JP7495363B2 - Personality information generation model, device and method using domain-independent RNN - Google Patents

Description

The present invention relates to a technology for estimating a user's personality and psychological state.

Personality, which is a characteristic that holistically captures mental indicators such as a person's character, personality, temperament, individuality, and even values, is known to have a significant impact on behavior in various behavioral domains, such as the purchase and selection of various products and services. For this reason, in recent years, there have been many attempts to understand users' personalities in order to predict their behavior and personalize the products and services they provide.

Today, human personality is generally understood according to a concept called "trait theory." This trait theory states that human personality is composed of multiple traits and can be expressed by the level (score) of each trait. For example, the "Big Five," a representative example of a personality model based on trait theory, states that human personality is composed of five traits: intellectual curiosity (O: Openness to experience), conscientiousness (C: Conscientiousness), extraversion (E: Extraversion), agreeableness (A: Agreeableness), and emotional stability (N: Neuroticism), and expresses human character as a set of scores for these five traits (OCEAN).

Questionnaire surveys are often used as a method of measuring such personality traits. For example, the "Big Five Scales," a representative questionnaire for measuring the Big Five, has subjects answer a set of questions consisting of a total of 60 items. When measuring intellectual curiosity (O), for example, subjects are asked to answer questions such as "creative" and "versatile" on a seven-point scale from "not at all applicable (1 point)" to "very applicable (7 points)," and the score for intellectual curiosity (O) is determined by the total score for all questions.

On the other hand, given the enormous cost involved in conducting questionnaire surveys of a huge number of users, there is also widespread research into technology that first administers questionnaire surveys to a small number of users and then uses the responses to these surveys and the behavioral data of those users to estimate their personalities.

Non-Patent Document 1 is a survey paper on such technologies, and introduces a technology that estimates a user's personality using various user behavioral data, such as the content posted to SNS (Social Networking Service), the number of friends connected on SNS, the call history of a mobile device, movement patterns derived based on device location information, and even the speed and intonation of speech during phone calls, as well as the volume of the voice.

A. Vinciarelli and G. Mohammadi, "A Survey of Personality Computing," IEEE Transactions on Affective Computing, vol. 5, no. 3, pp. 273-291, 2014, <https://doi.org/10.1109/TAFFC.2014.2330816> Donkers, T., Loepp, B., & Ziegler, J., "Sequential user-based recurrent neural network recommendations", RecSys'17 - Proceedings of the 11th ACM Conference on Recommender Systems, pp. 152-160, 2017, ＜https://doi.org/10.1145/3109859.3109877＞

As explained above, various techniques have been developed to measure personality, but these techniques share common issues that need to be resolved: (a) coarse measurement granularity (information granularity) and (b) low reliability.

First, (a) the coarseness of measurement granularity is a significant issue for technologies that understand personality based on "trait theory," including technologies that estimate a user's personality from the user's behavioral data. For example, the Big Five traits were originally determined by aggregating a huge number of personality-descriptive words extracted from dictionaries into five groups of words with similar concepts using statistical methods, and then based on the results of this aggregation, in a way that would be easy for humans to interpret.

However, in the process of aggregation, small or subtle differences in the original meanings of the personality expressions are lost. For example, conscientiousness (C), one of the Big Five, is aggregating personality expressions such as "serious," "hardworking," and "organized" into a single trait, and the measurement result of this conscientiousness (C) is merely a single score, and the differences in meaning of each personality expression before aggregation are lost from the measurement results.

Next, (b) low reliability is a significant problem for techniques that involve measuring personality through questionnaire surveys. For example, there is usually considerable variation in the way subjects respond to questions in a questionnaire. Specifically, it is not rare for subjects to answer based on how they "should be" rather than on their own original thoughts, or to consider answering itself to be a tedious task and give false answers. As a result, the reliability of personality measurement results is reduced.

Even if the expected (honest and serious) response attitude is guaranteed, the questions and response methods set in a questionnaire survey are unlikely to measure all of the areas of the characteristics to be measured, and the reliability of the measurement results is inevitably low. For example, there are more than 20 types of Big Five questionnaires, but the number of questions, question wording, and response methods (such as answering on a 5-point scale or yes/no) are significantly different for each. As a result, it is widely known that the measurement results of the same subject differ depending on the type of questionnaire answered, and there have been cases where the correlation r of the measurement results of agreeableness (A) between two questionnaires is below 0.40.

Furthermore, even in technology that estimates a user's personality from the user's behavioral data, the estimation models used are usually trained and constructed using measurement results from questionnaire surveys, so the low reliability described above (b) is a serious issue that needs to be resolved.

In light of this current situation, the inventors of the present application came up with the idea of realizing personality estimation that solves the above two problems by applying dynamic user representation learning (Sequential URL, Sequential User Representation Learning) technology, which is being researched in the field of product and service recommendation technology, without using measurement results from questionnaire surveys.

The Sequential URL technology is a technology that acquires distributed representations of user attributes (such as preferences for music or movies) from time-series behavioral data related to music playback, video viewing, ad clicks, item purchases, etc. For example, Non-Patent Document 2 discloses a technology for this Sequential URL technology in which time-series behavioral data on a user's movie viewing is input into a GRU (Gated Recurrent Unit), which is a type of Recurrent Neural Network (RNN), to acquire distributed representations of user attributes.

Specifically, the technology described in Non-Patent Document 2 discloses a GRU cell that receives as input the distributed representation of a user output from a distributed representation extraction unit that receives the user's identification information as input, and the distributed representation of a movie output from a distributed representation extractor that receives the identification information of a movie watched by the user as input. Non-Patent Document 2 claims that by training the GRU cell and the distributed representation extractor with a sufficient amount of data, the distributed representation of the user output from the distributed representation extraction unit reflects user attributes that are static and do not change easily, such as the user's movie preferences.

