KR20220155686A - System and method of recommending custmized course subjects based on deep learning - Google Patents

Abstract

The present invention relates to a technical idea for recommending customized courses based on deep learning and, more specifically, to technology for recommending customized courses by learning, based on deep learning, user behavior information including information related to a user job, qualifications, and language proficiency, and course attribute information including information related to courses and majors that a user can take, and learns the learned user behavior information and the learned course attribute information together to learn the interaction characteristics by having the participation of each user in an individual course as a target variable. According to an embodiment of the present invention, a deep learning-based user-customized course recommendation system comprises: a data collection unit which collects a plurality of pieces of user behavior data and a plurality of pieces of course attribute data; a deep learning model processing unit which learns the collected user behavior data to extract at least one user behavior characteristic, learns the collected course attribute data to extract at least one course attribute characteristic, and learns the extracted at least one user behavior characteristic and the extracted at least one course property characteristic together to extract at least one interaction characteristic related to the participation of the user in a plurality of courses; and a recommendation unit which recommends the at least one course based on the at least one extracted interaction characteristic. According to the present invention, the interest preference of the user can be sufficiently predicted.

Description

Deep learning-based user-customized course recommendation system and method {SYSTEM AND METHOD OF RECOMMENDING CUSTMIZED COURSE SUBJECTS BASED ON DEEP LEARNING}

The present invention relates to a technical concept of recommending courses customized for a user based on deep learning, and based on deep learning, user behavior information including information related to a user's job, license, language ability, etc., and courses and majors that the user can take Learning course attribute information including information related to etc., and learning the learned user behavior information and course attribute information together to learn the interaction characteristics by learning whether or not each user has taken an individual course as a target variable. It is about the art of recommending subjects.

Due to the development of information technology and the popularization of mobile devices, various types of structured and unstructured data have exploded, making it possible to produce and collect a lot of information.

However, due to the rapidly increasing amount of information, users are faced with the problem of information overload, which takes a lot of time and makes it difficult to make decisions in selecting necessary products or services.

Accordingly, the importance of personalized recommendation services is emerging, and recommendation systems are developed to provide various user-customized services in fields such as movies, news, and products.

Recently, demand for online education platforms is increasing due to COVID-19, which is spreading around the world.

However, it is difficult for a user to correctly select a subject suitable for him/herself among numerous subjects due to a difference in knowledge structure.

Therefore, a personalized service that provides customized subjects to users is essential because it plays an important role in improving the learning efficiency of users.

Currently, most online education platforms provide recommendation services by applying traditional Collaborative Filtering (CF) algorithms.

However, since the analysis logic of these existing recommendation algorithms is simple, it is difficult to build a model using various user information.

In addition, the recommendation service to which the collaborative filtering algorithm is applied has a "cold start" problem in which suitable items cannot be recommended for new users due to insufficient past information or data, and recommendations cannot be provided until the user's preferences are reflected. There are also "First Start" problems that don't exist.

If the user's preference data continuously increases, there is also a scalability problem that lowers the computational performance of the algorithm. Recently, research is being conducted to supplement problems such as data scarcity and scalability.

Recently, research on applying deep learning techniques to improve the limitations of existing personalized recommendation service research in various fields is being actively conducted.

However, in the current research on personalized recommendation services in the field of online education, there are not many studies that apply deep learning techniques to recommend subjects suitable for the user's knowledge structure.

Korean Patent Registration No. 10-2210504, "Customized Education Program System Using AI Education Platform Based on Big Data" Korean Patent Registration No. 10-2063183, "Apparatus and Method for Recommending Learning Contents" Korean Patent Registration No. 10-2070783, “A clothes handling device that recommends customized courses and options by learning patterns of courses and options performed by a user through artificial intelligence, a control method thereof, and an online system including the clothes handling device " Korean Patent Publication No. 10-2020-0103146, "Artificial Intelligence Learning Diagnosis System Using Ontology-based Knowledge Map"

An object of the present invention is to reduce information overload on curriculum resources and solve the problem of data scarcity based on a deep learning algorithm that models user behavior information and course attribute information.

An object of the present invention is to solve the problem of high-dimensional data sparsity and scalability, which are problems of existing collaborative filtering algorithms, by simultaneously considering user behavior information and course attribute information in order to solve the information overload problem in the field of online education.

The present invention constructs a high-level nonlinear deep learning algorithm by utilizing a multi-layer perceptron, reduces dimensionality in sparse data of an online education platform, and interacts with user behavior information and course attribute information. Its purpose is to sufficiently predict the user's interest preference by elaborately extracting characteristic information.

