Display order training method and display order estimation method

Abstract

In the present invention, a computer executes a training procedure for training a machine learning model in which attribute data of a user referring to contribution information on a certain commodity, attribute data of a contributor of the contribution information, and data indicating the degree of affirmation of the contribution information are inputted, so that an output from the machine learning model approaches a display order given as a correct answer to the contribution information. Accordingly, the contribution information on the commodity can be presented in a mode considering recognition processing of the user.

Description

Display order learning method and display order estimation method

[0001] The present invention relates to a display order learning method and a display order estimation method.

[0002] It is common for people considering purchasing a product online to refer to reviews (posted information about the product) by other buyers of the product before making a purchase.

[0003] These reviews are displayed in order by a system set up by the online shop, unless the user voluntarily changes the display order.

[0004] Meanwhile, research into cognitive processing has shown that there are gender differences in the types of opinions people refer to. It has been reported that men tend to refer to opinions of the same attributes or positive opinions to confirm whether their own opinions are correct, while women tend to refer to opinions of different attributes or negative opinions to confirm whether they are within their own acceptable range (Non-Patent Document 1).

[0005] Janne Chung et al., "Gender differences in information processing: An empirical test of the hypothesis-confirming strategy in an audit context", Accounting and Finance 38 (1998) pp.265-279

[0006] Even when it comes to posted information such as word-of-mouth, displaying it in a way that takes into account the user's cognitive processing may be able to further increase the user's purchasing motivation.

[0007] The present invention has been made in consideration of the above points, and has an object to make it possible to present posted information about products in a manner that takes into consideration the cognitive processing of a user.

[0008] To solve the above problem, a computer executes a learning procedure to train a machine learning model that receives input of attribute data of a user who references posted information about a certain product, attribute data of the poster of the posted information, and data indicating the degree of positivity of the posted information, so that the output from the machine learning model approaches the display order given as the correct answer for the posted information.

[0009] It is possible to present posted information about products in a manner that takes into consideration the cognitive processing of the user.

[0010] FIG. 1 is a diagram illustrating an example of a hardware configuration of the display order estimation device 10 according to the first embodiment. FIG. 2 is a diagram illustrating an example of a functional configuration of the display order estimation device 10 during learning according to the first embodiment. FIG. 3 is a diagram illustrating an example of a configuration of learning data of a display order estimation model m1 according to the first embodiment. FIG. 4 is a diagram illustrating an example of a functional configuration of the display order estimation device 10 during inference according to the first embodiment. FIG. 5 is a diagram illustrating an example of post-replacement word-of-mouth text data. FIG. 6 is a diagram illustrating an example of a functional configuration of the display order estimation device 10 during learning according to the second embodiment. FIG. 7 is a diagram illustrating an example of a configuration of learning data of a display order estimation model m1 according to the second embodiment. FIG. 8 is a diagram illustrating an example of a functional configuration of the display order estimation device 10 during inference according to the second embodiment.

[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. Fig. 1 is a diagram showing an example of the hardware configuration of a display order estimation device 10 according to a first embodiment. The display order estimation device 10 in Fig. 2 includes a drive device 100, an auxiliary storage device 102, a memory device 103, a processor 104, and an interface device 105, which are all interconnected by a bus B.

[0012] A program that realizes the processing in the display order estimation device 10 is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 storing the program is set in the drive device 100, the program is installed from the recording medium 101 to the auxiliary storage device 102 via the drive device 100. However, the program does not necessarily have to be installed from the recording medium 101, but may be downloaded from another computer via a network. The auxiliary storage device 102 stores the installed program as well as necessary files, data, etc.

[0013] When an instruction to start the program is received, the memory device 103 reads and stores the program from the auxiliary storage device 102. The processor 104 is a CPU or a GPU (Graphics Processing Unit), or a CPU and a GPU, and executes functions related to the display order estimation device 10 in accordance with the program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network.

