JP7160980B2 - INFORMATION PROVIDING DEVICE, INFORMATION PROVIDING METHOD AND PROGRAM - Google Patents

Description

The present invention relates to an information providing device, an information providing method, and a program.

Techniques for distributing content such as news articles and advertising images via the Internet are known. In relation to this, there is known a technique of estimating the degree of importance of content and giving higher priority to the content with higher degree of importance when recommending the content to the user (see, for example, Patent Literature 1).

JP 2017-59057 A

However, with the conventional technology, there are cases where content estimated to be of high importance is content of low interest to the user. As a result, content that is assumed to be of low interest to the user may be preferentially recommended.

The present invention has been made in view of the above problems, and aims to provide an information providing apparatus, an information providing method, and a program capable of accurately recommending content assumed to be of high interest to the user. purpose.

One aspect of the present invention is a derivation unit that derives the characteristics of the user based on the characteristics of the content viewed by the user; a deciding unit that decides content to be recommended to the user from among the plurality of content to be distributed, based on each of the characteristics, wherein the deriving unit determines the time course of each content browsed by the user. It is an information providing device for deriving characteristics of the user based on a change in the number of times of browsing according to.

According to one aspect of the present invention, it is possible to accurately recommend content that is assumed to be of high interest to the user.

BRIEF DESCRIPTION OF THE DRAWINGS It is a figure which shows an example of the information provision system 1 containing the information provision apparatus 100 in embodiment. It is a figure which shows an example of a structure of the information provision apparatus 100 in embodiment. 4 is a diagram showing an example of content information 132. FIG. 4 is a diagram showing an example of content vector information 134. FIG. 4 is a diagram showing an example of user information 136. FIG. 4 is a diagram showing an example of an action log 138; FIG. 4 is a diagram showing an example of user vector information 140. FIG. 4 is a flow chart showing a series of processes performed by the information providing apparatus 100; 4 is a flow chart showing a series of processes performed by the information providing apparatus 100; FIG. 10 is a diagram showing an example of the relationship between the parameter _bpv and the change over time in the number of views of content. FIG. 4 is a diagram schematically showing the contents of processing when determining parameters of a learning device; FIG. 4 is a diagram schematically showing the contents of processing when determining parameters of a learning device; It is a figure which shows an example of the hardware constitutions of the information provision apparatus 100 of embodiment.

An information providing apparatus, an information providing method, and a program to which the present invention is applied will be described below with reference to the drawings.

[Overview]
An information providing device is realized by one or more processors. The information providing device basically derives the user's characteristics based on the characteristics of the content viewed by the user, and based on the derived user's characteristics and the characteristics of each of the plurality of contents to be distributed. , determines the content to be recommended to the user from among the plurality of content to be distributed. Furthermore, when deriving the characteristics of the user, the information providing apparatus, among the characteristics of one or more pieces of content viewed by the user, has a characteristic of content with a smaller change in the number of times of viewing over time. To reflect. As a result, for example, even if a target user browses contents such as gossip articles temporarily browsed by an unspecified number of users, it is difficult to reflect the characteristics of the gossip articles in the user's characteristics. be able to. As a result, it becomes difficult to recommend content such as gossip articles that has been instantly viewed by an unspecified number of users, and content that is assumed to be of high interest to users can be recommended with high accuracy.

The content in this embodiment is, for example, an article posted on a blog, website, or the like, and is content containing text. Also, the content may be, for example, text, still images, moving images, audio, etc. posted as advertisements on media such as websites.

[overall structure]
FIG. 1 is a diagram showing an example of an information providing system 1 including an information providing device 100 according to an embodiment. The information providing system 1 in the embodiment includes, for example, one or more terminal devices 10 and an information providing device 100 . These devices are connected via a network NW.

