KR20210059845A - Apparatus for recommending context-aware content based on deep learning and a method therefor - Google Patents

Abstract

According to an embodiment of the present invention, a device for recommending context-aware content based on deep learning, comprises an artificial neural network including a cyclic layer for outputting a content vector representing content viewed after content in the last order among a plurality of contents with a predicted value through calculation applied with a weight on a plurality of content vectors when the content vectors representing the contents are input in a viewing order, and a merger layer for outputting, to an output layer, an output value calculated through calculation applied with a weight on the output content vector, and first and second viewing time vectors when the content vector output from the cyclic layer, the first viewing time vector representing a viewing time of content in the last order and the second viewing time vector representing a viewing time of next viewed content are input. Accordingly, various contents can be provided and content consumption can be activated.

Description

Apparatus for recommending context-aware content based on deep learning and a method therefor}

The present invention relates to a content recommendation technology, and more particularly, to an apparatus and a method for recommending context-aware content based on deep learning.

Deep learning 　 or 　 deep structured learning (deep learning or hierarchical learning) is defined as a set of 　machine learning 　algorithms that attempt a high level of abstraction through a combination of several nonlinear transformation techniques. It can be said that it is a branch of machine learning that teaches students. When there is any data, a lot of research is being conducted to express it in a form that can be understood by a computer and apply it to learning.As a result of these efforts, various deep learning techniques are applied to various fields to show cutting-edge results. have.

Korean Patent Laid-Open Patent No. 2012-0124107 published on November 13, 2012 (Name: Device and method for providing on-demand content recommendation information)

An object of the present invention is to provide an apparatus and method capable of dividing a user device having a problem in a specific function due to different implementations for each manufacturer into a device type and a software version, and individually controlling the corresponding user device.

In the apparatus for recommending context-aware content according to a preferred embodiment of the present invention for achieving the above-described object, when a plurality of content vectors representing a plurality of contents are input according to a viewing order, the plurality of contents A cyclic layer for outputting a content vector representing the next viewed content of the content in the last order among the plurality of contents as a predicted value through an operation in which a weight is applied to the vector, a content vector output from the cyclic layer, and a content vector in the last order. When a first viewing time vector representing the viewing time of the content and a second viewing time vector representing the viewing time of the next viewed content are input, the output content vector, the first viewing time vector, and the second viewing time vector It includes an artificial neural network including a merge layer that outputs an output value calculated through an operation to which a weight is applied to the output layer.

A plurality of content vectors representing a plurality of content according to the user's recent viewing order are input into the circulation layer, and a third viewing time vector representing the viewing time of the last order among the recent viewing order and a fourth viewing time representing the recommended viewing time And a content recommendation unit for inputting a time vector into the merge layer, wherein the circulatory layer of the artificial neural network performs a plurality of operations in which weights are applied to a plurality of content vectors input according to the latest viewing order. Through an operation in which a content vector is output as a predicted value, and the merging layer of the artificial neural network applies a weight to the content vector output as the predicted value, the third viewing time vector, and the fourth viewing time vector, the recommended viewing is performed. It is characterized in that the output value indicating the recommended content of time is output to the output layer.

The content recommendation unit may extract content according to an output value indicating recommended content at the recommended viewing time, and provide the extracted content to a user device.

The cyclic layer includes a plurality of computational models, and each of the plurality of computational models receives an input value and a predicted value of a previous computational model, and derives a content vector in the next order as a predicted value through one or more computations to which weights are applied. It is characterized by that.

The device includes a content vector generator for generating a content vector by embedding content included in the viewing history into a predetermined vector space, and a viewing time vector by embedding the viewing time of the content from the viewing history into a predetermined vector space. It further includes a viewing time vector generator for generating.

In a method for recommending context-aware content based on deep learning according to a preferred embodiment of the present invention for achieving the above-described object, the circulatory layer of the artificial neural network inputs a plurality of content vectors representing a plurality of contents according to the viewing order. Receiving and deriving a content vector as a predicted value through a plurality of calculations in which weights are applied to each of the plurality of content vectors by the circulatory layer of the artificial neural network, and the derived content by the merging layer of the artificial neural network. Receiving a vector, a first viewing time vector indicating a viewing time of a content in the last order among the plurality of contents, and a second viewing time vector indicating a viewing time of the next viewed content in the last order; and And deriving an output value through an operation in which a weight is applied to the derived content vector, the viewing rate vector, the first viewing time vector, and the second viewing time vector by the merging layer.

After the step of deriving the output value, the content recommendation unit inputs a plurality of content vectors representing a plurality of contents according to the user's recent viewing order into the circulation layer, and the content recommendation unit is viewed in the last order of the recent viewing order. Inputting a third viewing time vector representing time and a fourth viewing time vector representing recommended viewing time to the merge layer, and a plurality of circulatory layers of the artificial neural network to which weights are applied to the plurality of content vectors, respectively. Deriving a content vector as a predicted value through an operation of, and an operation in which a weight is applied to the content vector, the third viewing time vector, and the fourth viewing time vector output as the predicted value by the merging layer of the artificial neural network. And deriving an output value representing the recommended content of the recommended viewing time through.

