JP7041018B2 - Learning equipment, learning methods and learning programs - Google Patents

Description

The present invention relates to a learning device, a learning method and a learning program.

Conventionally, a technique for providing various types of information to a user has been known. As an example of such technology, based on the information related to the information selected by the user in the past, the information corresponding to the user such as the information highly related to the user and the information that the user is interested in is estimated. However, there is known a technique for providing the estimated information to the user.

Japanese Unexamined Patent Publication No. 2012-150561

However, with the above-mentioned conventional technology, there is a possibility that the information corresponding to the user cannot be estimated appropriately.

For example, in the above-mentioned conventional technology, it is not possible to appropriately estimate information corresponding to a user for a user who has not selected information in the past, such as a new user.

The present application has been made in view of the above, and an object thereof is to improve the estimation accuracy of the information corresponding to the user.

The learning device according to the present application includes an acquisition unit that acquires a plurality of distribution target vectors indicating distribution targets, and an acquisition unit.
When a plurality of distribution target vectors indicating a distribution target delivered to a predetermined user and a distribution candidate vector indicating a distribution candidate are input, a plurality of weights set based on the similarity with the distribution candidate vector are set. Based on the distribution target vector and the distribution candidate vector, a model for determining whether or not the distribution candidate corresponds to the predetermined user is learned using the distribution target vector acquired by the acquisition unit. It is characterized by having a learning unit.

According to one aspect of the embodiment, there is an effect that the estimation accuracy of the information corresponding to the user can be improved.

FIG. 1 is a diagram showing a configuration example of a network system according to an embodiment. FIG. 2 is a diagram showing an example of a learning process according to an embodiment. FIG. 3 is a diagram showing a configuration example of the learning device according to the embodiment. FIG. 4 is a diagram showing an example of a training data storage unit according to an embodiment. FIG. 5 is a flowchart showing a learning processing procedure by the learning device according to the embodiment. FIG. 6 is a diagram showing an example of a hardware configuration.

Hereinafter, the learning device, the learning method, and the embodiment for implementing the learning program (hereinafter referred to as “the embodiment”) according to the present application will be described in detail with reference to the drawings. It should be noted that this embodiment does not limit the learning device, learning method and learning program according to the present application. In addition, each embodiment can be appropriately combined as long as the processing contents do not contradict each other. Further, in each of the following embodiments, the same parts are designated by the same reference numerals, and duplicate description will be omitted.

[1. Network system configuration]
First, the configuration of the network system 1 according to the embodiment will be described with reference to FIG. FIG. 1 is a diagram showing a configuration example of the network system 1 according to the embodiment. As shown in FIG. 1, the network system 1 according to the embodiment includes a terminal device 10, a providing device 20, and a learning device 100. The terminal device 10, the providing device 20, and the learning device 100 are connected to the network N by wire or wirelessly, respectively.

Although not shown in FIG. 1, the network system 1 may include terminal devices 10 ₁ to 10 _n . In the present specification, when it is not necessary to distinguish the terminal devices 10 ₁ to 10 _n , the terminal devices 10 ₁ to 10 _n are collectively referred to as "terminal device 10". Further, the network system 1 may include a plurality of providing devices 20 and a plurality of learning devices 100.

The terminal device 10 is an information processing device used by the user. The terminal device 10 may be any type of information processing device including a smartphone, a desktop PC (Personal Computer), a notebook PC, a tablet PC, and a PDA (Personal Digital Assistant).

The providing device 20 is a server device that provides various information to the terminal device 10 and the learning device 100. For example, the providing device 20 provides content (for example, push notification) to the terminal device 10. Further, for example, the providing device 20 provides the learning device 100 with training data for generating a determination model related to the content.

The learning device 100 is a server device that learns a model for determining information suitable for the user. The learning device 100 communicates with the terminal device 10 and the providing device 20 by wire or wirelessly via the network N.

[2. Learning process]
Next, an example of the learning process according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram showing an example of a learning process according to an embodiment.

In the example of FIG. 2, the providing device 20 is _a push type and a plurality of "notifications N1 to Nn (n is an arbitrary natural number)" and "notification NM1" with respect to the terminal device 101 used by the "user U1". ~ NMm (m is an arbitrary natural number) ”is transmitted. The term "push" refers to the mechanism by which information is transmitted based on the trigger of the device that provides the information. In other words, the providing device 20 transmits the notifications N1 to Nn and the notifications NM1 to NMm to the terminal device 10 ₁ as push notifications.