Each device shown in FIG. 1 transmits and receives various information via the network NW. The network NW includes, for example, the Internet, WAN (Wide Area Network), LAN (Local Area Network), provider terminal, wireless communication network, wireless base station, leased line, and the like. Note that not all combinations of the devices shown in FIG. 1 need to be able to communicate with each other, and the network NW may partially include a local network.

The terminal device 10 is a terminal device including an input device, a display device, a communication device, a storage device, and an arithmetic device, such as a mobile phone such as a smart phone, a tablet terminal, and various personal computers. The communication device includes a network card such as a NIC (Network Interface Card), a wireless communication module, and the like. In the terminal device 10, a UA (User Agent) such as a web browser or an application program is activated, and a request corresponding to the contents input by the user is transmitted to the information providing device 100. FIG. In addition, the terminal device 10 on which the UA is activated causes the display device to display various images based on the information acquired from the information providing device 100 .

The information providing device 100 is, for example, a web server that provides web pages to the terminal device 10 in response to requests from web browsers. The content described above is posted on the web page. Further, the information providing device 100 may be an application server that provides contents such as images and sounds to the terminal device 10 in response to requests from application programs. For example, the information providing apparatus 100 provides a service (hereinafter referred to as a search service) for searching desired content from a certain database, such as document search and image search, via a website or an application to a user who uses the terminal device 10. provide to

[Configuration of Information Providing Device]
FIG. 2 is a diagram showing an example of the configuration of the information providing device 100 according to the embodiment. As illustrated, the information providing device 100 includes, for example, a communication unit 102, a control unit 110, and a storage unit .

The communication unit 102 includes, for example, a communication interface such as NIC. The communication unit 102 communicates with the terminal device 10 via the network NW and acquires various information from the terminal device 10 . For example, the communication unit 102 acquires login information from the terminal device 10 . The login information is information including a user ID (account ID) associated with an account for using the search service, a password, and the like.

The control unit 110 includes, for example, a content vector generation unit 112, a user vector generation unit 114, a recommended content determination unit 116, a content distribution unit 118, an information update unit 120, and a learning processing unit 122. The user vector generation unit 114 is an example of a "derivation unit", and the recommended content determination unit 116 is an example of a "determination unit".

These components are implemented by executing a program stored in the storage unit 130 by a processor such as a CPU (Central Processing Unit). Some or all of the components of the control unit 110 are hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or GPU (Graphics Processing Unit). (circuitry), or by cooperation of software and hardware.

The storage unit 130 is implemented by a storage device such as a HDD (Hard Disc Drive), a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), a ROM (Read Only Memory), and a RAM (Random Access Memory). The storage unit 130 stores content information 132, content vector information 134, user information 136, action log 138, user vector information 140, learning device information 142, etc., in addition to various programs such as firmware and application programs.

The content information 132 is information including various contents. For example, if the content is a news article, the content information 132 may include text information, images, audio, and other information included in the news article, and information such as the distributor and author of the news article.

FIG. 3 is a diagram showing an example of the content information 132. As shown in FIG. As shown in the illustrated example, the content information 132 may be information in which a content ID for identifying each content is associated with the total number of views per predetermined cycle _Ti . The predetermined period T _i may be any period of time, eg minutes, hours, days, weeks, months, years. In addition, the content information 132 may be information including various types of content. For example, when the content is a news article, text information contained in the news article, contents such as images, information such as the distributor and author of the news article. may contain

FIG. 4 is a diagram showing an example of the content vector information 134. As shown in FIG. The content vector information 134 is information in which the feature of each content included in the content information 132 is vectorized. For example, it is information in which a content vector obtained by vectorizing the content is associated with each content ID. .

FIG. 5 is a diagram showing an example of the user information 136. As shown in FIG. The user information 136 is, for example, information in which a user ID (account ID) for identifying each user is associated with a password to be collated at the time of login, a user attribute, and the like. User attributes include, for example, demographic attributes of the user, such as gender, age, residential area, income, occupation, and educational background. These demographic attributes are pre-registered, for example, when an account is created. Further, the user attribute includes, for example, the type of terminal device 10 indicating whether the terminal device 10 used by the user at the time of login is a mobile phone such as a smart phone, a tablet terminal, or a personal computer. Alternatively, various information such as the day of the week and the time zone at the time of login may be included as attributes.

