Influencer selection support system, influencer selection support method, and computer program

Abstract

To provide an influencer selection support system, a method and a computer program for supporting selection of an influencer who accurately predicts an evaluation value for submission information and who is suitable for requesting submission of the submission information.SOLUTION: An influencer selection support system 1 comprises: an influencer information storage unit 11A; a submission information input unit 23 that receives input of submission information scheduled to be submitted; a submission date input unit 25 that receives input of a submission scheduled date of the submission information; an influencer selection unit 27 that selects one or more influencers who request submission of the submission information; an evaluation value calculation unit 11B that calculates a prediction evaluation value for the submission information to output in a case where the submission information is submitted by the one or more influencers on the basis of the submission information, the submission scheduled date and attribute information of the selected one or more influencers; and a prediction evaluation value display unit 31 that displays the prediction evaluation value for each of the one or more influencers.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an influencer selection support system, an influencer selection support method, and a computer program. [Background technology]

[0002] As social networking services (SNS) become more widespread, social media marketing, which encourages consumer purchases through SNS, is becoming increasingly popular. Consequently, account management tools for managing both your own account and your competitors' accounts are becoming increasingly popular. For example, "Social Insight," offered by User Local Inc., allows you to manage follower growth rates and popular posts, and also analyzes which days and times of the week are most likely to generate engagement (responses from other SNS users).

[0003] In this context, influencer marketing, which utilizes the influence of so-called "influencers," people who have influence on social media, is gaining attention. For example, attempts are being made to predict the effect of information posted by influencers (specifically, the number of "likes," which are the reactions of viewers to the posted information), as shown in Non-Patent Document 1. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Konosuke Yamazaki, Taketoshi Ushiama, "Predicting the Number of Likes for Influencer Recommendation in Social Media Advertising," 13th Data Engineering and Information Management Forum (DEIM Forum 2021), March 3, 2021, I31-3 Summary of the Invention [Problem to be solved by the invention]

[0005] Even if the same information is posted, the effectiveness of the information posted will vary depending on who posts it. Therefore, when requesting information from an influencer, there is a need to compare the effectiveness of the information posted by the influencer.

[0006] An object of the present invention is to provide an influencer selection support system, an influencer selection support method, and a computer program for accurately predicting the evaluation value of posted information posted by influencers and supporting the selection of influencers who are suitable to be requested to post posted information. [Means for solving the problem]

[0007] In order to achieve the above-mentioned object, the influencer selection support system of the present disclosure is intended to support the selection of influencers who are suitable to be requested to post posted information, and is equipped with an influencer information storage unit that stores attribute information of multiple influencers who post on SNS, a post information input unit that accepts input of the posted information to be posted, a post date and time input unit that accepts input of the planned post date and time of the posted information, an influencer selection unit that selects one or more influencers to be requested to post the posted information from the multiple influencers stored in the influencer information storage unit, an evaluation value calculation unit that calculates and outputs a predicted evaluation value for the posted information when the posted information is posted by the one or more influencers based on the posted information, the planned post date and time, and the attribute information of the one or more influencers selected by the influencer selection unit, and a predicted evaluation value display unit that displays the predicted evaluation value for each of the one or more influencers. [Effects of the Invention]

[0008] According to the present disclosure, by referring to the predicted evaluation value displayed in the predicted evaluation value display section, it is possible to easily select an influencer suitable for requesting information transmission. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a block diagram illustrating an example of a configuration of an influencer selection support system according to an embodiment. [Figure 2] 1 is an example of a website of the influencer selection support system displayed on a display screen of a user terminal. [Figure 3] 10 is a flowchart showing a calculation process in an evaluation value calculation unit. [Figure 4] FIG. 1 is a diagram illustrating the branching structure of a trained model. [Figure 5] FIG. 1 is a diagram illustrating the structure of ensemble learning of trained models. [Figure 6] 10 is a display example of predicted evaluation values. [Figure 7] This is a graph (scatter plot) showing the correlation between the predicted number of likes and the actual number of likes, divided into categories based on trained models. [Figure 8] This is a graph (scatter plot) showing the correlation between the number of likes predicted by the base model and the actual number of likes. [Figure 9] 10 is a graph (scatter plot) showing the correlation between the predicted number of likes by the processing model and the actual number of likes. [Figure 10] This is a graph (scatter plot) showing the correlation between the predicted number of likes by the ensemble + processing model and the actual number of likes. [Figure 11] This is a graph (scatter plot) showing the correlation between the predicted number of likes by the branching model and the actual number of likes. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of an influencer selection support system, an influencer selection support method, and a computer program according to the present invention will be described in detail with reference to the accompanying drawings.

