extracted and classified. The reason for classifying users by topic is that most of the retweets are among users who mention keywords. Among retweet user's attribute information, retweet_count of retweeted user shows retweet_count, so these users must be classified separately. The process of classifying users by the user classification module 18 is specifically illustrated in FIG. 3. As shown in FIG. 3, users to be classified in this process are divided into: ① original author, ② retweeted user, ③ retweeted user in DB, and ④ retweeted user not in DB. First, it is determined whether the created tweet is retweeted or not. If it is not retweeted, the attribute information of the original author ID (①) is extracted. Conversely, if it is a retweet, it is determined whether the retweeted user exists in the tweeter DB (that is, the storage device 30) of the present invention. If it is in the Twitter database, the attribute information of the retweeted user (②) is extracted. However, if there is no retweeted user in the DB (that is, the storage device 30), the attribute information of the retweeted user (④) is retrieved from the Twitter data DB using the Twitter REST API. When collecting Twitter, the number of tweets that can be retrieved per hour with the REST API is limited to 350, so retweets are sometimes missing. However, since the retweeted user may be retweeted by other users, the data should not be excluded from the present invention as a user with a certain influence. Therefore, if a user does not exist in the Twitter DB, the attribute information of the user is separately retrieved from the Twitter data DB using the REST API. Finally, the retweeted user (③) in the DB extracts the attribute information in the Twitter DB (that is, the storage device 30). The extracted attribute information and the classified user are stored in the storage device 30.

Influence index and subject similarity calculation module 19 uses the'Follwers_count, RT_count, Favorite_count' of users in the attribute information to develop and apply an equation for obtaining an impact index as in Equation 4 below.

는 사용자 i가 주제 내에서 쓴 각 트윗에 해당하는 각 리트윗된 횟수를 모두 더해 사용자 i가 주제 내에서 쓴 모든 트윗 수로 나눈 사용자 i의 평균 리트윗된 수이고,

는

와 같이 사용자 i의 평균 favorite 수를 나타내며,

와

는 주제 내의 전체 사용자들의 평균 리트윗된 수와 평균 favorite 수를 나타낸 것이다. 리트윗과 favorite의 경우 사용자가 쓴 텍스트별로 값이 모두 다르므로 평균값을 구하는 것이고, 사용자별 리트윗과 favorite 수를 전체 평균으로 나누는 것은 해당 사용자의 텍스트에 대한 가중치를 의미하는 것으로 사용자가 쓴 텍스트가 전체 평균값보다 리트윗과 favorite의 수가 많을수록 그 주제에서의 영향력은 커지는 것이다.here,

Is the average number of retweets for user i divided by the number of all tweets written by user i within the topic by adding up each number of retweets corresponding to each tweet written by user i in the topic,

The

Like this, it represents the average number of favorite users i,

Wow

Is the average number of retweets and average number of favorite users across the subject. In the case of retweet and favorite, the value is different for each text written by the user, so the average value is obtained. Dividing the number of retweets and favorites by user by the total average means the weight for the user's text. The greater the number of retweets and favorites than the overall average, the greater the influence on the subject.

는 앞서

과

를 구한 것과 같이 사용자 i가 주제 내에서 쓴 각 트윗에 해당하는 각 팔로워 수를 모두 더해 사용자 i가 주제 내에서 쓴 모든 트윗 수로 나눈 평균 팔로워 수를 의미하는 것이고,

는 전체 사용자들의 평균 팔로워 수를 의미하는 것이다. 전체 사용자의 평균보다 사용자 i의 팔로워 수가 높으면 높을수록 다른 트위터 사용자들이 이 사용자를 믿고 따른다는 의미로 볼 수 있다.Also,

Ahead

and

It means the average number of followers divided by the number of tweets that user i wrote in the topic by adding all the number of each follower corresponding to each tweet that user i wrote in the topic,

Means the average number of followers of all users. The higher the number of followers of user i than the average of all users, the more Twitter users can trust and follow this user.

As a result, the impact index is a retweet, an indicator that texts written by one user are being spread and spread by other users, and a favorite that tells other users that they like the text by clicking'Like', and following the user. It is an indicator of how much influence this user has by finding the percentage value plus the number of followers that can tell the number of other users.

Meanwhile, the influence index and the subject similarity calculation module 19 obtains the subject similarity using the existing'Jaccard Index' in order to obtain the similarity between Twitter texts written by users.