However, the distributed representation of user attributes obtained by such Sequential URL technology is naturally highly dependent on the behavioral domain (area) to which the behavioral data used belongs, and is in no way interpretable as the user's personality (a characteristic that captures mental indicators holistically). For example, when behavioral data related to the behavioral domain of movie viewing is used, only user attributes related to movie viewing, such as the user's movie preferences, can be extracted. For this reason, no attempt has been made to estimate a user's personality using Sequential URL technology to date.

The present invention aims to provide a personality information generation model, a personality estimation device, and a personality estimation method that can generate personality information with higher information granularity and reliability.

According to the present invention, there is provided a personality information generation model for causing a computer to function to estimate a personality of a user from information related to each of a plurality of behavioral domains of the user, the personality information generation model comprising:
a plurality of behavioral expression generating units set for each of the behavioral domains, each generating behavioral expression information that is information expressing the behavior of the user in the behavioral domain from information that identifies the content of the behavior of the user in the behavioral domain;
a plurality of domain-specific regression neural network (RNN) cells set for each of the behavioral domains, each of which receives the behavioral expression information of the user in the behavioral domain, and generates new domain-specific hidden state information corresponding to the hidden state information of the RNN by reflecting the behavioral expression information of the user received this time on domain-specific hidden state information corresponding to the RNN hidden state information that the RNN cell itself generated when it previously received the behavioral expression information of the user ;
a user expression generating unit that generates user expression information that represents the user from information that identifies the user;
a computer is caused to function as a domain-independent RNN cell that receives the generated domain-specific hidden state information and the generated user expression information, and generates new domain-independent hidden state information corresponding to the RNN hidden state information by reflecting the currently received domain-specific hidden state information of the user and the user expression information on domain-independent hidden state information corresponding to the RNN hidden state information that the computer generated when it previously received the domain-specific hidden state information of the user ;
A user expression generating unit trained by the RNN backpropagation method in combination with a plurality of domain-specific RNN cells and domain-independent RNN cells generates user expression information of the user as information expressing the personality of the user in the sense of domain-independent and static information related to the user.
A personality information generation model is provided.

In the personality information generation model according to the present invention, it is also preferable that the domain-independent RNN cell receives the domain-specific hidden state information that has been recently generated or has just been generated in the domain-specific RNN cell, and generates the domain-independent hidden state information.

In addition, as one embodiment of the personality information generation model according to the present invention, it is also preferable that the model further causes the computer to function as a first conversion processing unit that performs a predetermined type of conversion processing on the domain-specific hidden state information output from the domain-specific RNN cell, and inputs the domain-specific hidden state information that has been subjected to the conversion processing to a domain-independent RNN cell.

Furthermore, in the personality information generation model according to the present invention, it is also preferable that the domain-specific RNN cell also receives the domain-independent hidden state information that was most recently generated from the time of receiving the behavioral expression information, and generates the domain-specific hidden state information using this domain-independent hidden state information as well.

In another embodiment of the personality information generation model according to the present invention, it is also preferable that the model further functions as a second conversion processing unit that performs a predetermined type of conversion processing on the domain-independent hidden state information output from the domain-independent RNN cell, and inputs the domain-independent hidden state information that has been subjected to the conversion processing to the domain-specific RNN cell.

Furthermore, in the personality information generation model according to the present invention, it is also preferable that the user expression generation unit outputs the generated user expression information as information expressing the personality of the user, and the domain-independent RNN cell outputs the generated domain-independent hidden state information as dynamic psychological state information, which is information related to the dynamic psychological state of the user.

It is also preferable that each of the multiple domain-specific RNN cells and domain-independent RNN cells of the present invention includes at least one of a forget gate, an update gate, and a reset gate that performs a predetermined process on the information input to the cell to generate new hidden state information.

According to the present invention, there is also provided a personality estimation device that generates and outputs information expressing the personality of a user using a learned user expression generation unit obtained by training the personality information generation model described above.

According to the present invention, further comprising:
The personality information generation model described above is trained using the RNN backpropagation method ,
A personality estimation method in a computer is provided, characterized in that, using the trained user expression generation unit obtained by the training, the user expression information of the user is information that expresses the personality of the user in the sense of domain-independent and static information related to the user , and is output.

The personality information generation model, personality estimation device, and personality estimation method of the present invention make it possible to generate personality information with higher measurement granularity and reliability.

FIG. 2 is a schematic diagram showing an embodiment of a personality information generation model according to the present invention. FIG. 2 is a functional block diagram showing an embodiment of the internal configuration of a domain-specific RNN cell and a domain-independent RNN cell according to the present invention. FIG. 1 is a schematic diagram for explaining characteristics of various pieces of information generated by a personality information generation model according to the present invention, and a functional block diagram showing an embodiment of a personality estimation device according to the present invention. FIG. 13 is a schematic diagram showing another embodiment of a personality information generation model according to the present invention.

The following describes in detail an embodiment of the present invention with reference to the drawings.

[Personality information generation model]
FIG. 1 is a schematic diagram showing an embodiment of a personality information generation model according to the present invention.

The personality information generation model 1 of the present embodiment shown in FIG.
(a) Using “behavioral domain information” which is information on a plurality of behavioral domains (two in FIG. 1: “item purchase” and “video viewing”) of a target user who is a personality estimation target,
(b) A machine learning model that generates “personality information”, which is information expressing the personality of the user to be estimated.

Specifically, the personality information generation model 1 is as follows:
(A) A plurality of domain-specific regression neural network (RNN) cells (two in FIG. 1 , a DSL (Domain Specific Layer) 1 cell 11 and a DSL 2 cell 12) are set for each of a plurality of behavioral domains, and receive “domain behavior information” of a user to be estimated in the behavioral domain, and generate new “domain-specific hidden state information” by reflecting the “domain behavior information” on “domain-specific hidden state information” that is hidden state information generated by the RNN cells themselves at a previous point in time;
(B) a user expression generation unit (in FIG. 1, a user distributed expression extraction unit 10u) that generates “user expression information” that is information expressing a user to be estimated from “user identification information” that is information that identifies a user to be estimated;
(C) This is a machine learning model that causes a computer to function as a domain-independent RNN cell (DIL (Domain Independent layer) 10 in FIG. 1 ) that receives the “domain-specific hidden state information” generated in (A) above and the “user expression information” generated in (B) above, and generates new “domain-independent hidden state information” by reflecting the “domain-specific hidden state information” and the “user expression information” on the “domain-independent hidden state information” that is the hidden state information that the computer itself generated at a previous point in time.