[0011] FIG. 1 is a block diagram showing the configuration of an embodiment of an influencer selection support system of the present invention, and FIG. 2 is an example of a website of the influencer selection support system displayed on the display screen of a user terminal.

[0012] As shown in FIG. 1, the influencer selection support system 1 of this embodiment is made up of a server system 3 and a user terminal 5 connected to a telecommunications line NW.

[0013] The server system 3 includes a front-end server 7, a back-end server 9, and an API server 11. The front-end server 7 acquires input data from a user terminal 5 and displays output data on the user terminal 5. The back-end server 9 routes appropriate requests to the API server 11 based on requests from the front-end server 7. The API server 11 includes an influencer information storage unit 11A and an evaluation value calculation unit 11B, and is configured to calculate a predicted evaluation value for posted information when the information is posted by an influencer based on a request from the back-end server 9, and return the predicted evaluation value to the back-end server 9.

[0014] The influencer information storage unit 11A stores attribute information of multiple influencers who post on SNS (Social Networking Service). Examples of SNS include "Twitter (registered trademark)" provided by Twitter Inc. (USA), "Instagram (registered trademark)" provided by Meta Platforms Inc. (USA), "YouTube (registered trademark)" provided by Google Inc. (USA), and "TikTok (registered trademark)" provided by ByteDance Inc. (China). In this embodiment, as an example, the influencer selection support system 1 targets "Instagram." Furthermore, "influencer" generally refers to "a person who has influence on SNS," but in this specification, it refers to "a public account with a history of posting on SNS."

[0015] In this embodiment, the "attribute information of influencers" stored in the influencer information storage unit 11A is the following information about influencers obtained by collecting and analyzing information about the influencers. (1) Number of followers: The number of SNS user accounts (followers) that follow the influencer (i.e., those who have registered to view the influencer's posts, etc.). (2) Number of Followers: The number of SNS user accounts that the influencer follows. (3) Average number of likes: The number obtained by dividing the total number of times that SNS users pressed the "Like button" (the specific name varies depending on the SNS, but it is a button that SNS users press to indicate that they like, enjoy, or support the content of the post) on past posts posted by the influencer on SNS (hereinafter referred to as "Likes") by the total number of posts by the influencer. (4) Trustworthiness (Fake Follower Rate): The number obtained by dividing the number of followers of the influencer who are inactive users who have not used SNS (Instagram) for a certain period of time by the number of followers of the influencer. (5) Engagement: The sum of the number of likes on the influencer's past posts and the number of comments from SNS users on the influencer's past posts. (6) Engagement rate: The number obtained by dividing "engagement" by the total number of posts by the influencer.

[0016] The evaluation value calculation unit 11B includes a trained model (regression model) LM obtained by training information of a plurality of influencers as training data. Specifically, the trained model LM is generated by performing machine learning using, as training data, attribute information of a plurality of influencers and posting information related to past posts of a plurality of influencers as explanatory variables, and evaluation information on the posting information of a plurality of influencers as a target variable.

[0017] The explanatory variable "attribute information of multiple influencers" is the same as above. "Post information about past posts of multiple influencers" is the following information.

[0018] (1) Number of hashtags: The number of hashtags in the influencer’s past posts (the number of keywords tagged with # (hash mark)) (2) Text length: Number of characters in the posted information (3) Posting date and time: The day of the week on which the post information was posted (Sunday is 0, Monday is 1, Tuesday is 2, etc., and the number is incremented by 1 for each day of the week), the posting time (0 to 23), and the posting minute (0 to 59).

[0019] The objective variable "evaluation information on posted information by multiple influencers" is the "number of likes" for the posted information.