In the example of FIG. 2, the providing device 20 receives the responses to the notifications N1 to Nn and the notifications NM1 to NMm. In _one example, the providing device 20 receives selection information from the terminal device 101 indicating that the notification N1 has been selected. In this case, the notification N1 is considered to be suitable for the user U1. In _another example, the providing device 20 receives selection information from the terminal device 101 indicating that the notification N1 has not been selected. In this case, the notification N1 is considered to be incompatible with the user U1.

In the example of FIG. 2, the providing device 20 transmits the training data to the learning device 100. The training data includes notification information of notifications N1 to Nn and notifications NM1 to NMm (for example, titles of notifications N1 to Nn and notifications NM1 to NMm), and whether or not the user has selected notifications N1 to Nn and notifications NM1 to NMm. Includes selection information (eg, selection flag) indicating.

In the example of FIG. 2, the learning device 100 uses the training data received from the providing device 20 to determine whether or not a predetermined notification is suitable for the user, and is a model (attention) with an attention mechanism. Also called a model). More specifically, the learning device 100 uses the notification information of the notifications N1 to Nn and the notifications NM1 to NMm as input of training data. Further, the learning device 100 uses the selection information indicating whether or not the notifications N1 to Nn are selected as the output of the training data.

In the example of FIG. 2, the notification information of the notifications N1 to Nn and the notifications NM1 to NMm includes "Bow vector (bag of words vector) BVN1 to BVNn" and "BoW vector BVNM1 to BVNMm". At least one non-zero component of the plurality of components of each BoW vector corresponds to at least one word contained in the notification information of each notification. Further, the notification information includes "embedded vectors NEN1 to NENn" and "embedded vectors NENM1 to NENMm" corresponding to the BoW vectors BVN1 to BVNn and the BoW vectors BVNM1 to BVNMm. In one example, the BoW vector corresponds to the title of the notification. For example, if the title of the notification contains the words "tonight", "movie", and "television", the embedded vector is the average of the three distributed representations corresponding to the words "tonight", "movie", and "television" respectively. It is a vector.

In the example of FIG. 2, the selection information indicating whether or not the notifications N1 to Nn are selected includes a selection flag of 0 or 1. The selection flag indicating that the notification N1 has been selected is "1". The selection flag indicating that the notification N1 has not been selected is "0". On the other hand, the selection information indicating whether or not the notifications NM1 to NMm are selected includes the selection flag of "1". That is, the notifications NM1 to NMm are all notifications selected in the past. As described above, the notifications NM1 to NMm correspond to the memory of the attention mechanism.

First, the learning device 100 refers to the training data and acquires the embedded vector NEN1 corresponding to the BoW vector BVN1.

Next, the learning device 100 refers to the training data and acquires the embedded vectors NENM1 to NENMm corresponding to the BoW vectors BVNM1 to BVNMm.

Next, the learning device 100 calculates the weights α ₁ to α _m between the embedded vector NEN1 and the embedded vectors NENM1 to NENMm. The learning device 100 calculates, for example, the weights α ₁ to α _m in the soft attention mechanism. The weights α ₁ to α _m indicate the importance of the embedded vectors NENM 1 to NENM m, respectively. In one example, the weights α ₁ to α _m are calculated based on an alignment model (eg, an alignment model parameterized as a forward propagation neural network). In another example, the weights α ₁ to α _m are calculated based on the inner product of the embedded vector NEN1 and each of the embedded vectors NENM1 to NENMm.

Next, the learning device 100 calculates the weighted average MEM1 of the embedded vectors NENM1 to NENMm using the calculated weights α ₁ to α _m .

Next, the learning device 100 concatenates the embedded vector NEN1 and the calculated weighted average MEM1 (that is, the weighted average vector) to generate a vector. Then, a hidden state vector is generated by applying an activation function (for example, a ReLU (Rectified Linear unit) function) to the generated vector.

Next, the learning device 100 calculates the relevance degree RS1 by multiplying the weight matrix by the generated hidden state vector.

Next, the learning device 100 uses the value of the selection flag included in the selection information indicating whether or not the notification N1 is selected and the calculated relevance degree RS1 to determine whether or not the predetermined notification is suitable for the user. Learn a model with an attention mechanism to judge. For example, the learning device 100 uses an error backpropagation method and uses an objective function (for example, an intersection) corresponding to the value of the selection flag included in the selection information indicating whether or not the notification N1 is selected and the calculated relevance RS1. Learn a model with a soft attention mechanism so that the entropy loss function) is minimized.

For notifications other than notification N1, the learning device 100 can learn a model with a soft attention mechanism using, for example, training data corresponding to notifications N2 to Nn. By repeating the learning process, the learning device 100 can learn a model with a soft attention mechanism so that the generalization error is minimized.