FIG. 6 is a diagram showing an example of the action log 138. As shown in FIG. The action log 138 is information based on the viewing history of content by the user. For example, for each user ID, information in which the content ID of the content viewed by the user and the number of times the content was viewed in a predetermined period are associated with each other. is. The predetermined period is, for example, a period from when the user logs in until several cycles T _i or several tens of cycles T _i elapse. Also, the predetermined period may be, for example, a period from when the user logs in until he/she logs out.

FIG. 7 is a diagram showing an example of user vector information 140. As shown in FIG. The user vector information 140 is information in which the characteristics of each user indicated by the user information 136 are vectorized. For example, it is information in which user IDs are associated with user vectors in which users are vectorized.

The learner information 142 is information that defines a learner (learning model). The learner may be realized by a recurrent neural network (RNN) of a type such as GRU (Gated Recurrent Unit) or LSTM (Long short-term memory), for example. When a learner is realized by these recursive neural networks, the learner information 142 includes an input layer, one or more intermediate layers (hidden layers), and an output layer that constitute the recursive neural network. It contains various information such as connection information indicating how neurons (units) included in are connected to each other, and connection coefficients given to data input/output between connected neurons. Connection information includes, for example, the number of neurons in each layer, information specifying the neurons to which each neuron connects, activation functions that realize each neuron, and information such as gates provided between neurons in the hidden layer. include. An activation function that implements a neuron may be, for example, a normalized linear function (ReLU function), a sigmoid function, a step function, or other functions. A gate selectively passes data transferred between neurons, for example, depending on the value (eg, 1 or 0) returned by the activation function. A coupling coefficient is a parameter of an activation function. For example, in a hidden layer of a neural network, a weight given to output data when data is output from a neuron in a certain layer to a neuron in a deeper layer. including. The coupling coefficients may also include bias components unique to each layer, and the like.

The content vector generation unit 112 vectorizes the feature of each content included in the content information 132 to generate a content vector. For example, when the content is a news article, the content vector generation unit 112 uses the news article as a corpus and vectorizes the news article based on the Countinuous Bag-of-Words and/or Skip-gram learning model. For example, the content vector generation unit 112 divides a news article into a plurality of morphemes, and among the plurality of morphemes, a specific part of speech (such as a noun) is used in both the Countinuous Bag-of-Words and Skip-gram learning models. is input as a word to a neural network (for example, word2vec) to which is applied, the word vector of each word output from the neural network is obtained. Then, the content vector generation unit 112 converts a value obtained by normalizing (for example, dividing by the number of words) the sum of a plurality of word vectors corresponding to each word to a news article vector (an example of a content vector) from which the word is extracted. derived as That is, the content vector generation unit 112 generates vectors of news articles that are represented by a plurality of word vectors in a distributed manner. This news article vector is a multidimensional vector whose dimensions (base vectors) are each of a plurality of word vectors.

In addition, the content vector generation unit 112 may vectorize each news article using a learning model configured by Bag-of-Words alone or another learning model such as Global Vector Prediction (GloVe). Moreover, when the content is a still image or a moving image, the content vector generation unit 112 may vectorize the content using, for example, a convolutional neural network.

The content vector generation unit 112 associates the generated content vector with the content ID of the content that is the source of the content vector, and stores this as content vector information 134 in the storage unit 130 . Note that the process of vectorizing the features of the content may be realized by requesting an external device to perform the process instead of being performed by the content vector generation unit 112 itself.