The present invention uses latent features learned from each of a user behavior network (UBN) layer, a course attribution network (CAN) layer, and a deep neural network (DNN) layer. The purpose of this study is to perform deep learning learning in a way that minimizes the pointwise error by using each user's participation in each course as a target variable.

An object of the present invention is to apply to an online education platform and efficiently recommend courses suitable for a user's knowledge structure.

In the case of a new user, the present invention solves the cold start problem in which suitable courses cannot be recommended due to lack of past information or data, and the first start in which recommendations cannot be provided until the user's preference is reflected. It aims to solve problems simultaneously.

A deep learning-based user-customized course recommendation system according to an embodiment of the present invention includes a data collection unit that collects a plurality of user behavior data and a plurality of course attribute data, and learns the collected plurality of user behavior data to provide at least one Extracting user behavioral characteristics, learning the collected plurality of course attribute data to extract at least one course attribute characteristic, and learning the extracted at least one user behavioral characteristic and the extracted at least one course attribute characteristic together A deep learning model processing unit extracting at least one interaction characteristic related to a user's participation in a plurality of courses and a recommendation unit recommending at least one course based on the extracted at least one interaction characteristic can do.

The data collection unit may collect the plurality of user behavior data including a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data.

The data collection unit may collect the plurality of course attribute data including a plurality of course data and a plurality of major data.

The deep learning model processing unit first embeds a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data included in the plurality of user behavior data. vectors, the converted first embedding vectors are input to a user behavior network (UBN) layer, and the at least one user behavior characteristic is extracted by learning the input first embedding vectors. have.

The deep learning model processing unit converts a plurality of course data and a plurality of major data included in the plurality of course attribute data into second embedding vectors, and converts the converted second embedding vectors into course attributes. The at least one course attribute feature may be extracted by inputting the input to a course attribution network (CAN) layer and learning the input second embedding vectors.

The deep learning model processing unit inputs the extracted at least one user behavioral characteristic and the extracted at least one course attribute characteristic to a deep neural network (DNN) layer, and inputs the at least one inputted user behavioral characteristic. and learning the input at least one course attribute characteristic together to extract the at least one interaction characteristic.

The deep learning model processing unit uses the user behavior network (UBN) layer and the course attribution network (CAN) layer as a parallel sub-network, and the parallel sub-network and The deep neural network (DNN) layer may be used as a prediction layer by connecting the deep neural network (DNN) layer.

The deep learning model processing unit controls a negative sampling rate of a point wise function in learning the at least one extracted user behavioral characteristic and the at least one extracted course attribute characteristic together to obtain the at least one In extracting interaction features, it is possible to reduce a point wise error of a deep learning model.

The recommendation unit may recommend the at least one course course based on implicit feedback that indirectly reflects the extracted at least one interaction characteristic to the recommendation of the at least one course course.

The recommendation unit may rank the at least one enrolled subject, list the ranked at least one enrolled subject, and provide a recommendation list including the listed at least one enrolled subject.

The recommendation unit may randomly select courses that the user has not interacted with among the at least one enrolled courses and rank the at least one enrolled course.

A deep learning-based course recommendation method customized for a user according to an embodiment of the present invention includes the steps of collecting a plurality of user behavior data and a plurality of course attribute data in a data collection unit; extracting at least one user behavioral characteristic by learning user behavioral data; extracting at least one course attribute characteristic by learning the collected plurality of course attribute data in the deep learning model processing unit; Extracting, in a processing unit, at least one interaction characteristic related to the user's participation in a plurality of courses by learning the extracted at least one user behavioral characteristic and the extracted at least one course attribute characteristic together; and recommending In the unit, it may include recommending at least one course to be taken based on the at least one extracted interaction characteristic.

The collecting of the plurality of user behavior data and the plurality of course attribute data includes a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data. The method may include collecting the plurality of user behavior data including a plurality of user behavior data and collecting the plurality of course attribute data including a plurality of course data and a plurality of major data.

The step of learning the collected plurality of user behavior data and extracting at least one user behavior characteristic includes a plurality of user data, a plurality of job data, and a plurality of user data included in the plurality of user behavior data. Certificate data and a plurality of language data are converted into first embedding vectors, the converted first embedding vectors are input to a user behavior network (UBN) layer, and the inputted first embedding vectors are input. 1 Learning embedding vectors to extract the at least one user behavior characteristic.

The step of learning the collected plurality of course attribute data and extracting at least one course attribute characteristic may include a plurality of course data and a plurality of major data included in the plurality of course attribute data as second Converting into embedding vectors, inputting the converted second embedding vectors to a course attribution network (CAN) layer, learning the input second embedding vectors to extract the at least one course attribute feature steps may be included.