[0020] The machine learning uses the LightGBM (Light Gradient Boosting Machine) algorithm. In this embodiment, the learning data is divided into three ranges of the average number of likes of influencers: (i) less than 1,000, (ii) between 1,000 and 10,000, and (iii) 10,000 or more, and multiple categories (in this embodiment, divided-trained models LM1 to LM3) are prepared that are trained independently.

[0021] Of course, to improve the efficiency of the learning process, a tuning process may be performed before the learning process. The tuning process is a process of tuning the hyperparameters of a model. Known parameter tuning methods include Bayesian optimization, grid search, random search, and metaheuristic search. A benchmark test can be performed to determine the tuning method taking into account the learning speed of the model.

[0022] The user terminal 5 is a computer terminal used by a user. Specific examples include, but are not limited to, mobile phones, smartphones, tablet devices, and personal computers (notebook PCs and desktop PCs). The user terminal 5 has a communication unit 13, an input unit 15, a data storage unit 17, a control unit 19, and a display screen 21. The communication unit 13 is connected to the telecommunications line NW and is used to communicate with the server system 3. The input unit 15 is used to input data, and various types of input units may be used depending on the user terminal, such as a keyboard, a mouse, or a touch panel using the display screen 21.

[0023] In this embodiment, by accessing the server system 3 using the web browser of the user terminal 5, the website shown in Fig. 2 is displayed on the display screen 21 of the user terminal 5, allowing the user to use the influencer selection support system 1. Of course, a computer program for using the influencer selection support system 1 may be installed in advance on the user terminal 5, and the influencer selection support system 1 may be made available by starting the computer program.

[0024] As shown in FIG. 2, the website of the influencer selection support system 1 includes a posting information input unit 23, a posting date and time input unit 25, an influencer selection unit 27, and an execute button 29. The posting information input unit 23 is a unit for inputting posting information to be posted. In this embodiment, the posting information includes a text document and an image. The text document to be posted is input into a text box, and an image is selected. The posting date and time input unit 25 is a unit for inputting the planned posting date and time (day of the week, hour, minute) of the posting information, i.e., the date and time (day of the week, hour, minute) at which the influencer is desired to post the posting information. Here, the date and time (day of the week, hour, minute) are input in a pull-down format. The influencer selection unit 27 is a unit for selecting one or more influencers to be requested to post posting information from among multiple influencers stored in the influencer information storage unit 11A. A desired influencer can be selected from a list of influencers displayed by pressing the "+" button.

[0025] After entering the posting information and posting date and time and selecting an influencer, pressing the execute button 29 inputs the posting information, the planned posting date and time, and the attribute information of the selected one or more influencers into the evaluation value calculation unit 11B, and calculates a predicted evaluation value for the posting information if the posting information is posted by one or more influencers.

[0026] In this embodiment, the predicted evaluation value is calculated as the "predicted number of likes" for the posted information when the posted information is posted by an influencer. This is based on the idea that posted information with a large number of "likes" spreads widely among SNS users and therefore has a high dissemination effect (posted information promoting a product or service has a high advertising effect).

[0027] 3 is a flowchart showing the calculation process in the evaluation value calculation unit 11B. When information is input as described above (step ST1), the evaluation value calculation unit 11B quantifies the input post information (step ST2). Specifically, the quantification is performed as follows (note that the definition of each numerical value is the same as in the above-mentioned "post information related to past posts by multiple influencers"). (1) Post information → (a) number of hashtags, (b) length of text (2) Posting date and time → (c) Posting day, (d) Posting time, (e) Posting amount

[0028] In addition, the following attribute information of the selected one or more influencers is read from the influencer information storage unit 11A. (3) Influencer demographic information: (f) number of followers, (g) number of followings, (h) average number of likes, (i) trustworthiness, (j) engagement, (k) engagement rate