After the step of deriving an output value representing the recommended content of the recommended viewing time, the content recommendation unit extracting content according to an output value representing the recommended content of the recommended viewing time, and the extracted content is a recommended content of the recommended viewing time. The method further includes providing to the user device.

In the step of deriving the content vector as the predicted value, each of the plurality of computational models of the cyclic layer receives the input value and the predicted value of the previous computational model, and derives the content vector in the next order as a predicted value through one or more computations to which weights are applied It characterized in that it comprises the step of.

Before the step of receiving the plurality of content vectors, the content vector generation unit embeds the content included in the viewing history into a predetermined vector space to generate a plurality of content vectors, and the viewing time vector generation unit generates the viewing history. And generating a viewing time vector by embedding the viewing time of the content in a predetermined vector space.

According to the present invention, it is possible to provide context-aware-based recommendations according to the user's content consumption time period by vectorizing the user's content consumption characteristics and the viewing time period characteristics and learning with a deep learning model. By configuring, it is possible to provide more diverse contents to users and activate contents consumption. In particular, the device and method for recommending context-aware content based on deep learning according to the present invention do not simply recommend content that the user prefers, but divide the content that the user prefers for each time period to suit the recommendation time. Recommended content can be derived and provided. Accordingly, it is possible to provide a user experience (UX) based on a higher understanding of the user.

1 is a diagram illustrating a system for recommending context-aware content based on deep learning according to an embodiment of the present invention.
2 is a block diagram illustrating a configuration of a content recommendation engine according to an embodiment of the present invention.
3 is a view for explaining the configuration of an artificial neural network according to an embodiment of the present invention.
4 is a diagram for explaining a computational model of an artificial neural network according to an embodiment of the present invention.
5 is a flowchart illustrating a method of generating data for recommending context-aware content based on deep learning according to an embodiment of the present invention.
6 is a flowchart illustrating a deep learning method for recommending context-aware content according to an embodiment of the present invention.
7 is a flowchart illustrating a method for recommending context-sensitive content using deep learning according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings in which a person of ordinary skill in the art can easily implement the present invention. However, when it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention in describing the operating principle of the preferred embodiment of the present invention in detail, the detailed description thereof will be omitted. This is to convey more clearly without obscuring the core of the present invention by omitting unnecessary description. In addition, the present invention is intended to illustrate specific embodiments in the drawings and describe in detail in the detailed description, as various changes may be made and various embodiments may be applied, but this is not intended to limit the present invention to specific embodiments, It is to be understood as including all changes, equivalents, and substitutes included in the spirit and scope of the present invention.

In addition, when a component is referred to as being "connected" or "connected" to another component, it means that it is logically or physically connected, or can be connected. In other words, it should be understood that a component may be directly connected or connected to another component, but another component may exist in the middle, or may be indirectly connected or connected.

In addition, terms used in the present specification are used only to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In addition, terms such as "comprises" or "have" described herein are intended to designate the existence of features, numbers, steps, actions, components, parts, or a combination thereof described in the specification. It is to be understood that the above other features or the possibility of the presence or addition of numbers, steps, actions, components, parts, or combinations thereof are not preliminarily excluded.

Now, a method for receiving channel state information and an apparatus supporting the same according to an embodiment of the present invention will be described in detail with reference to the drawings. In this case, the same reference numerals are used for portions having similar functions and functions throughout the drawings, and redundant descriptions thereof will be omitted. In addition, in order to avoid obscuring the concept of the present invention, well-known structures and devices may be omitted, or a block diagram centering on the core functions of each structure and device may be shown.

First, a system for recommending context-aware content based on deep learning according to an embodiment of the present invention will be described. 1 is a diagram illustrating a system for recommending context-aware content based on deep learning according to an embodiment of the present invention. Referring to FIG. 1, a system for recommending context-aware content based on deep learning according to an embodiment of the present invention includes a content server 10 and a user device 20.

The user device 20 can be applied to all information and communication devices including mobile communication terminals, multimedia terminals, wired terminals, fixed terminals, and Internet Protocol (IP) terminals. For example, the user device 20 may be a mobile phone, a portable multimedia player (PMP), a mobile internet device (MID), a smart phone, a tablet PC, a phablet PC, a notebook computer, a personal computer, and the like.

Here, the content server 10 is an entity existing on the network and serves as a web server, a database server, and an application server. According to a preferred embodiment, the content server 10 provides various content to the user device 20, stores the user's viewing history for the provided content, and recommends the user's preferred content based on the viewing history. can do. To this end, the content server 10 includes a content recommendation engine 100, an artificial neural network 200, a viewing history DB 300, and a content DB 400.