The Jaccard Index is a method of measuring the similarity between two sets. It has a value between 0 and 1, and if none of the same elements exist, it has a value of '0', and if they all match, it has a value of '1'. Equations 5 and 6 are as follows.

Here, A and B are a bunch of Twitter text (consisting only of nouns) written by each user. Accordingly, in one embodiment of the present invention, it is possible to obtain a similarity between the subjects among 172 users.

In the present invention, since the similarity between texts containing keywords is obtained, Jaccard Index is used, and as a result, nouns commonly appearing in Twitter texts between users are made into multiple sets, and then nouns common in multisets of all nouns through Jaccard Index It is possible to find the similarity of the topic through how many of them appear. Through the subject similarity, it is possible to check the degree of thematic similarity between users, and the more similar, the more interested the same subject.

Key player Detection (S80)

In the present invention, for the key player detection by subject, the key player detection module 20 defines KpRank (influence ranking of users by subject) that applies both the impact index and subject similarity, as shown in Equation 7 below.

The influence index and subject similarity in Equation 4 and Equation 5 are obtained and put into Equation 7. In addition, d means the probability that a Twitter user visits a Twitter page referring to a topic of interest within a specific topic, and the default value of 0.85 is used. The reason is that when it is not 0.85, it vibrates without convergence when calculating the KpRank value. For reference, as described above, the influence index already divides the influence of all users by the weight value.

In Equation 7, the previous term is a part for measuring the influence of the Twitter user i individual, and the latter term is a part for determining whether the Twitter user i is subjectively influential. Therefore, since d is 0.85, it can be seen that the subject influence ratio of Twitter user i is 0.85 and the individual influence ratio is 0.15.

Here, the influence index is used as a weight that increases the probability of other users visiting the Twitter page of p _i (Twitter user i), and the topic similarity increases the probability of visiting the Twitter page of users who are interested in the same topic. do.

Thus, p _i KpRank of (Twitter user i) is p _i (Twitter users i) and the weighted influence of the p _i (Twitter users i) and summation of the KpRank other user to refer to the same subject are connected by other users p _i (Twitter user i) is the value of the probability of staying on Twitter.

As a result, KpRank is an algorithm that finds relatively influential Twitter users within a specific topic. Finding relatively influential users within a specific topic is because the user has the most professional and knowledgeable information and knowledge in the topic. , KpRank refers to the user with the highest rank as a key player, and determines that this user is the most influential user in the subject.

In an embodiment of the present invention, to obtain KpRank values of each of ₁₇₂ Twitter users (p ₁ to p ₁₇₂ ) using Equation 7, the values of KpRank(p ₁ ) to KpRank(p ₁₇₂ ) are all used as initial values. It is set to 1/172, because the KpRank(p _i ) value is the sum of the normalized values of all other KpRank(p _j ) values, so adding the KpRank values of all users (p ₁ to p ₁₇₂ ) is 1 . Influence index of 172 users and user p _i The similarity of each subject between all other users (p _j ) is obtained, and the initial value of KpRank (1/172) of all users (p ₁ ~p ₁₇₂ ) is input to Equation (7). Then, KpRank values of 172 different from the initial value (1/172) are obtained, and Equation 7 is repeatedly calculated until the values of each of the 172 Twitter users KpRank converge to obtain the final KpRank values of all users.

Accordingly, the present invention is to find a key player on social media by using KpRank to which an influence index considering the similarity of a subject and a user's influence and ripple power acts as a weight for a relationship between people interested in a specific subject.

In addition, after the final key player for each subject is obtained, it can be visualized and displayed on a map, so that the key player corresponding to the subject that the user is looking for can be displayed for each region so that it can be easily recognized. For example, if you want to detect a key player in the theme of'Tteokbokki', you can detect the highest ranked key player by clicking on'View Rank' by region as shown in FIG.

Meanwhile, the series of processes may be performed by a program that directly codes an algorithm through a programming language (such as Python) or partially by a commercial program in order to perform it in a computer. In addition, in the term “text” in the present invention, all parts of speech are “nouns” as can be understood by the contents described in the examples.

The methodology of the present invention is applicable not only to Twitter but also to other social media (especially Instagram), and can be used for corporate influencer marketing, viral marketing, or policy analysis through big data analysis and public opinion formation. Through this, the present invention can be developed into an invention that finds a real-time key player when each subject becomes an issue in real time, including spatiotemporal elements.