[0029] Next, the evaluation value calculation unit 11B performs preprocessing on the above numerical values ​​(a) to (k) (step ST3). The preprocessing is a power normal transformation and / or normalization process selected according to each numerical value. "Power normal transformation" is a process that brings numerical values ​​closer to a normal distribution. "Normalization process" is a process that adjusts the scale of data to make it easier to handle. In this embodiment, the specific processes are as follows. (a) Number of hashtags → Min-Max normalization process using predetermined maximum and minimum values (b) Text length → Yeo-Johnson transformation using a predetermined lambda constant, followed by Min-Max normalization using predetermined maximum and minimum values (c) Posting day of the week, (d) Posting time, (e) Posting minutes → (Posting day of the week x 60 x 24) + (Posting time x 60) + Posting minutes = post_time, then Min-Max normalization using the specified maximum and minimum values (f) Number of followers, (g) Number of following → Yeo-Johnson transformation using a specified lambda constant, followed by Min-Max normalization using specified maximum and minimum values (h) Average number of likes, (i) Trust, (j) Engagement, (k) Engagement rate → Yeo-Johnson transformation using a specified lambda constant, followed by Min-Max normalization using specified maximum and minimum values

[0030] Here, the "Yeo-Johnson transformation" is a transformation process in which the following equation is applied to elements xi ∈ X (i = 1, 2, , n) in an arbitrary input data set X = {x1, x2 , xn} to convert them into an output data set Y = {y1, y2 , yn}.

number

[0031] Furthermore, "Min-Max normalization processing" refers to processing the following equation for elements xi ∈ X (i = 1, 2, , n) in an arbitrary input data set X = {x1, x2 , xn}, and normalizing the output data set Y = {y1, y2 , yn} so that the maximum value is 1 and the minimum value is 0.

number

[0032] Next, the evaluation value calculation unit 11B performs calculations using the preprocessed numerical values. In this embodiment, as described above, three types of categorized trained models LM1 to LM3 are prepared, and as shown in FIG. 4, calculations are performed using categorized trained models LM1 to LM3 branched according to the selected influencer's (h) average number of likes. Specifically, (i) categorized trained model LM1 is a trained model selected when the average number of likes is less than 1,000, (ii) categorized trained model LM2 is a trained model selected when the average number of likes is between 1,000 and 10,000, and (iii) categorized trained model LM3 is a trained model selected when the average number of likes is 10,000 or more, and a predicted evaluation value is output using the trained model for the category to which the selected influencer belongs.

[0033] 5, each of the segmented trained models LM1 to LM3 of this embodiment is an ensemble trained model constructed by connecting multiple (five in this embodiment) sub-trained models (sub-regression models) in parallel to improve accuracy. The average value of the outputs calculated by each sub-trained model is output as the calculation result.

[0034] Then, post-processing is performed on the calculation results (step ST5). The post-processing is inverse normalization and inverse power normalization. Specifically, the inverse normalization is Min-Max inverse normalization using predetermined maximum and minimum values, and the inverse power normalization is an inverse Yeo-Johnson transform using a predetermined lambda constant.

[0035] Here, the "Min-Max denormalization process" is a conversion process that converts the output data set Y into a data set X' using the following formula.

number

[0036] The "inverse Yeo-Johnson transformation" is a transformation process that transforms the output data set Y into a data set X' using the following equation.

number

[0037] FIG. 6 is an example of a display of predicted evaluation values. For each selected influencer, calculation (step ST4) and post-processing (step ST5) are performed, and the predicted evaluation value is output to the display screen 21 of the user terminal 5 and displayed on the predicted evaluation value display unit 31 (step ST6). In this embodiment, the predicted evaluation value is the "predicted number of likes" predicted when the influencer posts information. As shown in FIG. 6, the predicted evaluation value display unit 31 is provided for each selected influencer, and the user can compare and consider which influencer to request to post will have the greatest communication effect by comparing and considering the displayed "predicted number of likes," and select an influencer.

[0038] <Verification of prediction accuracy> For reference, the accuracy of the predicted number of likes of the trained model (hereinafter referred to as the "embodiment model") provided in the evaluation value calculation unit 11B of the influencer selection support system 1 is shown. FIG. 7 is a graph (scatter plot) showing the correlation between the predicted number of likes and the actual number of likes, divided into categorized trained models LM1 to LM3. In FIG. 7, the actual number of likes is plotted on the horizontal axis, and the predicted number of likes is plotted on the vertical axis. In the figure, a line with a slope of 1 passing through 0 indicates a prediction accuracy of 100% (a case where the relative error is 0), and the deviation from the line corresponds to the prediction error. The relative error (median) was 13% for categorized trained model LM1, 16% for categorized trained model LM2, and 15% for categorized trained model LM3, resulting in a favorable average result of 15%.