In this embodiment, the "behavior domain information" in (A) above is behavior expression information generated by a behavior expression generation unit (in FIG. 1, the item dispersed representation extraction unit 11i and the video dispersed representation extraction unit 12m) that receives information indicating the content of the behavior (such as purchase or viewing) of the estimated target user in the behavior domain (for example, identification information of a purchased item or a viewed video). Therefore, "behavior domain information" (behavior expression information) is generated each time a behavior occurs in each behavior domain, and as a result, the information as a whole forms a time series data group.

In addition, the user expression generation unit (user distributed expression extraction unit 10u) in (B) above is trained together with multiple domain-specific RNN cells (DSL1 cell 11 and DSL2 cell 12), which will be described in detail later, and as a result, after the training, it executes a domain-independent expression generation calculation (distributed expression extraction calculation) that is not biased toward a specific behavioral domain. As a result, the "user expression information" generated here is understood as the "personality information" of the estimated target user.

In this way, the personality information generation model 1 can generate more reliable "personality information" from the "behavioral domain information" of the estimated target user without using or relying on measurement results from questionnaire surveys, which have traditionally been used but have issues with reliability.

Furthermore, in this embodiment, such "personality information" is a multidimensional (e.g., tens to hundreds of dimensions) user expression vector output by a user expression generation unit (user distributed expression extraction unit 10u) that receives "user identification information" of the user to be estimated (e.g., a one-hot vector indicating the user), and is, for example, a list of numerous numerical values.

This user expression vector was created by the inventor of this application, inspired by well-known embedding techniques such as word embedding, and will hereafter be referred to as a personality expression vector or personality embedding vector.

Here, in word embedding representation, which is an essential technique in natural language processing using artificial intelligence, each word is represented by a word embedding vector of, for example, several hundred dimensions (a string of many numerical values), and the closer the words are in meaning, the smaller the distance between these vectors. In other words, a word embedding vector can be thought of as having the meaning of the corresponding word embedded in it. It is also possible to perform calculations between word embedding vectors, and one famous example is addition and subtraction, such as "king" - "man" + "woman" = "queen."

In a similar manner, the personality embedding vector of the present invention can be considered to have the personality characteristics of the corresponding user embedded in it, and therefore it is expected that by using the distance between vectors as a measure of personality similarity or by performing calculations between vectors, it will be possible to derive and generate even more useful information regarding personality.

In any case, the "personality information" (personality embedding vector) generated by the personality information generation model 1 is representation information based on distributed representation, so the information granularity is much higher than that of conventional personality indices such as the Big Five (which express personality in five dimensions).

Furthermore, although "personality information" as a distributed representation cannot be directly interpreted by humans, it is expected that it can be used, for example, as input data in machine learning models for various purposes, enabling artificial intelligence to understand a user's personality and predict the user's behavior and personalize services, etc.

In the personality information generation model according to the present invention, the number of domain-specific RNN cells, i.e., the number of behavioral domains handled, is naturally not limited to two as shown in FIG. 1, and can be three or more. In addition, the behavioral domains handled are not limited to the item purchases and video viewing shown in FIG. 1, and various domains related to behaviors that depend on or can be influenced by the personality of the behavior subject can be adopted as behavioral domains according to the present invention. For example, when time-series data of the behavioral domain "ad click" is acquired, a personality information generation model may be constructed by adding a domain-specific DNN cell that incorporates domain behavior information related to "ad click".

Here, the "personality information" and the above-mentioned "domain-independent hidden state information" are made domain-independent by receiving "domain-specific hidden state information" from cells related to multiple or many different behavioral domains. Of these, the "personality information" in particular can be considered as information related to the personality of the estimated target user as a static, domain-independent characteristic, as will be explained in detail later using Figure 3.

[Model construction, personality estimation method]
The configuration of the personality information generation model 1 of this embodiment will be described in more detail below.
(A) A DSL1 cell 11 as a domain-specific RNN cell related to the behavior domain “item purchase”, a user distributed representation extraction unit 11u, an item distributed representation extraction unit 11i, and an output layer 11o;
(a) A DSL2 cell 12 as a domain-specific RNN cell related to the behavior domain “video viewing”, a user distributed expression extraction unit 12u, a video distributed expression extraction unit 12m, and an output layer 12o;
(c) It has a DIL cell 10 as a domain-independent RNN cell, a user distributed expression extraction unit 10u as a user expression generation unit, and an output layer 10o as functional components embodied by a computer (a program installed in the computer).

Figure 1 shows the processing executed by the functional components (a) to (c) (the functional blocks on the left side of Figure 1) as described above, deployed on a time axis with time progression pointing to the right. Note that the three time axes set for each of (a) to (c) above are time axes for points in time when each has its own unique value.

As for the notation of time points on each time axis, for example, time point (t2+1) for DSL2 cell 12 means the time point at which DSL2 cell 12 next receives domain behavior information (video embedding vector) after receiving domain behavior information (video embedding vector) at time point t2. Here, the next reception time point is naturally (t2+2). Furthermore, from the perspective of time point (t2+1), time point t2 is the "previous time point." Furthermore, for the convenience of the following explanation, the DSL2 cell 12 that performs processing at this time point (t2+1) will be referred to as DSL2 cell 12(t2+1). Furthermore, similar notations of processing time points and notations including the processing time will be used hereafter for DSL1 cell 11 and DIL cell 10.