After that, the learning device 100 outputs a relevance score indicating the relevance between the user U1 and the notification by using the model with the learned attention mechanism. For example, the learning device 100 inputs the embedded vector corresponding to the predetermined notification and the embedded vector corresponding to the notifications NM1 to NMm (that is, the embedded vectors NENM1 to NENMm) into the determination model, so that the user U1 determines. The posterior probability of selecting the notification of is output as the relevance score between the user U1 and the notification.

As described above, the learning device 100 according to the embodiment inputs the training data of the embedded vector corresponding to the notification and the embedded vector corresponding to the specific notification (that is, the memory of the attention mechanism) selected in the past. Used as. Further, the learning device 100 uses the value of the selection flag of the notification as the output of the training data. Then, the learning device 100 learns a model with an attention mechanism so that the generalization error is minimized. As a result, the learning device 100 can learn a determination model for determining whether or not a predetermined notification is suitable for the user.

By the way, in the example of FIG. 2, the providing device 20 transmits the notification to the terminal device 10 as a push notification. In general, push notifications are seasonal information. For example, the providing device 20 transmits information with high breaking news (for example, tonight's program) only once as a push notification. Therefore, there may be no delivery history corresponding to such push notifications. Such a situation is associated with a cold start problem. In this regard, the learning device 100 can determine whether or not the push notification is suitable for the user by using the above-mentioned determination model even when the distribution history corresponding to the push notification does not exist. Therefore, the learning device 100 enables the user to receive push notifications of interest to the user in a timely manner. Hereinafter, the learning device 100 that realizes such a learning process will be described in detail.

[3. Configuration of learning device]
Next, a configuration example of the learning device 100 according to the embodiment will be described with reference to FIG. FIG. 3 is a diagram showing a configuration example of the learning device 100 according to the embodiment. As shown in FIG. 3, the learning device 100 includes a communication unit 110, a storage unit 120, and a control unit 130. The learning device 100 has an input unit (for example, a keyboard, a mouse, etc.) that receives various operations from an administrator or the like who uses the learning device 100, and a display unit (liquid crystal display, etc.) for displaying various information. You may.

(Communication unit 110)
The communication unit 110 is realized by, for example, a NIC (Network Interface Card) or the like. The communication unit 110 is connected to the network network by wire or wirelessly, and transmits / receives information to / from the terminal device 10 and the providing device 20 via the network network.

(Memory unit 120)
The storage unit 120 is realized by, for example, a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk. As shown in FIG. 3, the storage unit 120 has a training data storage unit 121 and a determination model storage unit 122.

(Training data storage unit 121)
FIG. 4 is a diagram showing an example of the training data storage unit 121 according to the embodiment. The training data storage unit 121 stores training data for generating a prediction model related to the content. The prediction model is, for example, a determination model for determining whether or not a predetermined notification is suitable for the user. In one example, the determination model is a model with an attention mechanism.

The training data storage unit 121 stores, for example, the training data received by the reception unit 131. In the example of FIG. 4, "training data" is stored in the training data storage unit 121 for each "user ID". The "user ID" indicates an identifier for identifying a user. By way of example, the "training data" includes the items "notification ID", "notification information" and "selection information".

The "notification ID" indicates an identifier for identifying the notification. "Notification information" indicates information regarding the content of the notification. The content of the notification is, for example, the title of the notification, the body of the notification (for example, the text of the text), the action when the notification is selected (for example, the screen transition to the link destination), and the like.

The notification information includes a BoW vector (also referred to as a BoW representation) corresponding to the notification. At least one non-zero component of the plurality of components of the BoW vector corresponds to at least one word contained in the notification information of the notification. Further, the notification information includes an embedded vector corresponding to the Bow vector. In one example, the BoW vector corresponds to the title of the notification. For example, if the title of the notification is "Go to Jim?", The embedded vector is the average vector of the two distributed representations corresponding to the words "Jim" and "Go" respectively.

"Selection information" indicates information regarding the selection of notifications. For example, the selection information indicates whether or not the notification has been selected. For example, the selection information includes a selection flag indicating whether or not the notification was selected. In the example of FIG. 4, the selection flag indicating that the notification has been selected is "1". In this example, the selection flag indicating that the notification was not selected is "0".

For example, FIG. 4 shows that the notification identified by the notification ID “N1” corresponds to the notification information NI1. Further, for example, FIG. 4 shows that the user identified by the user ID “U1” has selected the notification identified by the notification ID “N1”.

For example, FIG. 4 shows that a plurality of notifications corresponding to the notification IDs “NM1” to “NMm” have all been selected in the past by the user identified by the user ID “U1”. The plurality of notifications corresponding to the notification IDs "NM1" to "NMm" correspond to the memory of the attention mechanism.