[0039] To obtain the data, we used the K-fold cross-validation method, which is used to evaluate the generalization performance of a learning model. In K-fold cross-validation, the learning data is divided into K parts (K is an integer greater than or equal to 2), one of which is used as validation data and the remaining (K-1) as training data. After building a trained model using the training data, it is evaluated using the validation data. Generalization performance is evaluated by repeating the process from dividing the learning data to evaluation K times.

[0040] 8 to 11, it will be shown that the accuracy of the predicted number of likes of the trained model provided in the evaluation value calculation unit 11B of the influencer selection support system 1 is higher than that of other trained models. However, the example shown in Figs. 8 to 11 is not the prior art, but an example of an embodiment of the present invention that falls within the scope of the technical idea of ​​the present invention.

[0041] It is believed that the factors that contribute to improving the prediction accuracy of the trained model equipped in the evaluation value calculation unit 11B are (A) whether or not preprocessing (step ST3) and postprocessing (step ST5) are performed (hereinafter referred to as "preprocessing / postprocessing"), (B) whether or not ensemble learning is performed, and (C) whether or not a branching structure of the trained model based on the average number of likes (step ST4) (hereinafter referred to as "branching structure") is present.Therefore, in order to demonstrate the effect of each, we prepared the trained models for comparison below and calculated the predicted number of likes. (a) Base model: (A) Pre-processing and post-processing, (B) Ensemble learning, (C) Model without branching structure (b) Processing model: (A) Model in which only pre-processing and post-processing were performed (C) Ensemble + processing model: (A) Pre-processing and post-processing, and (B) a model that has undergone ensemble learning. (D) Branching model: (A) Pre-processing and post-processing, (B) Ensemble learning, and (C) Branching structure are implemented. In (C), the trained model is branched into three conditions based on the influencer's "number of followers": (i) less than 10,000, (ii) between 10,000 and 100,000, and (iii) 100,000 or more.

[0042] Similar to Figure 7, Figures 8 to 11 are scatter plots with the actual number of likes on the horizontal axis and the predicted number of likes on the vertical axis. To obtain the data, we used the K-fold cross-validation method, as in the case of the model of the embodiment. The relative error (median) of each model is as follows: (a) Base model: 28% (b) Treatment model: 17% (C) Ensemble + processing model: 16% (d) Branching model: The segmented trained model LM1 was 13%, the segmented trained model LM2 was 20%, and the segmented trained model LM3 was 19%, with an average of 17%.

[0043] By comparing the results of the (A) base model and the (B) processing model, it can be seen that (A) pre-processing and post-processing contribute to improving prediction accuracy. By comparing the results of the (B) processing model and the (C) ensemble + processing model, it can be seen that (B) ensemble learning contributes to improving prediction accuracy. By comparing the results of the (C) ensemble + processing model, the (D) branching model, and the embodiment model, it can be seen that (C) branching structure contributes to improving prediction accuracy depending on the branching conditions.

[0044] From the above, it can be seen that the embodiment model is a trained model suitable for calculating the predicted number of likes for information posted by an influencer.

[0045] Although the embodiments of the present invention have been specifically described above, the present invention is not limited to these embodiments, and it goes without saying that modifications are possible within the scope of the technical concept of the present invention.

