JP2017211898A - Learning system, feature learning device, method thereof, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a feature learning device with which it is possible to collect training data and execute learning by the training data and prevent learning by biased training data, without providing an explicit preregistration phase.SOLUTION: The feature learning device includes: a person's training data configuration unit for storing a lineage vector v(t) that includes feature information as an element in a person's queue table in correlation to the identifier IDof a user m, where the number of lineage vectors included in the person's queue table of the user m at time t is P; an other's training data configuration unit for extracting Plineage vectors one at a time from one of the person's queue tables of users other than the user m, and storing the extracted Plineage vectors in an other's queue table in correlation to the identifier ID; and a learning unit for performing supervised machine learning using the person's queue table and the other's queue table as training data.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a method for collecting data for use in biometric authentication using behavioral features.

  Non-Patent Document 1 is known as a conventional technique for biometric authentication using behavioral features. In Non-Patent Document 1, a plurality of feature information is acquired using a sensor provided in a mobile terminal such as a smartphone, and authentication is performed using the feature information.

  In biometric authentication using behavioral features, feature information extracted from behavior information in the pre-authentication phase (pre-registration phase) is input to the authentication system to create a template, and the created template is registered. In the authentication phase, authentication is realized by identifying the identity using the registered template.

  Non-Patent Document 1 proposes a method using a support vector machine. In the pre-registration phase, “binary class (user / other than user)” and feature vectors corresponding to each class (feature information) Is collected as teacher data, and learning is performed using the collected teacher data. In the authentication phase, the identity is identified by solving the class classification problem for the unknown feature vector (user feature information) using the learning result (template).

H. Witte, C. Rathgeb and C. Busch, "Context-Aware Mobile Biometric Authentication based on Support Vector Machines", 2013 Fourth International Conference on Emerging Security Technologies, pp.29-32, 2013.

However, when a machine learning algorithm typified by a support vector machine is applied to authentication, the following problems exist as authentication-specific problems.
(1) Timing of learning: It is necessary to learn from the teacher data before the authentication phase. In the pre-registration phase, registration related to collection and learning of feature information of the user (person) and other users (others) The operation leads to a decrease in user convenience.
(2) Creation of teacher data: When trying to solve a binary classification problem for a specific ID, the number of others is much larger than the number of users (1) -1), the number of teacher data in each class is not equal, and learning is performed with biased teacher data.

  The present invention can (i) collect teacher data without performing an explicit pre-registration phase and execute learning with the teacher data, and (ii) prevent learning with biased teacher data It is an object to provide a learning system, a feature learning device, a method thereof, and a program.

In order to solve the above problems, according to an aspect of the present invention, a learning system includes a feature collection device that collects feature information based on user behavior, and a feature learning device that learns using the collected feature information. Including. The feature collection device includes a feature collection unit that collects feature information based on the behavior of the user m, where t is an index representing time. The feature learning device sets P _{m, t} as the number of feature vectors included in the user m's own queue table at time t, a feature vector v _m (t) including feature information as an element, and an identifier ID _{m of the} user _m. Is stored in the principal queue table Table _{m, target} (t) = {v _m (t), v _m (t-1), ..., v _m (tP _{m, t} +1)} The component part and P _{m, t} feature vectors are extracted one by one from one of the user's own queue tables other than user _m, and P _{m, t} feature vectors and identifier ID _m are extracted. And the other person queue table Table _{m, others} (t) and the other person queue table Table _{m, target} (t) and the other person queue table Table _{m, others} (t) And a learning unit that performs supervised machine learning.

In order to solve the above-described problem, according to another aspect of the present invention, the feature learning device uses t as an index representing time, and calculates the number of feature vectors included in the personal queue table of the user m at time t. P _{m, t} , a feature vector v _m (t) that includes characteristic information obtained based on the behavior of user m as an element and an identifier ID _{m of} user m are associated with each other, and the user queue table Table _{m, target} (t ) = {v _m (t), v _m (t-1), ..., v _m (tP _{m, t} +1)} and P _{m, t} feature vectors Each one of the user's own queue tables other than user m is extracted, and P _{m, t} extracted feature vectors and identifier ID _m are associated with each other's queue table Table _{m, others} (t and others teacher data construction unit to be stored in), person-queue table table _{m, target} (t) and others queue table table _m, and _others (t) as teacher data , And a learning unit that performs a supervised machine learning.