The content DB 400 stores content or metadata having a location where the content is stored. Basically, the content recommendation engine 100 may search for the content desired by the user through the content DB 400 and then provide it to the user device 20.

The viewing history DB 300 stores a viewing history including content viewed by each user and a viewing time of viewing the content.

The artificial neural network 200 is for predicting each user's preferred content for each time period. The content recommendation engine 100 may deep learn the artificial neural network 200 using the viewing history stored in the viewing history DB 300 as learning data. In addition, when the user device 20 is connected after learning about the artificial neural network 200 is completed, the content recommendation engine 100 is a user preference corresponding to the time when the user device 20 is connected through the artificial neural network 200. Content can be derived and provided to the user device 20.

Then, the configuration of the above-described content recommendation engine 100 will be described in more detail. 2 is a block diagram illustrating the configuration of a content recommendation engine 100 according to an embodiment of the present invention. Referring to FIG. 2, the content recommendation engine 100 includes a viewing history preprocessing unit 110, a content vector generation unit 120, a viewing rate vector generation unit 130, a viewing time vector generation unit 140, and a deep learning model. It includes a learning unit 150 and a content recommendation unit 160.

The viewing history pre-processing unit 110 extracts the viewing history including the content viewed by each user for a predetermined period (eg, 6 months) and the viewing time of viewing the content from the viewing history DB 300 according to the time sequence. Align. For example, as shown in Table 1 below, for user A, the viewing history may be arranged according to the viewing time by mapping the viewed content and the viewing time.

contents C1 C3 C1 C7 … Viewing time October 10, 2019
13:45 October 10, 2019
19:30 October 11, 2019
18:45 2019.10.12.
20:03 …

The content vector generation unit 120 generates a plurality of content vectors by embedding a plurality of contents included in the viewing history of each user in a predetermined vector space. For example, the content vector is [0.11, 0.11, -0.20. It can be a 7-dimensional vector such as -0.27, 0.09, -0.19, 0.26]. That is, according to an embodiment of the present invention, a content vector is generated by embedding content in a vector space through a technique such as word2vec. Accordingly, the vector representation of the content vector is made of “Dense representation”. However, the present invention is not limited thereto. Those of ordinary skill in the art will appreciate that the content vector according to the embodiment of the present invention may be expressed as a "Sparse representation" such as, for example, one-hot encoding, if necessary. . The viewing rate vector generation unit 130 embeds the viewing history of the user provided from the viewing history preprocessing unit 110 in the vector space to generate a viewing rate vector representing the content viewing rate according to the viewing rate of the content viewed by the user. . At this time, the viewing rate vector generation unit 130 assumes the user's viewing history as a document, assumes the content viewed by the user as a word in the document, and generates a viewing rate vector from the viewing history according to Doc2Vec. can do. That is, such a viewing rate vector can be expressed as a proportion to the number (K) of topics defined for each user. For example, if a user has viewed a total of 100 contents and viewed the first to seventh contents 21 times, 15 episodes, 12 times, 5 times, 21 times, 23 times, and 3 times, respectively, the audience rate vector is [0.21] , 0.15, 0.12, 0.05, 0.21, 0.23, 0.03]. Such a viewing rate vector may be continuously updated as the number of viewings is accumulated. Therefore, the viewing rate vector is a topic rate for each user and is used as a long-term content consumption characteristic of users.

The viewing time vector generation unit 140 generates a viewing time vector by embedding the viewing time of the content in a predetermined vector space from the viewing history of the user provided from the viewing history preprocessing unit 110.

As an example, the viewing time vector generation unit 140 divides a week into weekdays and weekends, divides 24 hours into 6 hours in the morning, morning, afternoon, and evening, and generates a viewing time vector as shown in Table 2 below. can do.

Viewing time vector Viewing time Vector representation T1 Weekdays (Mon-Fri) early morning (0-6 o'clock) [1, 0, 0, 0, 0, 0, 0, 0] T2 Weekdays (Mon-Fri) morning (6-12 o'clock) [0, 1, 0, 0, 0, 0, 0, 0] T3 Weekdays (Mon-Fri) Afternoon (12-18:00) [0, 0, 1, 0, 0, 0, 0, 0] T4 Weekdays (Mon-Fri) Evening (18-24) [0, 0, 0, 1, 0, 0, 0, 0] T5 Weekend (Sat~Sun) Dawn (0~6) [0, 0, 0, 0, 1, 0, 0, 0] T6 Weekend (Sat-Sun) morning (6-12 o'clock) [0, 0, 0, 0, 0, 1, 0, 0] T7 Weekend (Sat~Sun) Afternoon (12~18:00) [0, 0, 0, 0, 0, 0, 1, 0] T8 Weekend (Sat-Sun) Evening (18-24) [0, 0, 0, 0, 0, 0, 0, 1]

As shown in Table 2, the viewing time vector may be expressed as an 8-dimensional one-hot encoding vector. For example, according to Table 3, when a specific content is viewed on Saturday, December 15, 2018 at 19:00, the viewing time vector may be T8 [0,0,0,0,0,0,0,1]. The deep learning model learning unit 150 is for training the artificial neural network 200. The deep learning model learning unit 150 inputs the training data into the artificial neural network 200 and compares the output value of the artificial neural network 200 with a target value corresponding to the training data, so that the difference between the output value and the target value is minimized. A plurality of weights of the neural network 200 are modified.