Based on the content vector information 134 and the action log 138, the user vector generation unit 114 vectorizes the characteristics of each user included in the user information 136 to generate a user vector. For example, the user vector generation unit 114 refers to the learner information 142 and generates (constructs) a recurrent neural network, which is an example of a learner. On the other hand, the user vector generation unit 114 refers to the action log 138, acquires the content ID of the content viewed by the user for each user ID from the storage unit 130, and then refers to the content vector information 134 to acquire the content ID. A content vector corresponding to the content ID is acquired from the storage unit 130 . Then, the user vector generation unit 114 inputs each content vector corresponding to the content ID of the content browsed by the user to the input layer of the generated recursive neural network, thereby allowing the user's Get the user vector with the features vectorized. That is, the user vector generation unit 114 generates user vectors that are distributed and represented by a plurality of content vectors. This user vector is a multi-dimensional vector whose dimension (base vector) is each of a plurality of content vectors. A user vector is an example of a "user characteristic."

In the example of FIG. 6, the action log 138 indicates that a user with a user ID of "USER_2" browsed contents with content IDs of "CONT_6" and "CONT_14". In such a case, the user vector generation unit 114 refers to the content vector information 134, the content vector V (CONT_6) corresponding to the content with the content ID "CONT_6" and the content vector V(CONT_6) corresponding to the content with the content ID "CONT_14". At least content vectors V(CONT_14) are obtained, and a vector set including at least these content vectors is input to the recurrent neural network. A recursive neural network assigns a coupling coefficient (weight) to each input content vector in a hidden layer, and averages (for example, weighted averaging) up to the output layer. Then, the user vector generation unit 114 derives one vector obtained by averaging a plurality of content vectors by the recursive neural network as the user vector of the user whose user ID is "USER_2".

The user vector generation unit 114 associates the generated user vector with the user ID of the user from which the user vector was generated, and stores this as user vector information 140 in the storage unit 130 .

The recommended content determination unit 116 selects a user vector obtained by vectorizing a target user (hereinafter referred to as a target user vector) from among a plurality of user vectors generated by the user vector generation unit 114, and a user vector generated by the content vector generation unit 112. A score (index value) based on the degree of similarity with each of the one or more content vectors obtained is derived, and content to be recommended for viewing by the target user (hereinafter referred to as recommended content) is determined according to the derived score. do. The target user is, for example, a user who has entered a query in a search window or the like when searching for desired information using a search service after login.

For example, the recommended content determining unit 116 determines, as recommended content, content corresponding to one content vector with the highest score. In addition, the recommended content determination unit 116 may determine, as recommended content, a set of content corresponding to each of a predetermined number of content vectors with the highest scores, or may select all content vectors with scores equal to or greater than a threshold value as recommended content. A set of contents corresponding to each may be determined as recommended contents.

The content distribution unit 118 uses the communication unit 102 to distribute one or a plurality of contents determined as recommended contents by the recommended content determination unit 116 to the terminal device 10 used by the target user to input the query. .

For example, when the recommended content distributed by the content distribution unit 118 is viewed by the target user, the information updating unit 120 adds the content ID of the recommended content to the user ID of the target user in the action log 138, or The action log 138 is updated by, for example, replacing with the content ID of content with an older browsing time. At this time, if the target user vector does not exist in the user vector information 140 , the user vector generation unit 114 may generate a new target user vector based on the updated action log 138 . Also, if the target user vector already exists in the user vector information 140, the user vector generation unit 114 may update the target user vector based on the updated action log 138. FIG.

Note that, in the action log 138, when the content ID exists in the user ID of the target user, i.e., when the target user vector already exists, the information updating unit 120 does not allow the target user to browse any content within a predetermined period of time. If not, in the target user's action log 138, all content IDs associated with the user ID may be deleted to initialize the browsing history. In this case, the user vector generation unit 114 may delete the already generated target user vector, for example.

The learning processing unit 122 causes the learner specified by the learner information 142 to learn based on the action log 138 . For example, when the learning device is a recursive _neural network, the learning processing unit 122 calculates the coupling coefficient (activity parameters of the function) are machine-learned. The details of the learning method of the learner will be described later.

