WO2020204468A1 - 액티브 러닝 기법을 적용한 머신 러닝 프레임워크 운용 방법, 장치 및 컴퓨터 프로그램 - Google Patents
액티브 러닝 기법을 적용한 머신 러닝 프레임워크 운용 방법, 장치 및 컴퓨터 프로그램 Download PDFInfo
- Publication number
- WO2020204468A1 WO2020204468A1 PCT/KR2020/004137 KR2020004137W WO2020204468A1 WO 2020204468 A1 WO2020204468 A1 WO 2020204468A1 KR 2020004137 W KR2020004137 W KR 2020004137W WO 2020204468 A1 WO2020204468 A1 WO 2020204468A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- user
- model
- data
- data analysis
- analysis model
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000010801 machine learning Methods 0.000 title abstract description 10
- 238000004590 computer program Methods 0.000 title 1
- 238000007405 data analysis Methods 0.000 claims abstract description 86
- 238000004458 analytical method Methods 0.000 claims abstract description 75
- 239000013598 vector Substances 0.000 claims description 103
- 238000012549 training Methods 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 2
- 230000007423 decrease Effects 0.000 claims 1
- 238000003745 diagnosis Methods 0.000 description 8
- 239000011159 matrix material Substances 0.000 description 8
- 238000010586 diagram Methods 0.000 description 6
- 239000000284 extract Substances 0.000 description 6
- 230000008569 process Effects 0.000 description 5
- 238000007796 conventional method Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000004941 influx Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/20—Education
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
Definitions
- the present invention relates to a method of providing user-customized content using a data analysis framework. More specifically, the present invention relates to a method of operating an expert model that generates an analysis model for a problem and/or a user using a large amount of user content consumption result data, and selects necessary data to efficiently learn the analysis model. For.
- the present invention aims to solve the above problems. More specifically, the present invention relates to a method of operating an expert model for selecting data necessary to efficiently generate a user and/or a problem model.
- a method of analyzing a user comprises constructing a problem database including a plurality of problems, collecting result data of a user's solution to the problem, and collecting the solution result data.
- D step of updating the expert model by applying a reward set in a direction in which prediction accuracy of the data analysis model is improved to update information of the data analysis model.
- the present invention it is possible to operate a data selection model to efficiently increase the performance of the analysis model separately from the data analysis model in the field of machine learning. According to this, since the data selection model proposes data for learning the data analysis model, it is possible to reduce the computer resources required for the data analysis model training, efficiently acquire the reliability of the data analysis model, and solve the problem of data selection. It works.
- 1 is a diagram for explaining a problem of a data set for machine learning
- FIG. 2 is a flowchart illustrating a method of operating a training data analysis model and a data coaching model in a data analysis framework according to an embodiment of the present invention.
- 3 is a diagram for explaining a relationship between a problem understanding degree X and a probability P of matching a problem
- FIG. 4 is a diagram for explaining a method of ending a data proposal for learning a data analysis model according to an embodiment of the present invention
- this method has a problem in that the tag information depends on the subjectivity of the person. There was a problem that the reliability of the result data could not be high because the tag information generated mathematically without human subjectivity was not given according to the concept of the problem.
- the present invention provides a method of applying a data analysis framework for big data processing and machine learning to exclude human intervention in the learning data processing process, and analyzing users and/or problems through the data analysis framework. It aims to do.
- the user's content application result log is collected, a multidimensional space composed of the user and/or problem is formed, and the result data of the user's consumption of content such as problems, explanations, and lectures, correct and incorrect answer data for each problem, or view of problems
- a user and/or problem may be modeled by calculating a vector for each user and problem, and a user modeling vector and a problem modeling vector may be calculated.
- the user modeling vector may be interpreted as representing the characteristics of all the problems of an individual user as a vector value
- the problem modeling vector may be interpreted as expressing the characteristics of all users of the individual problem as a vector value.
- a method of calculating the user modeling vector and/or the problem modeling vector is not limited, and a conventional technique applied to a big data analysis framework used to calculate the user modeling vector and/or the problem modeling vector may be used.
- the user modeling vector and the problem modeling vectors may not be interpreted as being limited to what attributes or features are included.
- a user modeling vector may be expressed to express characteristics of an individual user among all users
- a problem modeling vector may be expressed to express characteristics of an individual problem among all problems.
- the user modeling vector may include information on a degree to which the user understands a certain concept, that is, a degree of understanding of the concept.
- the problem modeling vector may include information on what concepts the problem is composed of, that is, a conceptual diagram.