10: Key player detection device by topic 11: News data collection module
12; News data pre-processing module 13: article text classification module
14: subject extraction module 15: Twitter data collection module
16: Twitter data pre-processing module 17: Twitter text classification module
18: User classification module
19: Impact Index and Subject Similarity Calculation Module
20: key player detection module 30: storage device

Claims (7)

(a) The news data collection module 11 collects news articles related to a specific category specified from the news data DB through crawling, and the news data pre-processing module 12 removes stop words from the'TEXT' field of the collected news articles And morphological analysis to extract the noun pre-processing step (S10) and;
(b) Article text classification module 13 selects features to be used for analysis for classifying news article texts, indexes texts expressing features by feature weights, and then extracts noun article texts in step (a). Dividing the training data and the test data at a constant rate to train the SVM through the training data, and classifying the text through the trained SVM model (S20);
(c) subject extracting module 14 extracts subject and subject-specific keywords by performing LDA on data found to be related to a specific category by text classification through the SVM model in step (b) (S30). Wow;
(d) The Twitter data collection module 15 collects Twitter data of the same period as the news article collection period from the Twitter data DB, and the Twitter data pre-processing module 16 collects Twitter data by subject extracted through the LDA. A pre-processing step (S40) of extracting nouns by removing idioms and analyzing morphemes in the'text' field of Twitter data including keywords;
(e) Twitter text classification module 17 trains the SVM through training data in which Twitter data is added to a certain number of article data, and Twitter data including subject-specific keywords in step (d) is used as test data. Classifying the text through the trained SVM model (S50);
(f) the user classification module 18 classifying users for data that is found to be related to the subject by text classification through the SVM model in step (e) (S60);
(g) an influence index and subject similarity calculation module 19 calculating an impact index using attribute information by the user classification, and calculating subject similarity using a'Jaccard Index' (S70), and
(h) characterized in that the key player detection module 20 comprises the step of selecting a user having the highest KpRank value among the users as the key player for each subject (S80) using the influence index and subject similarity (S80). How to detect key players by topic on social media.
According to claim 1,
The step (b),
(b1) labeling is performed on article texts extracted from nouns according to whether they are related to a category, and a feature selection step of selecting words to be used for classifying texts using a literature frequency;
(b2) generating an article text DTM applying TF-IDF as a feature weight using the selected features, and
(b3) Using the DTM, the training data and test data are divided into a certain ratio and applied to test data as an SVM model trained through the training data to classify article texts, and applied to test data as a trained SVM model to apply articles to test data. A method for detecting key players by topic on social media using KeyplayerRank, which is characterized by comparing the results of classifying the text with actual labeled results, and extracting the correctly classified ones.
According to claim 1,
In the step (c), the LDA model is generated by generating a data text in the DTM where both the labeled result and the result of classifying the article text through the SVM model are related to a specific category, and the parameters necessary to perform the LDA model A method of detecting key players by topic on social media using KeyplayerRank, characterized by extracting keywords for each topic by setting the'num_topics','passes', and'number of keywords'.
According to claim 1,
The step (e), by applying the test data to the trained SVM model, characterized by extracting the correct classification by comparing the results of the classification of the Twitter text and the actual labeled results, the key for each topic on social media using KeyplayerRank Player detection method.
According to claim 1,
The step (f) is classified as a retweet user and a retweet user to retrieve attribute information of Twitter users, but if the retweet user is not stored in the storage device 30, Twitter using the Twitter REST API Key player detection method for each topic on social media using KeyplayerRank, characterized in that the attribute information of the user is retrieved from the data DB.
According to claim 1,
In step (g),
The influence index is the following equation using'Follwers_count, RT_count, Favorite_count' among the attribute information of users,

(here,

Is the average number of retweets for user i divided by the number of all tweets written by user i within the topic by adding up each number of retweets corresponding to each tweet written by user i in the topic,

The

Like this, it represents the average number of favorite users i,

Wow

Is the average number of retweets and average favorites of all users in the subject,

The

and

The average number of followers divided by the total number of tweets that user i has written within the topic by adding up the number of each follower corresponding to each tweet that user i wrote within the topic, as obtained by

Is the average number of followers of all users),
The subject similarity is the following equation,

(Here, A and B is a bunch of Twitter text (consisting only of nouns) written by each user), characterized in that it is calculated, the key player detection method for each topic on social media using KeyplayerRank.
According to claim 1,
In step (h),
The KpRank value is the following equation,

(Where d is a probability that a Twitter user visits a Twitter page referring to a topic of interest within a specific topic), the method for detecting a key player by topic on social media using KeyplayerRank.

Images