[0014] 2 is a diagram showing an example of the functional configuration of the display order estimation device 10 during learning according to the first embodiment. In FIG. 2, the display order estimation device 10 includes a classification unit 11, a display order estimation model learning unit 12, and a display order estimation model m1. These units are realized by processes executed by a processor 104 in accordance with one or more programs installed in the display order estimation device 10.

[0015] The classification unit 11 receives learning word-of-mouth text data as input and outputs learning word-of-mouth classification data.

[0016] The word-of-mouth text data refers to word-of-mouth sentences (word-of-mouth text data) posted about a certain product on a certain shopping site. The word-of-mouth text data is an example of posted information about a product. The learning word-of-mouth text data refers to word-of-mouth text data collected from the word-of-mouth text data for training the display order estimation model m1.

[0017] Classification data refers to data that numerically represents the positivity / negativity of the content of a word-of-mouth (i.e., data that indicates the degree of positivity of a word-of-mouth (classifying the degree into a predetermined number). Word-of-mouth classification data expresses positivity / negativity (degree of positivity) using real numbers or 1-hot, etc., so that it becomes an n-level value. Learning word-of-mouth classification data is prepared in association with each learning word-of-mouth text data, and is classification data that expresses the positivity / negativity (degree of positivity) for the associated learning word-of-mouth text data.

[0018] For example, the classification unit 11 acquires the positivity / negativity of each piece of training word-of-mouth text data using an existing text classification model such as Mecab (Reference 1). At this time, the classification unit 11 converts the values ​​extracted by Mecab or the like into n levels, taking into account the overall variance, and outputs the converted values ​​as classification data. The output classification data is associated with the input training word-of-mouth text data as training word-of-mouth classification data.

[0019] If a rating value (such as "5 stars") is assigned to the learning word-of-mouth text, that value may be used as the learning word-of-mouth classification data, but to prevent discrepancies between the rating value and the actual text content, it is preferable to use the classification unit 11. Alternatively, the learning word-of-mouth classification data may be manually labeled for the learning word-of-mouth text data.

[0020] The display order estimation model learning unit 12 uses learning data of the display order estimation model m1 to learn the display order estimation model m1.

[0021] 3 is a diagram showing an example of the configuration of training data for the display order estimation model m1 in the first embodiment. As shown in Fig. 3, one piece of training data (one piece of training review text data corresponding to one training review text data) is a set of data including "training user attribute data" corresponding to the training review text data, "training poster attribute data" corresponding to the training review text data, "training review classification data" corresponding to the training review text data, and "training review order vector" corresponding to the training review text data.

[0022] The display order estimation model m1 is a machine learning model that inputs user attribute data, poster attribute data, and review classification data to output a review order vector that matches the user's cognitive characteristics. A conventional DNN such as that described in Reference 2 can be used to train the display order estimation model m1.

[0023] User attribute data is data that numerically represents the attributes (age, gender, place of residence, etc.) of users who refer to a collection of reviews. When quantified, it is expressed as real values ​​or 1-hot data, etc. Training user attribute data is user attribute data for training purposes. Training user attribute data can be obtained at the same time as conducting a subjective evaluation experiment to obtain a "training review order vector" by conducting a questionnaire survey of participants.

[0024] Contributor attribute data is data that numerically represents the attributes (age, gender, place of residence, etc.) of a person (contributor) who has already purchased a product and posted review text data. When quantified, it is expressed as real values ​​or 1-hot data, etc. Contributor attribute data for training is contributor attribute data for training purposes. Contributor attribute data for training can be obtained from user data, etc. registered on the shopping site to which the corresponding review text data for training is posted.

[0025] A review order vector is a vector that indicates the display order of review text data corresponding to each combination of poster attribute data and classification data, and each dimension expresses the display order of the combination (combination of poster attribute data and classification data) corresponding to that dimension using real values, 1-hot vectors, etc. A training review order vector is a vector that indicates the display order of correct reviews tailored to the user's cognitive characteristics for the (combination of poster attribute data and classification data) to which the corresponding training review text data belongs. A training review order vector can be obtained, for example, by a subjective evaluation experiment.