- a probability of a correct answer to a specific problem of a specific user may be estimated using the user modeling vector and the problem modeling vector.
- the problem vector can be extended to a problem-view vector, and the user modeling vector and the problem-view modeling The vector can be used to calculate the probability that a particular user will select a particular view of any problem.
- FIG. 1 is a diagram illustrating a problem of a data set applied to a conventional machine learning modeling.
- the problem model generated by learning based on the data set biased against the problem solving data for gerunds will dominate the inclusion of the gerund concept, not the whole concept constituting the English subject. In this case, it is difficult to evaluate the performance of the user/problem model highly. For example, there may be a large difference between a probability that a corresponding user will solve an infinitive problem calculated using the user model and a result of the corresponding user actually solving the same problem.
- each data constituting the entire data set is used as a machine learning input for generating an analysis model.
- the amount of data used to generate the analysis model is very large, and a problem arises that excessive resources are required for generating the data analysis model.
- a very large data set will be needed that contains enough data for other concepts.
- a very large data set needs to be collected and processed, so it takes a long time to learn and a large cost for operating the data analysis framework occurs.
- an object of the present invention is to provide a method of operating an expert model for coaching data necessary for learning the data analysis model separately from the data analysis model.
- the expert model may propose data necessary to update the data analysis model in a preset direction according to the state of the data analysis model at a corresponding time point. Furthermore, the data analysis model according to an embodiment of the present invention may learn result data of solving a problem having a modeling vector that is close to the proposed data of the expert model. In this case, since the data analysis model can learn based on the data most suitable for the state at a specific point in time in order to improve performance, it can process the minimum amount of data and quickly reach the required level of performance.
- the expert model may extract vectors having values for concepts other than the gerund and notify the data analysis model. Furthermore, the data analysis model selects problems with modeling vectors that are close to the vector and provides them to the user, and applies the resultant data to the user vector to reflect the understanding of the whole concept of the English subject in the user modeling vector. It can solve the problem of data bias.
- steps 210, 220, 225, 23, and 240 are for explaining a process of generating a user and/or problem analysis model by using data on a content consumption result in a data analysis framework according to an embodiment of the present invention.
- steps 260, 265, 270, 275, 280, and 285 describe the process of generating an expert model that recommends data necessary to efficiently generate the analysis model in a data analysis framework according to an embodiment of the present invention. For.
- the user and/or problem model and the expert model may be learned based on different data and may be referred to as software that perform separate functions. Connected, it can contribute to the performance improvement of the entire data analysis framework.
- step 210 of FIG. 2A all content and content consumption result data for all users are collected, and in step 220, an analysis model M for all users and/or content is used using the content consumption result data. Can be created.
- the data analysis server may configure a database for learning content such as text, image, audio, and/or video problems, commentary, and lecture, and collect data as a result of users' access to the content database. have.
- the data analysis server may collect problem solving result data, commentary inquiry data, or lecture video running data for all users. More specifically, the data analysis server builds a database for various problems on the market, provides a problem database to the user device, and collects the result of solving the problem by the user through the user device. Data can be collected.
- the data analysis server can organize the collected problem solving result data in the form of a list of users, problems, and results.
- Y (u, i) denotes the result of the user u solving the problem i, and may be assigned a value of 1 for the correct answer and 0 for the incorrect answer.
- the multiple-choice problem is composed of not only the fingerprint but also the viewing element. If only the correct or incorrect answer is reflected as the source of the analysis, if two students answer the same problem but choose a different option, it is equivalent to calculating the vector value of the two students. The effect of the problem is the same, so the effect of the problem on the analysis results may be diluted.
- the result of the student's solution is sufficiently reflected in the calculation of the vector value of the problem. It does not work, and it becomes practically diluted.
- the data analysis server may extend the collected problem solving result data by applying a viewing parameter selected by the user.
- the data analysis server may configure the collected solution result data in the form of a list of users, problems, and selection views.
- Y (u, i, j) denotes a result of the user u selecting the example j of the question i, and a value of 1 may be assigned when the user u selects it and 0 when the answer is an incorrect answer.
- the data analysis server configures a multidimensional space composed of a user and a problem, and assigns a value to the multidimensional space based on whether the user has a problem or is wrong,
- the modeling vector for can be calculated.
- the data analysis server configures a multidimensional space consisting of a user and a selection view of a problem, and assigns a value to the multidimensional space based on whether the user selects the view, and You can calculate the modeling vector for the problem view.