The step of extracting at least one interaction characteristic related to the user's participation in a plurality of courses by learning the extracted at least one user behavior characteristic and the extracted at least one course property characteristic together, At least one user behavioral characteristic and the extracted at least one course attribute characteristic are input to a deep neural network (DNN) layer, and the inputted at least one user behavioral characteristic and the inputted at least one course attribute characteristic are inputted. and extracting the at least one interaction feature by learning features together.

The present invention is based on a deep learning algorithm that models user behavior information and course attribute information, and can reduce information overload for curriculum resources and solve the problem of data scarcity.

The present invention can solve the problem of high-dimensional data sparsity and scalability, which are problems of existing collaborative filtering algorithms, by simultaneously considering user behavior information and course attribute information to solve the information overload problem in the field of online education.

The present invention constructs a high-level nonlinear deep learning algorithm by utilizing a multi-layer perceptron, reduces dimensionality in sparse data of an online education platform, and interacts with user behavior information and course attribute information. It is possible to sufficiently predict the user's interest preference by precisely extracting the characteristic information.

The present invention uses latent features learned from each of a user behavior network (UBN) layer, a course attribution network (CAN) layer, and a deep neural network (DNN) layer. Deep learning learning can be performed in a way that minimizes a pointwise error by setting each user's participation in each course as a target variable.

The present invention can be applied to an online education platform to efficiently recommend courses suitable for a user's knowledge structure.

In the case of a new user, the present invention solves the cold start problem in which suitable courses cannot be recommended due to lack of past information or data, and the first start in which recommendations cannot be provided until the user's preference is reflected. problems can be solved at the same time.

1 to 3 are diagrams illustrating a deep learning-based user-customized course recommendation system according to an embodiment of the present invention.
4 is a diagram illustrating a method for recommending a course course customized for a user based on deep learning according to an embodiment of the present invention.
5A to 7B are diagrams illustrating evaluation results for results recommended by a system for recommending courses customized to a user based on deep learning according to an embodiment of the present invention.

Hereinafter, various embodiments of this document will be described with reference to the accompanying drawings.

Examples and terms used therein are not intended to limit the technology described in this document to specific embodiments, and should be understood to include various modifications, equivalents, and/or substitutes of the embodiments.

In the following description of various embodiments, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the invention, the detailed description will be omitted.

In addition, terms to be described below are terms defined in consideration of functions in various embodiments, and may vary according to intentions or customs of users or operators. Therefore, the definition should be made based on the contents throughout this specification.

In connection with the description of the drawings, like reference numerals may be used for like elements.

Singular expressions may include plural expressions unless the context clearly dictates otherwise.

In this document, expressions such as "A or B" or "at least one of A and/or B" may include all possible combinations of the items listed together.

Expressions such as "first," "second," "first," or "second," may modify the corresponding components regardless of order or importance, and are used to distinguish one component from another. It is used only and does not limit the corresponding components.

When a (e.g., first) element is referred to as being "(functionally or communicatively) coupled to" or "connected to" another (e.g., second) element, that element refers to the other (e.g., second) element. It may be directly connected to the component or connected through another component (eg, a third component).

In this specification, "configured to (or configured to)" means "suitable for," "having the ability to," "changed to" depending on the situation, for example, hardware or software ," can be used interchangeably with "made to," "capable of," or "designed to."

In some contexts, the expression "device configured to" can mean that the device is "capable of" in conjunction with other devices or components.

For example, the phrase "a processor configured (or configured) to perform A, B, and C" may include a dedicated processor (eg, embedded processor) to perform the operation, or by executing one or more software programs stored in a memory device. , may mean a general-purpose processor (eg, CPU or application processor) capable of performing corresponding operations.

Also, the term 'or' means 'inclusive or' rather than 'exclusive or'.

That is, unless otherwise stated or clear from the context, the expression 'x employs a or b' means any one of the natural inclusive permutations.

Terms such as '..unit' and '..group' used below refer to a unit that processes at least one function or operation, and may be implemented by hardware or software, or a combination of hardware and software.

1 to 3 are diagrams illustrating a deep learning-based user-customized course recommendation system according to an embodiment of the present invention.

1 illustrates components of a deep learning-based user-customized course recommendation system according to an embodiment of the present invention.

Referring to FIG. 1 , a deep learning-based user-customized course recommendation system 100 according to an embodiment of the present invention may include a data collection unit 110, a deep learning model processing unit 120, and a recommendation unit 130. can

According to an embodiment of the present invention, the data collection unit 110 may collect a plurality of user behavior data and a plurality of course attribute data.

Specifically, the data collection unit 110 collects a plurality of user behavior data including a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data. can be collected

Also, the data collection unit 110 may collect a plurality of course attribute data including a plurality of course data and a plurality of major data.

According to an embodiment of the present invention, the deep learning model processing unit 120 may extract at least one user behavioral characteristic by learning a plurality of collected user behavioral data.