[Processing flow during operation]
The flow of a series of processes during operation of the information providing apparatus 100 will be described below with reference to flowcharts. During operation is a state in which a learning device that has been learned to some extent by the learning processing unit 122 is used. FIG. 8 is a flow chart showing a series of processes by the information providing apparatus 100. As shown in FIG. The processing of this flowchart is performed, for example, when a query for searching content is further acquired from the terminal device 10 from which the login information has been acquired by the communication unit 102 .

First, the user vector generation unit 114 refers to the content vector information 134 and the action log 138, and acquires the content vector of the content browsed by the target user from the storage unit 130 (S100).

Next, the user vector generation unit 114 inputs the acquired plurality of content vectors to the learner indicated by the learner information 142, thereby generating a target user vector in which the features of the target user are represented in a distributed manner by the plurality of content vectors. Generate (S102).

Next, the recommended content determination unit 116 determines the target user vector generated by the user vector generation unit 114, one or more content vectors generated by the content vector generation unit 112, and Based on the degree of similarity, a score to be referred to when determining recommended content is derived (S102).

u in Equation (1) represents a target user among a plurality of users whose user IDs are included in the action log 138, and a is selected as a score derivation target among a plurality of contents for which content vectors are generated. It represents one piece of content, and t represents the time when the content was published in the search service (the time when the article was submitted in the case of a news article). R(u, a, t) represents the score, V _cont (a) represents the content vector corresponding to the content a, and V _user (u) represents the target user vector corresponding to the target user u. there is T represents transposition. pv(a, t) represents an index value (for example, a logarithm of the number of times of viewing) based on the number of times content a is viewed per predetermined period T _i at an arbitrary time t. b _pv represents the bias component multiplied by pv(a,t). Since pv(a, t) does not include u as a variable, from the first term on the right side of Equation (1), the second term (b _pv pv(a, t)) is Treated as a constant term to add or subtract. The parameters b _pv are learned by the learning processing unit 122 when the coupling coefficients of the recurrent neural network are learned. The parameter b _pv is learned so that its value increases as the change in the number of views over time (for example, the time differential value of the number of views) of the content a selected as the score derivation target decreases.

For example, the recommended content determining unit 116 derives the cosine similarity between each content vector V _cont (a) and the target user vector V _user (u) as the first term on the right side of Equation (1). On the other hand, the recommended content determination unit 116 derives the product of the index value pv(a, t) based on the number of views of the content a at time t and the bias value b _pv as the second term on the right side of the formula (1). do. Then, the recommended content determining unit 116 derives the sum of these as the score R(u, a, t). In Equation (1), the first term V _cont (a) ^T ·V _user (u) on the right side is a term whose value varies depending on the browsing history of the content so far of the target user. , the second term on the right side, b _pv ·pv(a, t), is a term whose value does not change according to the target user but changes according to the number of times the content has been viewed at time t. Therefore, a higher score R(u, a, t) is derived for content that is more similar to content frequently viewed by the user and whose change in the number of views over time is smaller. In Expression (1), the score R(u, a, t) may be derived as the product of the first term and the second term. For improvement, the score R(u,a,t) is preferably derived as the sum of the first and second terms.

Next, the recommended content determining unit 116 determines recommended content to be recommended for viewing by the target user from among a plurality of sets of content corresponding to each content vector based on the derived score (S106).

Next, the content distribution unit 118 uses the communication unit 102 to distribute one or more pieces of content determined as recommended content by the recommended content determination unit 116 to the terminal device 10 of the target user (S108). This completes the processing of this flowchart.

[Processing flow during learning]
A series of processing flow during learning of the information providing apparatus 100 will be described below with reference to a flowchart. The time of learning is a state in which a learner used during operation learns. FIG. 9 is a flowchart showing a series of processes performed by the information providing apparatus 100. As shown in FIG. The processing of this flowchart may be performed repeatedly at a predetermined cycle, for example.