- a user and a problem are expressed as a modeling vector according to an embodiment of the present invention, whether a specific user will fit a specific problem or not, that is, a probability of a correct answer to a specific problem of a specific user can be mathematically calculated.
- the data analysis server may estimate a degree of understanding of a specific problem of a specific user using the user modeling vector and the problem modeling vector, and estimate a probability that a specific user will meet a specific problem using the degree of understanding.
- the data analysis system can mathematically connect the mutual relationship between the user and the problem through a low coefficient matrix if the user's concept understanding L and the problem concept inclusion R are estimated with sufficient reliability. .
- the total number of users to be analyzed is n and the total number of problems to be analyzed is m
- the number of unknown concepts constituting the subject is assumed to be r
- the understanding matrix L for each user's concept is n by It is defined as an r matrix
- the matrix R for the degree of inclusion for each concept of the problem can be defined as an m by r matrix.
- L is connected to the transposition matrix R T of R, the relationship between the user and the problem can be analyzed without separately defining the concept or the number of concepts.
- understanding and correct answer rate can be estimated by introducing methodologies used in psychology, cognitive science, and pedagogy.
- understanding and correct answer rate can be estimated by considering the M2PL (multidimensional two-parameter logistic) latent trait model devised by Reckase and McKinely.
- 3 is a two-dimensional graph of the result of an experiment using sufficiently large data on the degree of understanding of the problem X and the probability P of matching the problem.
- the X-axis indicates the degree of understanding and the Y-axis indicates the probability of a correct answer.
- a function ⁇ for estimating the probability P that the user will solve the problem can be derived as shown in the following equation.
- the probability P of the correct answer to the problem can be calculated by applying the degree of understanding of the problem X to the tn ⁇ .
- the present invention suffices if it is possible to calculate the probability of a correct answer to a user's problem by applying a conventional technique capable of estimating the relationship between understanding and correct answer rate in a reasonable manner, and the present invention is limited to the methodology for estimating the relationship between understanding and correct answer rate. It should be noted that it cannot be interpreted.
- the user modeling vector may be provided to mean the correct answer rate of a specific problem by using the relationship between the user modeling vector and the problem modeling vector.
- a rate of correct answers to a user's problem may be estimated by using a selection probability for each question. For example, if the first user has a probability of selecting an option for a specific question (0.1, 0.2, 0, 0.7), the user will select option 4 with a high probability, and when the correct answer to the corresponding question is 4, the first Users can expect a high probability of hitting the problem.
- the data analysis server configures a multidimensional space with the user and problem-view as variables, and assigns a value to the multi-dimensional space based on whether the user selects the problem-view, and We can compute vectors.
- the selectivity may be estimated by applying various algorithms to the user modeling vector and the problem-view modeling vector, and the algorithm for calculating the selectivity is not limited in interpreting the present invention. That is, by using the relationship between the user modeling vector and the problem-view modeling vector, the user modeling vector may be provided to mean a selection probability for a specific view of a specific problem.
- the user's problem-view selection rate can be estimated. (x is the problem-view vector, is the user vector)
- the data analysis server may estimate a correct answer rate of a problem using the user's view selection rate.
- a method of estimating the correct answer rate for the corresponding question may be considered by using a plurality of selection rates for the corresponding question.
- a method of comparing the choices of the correct answers to the choices of all the choices may be considered.
- the percentage of correct answers for the user's question will be calculated as 0.5 / (0.5+0.1+0.3+0.6).
- the user does not understand the problem by dividing it in units of views, but understands it in units of questions including the composition of the entire view and the intention to present the question, so the selection rate and the correct answer rate cannot be simply connected.
- the correct answer rate of a corresponding question can be estimated from the answer selection rate by averaging the selection rates of all the answers of a corresponding question and applying the averaged selection rates of the correct answers to the selection rates of all the answers.
- the service server may estimate the correct answer rate of a problem by using the user's problem-view selection probability, and through this, the user's understanding of a specific concept may be estimated.
- an expert model T for coaching the user and data necessary to efficiently update the problem analysis model M may be generated.
- the expert model T takes an action based on the state information of the analysis model M, the update information, and the data information that causes the update, and receives a reward for the action according to the change state of the model. It can be created through reinforcement learning that learns in the direction of maximizing the sum.
- the data analysis server may assign an initial value T int of the expert model T in an arbitrary form, and extract at least one or more arbitrary vector values for proposal to the analysis model M.