(Judgment model storage unit 122)
Returning to FIG. 3, the determination model storage unit 122 stores the determination model related to the distribution target. The determination model storage unit 122 stores, for example, a model learned by the learning unit 133 and determines whether or not the distribution candidate corresponds to a predetermined user.

(Control unit 130)
The control unit 130 is a controller, and for example, various programs stored in a storage device inside the learning device 100 by a processor such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit) store a RAM or the like. It is realized by being executed as a work area. Further, the control unit 130 is a controller, and may be realized by, for example, an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

(Receiver 131)
The receiving unit 131 receives training data from the providing device 20 for generating a determination model related to the content. The training data includes notification information of the notification and selection information indicating whether or not the user has selected the notification. The receiving unit 131 may store the received training data in the training data storage unit 121.

The receiving unit 131 receives a plurality of distribution target vectors indicating the distribution target from the providing device 20. For example, the receiving unit 131 receives a distribution target vector in which a predetermined value is stored in a dimension corresponding to a word included in a sentence to be distributed. In one example, the receiving unit 131 receives the embedded vector corresponding to the notification. The above-mentioned notification information may include a plurality of distribution target vectors indicating distribution targets. The receiving unit 131 may store a plurality of received distribution target vectors in the training data storage unit 121. The receiving unit 131 may receive the information included in the distribution target.

The above-mentioned distribution target is, for example, a notification (for example, push notification), a news article, a text-based message (for example, recomment information), or the like. When the delivery target is a push notification, the above-mentioned delivery target vector is, for example, a vector (for example, a distributed representation) in which a predetermined value is stored in a dimension corresponding to a word included in the title of the push notification.

(Acquisition unit 132)
The acquisition unit 132 acquires training data for generating a determination model related to the content. The acquisition unit 132 acquires, for example, the training data received by the reception unit 131. The acquisition unit 132 may acquire training data from the training data storage unit 121.

The acquisition unit 132 acquires a plurality of distribution target vectors indicating the distribution target. For example, the acquisition unit 132 acquires a distribution target vector in which a predetermined value is stored in a dimension corresponding to a word included in a sentence to be distributed.

The acquisition unit 132 acquires, for example, the distribution target vector received by the reception unit 131. The acquisition unit 132 may acquire the distribution target vector from the training data storage unit 121.

(Learning Department 133)
When a plurality of distribution target vectors indicating distribution targets distributed to a predetermined user and distribution candidate vectors indicating distribution candidates are input, the learning unit 133 sets weights based on the similarity with the distribution candidate vectors. A model for determining whether or not a distribution candidate corresponds to a predetermined user based on a plurality of distribution target vectors and a distribution candidate vector is learned using the distribution target vector acquired by the acquisition unit 132. do. For example, the learning unit 133 learns the co-occurrence between the distribution target vector indicating the distribution target selected by the predetermined user and the plurality of distribution target vectors indicating the distribution target distributed to the predetermined user as a model. ..

More specifically, the learning unit 133 selects the distribution target selected by the predetermined user based on the selection information (for example, the selection flag) indicating whether or not the distribution target delivered to the predetermined user is selected. The model learns the co-occurrence between the distribution target vector shown and a plurality of distribution target vectors indicating the distribution target delivered to a predetermined user. For example, the learning unit 133 uses a distribution target vector indicating a distribution target selected by a predetermined user and a plurality of distribution target vectors indicating a distribution target distributed to a predetermined user as input of training data, and is predetermined. By using the selection information indicating whether or not the distribution target distributed to the user has been selected as the output of the training data, the distribution target vector indicating the distribution target selected by the predetermined user and the distribution target vector indicating the distribution target selected by the predetermined user are distributed to the predetermined user. Let the model learn the distance related to co-occurrence with multiple distribution target vectors indicating the distribution target.

In one example, the model is a model with an attention mechanism. In this example, the plurality of distribution target vectors indicating the distribution targets delivered to a predetermined user are a plurality of embedded vectors (for example, distributed representation) corresponding to the plurality of distribution targets (for example, push notification). Further, in this example, the distribution target vector indicating the distribution target selected by the predetermined user is an embedded vector corresponding to the distribution target previously selected by the predetermined user (for example, a user). In other words, the distribution target vector indicating the distribution target selected by a predetermined user corresponds to the memory of the attention mechanism.

In the learning unit 133, the distribution candidate is a predetermined user from the weighted average by the above-mentioned weight (that is, the weight based on the similarity with the distribution candidate vector) of the distribution target vector indicating the distribution target selected by the predetermined user. Learn a model that determines whether or not it corresponds to. For example, in the learning unit 133, the delivery candidate corresponds to the predetermined user from the concatenated vector in which one of the plurality of delivery target vectors indicating the delivery target delivered to the predetermined user and the weighted average are concatenated. Learn the model to determine whether or not. In other words, the learning unit 133 learns the model from a vector that is a concatenation of one of the plurality of distribution target vectors and the weighted average. For example, the learning unit 133 learns a model that generates a value indicating whether or not the distribution candidate corresponds to a predetermined user by gradually convolving the connection vector so that the number of dimensions is reduced.