Each of the above-mentioned functional components will be specifically described below. In FIG. 1, the item distributed expression extraction unit 11i(t1), which is the item distributed expression extraction unit 11i at time t1, receives an event that "item item1 has been purchased by user UserA (the user to be estimated)" at time t1,
(a) Receive a one-hot vector i1 ^(t1) which is the item identification information of item item1,
(b) The received one-hot vector i1 ^(t1) is applied (multiplied) by the matrix W _i ^DSL1 as an item distributed representation extraction operator to generate an item representation vector (domain behavior information) r_i1 ^(t1) for item item1.

In addition, the user distributed expression extraction unit 11u(t1), which is the user distributed expression extraction unit 11u at the time point t1, receives the one-hot vector i1 ^(t1) in (a) above,
(c) receiving a one-hot vector u a ^(t1) which is the user identification information of user UserA;
(d) The received one-hot vector ua ^(t1) is applied (multiplied) by the matrix W _u ^DSL1 as a user distributed expression extraction operator to generate a user expression vector r_ua ^(t1) related to the behavioral domain “item purchase” of user UserA.

Also in FIG. 1, DSL1 cell 11(t1) at time t1 is
(e) receiving the item expression vector (domain behavior information) r_i1 ^(t1) generated in (b) above and the user expression vector r_ua ^(t1) generated in (d) above,
(f) (as will be described in detail later) also receiving domain-independent hidden state information h DIL,(tI) generated most recently (at time tI in FIG. 1) from time t1 when the domain action information (r_i1 ^{(t1 ))} of (e) is received;
(g) Generate new domain-specific hidden state information h ^DSL1,(t1) by reflecting the received item representation vector (domain behavior information) r_i1 ^(t1) , user representation vector r_ua ^(t1) , and domain-independent hidden state information h ^DIL,(tI) to the domain-specific hidden state information h ^{DSL1,(t1-1) generated by itself at the previous point in time (t1-1)} (this generation process will be explained in more detail later using Figure 2).

The above describes the processing at time t1 in the item shared representation extraction unit 11i, the user shared representation extraction unit 11u, and the DSL1 cell 11. Of course, the same processing as (a) to (g) above is also performed for processing at other times, for example, the processing at the item shared representation extraction unit 11i(t1+1), the user shared representation extraction unit 11u(t1+1), and the DSL1 cell 11(t1+1) relating to the next time (t1+1).

Furthermore, the same processes as those described above (a) to (g) are also executed for the video dispersed expression extraction unit 12m, the user dispersed expression extraction unit 12u, and the DSL2 cell 12 related to other behavioral domains (video viewing), for example, the video dispersed expression extraction unit 12m(t2), the user dispersed expression extraction unit 12u(t2), and the DSL2 cell 12(t2) related to time t2. However, in this case, for example, the video dispersed expression extraction unit 12m(t2) at time t2,
(b') The matrix W _m ^DSL2 as a video distributed expression extraction operator is applied (accumulated) to the received one-hot vector m1 ^(t2) (indicating the video Mov1 viewed by user UserA) to generate a video expression vector (domain behavior information) r_m1 ^(t2) of the video Mov1. The DSL2 cell 12 (t2) incorporates this domain behavior information r_m1 ^(t2) and performs processing.

Similarly, in FIG. 1, the user distributed expression extraction unit 10u(tI), which is the user distributed expression extraction unit (user expression generation unit) 10u at time tI,
(h) Receive a one-hot vector u a ^{(t I )} , which is the user identification information of user UserA,
(i) The received one-hot vector ua ^(tI) is applied (multiplied) by a matrix W _u ^DIL as a user distributed expression extraction operator to generate a user expression vector r_ua ^{(tI) that} is independent of the behavioral domain of user UserA.

Furthermore, DIL cell 10(tI), which is the DIL cell 10 at time tI, is
(j) receiving the user expression vector r_ua ^(tI) generated in (i) above; and
(k) also receiving domain-specific hidden state information h ^{DSL2,(t 2 )} generated most recently (at time t 2 in FIG. 1 ) from time t I ;
(l) Generate new domain-independent hidden state information h ^DIL,(tI) by reflecting the received user expression vector r_ua ^(tI) and the domain-specific hidden state information h ^DSL2,(t2) on the domain-independent hidden state information h ^DIL,(tI-1) generated by itself at the previous point in time (tI-1) (this generation process will also be explained in more detail later using Figure 2).

Note that the time tI related to the above-described processes (h) to (l) can be, for example, a time (one of) that is set periodically (every predetermined time has elapsed), but in this embodiment, a more preferable setting is the time immediately after domain-specific hidden state information is generated in one of the domain-specific RNN cells (in FIG. 1, either DSL1 cell 11 or DSL2 cell 12).

In other words, in the case described above, for example, when the DSL2 cell 12(t2) generates the domain-specific hidden state information h ^DSL2,(t2) at time t2, the above-mentioned processes (h) to (l) are initiated immediately thereafter (i.e., at time tI).Furthermore, as also shown in FIG. 1, when the DSL1 cell 11(t1) generates the domain-specific hidden state information h ^DSL1,(t1) at time t1, the immediately-generated domain-specific hidden state information h ^DSL1,(t1) is received immediately thereafter (at time tI+1), and the same processes as the above-mentioned processes (h) to (l) are executed.

In this manner, in this embodiment, each time domain-specific hidden state information is generated in any domain-specific RNN cell, the DIL cell 10 receives the domain-specific hidden state information and generates "domain-independent hidden state information."

In FIG. 1, the user UserA, who is the subject of estimation, performs the following actions in this order: "viewing video Mov1 (at time t2)" → "purchasing item item1 (at time t1)" → "viewing video Mov2 (at time t2+1)". The DIL cell 10 sequentially receives domain-specific hidden state information generated by the domain-specific RNN cell which sequentially receives the expression vectors (domain behavior information) related to these actions. This takes into account the order in which these domain-specific hidden state information was generated (received) and the contextual information, making it possible to generate "domain-independent hidden states" that are not dependent solely on the content of actions in a specific behavioral domain (domain-independent).