In addition, the deep learning model processing unit 120 may extract at least one course attribute characteristic by learning a plurality of collected course attribute data.

In addition, the deep learning model processing unit 120 learns at least one extracted user behavioral characteristic and at least one extracted course attribute characteristic together to determine at least one interaction characteristic related to the user's participation in a plurality of courses. can be extracted.

For example, the deep learning model processing unit 120 uses a user behavior network (UBN) layer and a course attribution network (CAN) layer as a parallel sub-network, and uses a parallel sub-network. The deep neural network (DNN) layer may be used as a prediction layer by connecting the network and the deep neural network (DNN) layer.

According to an embodiment of the present invention, the deep learning model processing unit 120 includes a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of user data included in the plurality of user behavior data. Language data is converted into first embedding vectors, the converted first embedding vectors are input to a user behavior network (UBN) layer, and the inputted first embedding vectors are learned to obtain at least one user. behavioral characteristics can be extracted.

For example, the deep learning model processing unit 120 converts a plurality of course data and a plurality of major data included in the plurality of course attribute data into second embedding vectors, and converts the converted second embedding vector. may be input to a course attribution network (CAN) layer, and at least one course attribute feature may be extracted by learning the input second embedding vectors.

According to an embodiment of the present invention, the deep learning model processing unit 120 inputs at least one extracted user behavioral characteristic and at least one extracted course attribute characteristic to a deep neural network (DNN) layer, and inputs At least one interaction characteristic may be extracted by learning at least one user behavior characteristic and at least one input course attribute characteristic together.

That is, the deep learning model processing unit 120 may extract at least one interaction characteristic by collectively learning at least one user behavior characteristic and at least one course attribute characteristic.

According to an embodiment of the present invention, the deep learning model processing unit 120 performs negative sampling of a point wise function in learning at least one extracted user behavioral characteristic and at least one extracted course attribute characteristic together. In extracting at least one interaction feature by controlling the ratio, a point wise error of the deep learning model may be reduced.

According to an embodiment of the present invention, the recommendation unit 130 may recommend at least one course course based on at least one interaction characteristic.

For example, the recommendation unit 130 may recommend at least one course to be taken based on implicit feedback that indirectly reflects the extracted at least one interaction characteristic to the recommendation of at least one course to be taken.

According to an embodiment of the present invention, the recommendation unit 130 ranks at least one course taken, lists the ranked at least one course taken, and creates a recommendation list including the listed at least one course taken. can provide

For example, the recommendation unit 130 may randomly select a course that the user has not interacted with among at least one course and rank the at least one course.

Here, the courses that the user did not interact with may be related to the courses that the user did not take.

Therefore, the present invention can solve the problem of data scarcity and reduction of information overload for curriculum resources based on a deep learning algorithm that models user behavior information and course attribute information.

In addition, the present invention can solve the problem of high-dimensional data scarcity and scalability, which are problems of existing collaborative filtering algorithms, by simultaneously considering user behavior information and course attribute information to solve the information overload problem in the field of online education.

2 illustrates a framework of a deep learning model used by a user-customized course recommendation system based on deep learning according to an embodiment of the present invention, and a framework of a deep learning model according to an embodiment of the present invention ( 200) may be related to the deep learning model processing unit described in FIG. 1 .

Referring to FIG. 2 , the deep learning model framework 200 according to an embodiment of the present invention includes an input network layer 210 into which embedding vectors are input, a user behavior network layer 220 which extracts user behavior characteristics, It includes a course attribute network layer 230 and a deep neural network layer 240. For example, the deep learning model may be referred to as a deep learning-based approach to course recommender (DECOR) model.

According to one embodiment of the present invention, the input network layer 210 is a hidden layer, and user behavior data related to a user, job, license, and language is converted into an embedding vector and input to the input network layer 210, Course attribute data related to courses and majors may be converted into embedding vectors and then input to the input network layer 210 .

According to one embodiment of the present invention, data input to the input network layer 210 may be transferred to the user behavior network layer 220 and the course attribute network layer 230.

For example, the user behavior network layer 220 may be referred to as a user behavior network (UBN) layer, and the course attribute network layer 230 may be referred to as a course attribute network (CAN) layer.

For example, the user behavior network layer 220 may use a feed-forward neural network (FFNN) to extract user behavior characteristics.

The first embedding vector input to the user behavior network layer 220 is a vector having a low-dimensional dense real value as it includes highly sparse and high-dimensional categorical and continuous values.

The first embedding vector may be fed to the hidden layer of the neural network in the forward pass.

The user behavior network layer 220 according to an embodiment of the present invention may provide a user behavior characteristic such as Equation 1 below as an output.