In one example, the learning unit 133 has a plurality of embedded vectors corresponding to a distribution target delivered to a predetermined user and a plurality of embedded vectors corresponding to a plurality of distribution targets selected by the predetermined user. Calculate the weight. Then, the learning unit 133 provides an intermediate layer corresponding to the weighted average (that is, the weighted average vector) by the plurality of weights of the plurality of embedded vectors corresponding to the plurality of distribution targets selected by the predetermined user. A model with a caution mechanism including, and a model with a caution mechanism for determining whether or not a delivery candidate corresponds to a predetermined user is learned.

For example, the learning unit 133 is weighted by a plurality of weights of an embedded vector corresponding to a distribution target delivered to a predetermined user and a plurality of embedded vectors corresponding to a plurality of distribution targets selected by the predetermined user. A first vector is generated by concatenating with the average. Next, the learning unit 133 generates a hidden state vector by applying an activation function to the generated first vector. Next, the learning unit 133 calculates the degree of relevance by multiplying the weight matrix by the hidden state vector. After that, the learning unit 133 uses the selection information (for example, a selection flag) indicating whether or not the distribution target distributed to the predetermined user has been selected, and the calculated relevance, and the distribution candidate uses the predetermined use. Learn a model with an attention mechanism that determines whether or not it corresponds to a person.

The learning unit 133 generates a vector that is the sum of the average values of the feature vectors indicating the characteristics of the information included in the distribution target indicated by one of the plurality of distribution target vectors from one of the plurality of distribution target vectors. Then, based on the generated vector, a model for determining whether or not the distribution candidate corresponds to a predetermined user is learned.

In one example, the feature vector that represents the feature of the information is a distributed representation. In this example, when the information included in the distribution target includes the first to third features, the embedded vector corresponding to the delivery target is the average vector of the three distributed representations corresponding to the first to third features, respectively. be.

The learning unit 133 is one of a distribution target vector indicating a distribution target selected by a predetermined user and a plurality of distribution target vectors indicating a distribution target distributed to a predetermined user, and is a predetermined use. When a distribution target vector indicating a distribution target selected by a user is input, it is output that the distribution candidate corresponds to a predetermined user, and a distribution target vector indicating a distribution target selected by the predetermined user and a distribution target vector. When one of a plurality of distribution target vectors indicating a distribution target delivered to a predetermined user and a distribution target vector indicating a distribution target not selected by the predetermined user is input, The model is learned so that it is output that the delivery candidate does not correspond to the predetermined user.

For example, the learning unit 133 has a distribution target vector indicating a distribution target selected by a predetermined user (for example, a distribution target vector which is test data), and a plurality of distribution targets indicating a distribution target distributed to a predetermined user. The value output by the model when one of the vectors and the distribution target vector indicating the distribution target selected by the predetermined user (for example, the memory of the attention mechanism) is input is the distribution target. The model is trained using the distribution target vector, which is the training data, so as to approach the value indicating that is selected (for example, the selection flag “1”). Further, for example, the learning unit 133 uses a regularization method (for example, L2 regularization, etc.) to learn the model so that the generalization error is minimized.

The learning unit 133 learns the model so that, for example, the selection information indicating whether or not the distribution target distributed to a predetermined user is selected and the objective function corresponding to the output of the model are minimized. In one example, the learning unit 133 uses an error backpropagation method to minimize the objective function corresponding to the value of the selection flag contained in the selection information and the output of the model with the attention mechanism (eg, the degree of relevance described above). So, learn a model with an attention mechanism.

(Generation unit 134)
The generation unit 134 generates a candidate vector in which a predetermined value is stored in a dimension corresponding to the information included in the distribution target that is a candidate for the distribution target. For example, the generation unit 134 acquires at least one vector (for example, a distributed expression) corresponding to a keyword included in the information included in the distribution target that is a candidate for distribution target from a predetermined server. Then, the generation unit 134 generates a candidate vector based on at least one vector corresponding to the keyword included in the information.

The generation unit 134 may generate a distribution target vector indicating a distribution target by using the information included in the distribution target. Further, the BoW vector corresponding to the distribution target may be generated by using the information included in the distribution target.