Furthermore, in this embodiment, as described in (f) above, such characteristic "domain-independent hidden states" are appropriately used as inputs to domain-specific RNN cells (DSL1 cell 11 and DSL2 cell 12). As a result, the domain-specific hidden state information later received by DIL cell 10 also reflects the above-mentioned characteristics, which promotes the domain independence of the "domain-independent hidden states." In other words, DIL cell 10 can further perform information processing that reflects the overall attributes of the estimated target user, independent of a specific behavioral domain.

Figure 2 is a functional block diagram showing one embodiment of the internal configuration of a domain-specific RNN cell and a domain-independent RNN cell according to the present invention.

2A shows the internal configuration of a DSL(p) cell (10+p)(tp) (p = 1, 2, ...), which is a domain-specific RNN cell at time tp. Specifically, the DSL(p) cell (10+p)(tp) is configured using a known GRU (Gated Recurrent Unit) capable of configuring an RNN, and includes the following functional components (a) to (f).
(a) [T]: Multiply multiple input vectors by a weight matrix, add a bias vector to the multiplication result, and output the resulting vector. Note that the two [T] in Fig. 2(A) each have an independent (different) weight matrix and bias vector.
(b) [σ]: Each element of the input vector is input to a sigmoid function, and the resulting vector is output. The two [σ] in Fig. 2(A) function as a reset gate r and an update gate u, respectively.
(c) [• in a circle]: Multiply the elements of two input vectors and output the resulting vector.
(d) [-1]: Multiply each element of the input vector by -1 and output the resulting vector.
(e) [tanh]: Input each element of the input vector into the tanh function, and output the resulting vector (hidden state candidate k).
(f) [+]: Add the elements of two input vectors together and output the resulting vector.

Here, the update gate u (one side [σ]) decides how much information to retain and take in from the previous ( ^last ) domain-specific hidden state information h ^DSLp,(tp-1) based on the information obtained by converting the input r_ua ^{(tp), r_ip(tp)} and h ^DIL,(tI-1) . Also, the reset gate r (the other side [σ]) decides how much information to discard from the previous (last) domain-specific hidden state information h ^DSLp,(tp-1) . Through such processing, the DSL(p) cell (10+p)(tp) can generate "domain-specific hidden state information" that reflects the dynamic (time-dependent) information in the domain behavior information, which is the time-series data of behavior related to the corresponding behavior domain.

Next, FIG. 2(B) shows the internal configuration of DIL cell 10(tI), which is a domain-independent RNN cell at time tI. This DIL cell 10(tI) has a configuration similar to the above-mentioned DSL(p) cell (10+p)(tp), and more specifically, it is configured to employ a GRU that includes the above-mentioned functional components (a) to (f).

This enables the DIL cell 10(tI) to generate "domain-independent hidden state information" that is dynamic (time-dependent) information about the attributes of the estimated target user by incorporating multiple domain-specific hidden state information that reflects dynamic (time-dependent) information in multiple domain behavior information and making it domain-independent.

In addition, in both the domain-specific RNN cell (DSL(p) cell (10+p)(tp)) shown in FIG. 2(A) and the domain-independent RNN cell (DIL cell 10(tI)) shown in FIG. 2(B), various RNN units other than GRUs can be adopted as the internal configuration, for example, a long-short term memory (LSTM) equipped with an input gate, an output gate, and a forget gate may be adopted. In any case, it is also preferable that each of the multiple domain-specific RNN cells and domain-independent RNN cells in the personality information generation model 1 is configured to have at least one of a forget gate, an update gate, and a reset gate that perform a predetermined process on the information input to the cell to generate new hidden state information.

<Training the personality information generation model>
Returning to FIG. 1, the training of the personality information generation model 1 will now be described.

First, in training the DSL1 cell 11, the item distributed representation extraction unit 11i, and the user distributed representation extraction unit 11u, time series data of one-hot vectors, which are identification information of items purchased by each user for a large number of users, are prepared and input sequentially to the item distributed representation extraction unit 11i. In addition, the one-hot vector, which is the user identification information of the user, is input to the user distributed representation extraction unit 11u. As a result, the DSL1 cell 11 updates the domain-specific hidden state information at the previous time point at each time point to generate domain-specific hidden state information at the current time point, and in response, the output layer 11o outputs a prediction result of the purchased item (identification information of the item predicted to be purchased by the user at the next time point, or the probability of each candidate item being purchased) (in this embodiment, the output layer 11o is configured to perform such output).

Here, the difference between the output item prediction result and the correct item identification information (the one-hot vector of the item actually purchased) is considered as the loss, and training is performed using the well-known RNN error backpropagation method to adjust each parameter in DSL1 cell 11 (such as parameters that determine the weight matrix and bias vector of [T]) and the parameters that determine the matrix W _i ^DSL1 and matrix W _u ^DSL1 .

Training of the DSL2 cell 12, the video distributed expression extraction unit 12m, and the user distributed expression extraction unit 12u can be performed in the same manner as the training of the DSL1 cell 11, etc. Specifically, in this case, the difference between the prediction result of the viewed video output from the output layer 12o (identification information of the video predicted to be viewed by the user at the next time point, or the viewing probability of each candidate video) and the correct video identification information (one-hot vector of the video actually viewed) is treated as a loss, and training is performed to adjust each parameter in the DSL2 cell 12 (parameters determining the weight matrix and bias vector of [T], etc.) and parameters determining the matrix W _m ^DSL2 and matrix W _u ^DSL2 using a known RNN backpropagation method. Furthermore, training can be performed in the same manner when a third or subsequent domain-specific RNN cell system is set.