In order to solve the above problems, according to another aspect of the present invention, a learning method uses a feature collection device that collects feature information based on a user's behavior and the collected feature information. In the learning method, t is an index representing time, and the feature collection device collects feature information based on the behavior of the user m, and the feature vector included in the user's own queue table at the time t the number of P _m, and _t, feature learning device, feature vector v _m (t) and the user himself queue table in association with the identifier ID _m of m table _m including characteristic information as _{elements, target} (t) = {v _m (t), v _m (t-1), ..., v _m (tP _{m, t} + 1)}
Feature learning device, P _m, the _t pieces of feature vectors taken out from either of the principal queue table of the user other than the user m one by one, P _{m, t} pieces of the extracted feature vector and the identifier ID _The other person teacher data composition step that associates _m with each other and stores it in the other person queue table Table _{m, others} (t), and the feature learning device, the person queue table Table _{m, target} (t) and the other person queue table Table _{m, others} ( and a learning step for performing supervised machine learning using t) as teacher data.

  According to the present invention, (i) it is possible to collect teacher data without performing an explicit pre-registration phase and execute learning using the teacher data, and (ii) prevent learning using biased teacher data. There is an effect that can be.

The functional block diagram of the learning system which concerns on 1st embodiment. The figure which shows the example of the processing flow of the learning system which concerns on 1st embodiment. The figure which shows the example of the processing flow of the learning system which concerns on 1st embodiment. The figure which shows the example of the processing flow of the learning system which concerns on 1st embodiment. The figure which shows the example of data of a principal queue table. The figure which shows the example of data of a stranger queue table. The figure which shows the example of data of a learning result table.

  Hereinafter, embodiments of the present invention will be described. In the drawings used for the following description, constituent parts having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted.

<Points of first embodiment>
Queues for storing the identity vector of the person (hereinafter also referred to as “personal queue”) and queues for temporarily storing the identity vector of other persons (hereinafter also referred to as “other person queue”) are used as user identifiers. Prepare for each. A “queue” is a data structure that holds data in a first-in first-out list structure.

  In this embodiment, first, authentication is performed by executing other authentication means different from the authentication means using the learning result of this embodiment and identifying the identity. When authentication is successful, the collected feature information (hereinafter also referred to as “feature information collected when authentication is successful”) is highly likely to be the identity of the person during authentication processing or until it is determined that authentication is successful. And the subsequent processing is executed in the background.

The feature learning device of the present embodiment acquires the identifier ID _{m of the} user m used by other authentication means and the feature information collected when the authentication is successful, and adds the feature vector to the principal queue corresponding to the identifier ID _m Store. Note that m is an index representing a user, and m is one of 1, 2,..., M, where M is the total number of users of the learning system of the present embodiment. If the personal queue is already full, the old feature vectors are deleted in order, and a storage area in the personal queue is secured.

When additional storage in the personal queue is completed, the other person's queue is initialized (cleared). Then, the same number of feature vectors as the number of stored principal queues are selected from the feature vectors stored in “the principal queue corresponding to the identifier ID _n other than the input identifier ID _m ” and stored in the other person queue. If the total number of users of the learning system of the present embodiment is M, n = 1, 2,..., M, and m ≠ n.

  When storage of feature vectors in the other person's queue is completed, machine learning learning processing is performed using all feature vectors stored in the "person's queue" and "other person's queue" corresponding to the ID as teacher data. To do.

  With such a configuration, a feature vector can be created based on feature information when authentication is successful in another authentication means. That is, feature information can be collected and a feature vector can be created in the course of executing the authentication phase. Since the identifier and the feature vector have been successfully authenticated by other authentication means, it is guaranteed that the feature vector and the identifier correspond to each other. With such a configuration, it is possible to collect and learn teacher data without providing an explicit pre-registration phase.

  In addition, since teacher data is collected and learned after successful authentication during the authentication phase, even if the amount of teacher data increases, there is no impact on the waiting time of the learning process for the user. It is possible to learn parameters, state variables, and the like (hereinafter also referred to as parameters) in machine learning in the ground. An identification function for identifying the identity can be configured using these parameters, etc., and the identity of the user is identified by giving an unknown feature vector (user feature information) to this identification function. can do.

  Even if the user's feature information changes over time, in order to collect and learn teacher data in the authentication phase, rather than in the pre-registration phase, an identification function for identifying the user feature information It can follow changes.

  Hereinafter, a configuration for realizing this process will be described.

<Learning system according to the first embodiment>
FIG. 1 is a functional block diagram of a learning system 10 according to the first embodiment, and FIGS. 2, 3 and 4 show the processing flow.