[Equation 1]

In Equation 1, e may indicate the inclusion of users, tasks, certificates, and languages, respectively, and a ⁽⁰⁾ may relate to a process supplied to the deep neural network layer 240.

A process supplied to the deep neural network layer 240 can be defined through Equation 2 below.

[Equation 2]

는 활성화 함수를 나타내고,

,

및

은 각각 l 번째 레이어의 출력, 모델 가중치 및 편향을 나타낼 수 있다.In Equation 2, l represents the depth of the layer

denotes the activation function,

,

and

may represent the output, model weight, and bias of the l-th layer, respectively.

For example, the course attribute network layer 230 may use a feed-forward neural network (FFNN) to extract user behavioral characteristics.

The second embedding vector input to the course attribute network layer 230 is a vector having a low-dimensional dense real value as it includes highly sparse and high-dimensional categorical and continuous values.

The second embedding vector may be fed to the hidden layer of the neural network in the forward pass.

The course attribute network layer 230 according to an embodiment of the present invention may provide a course attribute characteristic of Equation 3 below as an output.

[Equation 3]

In Equation 3, e may indicate the inclusion of courses and majors, respectively, and a ⁽⁰⁾ may be related to a process supplied to the deep neural network layer 240.

A process supplied to the deep neural network layer 240 can be defined through Equation 2 above.

According to an embodiment of the present invention, the deep neural network layer 240 may integrate the user behavior network layer 220 and the course attribute network layer 230.

The deep neural network layer 240 may extract the user behavior characteristics of the user behavior network layer 220 and the course attribute characteristics of the course attribute network layer 230 in an end-to-end process.

The deep neural network layer 240 may jointly learn a deep learning model based on Equation 4 below for all parameters and network parameters.

[Equation 4]

는 이진 클래스 레이블을 나타낼 수 있고,

는 시그 모이드(sigmoid) 함수를 나타낼 수 있으며, N은 사용자 행동 네크워크층(220)의 출력을 나타낼 수 있고, CAN은 코스 속성 네트워크층(230)의 출력을 나타낼 수 있다.In Equation 4,

can represent a binary class label,

may represent a sigmoid function, N may represent an output of the user behavior network layer 220, and CAN may represent an output of the course attribute network layer 230.

Therefore, the present invention constructs a high-level nonlinear deep learning algorithm by utilizing a multi-layer perceptron, reduces the dimension in sparse data of an online education platform, and combines user behavior information and course attribute information. It is possible to sufficiently predict the user's interest preference by precisely extracting the interaction characteristic information.

3 illustrates a course recommendation system customized for a user based on deep learning according to an embodiment of the present invention in more detail.

Referring to FIG. 3 , in the deep learning-based user-customized course recommendation system 300 according to an embodiment of the present invention, the data pre-processing unit 320 collects a plurality of user behavior data from the first database 310 and , A plurality of course attribute data is collected from the second database 311 .

For example, the data pre-processing unit 320 classifies and defines data related to user information, job information, qualification information, and language ability information in relation to a plurality of user behavior data, and classifies and defines data related to a plurality of course attribute data. Classify and define data related to major information.

According to one embodiment of the present invention, the data mapping unit 321 maps user information, job information, qualification information, and language ability information to user behavior data, and maps course information and major information to course attribute data.

For example, the training data set generation unit 322 generates the mapped user behavior data and course attribute data as a training data set.

According to one embodiment of the present invention, the data collection unit described in FIG. 1 may include a data pre-processing unit 320, a data mapping unit 321, and a training data set generation unit 322.

As an example, the training data set generated by the training data set generation unit 322 is transferred to the deep learning model processing unit 330, and the deep learning model processing unit 330 generates a plurality of user behavior data included in the training data set and a plurality of User behavioral characteristics and course attribute characteristics are extracted by learning the course attribute data of , and interaction characteristics related to the user's participation in the course are extracted by learning the user behavioral characteristics and course attribute characteristics together.

According to an embodiment of the present invention, the recommender 340 recommends courses to be taken based on the interaction characteristics, but may recommend the courses to be taken in a ranked list with respect to the recommended courses to be taken.

As an example, the test data set providing unit 350 provides a test data set, which may be defined as shown in Table 1 below.

For example, as a test dataset, a data set composed of user behavior data and course attribute data is collected from an online recruitment website and an online certification website, respectively, and the data set may be configured as shown in Table 1 below.

[Table 1]

Referring to Table 1, the data set was collected from 891 users in 11,099 courses with a sparsity of 99.73%. Included.

In addition, there are 25 job information, 15 language information, 235 qualification information, and 45 major information, and each interaction can indicate whether the user has taken a course or not.