Next, in training the DIL cell 10 and the user distributed expression extraction unit 10u, the above-mentioned loss in each domain-specific RNN cell (each of the DSL1 cell 11 and the DSL2 cell 12) is back-propagated to the DIL cell 10 and the user distributed expression extraction unit 10u through the link between each domain-specific RNN cell and the DIL cell 10 to perform training. For example, in Fig. 1, the loss, which is the difference between the predicted result of the purchase item based on the input up to time t1 in the DSL1 cell 11 and the identification information of the correct item, is back-propagated through the link that transmits the domain-independent hidden state information ^hDIL,(tI) and is used to train each parameter in the DIL cell 10 (parameters that determine the weight matrix and bias vector of [T], etc.) and parameters that determine the ^matrix _WuDIL of the user distributed expression extraction unit 10u.

In other words, in this embodiment, training of the DIL cell 10 and the user distributed expression extraction unit 10u is performed only by backpropagation of errors from the domain-specific RNN cells (DSL1 cell 11 and DSL2 cell 12) without using teacher data unique to the domain-independent RNN cell.

Figure 3 is a schematic diagram for explaining the characteristics of various pieces of information generated by the personality information generation model according to the present invention, and a functional block diagram showing one embodiment of a personality estimation device according to the present invention.

As shown in the upper part of Figure 3, user attributes in a broad sense, including personality, are as follows:
They can be classified into four groups from the viewpoint of whether they are (a) dynamic (time-dependent or easily changed over time) or static (time-independent or not easily changed over time) and (b) domain-dependent (affecting only a specific behavioral domain or relating to a specific behavioral domain) or domain-independent (affecting behaviors in various behavioral domains or not relating to only a specific behavioral domain). Specifically, in FIG. 3, the four groups are groups related to the first to fourth quadrants of a user attribute graph composed of the horizontal axis related to (a) and the vertical axis related to (b).

Here, personality, which is one of the user attributes, is static and domain-independent, and therefore belongs to the group related to the first quadrant. Therefore, when estimating the personality of a certain user using time-series data related to the behavioral domain of that user, it becomes necessary to separate and extract only the components corresponding to this first quadrant from the user attribute information during the estimation process. However, such separation and extraction could not be performed, for example, even with the Sequential URL technology described above as a conventional technology.

In contrast, the learned (trained) DIL cell 10 and user distributed expression extraction unit 10u in the personality information generation model 1 have been made domain-independent in the process in which the DIL cell 10 appropriately receives domain-specific hidden state information from multiple domain-specific RNN cells (DSL1 cell 11 and DSL2 cell 12) and further appropriately provides domain-independent hidden state information to each domain-specific RNN cell. In particular, the user distributed expression extraction unit 10u (its matrix W _u ^DIL ) is configured to receive user identification information and output information related to the static attributes of the user.

As a result, the user expression vector r_ua output by the user distributed expression extraction unit 10u (its matrix _Wu ^DIL ) becomes information on user attributes belonging to the first quadrant, which is static and domain-independent, and can therefore be regarded as a personality expression vector (personality embedding vector), which is "personality information". Here, the user distributed expression extraction unit 10u may of course output the generated personality expression vector (personality embedding vector) to the outside of the model as "personality information". Furthermore, it is also possible to take out this learned (trained) user distributed expression extraction unit 10u (its matrix _Wu ^DIL ) and use it as a personality expression extractor.

In addition, since the learned (trained) DIL cell 10 is domain-independent as described above, the domain-independent hidden state information h ^DIL output from this DIL cell 10 is information reflecting user attributes related to the processing time that may affect the behavioral domain in general or is independent of a specific behavioral domain. In other words, this domain-independent hidden state information h ^DIL is information of user attributes belonging to the second quadrant corresponding to dynamic and domain independence, and therefore can be regarded as dynamic psychological state information that is information related to the dynamic psychological state of the user. Here, the DIL cell 10 may of course output the generated main-independent hidden state information h ^DIL to the outside of the model as dynamic psychological state information.

Incidentally, this dynamic psychological state information includes, for example, the user's fatigue level, emotional state, and mood (busy/bored, feeling good/bad, etc.). The DIL cell 10 is also capable of generating user attributes that change dynamically and can affect behavior in various behavioral domains. Here, the dynamic psychological state information generated multiple times over a given period as time-series data can also be considered as information that represents the changes in the user's dynamic psychological state over that given period.

To summarize the above explanation, it can be understood that the personality information generation model 1 separates domain-dependent (third and fourth quadrants) user attributes from domain-independent (first and second quadrants) user attributes by operating the DIL cell 10 and multiple domain-specific RNN cells (DSL1 cell 11 and DSL2 cell 12) while exchanging hidden state information with each other.

Furthermore, in this exchange, by incorporating the hidden state information generated by oneself into the other cell in a manner that reflects its order and context (i.e., in a manner that reflects the chronological order of the occurrence of various actions), it is understood that this ultimately promotes the separation of dynamic (second quadrant) user attributes from static (first quadrant) user attributes between the DIL cell 10 and the user distributed expression extraction unit 10u connected to it.

Incidentally, the domain-specific hidden state information output by each domain-specific RNN cell (each of DSL1 cell 11 and DSL2 cell 12) is information reflecting dynamic (time-dependent or time-varying) user attributes for the behavior of the corresponding behavioral domain, that is, it is understood to be information related to the user attributes in the third quadrant. Specifically, the user attributes in the third quadrant can be understood as psychological state information related to item purchase history, video viewing history, etc., such as psychological state information arising from the user's history, such as "the feeling of being bored with a certain genre of movies that you have been watching", "the feeling of wanting to watch movies of a different genre", and even "the feeling of becoming interested in other movies that feature a favorite actor from among the movies you have been watching".

The user expression vector r_ua output by the user distributed representation extraction unit (11u, 12u) that has been trained in conjunction with the above is information that reflects static (time-independent or time-unchanging) user attributes for the behavior in the corresponding behavioral domain, i.e., it is understood to be information related to user attributes in the fourth quadrant (e.g., tendencies in purchasing items and video preferences).

Of course, this information relating to user attributes in both the third and fourth quadrants can be output outside the model and used in various situations (for example, in marketing activities at companies related to the corresponding behavioral domain).

Other embodiments of the model
FIG. 4 is a schematic diagram showing another embodiment of a personality information generation model according to the present invention.