  The learning system 10 includes a feature collection device 100 and a feature learning device 200.

  The feature collection device 100 includes a feature collection unit 101, a feature vector generation unit 102, and an authenticated unit 103.

  The feature learning device 200 includes a principal teacher data configuration unit 201, an authentication unit 202, another person teacher data configuration unit 203, a storage unit 204, and a learning unit 205.

  The feature collection device 100 and the feature learning device 200 can be connected via a network.

  In the feature collection device 100 and the feature learning device 200, for example, a special program is read into a known or dedicated computer having a central processing unit (CPU), a main storage device (RAM), and the like. It is a special device constructed. For example, the feature collection device 100 is incorporated in a client terminal used by a user, for example, a smartphone. The feature learning device 200 is incorporated in a server device (a computer that provides services to a plurality of users and programs, that is, clients on a network). The service requires a user identifier, and in this embodiment, the feature learning device 200 is incorporated in an authentication server that requires a user identifier.

  For example, the feature collection device 100 and the feature learning device 200 execute each process under the control of each central processing unit. The data input to the feature collection device 100 and the feature learning device 200 and the data obtained by each process are stored in each main storage device, for example, and the data stored in the main storage device is read as necessary And used for other processing. Further, at least a part of each processing unit of the feature collection device 100 and the feature learning device 200 may be configured by hardware such as an integrated circuit.

<Feature Collection Unit 101>
The feature collection unit 101 monitors the behavior of the user, collects data (hereinafter also referred to as “feature information”) based on the behavior of the user m (S1 in FIG. 2), and transmits the data to the feature vector generation unit 102. As a feature information acquisition method, a sensor mounted in a smartphone may be used. Any device may be used as long as it can collect feature information used when learning parameters and the like in machine learning. For example, an acceleration sensor or a gyro sensor mounted in a smartphone can be cited as the feature collection unit 101, and the output values thereof are collected as feature information. Further, a pressure sensor equipped with a touch panel or a GPS may be used as the feature collection unit 101, and the output value and position information of the pressure sensor may be collected as feature information. Further, these combinations may be collected as feature information.

<Feature Vector Generation Unit 102>
The feature vector generation unit 102 receives the feature information collected by the feature collection unit 101, converts the feature information into a multi-dimensional vector format, generates a feature vector (S2), and is included in the feature learning device 200. It transmits to the principal teacher data structure part 201. For example, when the feature collection unit 101 includes an acceleration sensor and a gyro sensor, these data may be converted into a two-dimensional vector format and used as a feature vector. Further, for example, by converting the output value of the time series N ₁ pieces of the acceleration sensor, by combining the output values of the time series of _two N gyro sensor (N ₁ + N ₂₎ in the form of a dimensional vectors, feature It may be a vector. For example, let d _m (t, e) be some feature information of user m at time t, and let v _m (t) = {d _m (t, 1), d _m (t, 2) , ..., d _m (t, E)}. Note that t is an index indicating discrete time, and m is an index indicating users. For example, M is the total number of users of the learning system 10, and m = 1, 2,. Note that the feature collection unit 101 may use values obtained by performing some processing as the feature information without using the output values of the acceleration sensor and the gyro sensor as the feature information as they are. For example, a total or average value of time-series output values for a predetermined time may be used as the feature information.

<Authenticated unit 103>
The authenticating unit 103 has a client function for authenticating the identifier ID _m of the user _m , and executes an authentication process with the authenticating unit 202 of the feature learning device 200 described later (S3). For example, the authenticating unit 103 requests the authenticating unit 202 to confirm the validity of the user _{m with} the identifier ID _m . Means for authentication is not limited, and any existing authentication technique may be used, but means corresponding to the authentication unit 202 is provided. For example, password authentication, authentication using an electronic certificate, or biometric authentication may be used. Moreover, you may utilize the one part authentication result in multistep authentication.

<Authentication unit 202>
The authentication unit 202 has a server function for authenticating the identifier ID _m of the user _m , receives an authentication request from the principal teacher data configuration unit 201, and performs an authentication process with the authenticated unit 103 of the feature collection device 100 Is executed (S3), and when the authentication is successful (when accepted), the identifier ID _{m of the} user m who has succeeded in the authentication is output to the principal teacher data construction unit 201. For example, the authentication unit 202 verifies the validity of the user _{m with} the identifier ID _m . Means for authenticating is not limited, and any existing authentication technique may be used, but means corresponding to the authenticated unit 103 is provided. For example, password authentication, authentication using an electronic certificate, or biometric authentication may be used. Moreover, you may utilize the one part authentication result in multistep authentication. The authentication result is transferred to the teacher data composition unit 201 together with the identifier ID _m .