In addition, the evaluation meter providing unit 351 provides an evaluation metric for performing a test on the test data set and the recommendation result.

For example, an evaluation metric may relate a hit ratio (HR) and a normalized discounted cumulative gain (NDCG).

According to an embodiment of the present invention, the evaluation result providing unit 360 may provide an evaluation result for a course recommended by the recommendation unit 340 using a test data set and an evaluation meter.

4 is a diagram illustrating a method for recommending a course course customized for a user based on deep learning according to an embodiment of the present invention.

Referring to FIG. 4 , in the deep learning-based course recommendation method customized for a user according to an embodiment of the present invention, in step 401, a plurality of user behavior data and a plurality of course attribute data are collected.

That is, the deep learning-based user-customized course recommendation method includes a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data. Data may be collected, and a plurality of course attribute data including a plurality of course data and a plurality of major data may be collected.

In step 402, the deep learning-based course recommendation method customized for a user according to an embodiment of the present invention learns user behavior data to extract user behavior characteristics, learns course attribute data to extract course attribute characteristics, and extracts course attribute characteristics. Interaction characteristics can be extracted by learning the user behavior characteristics and course attribute characteristics together.

That is, the deep learning-based course recommendation method customized for a user according to an embodiment of the present invention includes a plurality of user data, a plurality of job data, and a plurality of certificates included in a plurality of user behavior data. Data and a plurality of language data are converted into first embedding vectors, the converted first embedding vectors are input to a user behavior network (UBN) layer, and the input first embedding vectors are learned. At least one user behavior characteristic may be extracted.

In addition, in the deep learning-based user-customized course recommendation method according to an embodiment of the present invention, a plurality of course data and a plurality of major data included in a plurality of course attribute data are used as second embedding vectors. At least one course attribute feature may be extracted by performing conversion, inputting the converted second embedding vectors to a course attribution network (CAN) layer, and learning the input second embedding vectors.

In addition, in the deep learning-based course recommendation method customized for a user according to an embodiment of the present invention, at least one user behavioral characteristic and at least one course attribute characteristic are input to a deep neural network (DNN) layer, and the input At least one interaction characteristic may be extracted by learning at least one user behavior characteristic and at least one input course attribute characteristic together.

In step 403, the deep learning-based course recommendation method customized for a user according to an embodiment of the present invention may recommend a user-customized course course based on the interaction characteristics extracted in step 402.

That is, the deep learning-based user-customized course recommendation method according to an embodiment of the present invention ranks at least one course course recommended based on interaction characteristics, lists the ranked at least one course course, A recommendation list including at least one listed course may be provided.

Therefore, the present invention can be applied to an online education platform to effectively recommend courses suitable for the user's knowledge structure.

In addition, in the case of a new user, the present invention provides a cold start problem in which suitable courses cannot be recommended due to lack of past information or data, and a first start problem in which recommendations cannot be provided until the user's preference is reflected. start) aims to solve the problem at the same time.

5A to 7B are diagrams illustrating evaluation results for results recommended by a system for recommending courses customized to a user based on deep learning according to an embodiment of the present invention.

5A and 5B show the number of user behavioral characteristics, the number of course attribute characteristics, and the number of interaction characteristics corresponding to predictive factors in the deep learning-based user-customized course recommendation system according to an embodiment of the present invention. The results according to the change will be compared with a course recommendation system based on a recommendation algorithm according to the prior art.

The graph 500 of FIG. 5A may represent a hit ratio (HR), and the graph 510 of FIG. 5B may represent a normalized discounted cumulative gain (NDCG).

The graph 500 compares the first prior art 501, the second prior art 502, and the third prior art 503 with the present invention 504, and the graph 510 shows the first prior art ( 511), the second prior art 512 and the third prior art 513 and the present invention 514 will be compared and described.

The first prior art 501 and the first prior art 511 may represent an item-based collaborative filtering (itemKNN) recommendation algorithm, and the second prior art 502 and the second prior art 512 may represent an MF (matrix factorization) recommendation algorithm, and the third prior art 503 and the third prior art 513 may indicate a neural collaborative filtering (NCF) recommendation algorithm.

According to the graphs 500 and 510, when the predictive factor increases, the representation of the model improves, but at the same time, more noise is involved and the performance of the model deteriorates due to overfitting.

As a result of measuring the model performance according to the change in the predictive factor, it can be confirmed that the best performance is shown when the value is 32.

6A and 6B illustrate a result according to a change in the number of recommendations in a deep learning-based user-customized course recommendation system according to an embodiment of the present invention, compared with a course recommendation system based on a recommendation algorithm according to the prior art. Here, the change in the number of recommendations is related to the size of the recommendation list, and the size of the recommendation list increases as the number of recommendations increases, and the size of the recommendation list decreases as the number of recommendations decreases.