As shown in FIG. 4, the personality information generation model 1′ of this embodiment is as follows:
(a1) a conversion processing unit 11f1out as a first conversion processing unit, which performs a predetermined type of conversion processing (f1out) on the domain-specific hidden state information h ^DSL1 output from the DSL1 cell 11, and inputs the domain-specific hidden state information f1out(h ^DSL1 ) that has been subjected to the conversion processing (f1out) to the DIL cell 10;
(a2) A conversion processing unit 12f2out as a first conversion processing unit is provided, which performs a predetermined form of conversion processing (f2out) on the domain-specific hidden state information h ^DSL2 output from the DSL2 cell 12, and inputs the domain-specific hidden state information f2out(h ^DSL2 ) that has been subjected to the conversion processing (f2out) to the DIL cell 10.

Incidentally, FIG. 4 shows the situation when the conversion processing unit 11f1out and the conversion processing unit 12f2out are deployed at time t1 and time (t2+1), respectively. It is also preferable to prepare as many such first conversion processing units as there are configured domain-specific RNN cells (two in FIG. 4).

Furthermore, each conversion process (f1out, f2out) may be, for example, a process of multiplying a weight matrix by domain-specific hidden state information, or may be a process of further adding a bias vector to it. Furthermore, it is also possible to input it into a sigmoid function, a tanh function, a ReLu function, or the like, and output the obtained value. In any case, by setting the first conversion processing unit (conversion processing unit 11f1out and conversion processing unit 12f2out) and executing such a conversion process, it is possible to adjust the influence that the user's behavior in each behavior domain has on the domain-independent dynamic user attributes (user attributes in the second quadrant in Figure 3).

Here, these domain-independent dynamic user attributes are dynamic psychological characteristics such as fatigue level, drowsiness, emotions, and mood, and the extent to which user behavior in each behavioral domain affects such characteristics is thought to differ for each behavioral domain. For example, video viewing behavior on a video distribution service is thought to have a greater impact on fatigue level, emotions, and mood than purchasing behavior on a shopping site for daily necessities. Therefore, by setting a first conversion processing unit (which executes different conversion processing) for each domain-specific RNN cell, it is possible to adjust such influences.

In addition, the personality information generation model 1′ of this embodiment further includes:
(b1) a conversion processing unit 11f1in as a second conversion processing unit, which performs a predetermined type of conversion processing (f1in) on the domain-independent hidden state information h ^DIL output from the DIL cell 10, and inputs the domain-independent hidden state information f1in(h ^DIL ) subjected to the conversion processing (f1in) to the DSL1 cell 11;
(b2) a conversion processing unit 12f2in as a second conversion processing unit, which performs a predetermined type of conversion processing (f2in) on the domain-independent hidden state information h ^DIL output from the DIL cell 10, and inputs the domain-independent hidden state information f2in(h ^DIL ) subjected to the conversion processing (f2in) to the DSL2 cell 12;
It is equipped with:

Incidentally, FIG. 4 also shows the situation when the conversion processing unit 11f1in and the conversion processing unit 12f2in are deployed at time t1 and time (t2+1), respectively. It is also preferable to prepare as many such second conversion processing units as there are configured domain-specific RNN cells (two in FIG. 4).

Furthermore, each conversion process (f1in, f2in) can be a process similar to the conversion process (f1in and f2in) in the first conversion processing unit. In any case, by setting the second conversion processing unit (conversion processing unit 11f1in and conversion processing unit 12f2in) and executing such a conversion process, it is possible to adjust the influence of user attributes that are independent of the behavioral domain on dynamic user attributes that are dependent on each behavioral domain (user attributes in the third quadrant in Figure 3, dynamic psychological state information).

It is also preferable that the first conversion processing unit and the second conversion processing unit described above are trained together with the DSL1 cell 11 and the DSL2 cell 12. This ensures that the effects of the behavior of each behavioral domain as described above can be adjusted more reliably.

[Personality estimation device, personality estimation program]
3, a personality estimation device 9 that is equipped with the personality information generation model 1 or the personality information generation model 1' (FIG. 4) as described above and estimates the personality of a user to be estimated will be described below. Incidentally, the following description will be given assuming that this device is equipped with the personality information generation model 1.

The personality estimation device 9 of this embodiment shown in the lower part of Figure 3 generates and outputs "personality information" of the user to be estimated using a learned user distributed expression extraction unit 10u (user expression generation unit) obtained by training the installed personality information generation model 1.

Specifically, in FIG. 3, the input unit 91 of the personality estimation device 9 has a communication function, and acquires time-series information on the behavior of a large number of users in each behavioral domain from, for example, an externally installed behavioral domain-related management server (e.g., a web shopping management server or a video distribution management server), and outputs the information to the training unit 92. It also receives information on the estimation target user, and outputs the information to the personality estimation unit 93.

The training unit 92 generates teacher data from the received time-series information relating to the behavior of the multiple users in each behavioral domain, and uses this to train the personality information generation model 1.

The personality estimation unit 93 generates user identification information (one-hot vector) of the estimated user from the received information related to the estimated user, inputs this to the learned user distributed expression extraction unit 10u (user expression generation unit) in the trained personality information generation model 1, and obtains the user expression vector of the estimated user from the user distributed expression extraction unit 10u. As already explained, this user expression vector can be regarded as a distributed expression of the estimated user's personality, that is, as a personality expression vector (personality embedding vector).

The output unit 94 transmits the received personality embedding vector to the outside as "personality information" of the estimated target user (if it has a communication function) or displays it (if it has a display function).

The training unit 92 and the personality estimation unit 93 are the main functional components that implement one embodiment of the personality estimation method according to the present invention, and can also be considered as functions of a processor/memory that stores one embodiment of the personality estimation program according to the present invention. As a result, the personality estimation device 9 may be a dedicated device for personality estimation, but it can also be, for example, a cloud server, a non-cloud server device, a personal computer (PC), a notebook or tablet computer, or a smartphone, etc., that is equipped with the personality estimation program according to the present invention.