[0026] In the subjective evaluation experiment, each participant is presented with n pieces of learning review text data and the corresponding learning poster attribute data and learning review classification data, and is asked to rate their willingness to purchase the product using either a binary value of "want to purchase" or "don't want to purchase," or a seven-point scale. However, n must be 10 or greater. It is not necessary to present all of the learning review text data to one participant, but all of the review text data evaluated by one participant must be for the same product. It is also desirable to present reviews of all ages and genders to one participant.

[0027] All n pieces of learning review text data are presented at once in a random order, with the exception that at least one of slightly negative reviews from people of the opposite sex or different age (2 on a 7-point scale) or slightly positive or more positive reviews from people of the same sex or age (6 on a 7-point scale) must be presented within the top five.

[0028] After the experiment is completed, the display order of the learning review text data for which the purchasing intent rating was "I want to buy" or rated 5 or higher out of 7 is obtained as a learning review order vector, which is associated with the learning user attribute data, learning poster attribute data, and learning review classification data. Therefore, n pieces of learning data are obtained from one participant.

[0029] The learning word-of-mouth classification data is obtained by inputting the learning word-of-mouth data into the classification unit 11 .

[0030] The display order estimation model learning unit 12 inputs the training user attribute data, training poster attribute data, and training review classification data into the display order estimation model m1 for each training data, and learns the display order estimation model m1 (updates the parameters of the display order estimation model m1) so that the output from the display order estimation model m1 approaches the training review order vector.

[0031] Next, the inference process will be described. Fig. 4 is a diagram showing an example of the functional configuration of the display order estimation device 10 at the time of inference according to the first embodiment. In Fig. 4, the same or corresponding parts as those in Fig. 2 are denoted by the same reference numerals, and their description will be omitted as appropriate.

[0032] 4, the display order estimation device 10 does not have a learning unit but has a display order rearrangement unit 13. The display order rearrangement unit 13 is realized by processing that is executed by the processor 104 of one or more programs installed in the display order estimation device 10. Furthermore, the display order estimation model m1 in FIG. 4 is a trained display order estimation model m1.

[0033] The display order rearrangement unit 13 receives as input poster attribute data and any user attribute data corresponding to each set of multiple word-of-mouth text data that are to be the subject of display order estimation (subject of display order rearrangement), estimates a display order for each of the multiple word-of-mouth text data that matches the cognitive characteristics based on information obtained using the classification unit 11 and the display order estimation model m1, and outputs text data in which the multiple word-of-mouth text data are rearranged in the display order (hereinafter referred to as "rearranged word-of-mouth text data").

[0034] Specifically, the display order rearrangement unit 13 performs the following steps (1) to (3) for each word-of-mouth text data included in a set of multiple word-of-mouth text data to be rearranged in display order. Hereinafter, one word-of-mouth text data to be processed is referred to as "target word-of-mouth text data."

[0035] (1) By inputting the target word-of-mouth text data into the classification unit 11, the classification unit 11 generates word-of-mouth classification data for the target word-of-mouth text data.

[0036] (2) The review classification data generated by the classification unit 11, the input user attribute data, and the input poster attribute data are input to the trained display order estimation model m1, and the review order vector output by the display order estimation model m1 is obtained.

[0037] (3) The word-of-mouth order vector obtained in (2) is assigned to the target word-of-mouth text data (that is, the display order indicated by the word-of-mouth order vector is set as the display order of the target word-of-mouth text data).

[0038] When steps (1) to (3) have been executed for all the word-of-mouth text data, the display order rearrangement unit 13 executes step (4) below.

[0039] (4) The set of word-of-mouth text data is sorted (arranged) based on the display order indicated by the corresponding word-of-mouth order vector, thereby generating rearranged word-of-mouth text data.

[0040] In this embodiment, the display order estimation model m1 estimates the display order only from poster attribute data, word-of-mouth classification data, and user attribute data. Therefore, even if the learning data and the word-of-mouth text data at the time of inference are different, the display order can be determined according to the user attribute data.