- the vector is used to collect data necessary to increase the reliability of the performance of the user vector calculated by the analysis model M, for example, the probability that a random user calculated according to the analysis model M will fit a random problem. Can mean a problem.
- the vector value extracted from the expert model T int can be proposed to the analysis model M.
- the analysis model M checks at least one problem having a modeling vector close to the vector value, (Step 225) provides the corresponding problems to the user (Step 230) and collects problem solving result data. , We will update to reflect this. (Step 240)
- the expert model status information of the updated analysis model based on the suggestions of T int can be used to train the professional model T int. More specifically, the expert model T compares and learns the prediction performance of the analysis model M before the update and the analysis model M after the update, and receives an evaluation and rewards for their proposal based on the changed value of the prediction performance of the analysis model. Can be obtained. (Step 270). And the expert model T can be updated in the direction of maximizing rewards. (Step 275)
- the reward according to the exemplary embodiment of the present invention may be interpreted as indicating a learning direction or orientation point of the analysis model M.
- the reward is set to increase the prediction accuracy of the analysis model M by updating the analysis model M in a direction in which the probability of correct answer or the probability of selecting an answer for a specific problem predicted by analysis model M coincides with the actual solution result. Can be.
- a modeling vector U A of user A created by applying data obtained by solving problems 1, 2, 3, 4, and 5 by user A may be considered.
- U A is a vector representing the probability of the correct answer to the entire problem of user A
- the data analysis model M is the prediction accuracy of U A , that is, the result of the actual solution by user A for each problem and the data analysis model M
- the expert model T should be updated to recommend data necessary for the data analysis model M to be updated in this direction.
- the expert model T may recommend a vector for a problem that requires solution result data to increase the prediction accuracy of U A at the time point.
- the data analysis server may extract problem 6 having a modeling vector close to the recommended vector value, provide problem 6 to user A, and collect data as a result of solving user A's 6 problem.
- the solution result data may include the view selected by the user A while solving the 6 problem, the correct answer view of the 6 problem, and information on the time point of the solution, and the data analysis model M will be updated by applying the solution result data.
- the expert model T is for ⁇ U A , U A ', and the modeling vector Q 6 of problem 6 from the data analysis model M. You will receive information.
- T will be trained to extract a vector different from Q 6 when the analytic model M is in the state of U A.
- user A has a user modeling vector U A formed based on the data obtained by solving problems 1, 2, 3, 4, and 5.
- the difference between the probability of the correct answer estimated by the vector Q 6 of the 6 problem and the result of the user A solving the 6 problem is based on the data obtained by the user A solving the 1, 2, 3, 4, 5, 6
- the expert model T will be updated by applying a negative reward to ( ⁇ U A , U A ', Q 6 ).
- the expert model T will be trained to extract a vector that is not similar to Q 6 in the direction of recommending 6 dissimilar problems in the state of similar data analysis model M.
- T will be trained to extract a vector similar to Q 6 when the analysis model M is in the state of U A.
- the user modeling vector U A and The difference between the probability of the correct answer estimated by the vector Q 6 of problem 6 and the result of solving the problem 6 by user A is U A '
- the difference between the probability of the correct answer estimated by Q 6 and the result of actually solving the 6 problem by user A is greater than the difference between the result of solving the 6 problem, it can be interpreted that the prediction accuracy of the analysis model M is improved by applying the data for solving the 6 problem.
- the expert model T will be updated by applying a positive reward to ( ⁇ U A , U A ', Q 6 ).
- the expert model T will be trained to extract a vector similar to Q 6 in the direction of recommending the problem similar to 6 in the state of the similar data analysis model M.
- the reward applied to the training of the expert model T can be set in a direction to increase the prediction accuracy of the data analysis model M, as described above, but according to another embodiment of the present invention, it is also set in a direction to narrow the prediction score variance range. I can.
- the expert model T will be formed in the direction of extracting data for learning in a direction in which prediction of the analysis model M is elaborated.
- step 275 the expert model T will be updated by learning the data ( ⁇ U A , U A ', Q) received from the analysis model M according to the reward.
- Step 280 is a step for learning the analysis model M and/or the expert model T to an optimized level.
- Expert model T will continue to propose data for learning analysis model M if the performance of the analysis model M formed at that time is insufficient, but if the performance of the analysis model M is sufficient, the data proposal ends, and the data analysis server It would be appropriate to analyze users and/or content with the analysis model M of.
- FIG. 4 is a diagram for describing a case in which the update of the analysis model M and/or the expert model T is ended.