The content recommendation unit 160 is to derive recommended content that the user prefers at a recommendation time, which is a time that the user wants to watch, using the artificial neural network 200 on which the learning has been completed. The content recommendation unit 160 inputs the recent viewing history into the artificial neural network 200 and then derives recommended content of the recommended viewing time through the output value of the artificial neural network 200.

Then, the artificial neural network 200 according to an embodiment of the present invention will be described in more detail. 3 is a view for explaining the configuration of an artificial neural network according to an embodiment of the present invention. 4 is a diagram for explaining a computational model of an artificial neural network according to an embodiment of the present invention. Referring to FIG. 3, an artificial neural network 200 according to an embodiment of the present invention includes a circulatory layer 210, a merge layer 220, and an output layer 230.

The cyclic layer 210 receives a plurality of content vectors arranged according to the viewing order, and outputs a content vector as a predicted value through a plurality of operations in which a weight is applied to each of the plurality of input content vectors. The circulatory layer 210 may be implemented as a recurrent neural network (RNN), long short-term memory models (LTSM), a gated recurrent unit (GRU), or the like.

The circulation layer 210 includes a plurality of computational models 211. The calculation model 211 consists of one or more calculations to which a weight (W) is applied. Here, the operation means an operation using an activation function. The activation function is a sigmoid function, a hyperbolic tangent (tanh) function, a ReLU (Rectified Linear Unit) function, a Leakly ReLU function, a PReLU function, an ELU (Exponential Linear Unit) function, a Softmax function, and a Maxout function. , Minout function, etc. can be illustrated. In addition, the weight can be applied to the input or output of the above-described activation function.

The computational model 211 receives a content vector that is a predicted value of the previous computational model 211 and a content vector in the current sequence, and calculates a content vector in the next sequence as a predicted value through one or more computations to which weights are applied.

The plurality of calculation models 211 are connected in sequence, and each of the plurality of calculation models 211 connected in sequence receives each of a plurality of content vectors having an order as an input value. For example, as shown in FIG. 3, each of a plurality of computational models 211 connected in sequence is a plurality of content vectors having an order [-0.01, 0.26, -0.264, 0.12, 0.269, 0, 0.032], [0.34, Enter 0.202, 0.04, 0.31,-0.41, 0.303, 0.108], [0.211, 0.12,-0.24, 0.31, 0.51,-0.43, 0.108], [0.11, 0.02, 0.22, 0.31,-0.39, 0.03, 0.08] I can receive it.

In particular, the computational model 211 receives an input value that is a content vector and a predicted value of the previous computational model 211, and derives a content vector in the next order as a predicted value through one or more computations to which weights are applied. For example, as shown in FIG. 4, the computational model 211 includes a content vector [0.211, 0.12,-0.24, 0.31, 0.51,-0.43, 0.108], which is an input value Xt, and a predicted value Ht- of the previous computational model 211. One content vector [0.01, -0.11, 0.17, -0.21, 0.39, 0.231, 0.19] is input, and the next order of content vectors [0.11, 0.25, 0.222, 0.15] as the predicted value Ht through one or more operations to which weights are applied. , -0.112, 0.24, -0.03].

Each of the plurality of calculation models 211 derives a predicted value by performing an operation to which a weight is applied as illustrated in FIG. 4. Accordingly, the last computational model 211 derives a content vector next to the last content vector among a plurality of content vectors as predicted values. For example, as shown in FIG. 3, the next content vector of the last content vector [0.11, 0.02, 0.22, 0.31,-0.39, 0.03, 0.08] is predicted. For example, as shown in FIG. 3, the derived predicted value may be a content vector [0.01, 0.35, 0.442, 0.07,-0.39,-0.433, 0.108]. The derived predicted value is input to the merge layer 220.

The merging layer 220 receives a content vector, which is a predicted value derived by the cyclic layer 210, a viewing rate vector, and a viewing time vector, and derives a content vector as an output value through an operation to which a weight is applied.

Here, the viewing time vector includes a last viewing time vector and a recommended viewing time vector. The last viewing time vector represents the viewing time of the content of the last content vector input to the circulation layer 210. The recommended viewing time vector represents the recommended viewing time of the recommended content to be derived through the artificial neural network 200.