[0041] Note that the review order vector output by the display order estimation model m1 is the display order for each combination of poster attribute data and classification data, so there is a possibility that the same review order vector indicating the nth place may be assigned to multiple reviews. Therefore, a limit on the number of reviews assigned the same review order vector to be displayed (hereinafter referred to as a "limit on the number of displays in the same order") may be set. For example, if the number of reviews with the same review order vector to be displayed is limited to three or less, the order (display order) assigned to each of the review text data after the swapping is completed will be 1st, 1st, 2nd, 3rd, 3rd, 3rd...

[0042] The display order rearrangement unit 13 may randomly determine the display order of review text data assigned the same review order vector. The display order rearrangement unit 13 may exclude review text data that has been truncated due to the limit on the number of reviews displayed in the same order from the display targets, or may sort the truncated reviews for the second week after the first week of processing for displaying the same order is completed.

[0043] 5 is a diagram showing an example of rearranged word-of-mouth text data. In FIG. 5, (1) shows a set of word-of-mouth text data before rearrangement of the display order. (2) shows the rearranged word-of-mouth text data for (1).

[0044] As described above, according to the first embodiment, the display order of a collection of reviews is determined using a display order estimation model m1 that has learned the correspondence between the attributes of users who refer to reviews (user cognitive characteristics), the attributes of posters, and the classification of reviews and the display order. Therefore, it is possible to present posted information about products in a manner that takes into account the cognitive processing of users. As a result, it is expected that, for example, it will be possible to increase users' purchasing motivation.

[0045] Next, a second embodiment will be described. In the second embodiment, differences from the first embodiment will be described. Points not specifically mentioned in the second embodiment may be the same as those in the first embodiment.

[0046] In the first embodiment, the display of word-of-mouth information is changed in accordance with the cognitive characteristics of the user. However, users do not necessarily perform the same cognitive processing for all products.

[0047] For example, it is known that the perceived financial risk and risk to product quality (perceived risk) of a product affect purchasing intentions (Reference 3). For products with high perceived risk, it may be possible to reduce the perceived risk by displaying an acceptable level of negative reviews, regardless of gender differences in cognitive processing.

[0048] In other words, users are likely to know that any product has some drawbacks, large or small. Therefore, if there are no negative reviews at all, users may suspect that the reviews are from fake buyers hired by the seller, which is likely to lower not only the reviews themselves but also the credibility of the product and store. For products with low perceived risk to begin with (e.g., tissue paper), users do not mind even if the product is of poor quality, but for products with high perceived risk (e.g., cameras, children's snacks), users tend to want to avoid being sold a poor quality product. Therefore, for products with high perceived risk to begin with, if there are negative reviews, and the reviews are tolerable, users will judge the seller to be trustworthy and the product itself to be safe to purchase (the problems are tolerable), which is likely to lower the perceived risk compared to when there are no negative reviews at all.

[0049] Therefore, in the second embodiment, an example will be described in which the perceived risk of a product is acquired and the order in which word-of-mouth information is displayed is changed depending on the product.

[0050] 6 is a diagram showing an example of the functional configuration of the display order estimation device 10 during learning according to the second embodiment. In Fig. 6, parts that are the same as or correspond to those in Fig. 2 are given the same reference numerals, and descriptions thereof will be omitted as appropriate.

[0051] 6 , the display order estimation device 10 further includes a perceptual risk model learning unit 14, a perceptual risk model m2, and a perceptual risk estimation unit 15. Each of these units is realized by processing executed by the processor 104 of one or more programs installed in the display order estimation device 10.

[0052] The perceived risk model m2 is a machine learning model (for example, a neural network) that receives product data of a certain product as input and outputs perceived risk data regarding the product.

[0053] Product data is information that indicates a product, such as product price, sales, product category, and product provider evaluation information. In other words, product data is information for each product. Product data is expressed as real numbers, 1-hot, etc. Product data may be defined using a product classification method such as that described in Reference 4.

[0054] The perceived risk data is data that expresses the perceived risk of a product, for example, using seven-level real numbers or 1-hot vectors.