First, the learning processing unit 122 learns the parameter b _pv included in the score derivation formula (1) (S200). For example, the learning processing unit 122 increases the value of the parameter bpv for content a whose change in the number of views over time is small, and increases the value of the parameter _bpv for content a whose change in the number of views over time is large. Learn so that the value of b _pv becomes small.

FIG. 10 is a diagram showing an example of the relationship between the parameter _bpv and the change over time in the number of views of content. For example, the parameter b _pv may be represented by a step function that decreases as the number of times the content is viewed increases over time, as shown. For example, in the case of content (hereinafter referred to as "buzz content") that has a high number of views (hereinafter referred to as "buzz content"), such as the content with the content ID "CONT_1" or "CONT_3" in FIG. 3, the value of the parameter _bpv becomes small. Cheap. Therefore, even if the content a for which the score is to be derived is buzz content and the value of the parameter pv(a, t) increases, the value of the parameter b _pv decreases. 2 terms become smaller. For example, the learning processing unit 122 determines the degree of reduction of the parameter _bpv according to the number of views by machine learning. Note that the parameter b _pv is not limited to a step function, and may be any function as long as it decreases as the time change in the number of views of the content increases.

Next, the learning processing unit 122 learns parameters of the learning device defined by the learning device information 142 (S204). For example, when the user vector output by the learner is represented by a weighted average of a plurality of content vectors input to the learner, the learning processing unit 122 determines the weight given to each content vector through learning. . Equation (2) is an example of a user vector derivation equation.

In Expression (2), w ₁ to w _k are weights given to the content vectors V _cont (1) to V _cont (k) input to the learner, respectively. These weights w ₁ to w _k become coupling coefficients (activation function parameters) of the neural network when the learner is a recurrent neural network. For example, as illustrated in FIG. 6, when the user with the user ID "USER_1" is set as the target user, the learning processing unit 122 derives the user vector V _USER (1) of the user when the content ID is The weights w ₁ to w ₃ given to the content vectors V _cont (1) to V _cont (3) of the contents "CONT_1", "CONT_2", and "CONT_3" are compared with the weights given to the other content vectors. is updated by error backpropagation so that

At this time, the learning processing unit 122 learns so that the weight w is smaller for content with a larger time change in the number of views. For example, in FIG. 3, the content with the content IDs "CONT_1" and "CONT_3" has a larger change in the number of views per predetermined period Ti than the content with the content ID " _{CONT_2} ". Indicates that it is buzz content. Therefore, since the user whose user ID is "USER_1" originally browsed the contents of "CONT_1", "CONT_2", and "CONT_3", the learning processing unit 122 determines that these content vectors V _cont (1) ∼V _cont (3) is determined by machine learning so that the weights w ₁ to w ₃ given to (3) are large. Both the weight w ₁ given to the vector V _cont (1) and the weight w ₃ given to the content vector V _cont (3) of the content “CONT_3” are the content vector V of the content “CONT_2” that is not the buzz content. It is determined by machine learning so as to be smaller than the weight w ₂ given to _cont (2). As a result, when deriving the target user vector, among the content vectors of the plurality of contents viewed by the target user, the content vector of the content with the smaller temporal change in the number of views (for example, V _cont (2)) is the target user vector. It is possible to increase the degree of contribution to the vector, and make the degree of contribution to the target user vector lower for a content vector (for example, V _cont (1) or V _cont (2)) of a content with a larger temporal change in the number of views. As a result, since the degree of similarity of the user vector to the buzz content is reduced, the buzz content can be excluded from the recommended content.

Next, after updating the learning device parameters, the learning processing unit 122 updates the learning device information 142 stored in the storage unit 130 (S204), and ends the processing of this flowchart.