The merged layer 220 may be, for example, a fully-connected layer. Accordingly, the merge layer 220 includes a plurality of nodes corresponding to each of the content vector, the viewing rate vector, and the viewing time vector. For example, the content vector and the viewing rate vector are a vector of length 7, and two viewing time vectors (last If the viewing time vector and the recommended viewing time vector) are vectors of length 8, respectively, the merge layer 220 may consist of 30 computation nodes. A content vector is derived as an output value through the operation by the 30 operation nodes of the merge layer 220.

The output layer 230 stores an output value of the merge layer 220. The output layer 230 includes a plurality of output nodes, and each of the plurality of output nodes corresponds to different contents. In addition, the values of each of the plurality of output nodes indicate a probability that the corresponding content is a content preferred by a user at a recommended viewing time corresponding to the recommended viewing time vector. For example, assuming that the content is seven contacts, the output layer 230 may include seven output nodes, that is, first to seventh output nodes. Accordingly, the first to seventh output nodes respectively correspond to the first to seventh content, and if the values of the first to seventh output nodes are 0.01, 0.02, 0.04, 0.31, 0.51, 0.03, 0.08, recommended viewing time The probability of the content that the user prefers at the viewing time corresponding to the vector is 1% for the first content, 2% for the second content, 4% for the third content, 31% for the fourth content, 51% for the fifth content, The 6th content is 3% and the 7th content is 8%.

Next, a method for recommending content based on deep learning according to an embodiment of the present invention will be described.

First, a method of generating data for recommending context-aware content based on deep learning according to an embodiment of the present invention will be described. 5 is a flowchart illustrating a method of generating data for recommending context-aware content based on deep learning according to an embodiment of the present invention.

Referring to FIG. 5, the viewing history preprocessor 110 includes content viewed by each user for a predetermined period (eg, 6 months) from the viewing history DB 300 in step S110 and a viewing time for viewing the content. The viewing history is extracted and sorted in chronological order. For example, as shown in Table 1 above, for user A, the viewing history may be sorted according to the viewing time by mapping the viewed content and viewing time.

Next, the content vector generation unit 120 generates a content vector by embedding the content included in the user's viewing history provided from the viewing history preprocessor 110 in a predetermined vector space in step S120. For example, the content vector generation unit 120 may select [0.11, 0.11, -0.20. -0.27, 0.09, -0.19, 0.26] can generate 7-dimensional content vectors.

In addition, the viewing rate vector generation unit 130 embeds the viewing history of the user provided from the viewing history preprocessing unit 110 in the vector space in step S130 to indicate the content viewing rate according to the viewing rate of the content viewed by the user. Create a vector. That is, the viewing rate vector generation unit 130 may generate a viewing rate vector indicating a rate at which each user has viewed content belonging to each topic.

Subsequently, the viewing time vector generation unit 140 embeds the viewing time of the content in a predetermined vector space from the viewing history of the user provided from the viewing history preprocessing unit 110 in step S140 to generate a viewing time vector. . For example, it is possible to generate a viewing time vector having a predetermined length according to the time division unit (eg, month, day, day of the week, time, etc.) as shown in Table 3 in one-hot encoding according to "Sparse representation". .

The deep learning of the present invention is to recommend content that a user prefers at a recommendation time, which is a time the user wants to watch. To this end, deep learning is performed using data generated based on the user's viewing history described above in FIG. 5 as learning data. Then, a deep learning method for recommending context-aware content according to an embodiment of the present invention will be described. 6 is a flowchart illustrating a deep learning method for recommending context-aware content according to an embodiment of the present invention.

Referring to FIG. 6, the deep learning model learning unit 150 inputs a plurality of content vectors into the circulation layer 210 of the artificial neural network 200 according to the viewing order of the plurality of contents in step S210. For example, it is assumed that the viewing order is the order of first, second, first, fourth, and third content. Then, as shown in FIG. 3, the deep learning model learning unit 150 can input a content vector corresponding to the first, second, first, and fourth contents to the circulation layer 210 according to the viewing order. have. Specifically, the deep learning model learning unit 150 calculates the first, second, first, and fourth content vectors corresponding to the first, second, first, and fourth contents, respectively, from the first computational model 211 to the fourth. It is input to the calculation model 211.

Then, the plurality of calculation models 211 of the circulatory layer 210 of the artificial neural network 200 outputs a content vector as a predicted value through a plurality of calculations in which a weight is applied to each of the plurality of content vectors in step S220. Here, the operation means an operation using an activation function. For example, as shown in FIG. 3, each of the plurality of calculation models 211 performs an operation to which a weight is applied, and as a predicted value, the next content vector of the fourth content vector, which is the last content vector, [0.01, 0.35, 0.442, 0.07, and -0.39,-0.433, 0.108] can be derived.