[0055] The perceptual risk model learning unit 14 learns the perceptual risk model m2 using a plurality of learning data, each set of which is product data for the perceptual risk model and perceptual risk data for the perceptual risk model. That is, the perceptual risk model learning unit 14 updates the parameters of the perceptual risk model m2 so that the output from the perceptual risk model m2 when product data for the perceptual risk model is input to the perceptual risk model m2 approaches the perceptual risk data for the perceptual risk model. A conventional DNN such as that in Reference 2 can be used to learn the perceptual risk model m2.

[0056] The product data for the perceptual risk model is product data prepared for training the perceptual risk model m2. The set of training data for the perceptual risk model m2 must at least include the product data for the perceptual risk model as the product data for the perceptual risk model of any of the training data, but may include more data than the product data for the perceptual risk model.

[0057] The perceived risk data for the perceived risk model is prepared for training the perceptual risk model m2, and is data that indicates the perceived risk as a correct answer for a product related to the product data for the perceptual risk model that is paired in the training data. The perceived risk data for the perceptual risk model corresponding to a certain product data for the perceptual risk model may be obtained, for example, by displaying the price etc. of a product related to the product data for the perceptual risk model and then conducting a subjective evaluation experiment to evaluate the perceived risk, or may be generated on a rule-based basis for conditions that are generally considered to be high risk, such as a high product price or a low product provider rating.

[0058] The perceived risk estimation unit 15 receives product data as input and outputs perceived risk data for the product data using the trained perceptual risk model m2. Specifically, the perceived risk estimation unit 15 inputs the input product data to the trained perceptual risk model m2 and obtains the perceived risk data output by the perceptual risk model m2. In Figure 6, for each piece of product data of a product related to each piece of training data used for training by the display order estimation model training unit 12 (hereinafter referred to as "training product data"), the perceived risk estimation unit 15 inputs the training product data to the perceptual risk model m2 and obtains the perceived risk data output by the perceptual risk model m2 as training perceptual risk data.

[0059] The display order estimation model learning unit 12 learns the display order estimation model m1 using the learning data of the display order estimation model m1, as in the first embodiment, but the configuration of the learning data of the display order estimation model m1 differs from that in the first embodiment.

[0060] 7 is a diagram showing an example of the configuration of training data for the display order estimation model m1 in the second embodiment. As shown in FIG. 7, one training data further includes training perceived risk data as input data. The training perceived risk data of a certain training data is the perceived risk data output by the perceived risk estimation unit 15 for the training product data of the product corresponding to the training data.

[0061] That is, in the second embodiment, the perceived risk model m2 also receives perceived risk data as an input. Therefore, the display order estimation model learning unit 12 inputs the training user attribute data, training poster attribute data, training review classification data, and training perceived risk data to the display order estimation model m1 for each training data, and learns the display order estimation model m1 (updates the parameters of the display order estimation model m1) so that the output from the display order estimation model m1 approaches the training review order vector as the correct answer.

[0062] Next, the inference process will be described. Fig. 8 is a diagram showing an example of the functional configuration of the display order estimation device 10 at the time of inference according to the second embodiment. In Fig. 8, the same or corresponding parts as those in Fig. 6 or Fig. 4 are denoted by the same reference numerals, and their description will be omitted as appropriate.

[0063] In the second embodiment, the display order rearrangement unit 13 receives, for each word-of-mouth text data to be rearranged in the display order, product data of the product corresponding to the word-of-mouth text data, poster attribute data corresponding to the word-of-mouth text data, and any user attribute data as input, and outputs rearranged word-of-mouth text data in which the word-of-mouth text data is rearranged in an order that matches the cognitive characteristics based on information obtained using the perceived risk estimation unit 15, the classification unit 11, and the display order estimation model m1.

[0064] Specifically, the display order rearrangement unit 13 executes the following steps (1-1), (1-2) to (3) for each word-of-mouth text data included in the set of word-of-mouth text data to be rearranged in display order. Hereinafter, one word-of-mouth text data to be processed is referred to as "target word-of-mouth text data." Note that, among the steps (1) to (3) in the first embodiment, (1) is the same as the following step (1-1), and (3) is the same as the following step (3). Therefore, in the second embodiment, step (1-2) is added, and (2) in the first embodiment is changed to (2').