11 and 12 are diagrams schematically showing the contents of the processing when determining the parameters of the learning device. For example, in the example of FIG. 11, a content A related to "election" and a content B related to "baseball" are distributed at _a certain period T1. and B are viewed. It is also assumed that at the time of this period T1, the time change in the number of times content _A is viewed is 0.01, and the time change in the number of times content B is viewed is 0.9. In such a case, the learning processing unit 122 increases the weight of the content vector V _cont (B) of the content B, among the contents viewed by the target user k, for which the change in the number of views over time is large. The parameters of the learner are determined so that the weight of the content vector V _cont (A) of small content A is reduced.

_The example of FIG. 12 shows how the contents are delivered at the _time of the cycle T2, which is further advanced from the time of the cycle T1. _At the time of cycle T2, content A related to "election" and content B related to "baseball" continue to be delivered, and content C related to "disaster" is newly delivered. It is assumed that target user k viewed contents B and C among these contents. Also, at the _time of this period T2, the change over time in the number of views of content A is 0.01, the change over time in the number of views of content B is 0.9, and the change over time in the number of views of content C is 0. .0.3. In such a case, the learning processing unit 122 increases the weight of the content vector V _cont (B) of the content B, among the contents viewed by the target user k, for which the change in the number of views over time is large. The learner parameters are determined so that the weight of the content vector V _cont (C) of small content C is reduced. As a result, when the target user k browses a buzz content that is temporarily viewed more frequently, such as content such as "election" or "disaster", in the target user vector distributedly represented by a plurality of content vectors , the contribution of buzz content is lower than that of other content. On the other hand, when the target user k browses a content such as "baseball" that has a substantially constant number of views regardless of time, the contribution of the content such as "baseball" relatively increases in the target user vector. . In this way, even if the user browses the buzz content that has temporarily increased the number of views due to a temporary event such as an election being held at that time, the browsed buzz content After reducing the contribution of , the user vector can be generated. As a result, it is possible to recommend content that is assumed to be of high interest to the user with higher accuracy than simply increasing the degree of contribution of the content vector of content that has been viewed a large number of times.

In the above example, the learning processing unit 122 determines the weight of each content vector according to the time change in the number of content views when generating the user vector, but the present invention is not limited to this. For example, when content is posted on a medium such as a web page or application, the learning processing unit 122 may determine the weight of the content vector according to the placement position of the content on the medium. For example, when the medium is a web page, the learning processing unit 122 decreases the weight w as the content browsed by the user is posted higher on the web page, and the content is posted lower on the web page. The weight w may be increased as As a result, for example, it is possible to appropriately generate a user vector taking into consideration that content at a higher publication position is more likely to be browsed, and content at a lower publication position is less likely to be browsed.

According to the embodiments described above, the user vector generating unit 114 generates a user vector obtained by vectorizing the user's characteristics based on the content vector obtained by vectorizing the characteristics of the content browsed by the user; A recommended content determination unit 116 that determines recommended content to be recommended to a user from among a plurality of contents to be distributed, based on the user vector generated by the unit 114 and the content vectors of the plurality of contents to be distributed. and, when the user vector generation unit 114 generates a user vector, among the content vectors of one or more pieces of content viewed by the user, the feature of the content whose change in the number of times of viewing over time is small. By largely reflecting the user vector, it is possible to accurately recommend content that is assumed to be of high interest to the user.

In general, there are users who have a tendency to view content that is often viewed in the world or content published as advertisements, in addition to content that they themselves are interested in. In the case of such a user, if the user vector is generated by simply increasing the contribution of the content vector of content that has been viewed many times, content similar to that content will be recommended to the user. However, content that is often viewed by the public and content published as advertisements are not necessarily content that interests the user, and content that has little interest for the user may be recommended.

On the other hand, in the present embodiment, the degree of contribution of the content vector is adjusted according to the degree of change in the number of times of viewing over time, regardless of whether the number of times of viewing is large or small. Instead of uniformly judging that users are more interested in content that has a higher number of digits, if the change in the number of views over time is small even for content with a low number of views, that content is essentially It can be determined that the content is of interest. This allows a suitable user vector to be generated. As a result, for example, even if a user accidentally browses a content such as a gossip article that has been browsed by an unspecified number of users in an instant, the user cannot read the gossip article. It is possible to make it difficult to recommend such content, and it is possible to recommend content that is assumed to be of real interest to the user.