Subsequently, the merging layer 220 of the artificial neural network 200 includes a content vector, which is a predicted value derived by the circulatory layer 210 from the circulatory layer 210 in step S230, and a viewing rate vector from the deep learning model learning unit 150, The first viewing time vector and the second viewing time vector are input. Here, the viewing rate vector is generated by the viewing rate vector generation unit 130, and embeds the user's viewing history in the vector space to indicate the content viewing rate according to the viewing rate of the content viewed by the user. As shown in FIG. 3, if the viewing rate vector is [0.21, 0.15, 0.12, 0.05, 0.21, 0.23, 0.03], the viewing rates of the first to seventh contents of the corresponding user are 21%, 15%, respectively, 12%, 5%, 21%, 23%, 3%. In addition, the first viewing time vector and the second viewing time vector represent viewing times of the last-order content according to a plurality of content vectors input to the circulation layer 210. Further, the second viewing time vector represents the viewing time of the next content vector in the last order. For example, referring to Table 3, according to a plurality of content vectors input to the circulation layer 210, the user's last-order content, for example, the fourth content, was viewed at 19:00 on Saturday, and then the content, for example, the third If the content is watched at 20:00 on Sunday, both the first viewing time vector and the second viewing time vector may be T8 [0, 0, 0, 0, 0, 0, 0, 1]. As another example, according to the plurality of content vectors input to the circulation layer 210, if the user's last-order content is viewed at 23:00 on Saturday, and then the next content is viewed at 1 on Sunday, the first viewing time vector is T8 [ 0, 0, 0, 0, 0, 0, 0, 1], and the second viewing time vector may be T5 [0, 0, 0, 0, 1, 0, 0, 0].

Then, the merging layer 220 of the artificial neural network 200 has a weight for the content vector, the viewing rate vector, the first viewing time vector, and the second viewing time vector, which are predicted values derived by the circulatory layer 210 in step S240. The content vector is derived as an output value through the applied operation. Here, the operation means an operation using an activation function. The activation function may be a Sigmoid function, a tanh function, a ReLU function, a Leakly ReLU function, a PReLU function, an ELU function, a Softmax　 function, a Maxout function, a Minout function, and the like. In addition, the weight can be applied to the input or output of the above-described activation function. In addition, the output values derived by the merge layer 220 are stored in a plurality of output nodes of the output layer 230. For example, as shown in FIG. 3, if the output value is (0.01, 0.02, 0.04, 0.31, 0.51, 0.03, 0.08), 0.01, 0.02, 0.04, respectively, to the first to seventh output nodes of the output layer 230 0.31, 0.51, 0.03, and 0.08 are stored.

Next, the deep learning model learning unit 150 calculates a weight through back propagation so that the difference between the output value and the target value (LABEL) is minimized in step S250.

Here, it is assumed that the output values are stored in the first to seventh output nodes of the output layer 230 and are stored as 0.01, 0.02, 0.04, 0.31, 0.51, 0.03, and 0.08, respectively. In addition, the target value (LABEL) is the content viewed after the content corresponding to the last content vector from the plurality of content vectors input to the circulation layer 210, and is viewed by the user at the viewing time corresponding to the second viewing time vector. It is an index indicating the content. It is assumed that the content viewed next to the content corresponding to the last content vector and the content viewed by the user at the viewing time corresponding to the second viewing time vector is the third content. Then, the target value is (0, 0, 1, 0, 0, 0, 0), and the value of the third output node corresponding to the third content is 1.00 (100%), and all other output nodes are 0.00 ( 0%). Then, the deep learning model learning unit 150 outputs (0.01, 0.02, 0.04, 0.31, 0.51, 0.03, 0.08) and target values (0, 0, 1, 0, 0, 0) through back propagation. A weight applied to a plurality of calculations of the artificial neural network 200 to minimize the difference between, 0) is calculated.

The deep learning model learning unit 150 uses a plurality of different learning data, that is, a plurality of content vectors, viewing rate vectors, and first and second viewing time vectors, so that the difference between the output value and the target value is within a preset range. The above-described steps S210 to S250 are repeated until they become. Therefore, the deep learning model learning unit 150 determines whether or not to end learning by determining whether the difference between the output value and the target value is within a preset range in step S260. As a result of the determination in step S260, if the difference between the output value and the target value is not within a preset range, the process proceeds to step S210 and steps S210 to S250 described above are repeated. Further, as a result of the determination in step S260, if the difference between the output value and the target value is within a preset range, the learning is terminated.

Then, a method of recommending a content preferred by a user at a recommendation time, which is a time that the user wants to watch, using the artificial neural network 200 learned as described above will be described. 7 is a flowchart illustrating a method for recommending context-sensitive content using deep learning according to an embodiment of the present invention.