As described above in detail, according to the present invention, it is possible to generate personality information with higher information granularity and reliability from behavioral domain information of the estimated target user without using or relying on measurement results from, for example, a questionnaire survey that has been used in the past. Furthermore, by using the user's personality information generated in this way, it is possible to personalize the products and services provided, for example, in the field of marketing, and to implement suitable or effective recommendations, etc.

In addition, for example, the personality information of children (estimated from their behavior) generated by the present invention can be used to provide them with a high-quality education that is tailored to their individual personalities. In other words, the present invention can contribute to Goal 4 of the United Nations-led Sustainable Development Goals (SDGs), which is to "Ensure inclusive and equitable quality education and promote lifelong learning opportunities for all."

Furthermore, the personality information of adults (estimated from their behavior) generated by the present invention can be used to provide them with decent work that does not harm the environment, and high-quality work that suits their individual personalities. In other words, the present invention can contribute to Goal 8 of the United Nations-led SDGs, "Promote inclusive and sustainable economic growth, employment and decent work for all."

Furthermore, for example, the personality information (estimated from the consumer's behavior) generated by the present invention, as well as the consumer's lifestyle history and consumption activity history, can be utilized to provide the consumer with education on sustainable consumption and lifestyles that are in line with the consumer's personality and current lifestyle. In other words, the present invention can contribute to the achievement of Goal 12 of the United Nations-led SDGs, "Ensure sustainable consumption and production patterns."

With respect to the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical ideas and viewpoints of the present invention can be easily made by a person skilled in the art. The above description is merely an example and is not intended to be restrictive in any way. The present invention is limited only by the scope of the claims and their equivalents.

1, 1' Personality information generation model 10 DIL cell (domain-independent RNN model)
10o, 11o, 12o Output layer 10u User distributed expression extraction unit (user expression generation unit)
11 DSL1 cell (domain specific RNN model)
11f1in, 11f1out, 12f2in, 12f2out Conversion processing unit 11i Item distributed representation extraction unit 11u, 12u User distributed representation extraction unit 12 DSL2 cell (domain-specific RNN model)
12m Video distributed expression extraction unit 9 Personality estimation device 91 Input unit 92 Training unit 93 Personality estimation unit 94 Output unit

Claims (10)

A personality information generation model that causes a computer to function to estimate a personality of a user from information related to each of a plurality of behavioral domains of the user, the personality information generation model comprising:
a plurality of behavioral expression generating units set for each of the behavioral domains, each generating behavioral expression information that is information expressing the behavior of the user in the behavioral domain from information that identifies the content of the behavior of the user in the behavioral domain;
a plurality of domain-specific regression neural network (RNN) cells set for each of the behavioral domains, each of which receives the behavioral expression information of the user in the behavioral domain, and generates new domain-specific hidden state information corresponding to the hidden state information of the RNN by reflecting the behavioral expression information of the user received this time on domain-specific hidden state information corresponding to the RNN hidden state information that the RNN cell itself generated when it previously received the behavioral expression information of the user ;
a user expression generating unit that generates user expression information that represents the user from information that identifies the user;
a computer is caused to function as a domain-independent RNN cell that receives the generated domain-specific hidden state information and the generated user expression information, and generates new domain-independent hidden state information corresponding to the RNN hidden state information by reflecting the currently received domain-specific hidden state information of the user and the currently received user expression information on domain-independent hidden state information corresponding to the RNN hidden state information that the computer generated when the domain-specific hidden state information of the user was previously received ;
The user expression generating unit, trained by the backpropagation method of the RNN together with the plurality of domain-specific RNN cells and the domain-independent RNN cells, generates the user expression information of the user as information expressing the personality of the user in the sense of domain-independent and static information related to the user.
A personality information generation model characterized by : The personality information generation model according to claim 1, characterized in that the domain-independent RNN cell receives domain-specific hidden state information that has been recently generated or has just been generated in the domain-specific RNN cell, and generates the domain-independent hidden state information. The personality information generation model according to claim 1 or 2, further comprising a computer that functions as a first conversion processing unit that performs a predetermined type of conversion processing on the domain-specific hidden state information output from the domain-specific RNN cell, and inputs the domain-specific hidden state information that has been subjected to the conversion processing to the domain-independent RNN cell. The personality information generation model according to any one of claims 1 to 3, characterized in that the domain-specific RNN cell also receives the domain-independent hidden state information that was generated most recently from the time of receiving the behavioral expression information, and generates the domain-specific hidden state information using the domain-independent hidden state information. The personality information generation model according to claim 4, further comprising a computer that functions as a second conversion processing unit that performs a predetermined type of conversion processing on the domain-independent hidden state information output from the domain-independent RNN cell and inputs the domain-independent hidden state information that has been subjected to the conversion processing to the domain-specific RNN cell. 6. The personality information generation model according to claim 1, wherein the user expression generation unit outputs the generated user expression information as information expressing the personality of the user. The personality information generation model according to any one of claims 1 to 6, characterized in that the domain-independent RNN cell outputs the generated domain-independent hidden state information as dynamic psychological state information, which is information related to the dynamic psychological state of the user. The personality information generation model according to any one of claims 1 to 7, characterized in that each of the plurality of domain-specific RNN cells and the domain-independent RNN cell has at least one of a forget gate, an update gate, and a reset gate that perform a predetermined process on information input to the cell to generate new hidden state information. A personality estimation device, characterized in that it generates and outputs information expressing the personality of a user using the learned user expression generation unit obtained by training a personality information generation model described in any one of claims 1 to 8. The personality information generation model according to any one of claims 1 to 8 is trained by a backpropagation method of an RNN ;
A personality estimation method in a computer, characterized in that the trained user expression generation unit obtained by the training is used to generate and output user expression information of the user, the user expression information being information that expresses the personality of the user in the sense of being domain-independent and static information related to the user .

Images