<Hardware configuration>
The information providing apparatus 100 of the embodiment described above is implemented by, for example, a hardware configuration as shown in FIG. 13 . FIG. 13 is a diagram showing an example of the hardware configuration of the information providing device 100 of the embodiment.

The information providing device 100 includes a NIC 100-1, a CPU 100-2, a RAM 100-3, a ROM 100-4, a secondary storage device 100-5 such as a flash memory and an HDD, and a drive device 100-6 connected to an internal bus or a dedicated communication line. are interconnected by A portable storage medium such as an optical disc is mounted in the drive device 100-6. A program stored in the secondary storage device 100-5 or a portable storage medium attached to the drive device 100-6 is developed in the RAM 100-3 by a DMA controller (not shown) or the like, and executed by the CPU 100-2. Thus, the control unit 110 is realized. The program referred to by control unit 110 may be downloaded from another device via network NW.

As described above, the mode for carrying out the present invention has been described using the embodiments, but the present invention is not limited to such embodiments at all, and various modifications and replacements can be made without departing from the scope of the present invention. can be added.

DESCRIPTION OF SYMBOLS 1... Information provision system 10... Terminal device 100... Information provision apparatus 102... Communication part 110... Control part 112... Content vector generation part 114... User vector generation part 116... Recommended content determination part 118... Content distribution unit 120 Information updating unit 122 Learning processing unit 130 Storage unit

Claims (6)

a derivation unit that derives the features of the user based on the features of the content viewed by the user;
a determination unit that determines content to be recommended to the user from among the plurality of contents to be distributed, based on the characteristics of the user derived by the derivation unit and the characteristics of each of the plurality of contents to be distributed; , and
The derivation unit is
Deriving characteristics of the user based on changes in the number of views of each piece of content viewed by the user over time,
Weighting one or more input features of content, and based on the weighted one or more content features, a learner configured to output features of a user who viewed the content, Deriving characteristics of the user by inputting characteristics of the content;
Information provider. The decision unit
deriving a degree of similarity between the characteristics of the user derived by the derivation unit and the characteristics of each of the plurality of contents to be distributed;
of the plurality of contents to be distributed based on similarity derived for each feature of the plurality of contents to be distributed and an index value based on the number of views of each of the plurality of contents to be distributed; determine which content to recommend to the user;
The information providing device according to claim 1 . Further comprising a learning processing unit that determines the index value by machine learning according to a change in the number of views of the content over time,
3. The information providing device according to claim 2 . The learning processing unit determines the index value to be smaller as the change in the number of times the content is viewed over time increases, and to increase as the change in the number of times the content is viewed over time decreases.
The information providing device according to claim 3 . the computer
deriving features of the user based on features of content viewed by the user;
determining content to be recommended to the user from among the plurality of contents to be distributed based on the derived characteristics of the user and the characteristics of each of the plurality of contents to be distributed;
Deriving characteristics of the user based on changes in the number of views of each piece of content viewed by the user over time,
Weighting one or more input features of content, and based on the weighted one or more content features, a learner configured to output features of a user who viewed the content, Deriving characteristics of the user by inputting characteristics of the content;
How to provide information. to the computer,
a process of deriving features of the user based on features of content viewed by the user;
a process of determining content to be recommended to the user from among the plurality of contents to be distributed based on the derived characteristics of the user and the characteristics of each of the plurality of contents to be distributed;
a process of deriving characteristics of the user based on changes in the number of views of each piece of content viewed by the user over time;
Weighting one or more input features of content, and based on the weighted one or more content features, a learner configured to output features of a user who viewed the content, a process of deriving the features of the user by inputting the features of the content;
program to run the

Images