Referring to FIG. 7, the content recommendation unit 160 inputs a plurality of content vectors, a viewing rate vector, a third viewing time vector, and a fourth viewing time vector into the artificial neural network 200 in step S310. That is, the content recommendation unit 160 inputs the content recently viewed by the user into the circulation layer 210 of the artificial neural network 200 as a plurality of content vectors according to the viewing order, and the merge layer 220 of the artificial neural network 200 In ), input the user's viewing rate vector, the third viewing time vector, and the fourth viewing time vector. For example, it is assumed that the latest viewing order is the third, first, second, and fourth contents. Then, the deep learning model learning unit 150 transfers the third, first, second, and fourth content vectors to the first calculation model 211 to the fourth calculation model 211, respectively, in the circulation layer 210 according to the viewing order. Enter. In addition, the third viewing time vector represents the viewing time of the last viewed content in the latest viewing sequence. For example, when the latest viewing order is the third, first, second, and fourth content, the third viewing time vector represents the viewing time of the fourth content. In addition, the fourth viewing time vector represents the recommended viewing time. According to an embodiment, the fourth viewing time vector may represent a time when the user device 20 accesses the content server 10. Alternatively, the fourth viewing time vector may represent a time at which the user requests a recommendation.

Next, the plurality of computational models 211 of the circulatory layer 210 of the artificial neural network 200 derives a content vector as a predicted value through a plurality of computations in which weights are applied to each of the plurality of content vectors in step S320. . Here, the operation means an operation using an activation function. For example, as shown in FIG. 3, each of the plurality of calculation models 211 performs an operation to which a weight is applied, and as a predicted value, the next content vector of the fourth content vector, which is the last content vector, [0.01, 0.05, 0.02, 0.18, 0.69, 0.03, 0.02] can be derived.

Subsequently, the merging layer 220 of the artificial neural network 200 is weighted to the content vector, the viewing rate vector, the third viewing time vector, and the fourth viewing time vector, which are predicted values derived by the circulatory layer 210 in step S330. The content vector is derived as an output value through an operation that is performed. Here, the operation means an operation using an activation function. In addition, the weight can be applied to the input or output of the above-described activation function. In addition, the output values derived by the merge layer 220 are stored in a plurality of output nodes of the output layer 230. For example, if the output value is (0.01, 0.02, 0.04, 0.01, 0.81, 0.03, 0.08), 0.01, 0.02, 0.04, 0.01, 0.81, 0.03, and 0.08 are respectively assigned to the first to seventh output nodes of the output layer 230. Is saved. The content vector, which is an output value, represents the recommended content of the recommended viewing time corresponding to the fourth viewing time vector.

Next, the content recommendation unit 160 recognizes the recommended content of the recommended viewing time from the output value in step S340 and extracts the content corresponding to the recommended content of the recommended viewing time from the content DB 400. More specifically, as described above, each of the first to seventh output nodes corresponds to the first to seventh content, and each value of the first to seventh output nodes is a recommendation corresponding to the fourth viewing time vector. It represents the probability of the content that the user prefers at the viewing time. Accordingly, the probability of the content that the user prefers at the recommended viewing time corresponding to the fourth viewing time vector is 1%, 2%, 4%, 1%, 81%, 3%, 8%, respectively. to be. Accordingly, the content recommendation unit 160 recognizes the fifth content having the highest probability as the recommended content of the recommended viewing time. Accordingly, the content recommendation unit 160 extracts content corresponding to the fifth content from the content DB 400. Then, the content recommendation unit 160 provides the extracted content to the user device 20 as recommended content of the recommended viewing time in step S350.

The method according to an embodiment of the present invention may be provided in the form of a computer-readable medium suitable for storing computer program instructions and data. Such a computer-readable recording medium may include program commands, data files, data structures, etc. alone or in combination, and includes all types of recording devices storing data that can be read by a computer system. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROM (Compact Disk Read Only Memory), and DVD (Digital Video Disk). Optical Media), Magnetic-Optical Media such as a floptical disk, and program instructions such as ROM (Read Only Memory), RAM (RAM, Random Access Memory), flash memory, etc. And a hardware device specially configured to perform. Further, the computer-readable recording medium is distributed over a computer system connected through a network, so that computer-readable codes can be stored and executed in a distributed manner. In addition, functional programs, codes, and code segments for implementing the present invention can be easily inferred by programmers in the technical field to which the present invention belongs.

As described above and shown in connection with a preferred embodiment for illustrating the technical idea of the present invention, the present invention is not limited to the configuration and operation as shown and described as described above, without departing from the scope of the technical idea It will be well understood by those skilled in the art that many changes and modifications are possible to the present invention. Accordingly, all such appropriate changes and modifications and equivalents should be considered to be within the scope of the present invention.

The apparatus and method for recommending context-aware content based on deep learning according to the present invention do not simply recommend the content that the user prefers, but the recommended content suitable for the recommended time by classifying the content that the user prefers for each time period. Can be provided by deriving. Accordingly, it is possible to provide a user experience (UX) based on a higher understanding of the user. The present invention is commercially available or commercially feasible, as well as a degree that can be practically clearly implemented, and thus has industrial applicability.