A graph 600 of FIG. 6A may represent a hit ratio (HR), and a graph 610 of FIG. 6B may represent a normalized discounted cumulative gain (NDCG).

The graph 600 compares the first prior art 601, the second prior art 602, and the third prior art 603 with the present invention 604, and the graph 610 shows the first prior art ( 611), the second prior art 612 and the third prior art 613 and the present invention 614 will be compared and described.

The first prior art 601 and the first prior art 611 may represent an item-based collaborative filtering (itemKNN) recommendation algorithm, and the second prior art 602 and the second prior art 612 may represent an MF (matrix factorization) recommendation algorithm, and the third prior art 603 and the third prior art 613 may indicate a neural collaborative filtering (NCF) recommendation algorithm.

According to the graphs 600 and 610, when the number of recommendation lists corresponding to the number of recommendations increases, it can be confirmed that the present invention provides improved performance compared to the prior art.

The traditional neighbor-based prior art 1 has lower performance than the model-based approach, which confirms the need for a model-based approach rather than a memory-based approach in personalized recommendation.

7A and 7B illustrate a result of a change in negative sampling in a deep learning-based user-customized course recommendation system according to an embodiment of the present invention, compared with a course recommendation system based on a recommendation algorithm according to the prior art.

A graph 700 of FIG. 7A may represent a hit ratio (HR), and a graph 710 of FIG. 7B may represent a normalized discounted cumulative gain (NDCG).

The graph 700 compares the first prior art 701 and the second prior art 702 with the present invention 703, and the graph 710 shows the first prior art 711 and the second prior art ( 712) and the present invention (713) will be compared and described.

The first prior art 701 and the first prior art 711 may indicate a matrix factorization (MF) recommendation algorithm, and the second prior art 702 and the second prior art 712 may represent neural collaborative filtering (NCF) It can represent a recommendation algorithm.

According to the graphs 700 and 710, it can be confirmed that the advantage of the point wise function compared to the pair wise function lies in flexible sampling for the negative instance.

For example, it can be confirmed that a pair wise function can be paired with only one negative case, and a point wise function can set a sampling rate suitable for a model.

Checking the performance for the change in the negative sampling rate in the graphs 700 and 710, it can be seen that the optimal sampling rate is about 3 to 6.

That is, in the deep learning-based user-customized course recommendation system according to an embodiment of the present invention, in extracting at least one interaction feature by controlling the negative sampling rate of a point wise function, the deep learning model It has the advantage of reducing point wise errors.

Therefore, the present invention provides latent features learned in each of the user behavior network (UBN) layer, course attribution network (CAN) layer, and deep neural network (DNN) layer. It is possible to perform deep learning learning in a way that minimizes the pointwise error by setting each user's participation in each course as a target variable.

The devices described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. A processing device may run an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of software. For convenience of understanding, there are cases in which one processing device is used, but those skilled in the art will understand that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it can include. For example, a processing device may include a plurality of processors or a processor and a controller. Other processing configurations are also possible, such as parallel processors.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program commands recorded on the medium may be specially designed and configured for the embodiment or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Software may include a computer program, code, instructions, or a combination of one or more of the foregoing, which configures a processing device to operate as desired or processes independently or collectively. You can command the device. Software and/or data may be any tangible machine, component, physical device, virtual equipment, computer storage medium or device, intended to be interpreted by or provide instructions or data to a processing device. , or may be permanently or temporarily embodied in a transmitted signal wave. Software may be distributed on networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer readable media.

As described above, although the embodiments have been described with limited drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques may be performed in an order different from the method described, and/or components of the described system, structure, device, circuit, etc. may be combined or combined in a different form than the method described, or other components may be used. Or even if it is replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents of the claims are within the scope of the following claims.

100: Deep learning-based user-customized course recommendation system
110: data collection unit 120: deep learning model processing unit
130: recommendation unit

Claims (16)