(Judgment unit 135)
The determination unit 135 determines whether or not to distribute the candidate to be distributed to a predetermined user from the candidate vector using the model learned by the learning unit 133. Further, when the model outputs that the distribution candidate indicated by the candidate vector corresponds to a predetermined user when the candidate vector is input, the determination unit 135 distributes the distribution candidate to the predetermined user. judge.

In one example, the determination unit 135 uses a learned model (for example, a model with an attention mechanism) to output a relevance score indicating a relevance between a predetermined user and a candidate to be delivered. More specifically, the determination unit 135 inputs a distribution target vector indicating a distribution target selected by a predetermined user and a candidate vector corresponding to the candidate of the distribution target into the model, so that the predetermined user Outputs the probability of selecting a candidate for distribution as a relevance score between a predetermined user and the candidate for distribution.

(Providing Department 136)
The providing unit 136 provides a model learned by the learning unit 133. For example, the providing unit 136 provides the trained model to the providing device 20. The providing unit 136 may provide the determination result by the determination unit 135 to the providing device 20.

[4. Learning process flow]
Next, the procedure of the learning process by the learning device 100 according to the embodiment will be described. FIG. 5 is a flowchart showing a learning processing procedure by the learning device 100 according to the embodiment.

As shown in FIG. 5, first, the learning device 100 acquires a plurality of distribution target vectors indicating distribution targets (step S101). For example, the learning device 100 acquires a distribution target vector in which a predetermined value is stored in a dimension corresponding to a word included in a sentence to be distributed.

Next, when a plurality of distribution target vectors indicating distribution targets distributed to a predetermined user and distribution candidate vectors indicating distribution candidates are input, the learning device 100 weights based on the similarity with the distribution candidate vectors. A model for determining whether or not a distribution candidate corresponds to a predetermined user based on a plurality of distribution target vectors in which is set and a distribution candidate vector is learned using the acquired distribution target vector (. Step S102). For example, the learning device 100 makes a model learn the co-occurrence between a distribution target vector indicating a distribution target selected by a predetermined user and a plurality of distribution target vectors indicating a distribution target distributed to a predetermined user. ..

More specifically, the learning device 100 uses a distribution target vector indicating a distribution target selected by a predetermined user and a plurality of distribution target vectors indicating a distribution target distributed to a predetermined user as input of training data. In addition, by using the selection information indicating whether or not the distribution target delivered to the predetermined user has been selected as the output of the training data, the distribution target vector indicating the distribution target selected by the predetermined user and the predetermined use Let the model learn the distance related to the co-occurrence with a plurality of delivery target vectors indicating the delivery target delivered to the person.

[5. others〕
Further, among the processes described in the above-described embodiment, a part of the processes described as being automatically performed can also be performed manually. Alternatively, all or part of the process described as being performed manually can be automatically performed by a known method. In addition, information including processing procedures, specific names, various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified. For example, the various information shown in each figure is not limited to the information shown in the figure.

Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of them may be functionally or physically distributed / physically in any unit according to various loads and usage conditions. Can be integrated and configured.

For example, a part or all of the storage unit 120 shown in FIG. 3 may not be held by the learning device 100, but may be held by a storage server or the like. In this case, the learning device 100 acquires various information such as training data by accessing the storage server.

[6. Hardware configuration]
Further, the learning device 100 according to the above-described embodiment is realized by, for example, a computer 1000 having a configuration as shown in FIG. FIG. 6 is a diagram showing an example of a hardware configuration. The computer 1000 is connected to the output device 1010 and the input device 1020, and the arithmetic unit 1030, the primary storage device 1040, the secondary storage device 1050, the output IF (Interface) 1060, the input IF 1070, and the network IF 1080 are connected by the bus 1090. Have.

The arithmetic unit 1030 operates based on a program stored in the primary storage device 1040 or the secondary storage device 1050, a program read from the input device 1020, or the like, and executes various processes. The primary storage device 1040 is a memory device such as a RAM that temporarily stores data used by the arithmetic unit 1030 for various operations. Further, the secondary storage device 1050 is a storage device in which data used by the arithmetic unit 1030 for various calculations and various databases are registered, and is realized by a ROM (Read Only Memory), an HDD, a flash memory, or the like.

The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various information such as a monitor and a printer. For example, USB (Universal Serial Bus), DVI (Digital Visual Interface), and the like. It is realized by a connector of a standard such as HDMI (registered trademark) (High Definition Multimedia Interface). Further, the input IF 1070 is an interface for receiving information from various input devices 1020 such as a mouse, a keyboard, a scanner, and the like, and is realized by, for example, USB.

The input device 1020 is, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), a PD (Phase change rewritable Disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), or a tape. It may be a device that reads information from a medium, a magnetic recording medium, a semiconductor memory, or the like. Further, the input device 1020 may be an external storage medium such as a USB memory.