  For example, when password authentication is performed between the authenticated unit 103 and the authenticating unit 202, SSL (Secure Sockets Layer) / TLS (Transport Layer Security) communication or the like is performed between the authenticated unit 103 and the authenticating unit 202. The authenticated unit 103 transmits the encrypted password to the authenticating unit 202. For example, a control signal indicating that a connection is requested to the authentication target unit 103 is transmitted to the authentication unit 202. The authentication unit 202 transmits the server certificate to the authenticated unit 103. If the server certificate is trustworthy, the authenticated unit 103 encrypts the common key with the public key included in the server certificate, and transmits the encrypted common key to the authentication unit 202. The authentication unit 202 decrypts the common key encrypted with the secret key corresponding to the public key. Thereafter, data encrypted using the common key is transmitted and received between the authenticated unit 103 and the authenticating unit 202. For example, the to-be-authenticated unit 103 transmits an encrypted user identifier and password, and the authentication unit 202 decrypts it and refers to the user identifier and password stored in a storage unit (not shown) to authenticate the user. Execute the process.

  The subsequent processing is executed when authentication is successful in the authentication unit 202, and is allowed to be executed in the background. Further, for example, when authentication is successful (accepted), the user m may be notified that the authentication was successful, and a predetermined service may be provided. Further, when the authentication fails (when rejected), the user m may be requested to authenticate again or may be terminated without providing the service.

<Principal Teacher Data Configuration Unit 201>
The principal teacher data construction unit 201 is called from the feature vector generation unit 102 together with the feature vector v _m (t). The principal teacher data construction unit 201 receives the feature vector v _m (t) as an input, and calls the authentication unit 202 to associate the received feature vector v _m (t) with the user m. After successful authentication by the authentication unit 202, the principal teacher data configuration unit 201 receives an authentication result indicating that the authentication is successful and an identifier ID _m (identifier subjected to authentication processing) of the user m, and the received feature It assumes that brute string vector v _m (t) is the identifier ID _m of authentication success, in association with the feature vector v _m (t) and the identifier ID _m, principal queue table table _m of the storage unit _{204, target} (t) = {v _m (t), v _m (t−1),..., v _m (tP _{m, t} +1)} (S4). FIG. 5 shows an example of data in the personal queue table. Here, P _{m, t} indicates the number of feature vectors included in the personal queue table of the user m at time t, and is also referred to as the stored number. Thereafter, the other person teacher data construction unit 203 is called together with the stored number P _{m, t of} the principal queue table. As shown in FIG. 5, the personal queue table is prepared for each user. Note that the number of feature vectors stored in the principal queue table at time t differs depending on the number of users because it differs depending on the number of authentications. However, the maximum number of feature vectors stored in the principal queue table is the same, for example, P _max . The principal teacher data construction unit 201 already has the principal queue table Table _{m, target} (t) = {v _m (t), v _m (t-1), ..., v _m (tP _{m, t} +1)}. When P _max feature vectors are stored, the oldest feature vector v _m (tP _max ) is deleted, a new feature vector v _m (t) is stored in the user queue table of the storage unit 204, and the user queue is stored. Let the table be Table _{m, target} (t) = {v _m (t), v _m (t-1),..., V _m (tP _max +1)}.

<Other teacher data configuration unit 203>
The other person teacher data construction unit 203 is called from the principal teacher data construction unit 201 together with the identifier ID _m and the stored number P _{m, t of} the principal queue table. The other person teacher data configuration unit 203 receives the identifier ID _m and the stored number P _{m, t} of the personal queue, and first, for each identifier ID _n other than the received identifier ID _m , each personal queue table Table _{n, target} A feature vector is sampled from (t). However, n = 1, 2,..., M and n ≠ m. The number of samplings is the same as the stored number P _{m, t of} the principal queue table Table _{m, target} (t) received when this function unit is called. Thereafter, the set of sampled feature vectors is regarded as a feature vector of another person corresponding to the identifier ID _m , and the other person queue table Table _{m, others} (t) = {v _{m, others} (t), v _{m, others} (t-1), ..., v _{m, others} (tP _{m, t} +1)} (S5 in FIG. 3). FIG. 6 shows an example of data in the other person queue table.