10: content server 20: user device
100: content recommendation engine 110: viewing history pre-processing unit
120: content vector generation unit 130: viewing rate vector generation unit
140: viewing time vector generation unit 150: deep learning model learning unit
160: content recommendation unit 200: artificial neural network
210: cyclic layer 211: computational model
220: merge layer 230: output layer
300: viewing history DB 400: content DB

Claims (10)

When a plurality of content vectors representing a plurality of contents are input according to the viewing order, weights are applied to the plurality of content vectors, and a predicted value is a content indicating the next viewed content of the last content among the plurality of contents. A cyclic layer that outputs a vector,
When a content vector output from the circulation layer, a first viewing time vector representing a viewing time of the last-order content, and a second viewing time vector representing a viewing time of the next viewed content are input, the output content vector, A merge layer that outputs an output value calculated through an operation in which a weight is applied to the first viewing time vector and the second viewing time vector to an output layer
Artificial neural network comprising a; Characterized in that it comprises a
A device for recommending context-aware content based on deep learning. The method of claim 1,
A plurality of content vectors representing a plurality of content according to the user's recent viewing order are input into the circulation layer, and a third viewing time vector representing the viewing time of the last order among the recent viewing order and a fourth viewing time representing the recommended viewing time Further comprising; a content recommendation unit for inputting a time vector into the merge layer,
The circulatory layer of the artificial neural network outputs a content vector as a predicted value through a plurality of calculations in which a weight is applied to a plurality of content vectors input according to the latest viewing order,
The merging layer of the artificial neural network outputs an output value representing the recommended content of the recommended viewing time through an operation in which a weight is applied to the content vector output as the predicted value, the third viewing time vector, and the fourth viewing time vector. Characterized in that output as a layer
A device for recommending context-aware content based on deep learning. The method of claim 2,
The content recommendation unit
Extracting content according to an output value representing the recommended content of the recommended viewing time, and providing the extracted content to a user device
A device for recommending context-aware content based on deep learning. The method of claim 1,
The circulation layer includes a plurality of computational models,
Each of the plurality of calculation models
Characterized in that, receiving an input value and a predicted value of a previous computational model, and deriving a content vector in a next order as a predicted value through one or more operations to which a weight is applied.
A device for recommending context-aware content based on deep learning. The method of claim 1,
A content vector generator for generating a content vector by embedding the content included in the viewing history into a predetermined vector space; And
And a viewing time vector generator for generating a viewing time vector by embedding the viewing time of the content from the viewing history into a predetermined vector space.
A device for recommending context-aware content based on deep learning. Receiving, by a circulatory layer of the artificial neural network, a plurality of content vectors representing a plurality of contents according to a viewing order;
Deriving, by a circulatory layer of the artificial neural network, a content vector as a predicted value through a plurality of operations in which weights are applied to the plurality of content vectors, respectively;
The content vector obtained by the merging layer of the artificial neural network, a first viewing time vector indicating a viewing time of a last-order content among the plurality of contents, and a second viewing time indicating a viewing time of the next viewed content in the last order Receiving a vector; And
Deriving, by the merging layer of the artificial neural network, an output value through an operation in which a weight is applied to the derived content vector, the first viewing time vector, and the second viewing time vector;
Characterized in that it comprises a
A method for recommending context-aware content based on deep learning. The method of claim 6,
After the step of deriving the output value,
Inputting, by a content recommendation unit, a plurality of content vectors representing a plurality of contents according to a user's recent viewing order, into the circulation layer;
Inputting, by the content recommendation unit, a third viewing time vector indicating a viewing time of a last order of the recent viewing order and a fourth viewing time vector indicating a recommended viewing time to the merging layer;
Deriving, by a circulatory layer of the artificial neural network, a content vector as a predicted value through a plurality of operations in which weights are applied to the plurality of content vectors, respectively;
The merging layer of the artificial neural network derives an output value representing the recommended content of the recommended viewing time through an operation in which a weight is applied to the content vector output as the predicted value, the third viewing time vector, and the fourth viewing time vector. Step; characterized in that it further comprises
A method for recommending context-aware content based on deep learning. The method of claim 7,
After the step of deriving an output value representing the recommended content of the recommended viewing time,
Extracting, by the content recommendation unit, content according to an output value indicating recommended content of the recommended viewing time; And
And providing the extracted content as recommended content for a recommended viewing time to a user device;
A method for recommending context-aware content based on deep learning. The method of claim 6,
The step of deriving a content vector from the predicted value
Each of the plurality of computational models of the circulation layer
And receiving an input value and a predicted value of a previous computational model, and deriving a content vector in a next order as a predicted value through one or more operations to which a weight is applied.
A method for recommending context-aware content based on deep learning. The method of claim 6,
Before the step of receiving the plurality of content vectors,
Generating a plurality of content vectors by embedding, by the content vector generator, the content included in the viewing history in a predetermined vector space; And
And generating, by a viewing time vector generation unit, a viewing time vector by embedding the viewing time of the content from the viewing history into a predetermined vector space.
A method for recommending context-aware content based on deep learning.

Images