[0065] (1-1) By inputting the target word-of-mouth text data into the classification unit 11, the classification unit 11 generates word-of-mouth classification data of the target word-of-mouth text data.

[0066] (1-2) By inputting the input product data into the perceived risk estimation unit 15, the perceived risk estimation unit 15 (perceived risk model m2) generates perceived risk data corresponding to the product data. However, the perceived risk data may also be obtained by the rule base described above.

[0067] (2') The word-of-mouth classification data generated by the classification unit 11, the perceived risk data generated by the perceived risk estimation unit 15, the input user attribute data, and the input poster attribute data are input into the trained display order estimation model m1, and the word-of-mouth order vector output by the display order estimation model m1 is obtained. (3) The word-of-mouth order vector obtained in (2) is assigned to the target word-of-mouth text data (i.e., the display order indicated by the word-of-mouth order vector is set as the display order of the target word-of-mouth text data).

[0068] When steps (1) to (3) have been executed for all the word-of-mouth text data, the display order rearrangement unit 13 executes the process of step (4) described in the first embodiment.

[0069] As described above, the second embodiment can achieve the same effects as the first embodiment. Furthermore, the second embodiment can present a collection of reviews of a certain product in a display order according to the perceived risk of the product.

[0070] [Reference 1] Taku Kudo et al., "Applying Conditional Random Fields to Japanese Morphological Analysis", Association for Computational Linguistics, 2004 [Reference 2] Han, K., Yu, D. and Tashev, I., "Speech Emotion Recognition Using Deep Neural Network and Extreme Learning Machine", Proc. of INTERSPEECH, pp. 223-227, 2014 [Reference 3] Hyejeong Kim et al., "E-atmosphere, emotional, cognitive, and behavioral responses", Journal of Fashion Marketing and Management, 2010 [Reference 4] Shiu-Li Huang et al., "Exploring consumer online purchase and search behavior: An FCB grid perspective", Asia Pacific Management Review, 2022 The above describes in detail the embodiments of the present invention, but the present invention is not limited to such specific embodiments, and various modifications and variations are possible within the scope of the gist of the present invention described in the claims.

[0071] REFERENCE SIGNS LIST 10 display order estimation device 11 classification unit 12 display order estimation model learning unit 13 display order rearrangement unit 14 perceived risk model learning unit 15 perceived risk estimation unit 100 drive device 101 recording medium 102 auxiliary storage device 103 memory device 104 processor 105 interface device B bus m1 display order estimation model m2 perceived risk model

Claims

1. A display order learning method characterized by a computer executing a learning procedure to train a machine learning model to which attribute data of users who refer to posted information about a certain product, attribute data of the poster of the posted information, and data indicating the degree of positivity of the posted information are input, so that the output from the machine learning model approaches the display order given as the correct answer for the posted information.

2. The display order learning method according to claim 1, characterized in that the learning procedure involves training the machine learning model so that the output from the machine learning model, to which the perceived risk of the product has been further input, approaches the display order.

3. A display order estimation method characterized by being executed by a computer: an estimation procedure for estimating the display order for each of a plurality of posted information about a certain product based on the attribute data of the users who refer to the posted information, the attribute data of the poster of the posted information, and the data indicating the degree of positivity of the posted information, using a machine learning model trained so that when attribute data of users who refer to posted information about the certain product, attribute data of the poster of the posted information, and data indicating the degree of positivity of the posted information are input, the output approaches the display order given as the correct answer for the posted information.

4. The display order estimation method described in claim 3, characterized in that the machine learning model is trained so that when a perceived risk of the product is further input, the output approaches the display order given as the correct answer, and the estimation procedure estimates the display order for each of multiple pieces of posted information about a certain product based on attribute data of users who refer to the posted information, attribute data of the poster of the posted information, data indicating the degree of positivity of the posted information, and the perceived risk of the product.

Images