If the stored data of the queue corresponding to the identifier ID _m is already stored in the other person queue table, the old data is cleared and stored. Thereafter, the learning unit 205 is called together with the identifier ID _m .

The sampling method is not limited. For example, the personal queue table to be sampled may be randomly selected from the personal queue tables Table _{n, target} (t) corresponding to M-1 identifiers ID _n for each storage order of the other person queues. In other words, the p-th queue of any one of the individual queue tables Table _{n, target} (t) is selected as the p-th queue stored in the other-person queue table.

Further, the necessary number of feature vectors may be randomly sampled from the principal queue table Table _{n, target} (t) corresponding to M-1 identifiers ID _n without considering the queue storage order. . In other words, as the p-th queue to be stored in the other-person queue table, the q-th queue of any one of the individual queue tables Table _{n, target} (t) is selected. However, q is one of 1, 2,..., P _{t, n , where} P _{t, n} is the number of stored user n's personal queue table Table _{n, target} (t).

That is, the other person teacher data configuration unit 203 extracts P _{m, t} feature vectors one by one from any one of the user's own queue tables Table _{n, target} (t) other than the user m, and P _{The m, t} extracted feature vectors and the identifier ID _m are associated with each other and stored in the other person queue table Table _{m, others} (t) (S5).

<Storage unit 204>
The storage unit 204 receives an identifier ID _m , a feature vector v _m (t), a set of feature vectors {v _{m, others} (t), v _{m, others} (from the principal teacher data configuration unit 201 and the other teacher data configuration unit 203. t-1),..., v _{m, others} (tP _{m, t} +1)} are received, and reading / writing is performed on the personal queue table (see FIG. 5) and the other person queue table (see FIG. 6). Further, the learning unit 205 receives an instruction together with the ID and the learning result w _m (t), and reads / writes the learning result table.

<Learning unit 205>
The learning unit 205 is called from the other person teacher data configuration unit 203 together with the identifier ID _m . The learning unit 205 receives the identifier ID _m as input, refers to the storage unit 204, and stores it from each queue corresponding to the identifier ID _m in the principal queue table (see FIG. 5) and the other person queue table (see FIG. 6). Set of feature vectors {v _m (t), v _m (t-1),…, v _m (tP _{m, t} +1)}, {v _{m, others} (t), v _{m, others} (t -1), ..., v _{m, others} (tP _{m, t} +1)}. At this time, a label of “principal” is assigned to the feature vector {v _m (t), v _m (t−1),..., V _m (tP _{m, t} +1)} extracted from the principal queue table. Feature vectors taken from other person's queue table {v _m (t), v _m (t-1),…, v _m (tP _{m, t} +1)}, {v _{m, others} (t), v _{m, others} (t-1), ..., v _{m, others} (tP _{m, t} +1)} are given the label “other”. Thereafter, a learning algorithm such as supervised machine learning is executed by the learning unit 205 using a set of these labels and feature vectors as teacher data (S6 in FIG. 4), and a learning result w _m (t) is generated. Examples of the generated learning result include parameters and state variables in machine learning. Using this parameter or the like, an identification function for identifying the identity can be configured. These pieces of information are stored in the learning result table of the storage unit 204 as learning results w _m (t) corresponding to the identifier ID _m (see FIG. 7).

The learning result w _m (t) stored in the learning result table can be used in an authentication unit different from the authentication performed between the authenticating unit 103 and the authenticating unit 202. Output in accordance with For example, another authentication means uses a parameter of the learning result w _m (t) to configure an identification function for identifying the identity, and an unknown feature vector (user characteristic information) is used for this identification function. The identity of the user can be identified.

<Effect>
With the above configuration, (i) teacher data can be collected and learning using the teacher data can be executed without providing an explicit pre-registration phase. Further, since the other person teacher data construction unit 203 stores the same number of feature vectors as the other person queue table in the other person queue table, (ii) learning with biased teacher data can be prevented. Moreover, since a learning result can be produced | generated for every authentication process, a learning result can be made to follow the change of a user's feature information.

<Modification>
In the present embodiment, the feature learning device 200 is incorporated in the authentication server, but any server can be applied as long as it uses a user identifier. In other words, any server may be used as long as the server can acquire the user identifier when providing the service and can receive the feature vector at that time.