The network IF 1080 receives data from another device via the network N and sends it to the arithmetic unit 1030, and also transmits the data generated by the arithmetic unit 1030 to the other device via the network N.

The arithmetic unit 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic unit 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040, and executes the loaded program.

For example, when the computer 1000 functions as the learning device 100, the arithmetic unit 1030 of the computer 1000 realizes the function of the control unit 130 by executing the program loaded on the primary storage device 1040.

[7. effect〕
As described above, the learning device 100 according to the embodiment has an acquisition unit 132 and a learning unit 133. The acquisition unit 132 acquires a plurality of distribution target vectors indicating the distribution target. When a plurality of distribution target vectors indicating distribution targets distributed to a predetermined user and distribution candidate vectors indicating distribution candidates are input, the learning unit 133 sets weights based on the similarity with the distribution candidate vectors. A model for determining whether or not a distribution candidate corresponds to a predetermined user based on a plurality of distribution target vectors and a distribution candidate vector is learned using the distribution target vector acquired by the acquisition unit 132. do.

Further, in the learning device 100 according to the embodiment, the acquisition unit 132 acquires a distribution target vector in which a predetermined value is stored in a dimension corresponding to a word included in a sentence to be distributed.

Further, in the learning device 100 according to the embodiment, the learning unit 133 includes a distribution target vector indicating a distribution target selected by a predetermined user, and a plurality of distribution target vectors indicating a distribution target distributed to the predetermined user. Let the model learn the co-occurrence of.

Further, in the learning device 100 according to the embodiment, the learning unit 133 is a distribution candidate from a weighted average of the distribution target vectors indicating the distribution target selected by a predetermined user by the weight based on the similarity with the distribution candidate vector. Learns a model that determines whether or not corresponds to a given user.

Further, in the learning device 100 according to the embodiment, the learning unit 133 distributes from a concatenated vector in which one of a plurality of distribution target vectors indicating a distribution target distributed to a predetermined user and a weighted average are concatenated. Learn a model that determines whether a candidate corresponds to a given user.

Further, in the learning device 100 according to the embodiment, the learning unit 133 gradually convolves the connection vector so that the number of dimensions is reduced, so that a value indicating whether or not the distribution candidate corresponds to a predetermined user is set. Learn the model to generate.

Further, in the learning device 100 according to the embodiment, the learning unit 133 is a feature showing the characteristics of the information included in the distribution target indicated by one of the plurality of distribution target vectors from one of the plurality of distribution target vectors. A vector to be the sum of the average values of the vectors is generated, and a model for determining whether or not the distribution candidate corresponds to the predetermined user is learned based on the generated vector.

Further, in the learning device 100 according to the embodiment, the learning unit 133 has a distribution target vector indicating a distribution target selected by a predetermined user, and a plurality of distribution target vectors indicating a distribution target distributed to the predetermined user. When one of them and the distribution target vector indicating the distribution target selected by the predetermined user is input, the distribution candidate outputs that the distribution candidate corresponds to the predetermined user, and the predetermined user Indicates a distribution target vector indicating a selected distribution target and a distribution target that is one of a plurality of distribution target vectors indicating a distribution target distributed to a predetermined user and is not selected by the predetermined user. The model is learned so that when the distribution target vector is input, it is output that the distribution candidate does not correspond to the predetermined user.

Further, the learning device 100 according to the embodiment has a generation unit 134 that generates a candidate vector in which a predetermined value is stored in a dimension corresponding to the information included in the distribution target that is a candidate for the distribution target. Further, the learning device 100 according to the embodiment has a determination unit 135 that determines whether or not to distribute the candidate to be distributed to a predetermined user from the candidate vector using the model learned by the learning unit 133. ..

Further, in the learning device 100 according to the embodiment, when the candidate vector is input, the determination unit 135 distributes when the model outputs that the distribution candidate indicated by the candidate vector corresponds to a predetermined user. It is determined that the candidate will be delivered to a predetermined user.

By each of the above-mentioned processes, the learning device 100 can learn a determination model for determining whether or not a predetermined notification is suitable for the user.

Although some of the embodiments of the present application have been described in detail with reference to the drawings, these are examples, and various modifications are made based on the knowledge of those skilled in the art, including the embodiments described in the disclosure column of the invention. It is possible to carry out the present invention in other modified forms.

Further, the above-mentioned learning device 100 may be realized by a plurality of server computers, and depending on the function, it may be realized by calling an external platform or the like by API (Application Programming Interface), network computing, or the like. It can be changed flexibly.

Further, the above-mentioned "section, module, unit" can be read as "means" or "circuit". For example, the receiving unit can be read as a receiving means or a receiving circuit.