[0046] For example, the trained model does not necessarily have to include all of the above information as explanatory variables in the training data, and other attribute information or post information of influencers may also be included as explanatory variables. Furthermore, pre-processing and post-processing do not necessarily have to be performed, and pre-processing and post-processing using other methods may of course be performed. Furthermore, the trained model does not necessarily have to be segmented, and it is of course possible to prepare a segmented trained model based on other segments. [Explanation of symbols]

[0047] 1. Influencer selection support system 3. Server System 5. User terminal 7 Front-end Server 9 Backend Server 11 API Server 11A Influencer information storage unit 11B Evaluation value calculation unit 13 Communications Department 15 Input section 17 Data storage unit 19 Control Unit 21 Display screen 23 Posting information input section 25 Posting date and time input section 27 Influencer Selection Department 29 Execute button 31 Predicted evaluation value display section LM pre-trained model LM1~LM3 Partition-trained models

Claims

1. An influencer selection support system for supporting the selection of an influencer suitable for requesting posting of posting information, an influencer information storage unit that stores attribute information of a plurality of influencers who post on the SNS; a posting information input unit that accepts input of the posting information to be posted; a posting date and time input unit that accepts input of a scheduled posting date and time of the posted information; an influencer selection unit that selects one or more influencers to be requested to post the posted information from among the plurality of influencers stored in the influencer information storage unit; an evaluation value calculation unit that calculates and outputs a predicted evaluation value for the posted information when the posted information is posted by the one or more influencers, based on the posted information, the scheduled posting date and time, and attribute information of the one or more influencers selected by the influencer selection unit; a predicted evaluation value display unit that displays the predicted evaluation value for each of the one or more influencers, An influencer selection support system characterized by:

2. 2. The influencer selection support system according to claim 1, the evaluation value calculation unit is provided with a trained model generated by performing machine learning using, as learning data, attribute information of the plurality of influencers and post information related to past posts of the plurality of influencers as explanatory variables, and evaluation information on the post information of the plurality of influencers as a target variable; An influencer selection support system characterized by:

3. 3. The influencer selection support system according to claim 2, The attribute information of the plurality of influencers includes at least one of the number of followers, the number of followings, the average number of likes, the trustworthiness, the engagement, and the engagement rate of the plurality of influencers; The post information regarding past posts of the plurality of influencers includes at least one of the number of hashtags included in the post, the number of characters included in the post, and the posting date and time of the post. An influencer selection support system characterized by:

4. 4. The influencer selection support system according to claim 2, The evaluation value calculation unit, before calculation, performs a power normal transformation to bring the digitized content of the posted information, the scheduled posting date and time, and the attribute information of the one or more influencers selected by the influencer selection unit closer to a normal distribution, and then performs a normalization process; and after calculation, performs an inverse power normal transformation on the calculation result, and outputs the result as the predicted evaluation value. An influencer selection support system characterized by:

5. 5. The influencer selection support system according to claim 4, The power normal transformation is a Yeo-Johnson transformation using a predetermined lambda constant, the normalization process is a Min-Max normalization process using predetermined maximum and minimum values, The denormalization process is a Min-Max denormalization process using predetermined maximum and minimum values, The inverse power normal transformation is the inverse Yeo-Johnson transformation using a predetermined lambda constant An influencer selection support system characterized by:

6. In the influencer selection support system described in any one of claims 2 to 5, the trained model of the evaluation value calculation unit includes a plurality of segmented trained models that are trained independently by dividing the training data into numerical ranges of attribute information of the plurality of influencers, The evaluation value calculation unit outputs the predicted evaluation value using the category-trained model of the category to which the attribute information of the one or more influencers selected by the influencer selection unit belongs. An influencer selection support system characterized by:

7. 7. The influencer selection support system according to claim 6, The plurality of segmented trained models are segmented by numerical ranges of the average number of likes of the plurality of influencers. An influencer selection support system characterized by:

8. An influencer selection support method using an influencer selection support system for supporting selection of an influencer suitable for requesting posting of posted information, the influencer selection support system including an influencer information storage unit that stores attribute information of a plurality of influencers who post on SNS, A step of accepting input of the posting information to be posted; receiving an input of a scheduled posting date and time of the posted information; receiving a selection of one or more influencers to be requested to post the posted information from among the plurality of influencers stored in the influencer information storage unit; calculating a predicted evaluation value for the posted information when the posted information is posted by the one or more influencers, based on the posted information, the scheduled posting date and time, and attribute information of the selected one or more influencers; displaying the predicted evaluation value for each of the one or more influencers; having 1. A method for supporting influencer selection, comprising:

9. A computer program for causing a computer to execute the influencer selection support method according to claim 8.

Images