In the present embodiment, the learning result w _m (t) stored in the learning result table is provided to an authentication unit different from the authentication performed between the authenticated unit 103 and the authenticating unit 202. The authentication unit 103 and the authentication unit 202 may be used for authentication. That is, the authentication unit 202 verifies the validity of the identifier ID _m using the learning result w _m (t). In this case, as the number of authentications increases, the feature vectors stored in the principal queue table and the other person queue table increase, so the accuracy of authentication using the learning result w _m (t) obtained from those feature vectors also improves. It is thought to go. With such a configuration, multi-step authentication can be realized without making the user feel bothersome.

<Other variations>
The present invention is not limited to the above-described embodiments and modifications. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

<Program and recording medium>
In addition, various processing functions in each device described in the above embodiments and modifications may be realized by a computer. In that case, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its storage unit. When executing the process, this computer reads the program stored in its own storage unit and executes the process according to the read program. As another embodiment of this program, a computer may read a program directly from a portable recording medium and execute processing according to the program. Further, each time a program is transferred from the server computer to the computer, processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program includes information provided for processing by the electronic computer and equivalent to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In addition, although each device is configured by executing a predetermined program on a computer, at least a part of these processing contents may be realized by hardware.

Claims (6)

A learning system including a feature collection device that collects feature information based on user behavior and a feature learning device that learns using the collected feature information,
The feature collection device includes:
including a feature collection unit that collects feature information based on the behavior of user m, where t is an index representing time,
The feature learning device includes:
The number of feature vectors included in the user's m queue table at time t is P _{m, t,} and the feature vector v _m (t) including the feature information as an element corresponds to the identifier ID _{m of the} user m In addition, the principal teacher data configuration unit stored in the principal queue table Table _{m, target} (t) = {v _m (t), v _m (t-1), ..., v _m (tP _{m, t} +1)} ,
P _{m, t} feature vectors are extracted one by one from the user's own queue table other than user _m, and P _{m, t} feature vectors are associated with the identifier ID _m. In addition, the other person teacher data composition part to be stored in the other person queue table Table _{m, others} (t),
A learning unit that performs supervised machine learning using the person queue table Table _{m, target} (t) and the other person queue table Table _{m, others} (t) as teacher data;
Learning system. The learning system of claim 1,
The feature collection device includes:
An authenticated unit that requests confirmation of the validity of the user _m of the identifier ID _m ,
The feature learning device includes:
An authentication unit for verifying the validity of the identifier ID _m ,
The processing in the principal teacher data configuration unit, the other person teacher data configuration unit, and the learning unit is executed when authentication is successful in the authentication unit.
Learning system. The learning system according to claim 2,
In the authentication unit, the validity of the identifier ID _m is verified using the learning result of the learning unit.
Learning system. Let t be an index representing time, let P _{m, t} be the number of feature vectors included in the user's m personal queue table at time t _, and feature vector v containing feature information obtained based on the behavior of user m as an element _m (t) and the user himself queue table in association with the identifier ID _m of _{m table m, target (t)} = {v m (t), v m (t-1), ..., v m (tP _{m, t} +1)} and the teacher data composition part
P _{m, t} feature vectors are extracted one by one from the user's own queue table other than user _m, and P _{m, t} feature vectors are associated with the identifier ID _m. In addition, the other person teacher data composition part to be stored in the other person queue table Table _{m, others} (t),
A learning unit that performs supervised machine learning using the person queue table Table _{m, target} (t) and the other person queue table Table _{m, others} (t) as teacher data;
Feature learning device. A feature collection device that collects feature information based on user behavior, and a learning method that uses the collected feature information,
t is an index representing time, and the feature collection device collects feature information based on the behavior of the user m; and
Let P _{m, t} be the number of feature vectors included in the user's m personal queue table at time t _, and the feature learning device may identify the feature vector v _m (t) including the feature information as an element and the user m's The identifier ID _m is associated and stored in the principal queue table Table _{m, target} (t) = {v _m (t), v _m (t-1), ..., v _m (tP _{m, t} +1)} Personal teacher data composition step;
The feature learning device takes out P _{m, t} feature vectors one by one from any of the user's own queue tables other than the user _{m, and} extracts the P _{m, t} feature vectors and the The other person teacher data configuration step of associating the identifier ID _m with each other and storing it in the other person queue table Table _{m, others} (t),
The feature learning device includes a learning step of performing supervised machine learning using the person queue table Table _{m, target} (t) and the other person queue table Table _{m, others} (t) as teacher data.
Learning method.   A program for causing a computer to function as the feature learning apparatus according to claim 4.

Images