1 Network system 10 Terminal device 20 Providing device 100 Learning device 120 Storage unit 121 Training data storage unit 122 Judgment model storage unit 130 Control unit 131 Reception unit 132 Acquisition unit 133 Learning unit 134 Generation unit 135 Judgment unit 136 Providing unit

Claims (9)

A first distribution target vector indicating a first distribution target distributed to a predetermined user, and a plurality of second distribution targets distributed to the predetermined user and selected by the predetermined user. An acquisition unit that acquires a plurality of second distribution target vectors, respectively, and selection information indicating whether or not the predetermined user has selected the first distribution target .
Using the first distribution target vector and the plurality of second distribution target vectors as input of training data and using the selection information as output of the training data, the predetermined user becomes a distribution candidate . It is a model for determining whether or not it matches, and uses the distribution candidate vector indicating the distribution candidate and the plurality of second distribution target vectors as inputs, and the distribution candidate vector and the plurality of the second distribution targets. A learning device characterized by having a learning unit for learning a model in which a plurality of weights are set between the distribution target vector and the vector . The acquisition unit
The first distribution target vector corresponding to the vector in which a predetermined value is stored in the dimension corresponding to the word included in the first distribution target sentence is acquired , and the corresponding second distribution target sentence is obtained. Acquires each second distribution target vector corresponding to the vector in which a predetermined value is stored in the dimension corresponding to the word contained in.
The learning device according to claim 1. The learning unit
From the weighted average of the plurality of second distribution target vectors calculated by using the weights between the first distribution target vector and the plurality of second distribution target vectors, the predetermined user is said to be the said. The predetermined probability of selecting the first distribution target is calculated, and the selection information and the calculated probability are used to determine whether or not the predetermined user is suitable for the distribution candidate. The learning device according to claim 1 , wherein the user learns the model for calculating the probability of selecting the delivery candidate . The learning unit
From the connection vector generated by concatenating the first distribution target vector and the weighted average, the probability that the predetermined user selects the first distribution target is calculated.
The learning device according to claim 3 , wherein the learning device is characterized by the above. The learning unit
By applying an activation function to the concatenated vector, a hidden state vector is generated, and by multiplying the generated hidden state vector by a weight matrix, the probability that the predetermined user selects the first delivery target. To calculate
The learning device according to claim 4 , wherein the learning device is characterized in that. The learning unit
An average vector of a plurality of distributed expressions corresponding to a plurality of words included in the first distribution target sentence is generated as the first distribution target vector, and the corresponding second distribution target sentence is generated. Generate an average vector of multiple distributed representations corresponding to each of the included words as a second distribution target vector for each.
The learning device according to any one of claims 1 to 5 , wherein the learning device is characterized in that. A generation unit that generates the distribution candidate vector using the information included in the distribution candidate, and
By inputting the distribution candidate vector generated by the generation unit and the plurality of second distribution target vectors into the model learned by the learning unit, the predetermined user selects the distribution candidate. The learning device according to any one of claims 1 to 6 , further comprising a determination unit that outputs a probability as a relevance score indicating a relevance between the predetermined user and a delivery candidate. .. The way the learning device performs,
A first distribution target vector indicating a first distribution target distributed to a predetermined user, and a plurality of second distribution targets distributed to the predetermined user and selected by the predetermined user. An acquisition process for acquiring a plurality of second distribution target vectors, respectively, and selection information indicating whether or not the predetermined user has selected the first distribution target .
Using the first distribution target vector and the plurality of second distribution target vectors as input of training data and using the selection information as output of the training data, the predetermined user becomes a distribution candidate . It is a model for determining whether or not it matches, and uses the distribution candidate vector indicating the distribution candidate and the plurality of second distribution target vectors as inputs, and the distribution candidate vector and the plurality of the second distribution targets. A learning method comprising a learning process of learning a model in which multiple weights are set with and to a delivery target vector of . A first distribution target vector indicating a first distribution target distributed to a predetermined user, and a plurality of second distribution targets distributed to the predetermined user and selected by the predetermined user. An acquisition procedure for acquiring a plurality of second distribution target vectors, respectively, and selection information indicating whether or not the predetermined user has selected the first distribution target, and an acquisition procedure.
Using the first distribution target vector and the plurality of second distribution target vectors as input of training data and using the selection information as output of the training data, the predetermined user becomes a distribution candidate . It is a model for determining whether or not it is compatible, and uses the distribution candidate vector indicating the distribution candidate and the plurality of second distribution target vectors as inputs, and the distribution candidate vector and the plurality of the second distribution targets. A learning program for a computer to perform a learning procedure to learn a model in which multiple weights are set between the delivery target vector and the .

Images