a data collection unit that collects a plurality of user behavior data and a plurality of course attribute data;
At least one user behavior characteristic is extracted by learning the plurality of collected user behavior data, at least one course attribute characteristic is extracted by learning the plurality of collected course attribute data, and at least one user behavior characteristic is extracted. a deep learning model processor configured to extract at least one interaction characteristic related to whether or not a user has taken a plurality of subjects by learning the characteristic and the extracted at least one course attribute characteristic together; and
Characterized in that it comprises a recommendation unit for recommending at least one course course based on the at least one extracted interaction characteristic
Deep learning based course recommendation system customized for users. According to claim 1,
Characterized in that the data collection unit collects the plurality of user behavior data including a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data doing
Deep learning based course recommendation system customized for users. According to claim 1,
Characterized in that the data collection unit collects the plurality of course attribute data including a plurality of course data and a plurality of major data
Deep learning based course recommendation system customized for users. According to claim 1,
The deep learning model processing unit first embeds a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data included in the plurality of user behavior data. converting into vectors, inputting the converted first embedding vectors to a user behavior network (UBN) layer, learning the input first embedding vectors, and extracting the at least one user behavior characteristic. characterized
Deep learning based course recommendation system customized for users. According to claim 4,
The deep learning model processing unit converts a plurality of course data and a plurality of major data included in the plurality of course attribute data into second embedding vectors, and converts the converted second embedding vectors into course attributes. Characterized in that the at least one course attribute feature is extracted by inputting the input to a course attribution network (CAN) layer and learning the input second embedding vectors.
Deep learning based course recommendation system customized for users. According to claim 5,
The deep learning model processing unit inputs the extracted at least one user behavioral characteristic and the extracted at least one course attribute characteristic to a deep neural network (DNN) layer, and inputs the at least one inputted user behavioral characteristic. and extracting the at least one interaction feature by learning the input at least one course attribute feature together.
Deep learning based course recommendation system customized for users. According to claim 6,
The deep learning model processing unit uses the user behavior network (UBN) layer and the course attribution network (CAN) layer as a parallel sub-network, and the parallel sub-network and Characterized in that the deep neural network (DNN) layer is connected and the deep neural network (DNN) layer is used as a prediction layer.
Deep learning based course recommendation system customized for users. According to claim 1,
The deep learning model processing unit controls a negative sampling rate of a point wise function in learning the at least one extracted user behavioral characteristic and the at least one extracted course attribute characteristic together to obtain the at least one In extracting interaction features, it is characterized by reducing the point wise error of the deep learning model
Deep learning based course recommendation system customized for users. According to claim 1,
Characterized in that the recommendation unit recommends the at least one course course based on implicit feedback that indirectly reflects the extracted at least one interaction characteristic to the recommendation of the at least one course course
Deep learning based course recommendation system customized for users. According to claim 1,
Characterized in that the recommendation unit ranks the at least one enrolled subject, lists the ranked at least one enrolled subject, and provides a recommendation list including the listed at least one enrolled subject.
Deep learning based course recommendation system customized for users. According to claim 10,
Characterized in that the recommendation unit randomly selects courses that the user has not interacted with among the at least one enrolled courses and ranks the at least one enrolled course.
Deep learning based course recommendation system customized for users. collecting a plurality of user behavior data and a plurality of course attribute data by a data collection unit;
extracting at least one user behavioral characteristic by learning the collected plurality of user behavioral data in a deep learning model processing unit;
extracting at least one course attribute characteristic by learning the collected plurality of course attribute data in the deep learning model processing unit;
In the deep learning model processing unit, at least one interaction characteristic related to the user's participation in a plurality of courses is extracted by learning the extracted at least one user behavioral characteristic and the extracted at least one course attribute characteristic together. doing; and
In a recommendation unit, recommending at least one course course based on the at least one extracted interaction characteristic
A method for recommending customized courses based on deep learning. According to claim 12,
Collecting the plurality of user behavior data and the plurality of course attribute data,
collecting the plurality of user behavior data including a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data; and
Collecting the plurality of course attribute data including a plurality of course data and a plurality of major data
A method for recommending customized courses based on deep learning. According to claim 12,
The step of extracting at least one user behavioral characteristic by learning the collected plurality of user behavior data,
converting a plurality of user data, a plurality of job data, a plurality of certificate data, and a plurality of language data included in the plurality of user behavior data into first embedding vectors; and inputting the converted first embedding vectors to a user behavior network (UBN) layer, learning the input first embedding vectors, and extracting the at least one user behavior characteristic. doing
A method for recommending customized courses based on deep learning. According to claim 14,
The step of extracting at least one course attribute characteristic by learning the plurality of collected course attribute data,
Converting a plurality of course data and a plurality of major data included in the plurality of course attribute data into second embedding vectors, and converting the converted second embedding vectors into a course attribution network, CAN) layer, and extracting the at least one course attribute feature by learning the input second embedding vectors.
A method for recommending customized courses based on deep learning. According to claim 15,
The step of extracting at least one interaction characteristic related to the user's participation in a plurality of courses by learning the extracted at least one user behavioral characteristic and the extracted at least one course property characteristic together,
The extracted at least one user behavioral characteristic and the extracted at least one course attribute characteristic are input to a deep neural network (DNN) layer, and the inputted at least one user behavioral characteristic and the inputted at least one course attribute characteristic are inputted. And extracting the at least one interaction feature by learning the course attribute characteristics of
A method for recommending customized courses based on deep learning.

Images