- the first is when the analysis model M at that time can sufficiently diagnose the user and/or the problem. For example, if analytic model M can estimate the probability of correct answer to user A's entire problem with sufficient accuracy through user vector U A without additional learning of user A's problem-solving result data, or User A's external test score can be estimated. This can be determined by checking whether the accuracy of the estimated value calculated by the analysis model at the time point is greater than or equal to a threshold value. (450 in Fig. 4)
- the second is when the characteristics of the user and/or the problem can no longer be grasped even if additional learning of the result data is performed.
- this is a case where there is no effect of learning, and it is expected that no change in the analysis model M is expected even if data is additionally learned according to the recommendation of the expert model T.
- this may be the case when the accuracy of the estimated value calculated through the user vector U A is not changed and is maintained within a certain range despite the addition of user A's solution result data. (460 in Fig. 4)
- the data recommended by the expert model T is already reflected in the analysis model M.
- the recommendation problem calculated by the expert model T is any one of the first to twentieth problems Can be considered.
- the expert model T ends the data proposal, and the training of the expert model T and the analysis model M will also end.
- the expert model T may extract data necessary for learning the analysis model M at a corresponding time point and recommend it to the analysis model M.
- the state information, update information, and problem modeling vector information of the analysis model M acquired in the expert model T in step 245 are used for learning the expert model T as well as ( Step 275), there is a feature that is used as an input for determining the next proposed data in the updated expert model. (Step 285)
- the expert model T may propose the next data necessary to increase the performance of the analysis model M by referring to the state information of the analysis model changed according to the conventional proposal.
- the next vector value to be proposed to increase the performance of U A ' may be calculated.
- the vector refers to a problem for collecting necessary data to increase the performance of the user vector U A calculated by the analysis model M, for example, the reliability of the probability that the user A calculated according to the analysis model M will fit a random problem. can do.
- the analysis model M extracts the vector received from the expert model T and a problem vector having a degree of similarity in a preset range and provides it to the user, and learns the result data of solving the problem.
- the data analysis server randomly extracts at least one problem from the problem database and provides it to a new user, and applies the solution result data to determine the user modeling vector U int of the new user. You can set it up and point it out to Expert Model T.
- the initial modeling vector of the new user u new can be calculated.
- the expert model T may recommend at least one or more arbitrary problem vectors necessary for diagnosis of a new user.
- the data analysis server will provide the new user with problems following the recommendation of the expert model T as diagnostic problems.
- the analysis model M updates the user vector by applying the result data of the user's solution to the diagnosis problem, and notifies the expert model T with information about the updated user vector, the changed value of the user vector, and the diagnosis problem vector.
- the expert model T learns information by applying a positive reward.
- the expert model T learns information by applying a negative reward. Thereafter, the expert model T may determine whether the performance of the user model U is sufficient, and may recommend a problem vector necessary for improving the performance of U until the performance of the user model exceeds a preset range.
- FIG. 2A is a case in which the analysis model M and the expert model T are updated to reflect the data collection result while providing a recommendation problem to the user.
- a framework for operating an analysis model and a framework for learning an expert model may be implemented in logically and/or physically separated computing devices in order to recommend a problem to a user. More specifically, a system for recommending a problem to a user and a system for learning an expert model may be logically and physically separated and operated.
- FIG. 2B is a flow chart for explaining the embodiment of the present invention. In the description of FIG. 2B, descriptions of portions overlapping with those of FIG. 2A will be omitted.
- the framework for operating the expert model T may record a history of update information of the analysis model. That is, the history of the state information of the analysis model M, the update M'information, and the problem modeling vector information causing the update may be recorded. Furthermore, unlike FIG. 2A, as long as the expert model T is not updated and the termination condition is not satisfied (step 280), a problem vector may be proposed to the analysis model M by using the expert model T. (Step 265)
- the expert model T operation framework may update the expert model T by reflecting the update history information of the analysis model at an arbitrary point in time.
- Step 275 the reward for setting the update direction of the expert model T is applied (Step 270), which may be substantially the same as the embodiment of FIG. 2A.
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Tourism & Hospitality (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Marketing (AREA)
- Economics (AREA)
- General Engineering & Computer Science (AREA)
- General Business, Economics & Management (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Strategic Management (AREA)
- Primary Health Care (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Educational Technology (AREA)
- Educational Administration (AREA)
- Medical Informatics (AREA)
- Computational Linguistics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Electrically Operated Instructional Devices (AREA)
Abstract
Description
Claims (6)
- 데이터 분석 서버에서, 사용자를 분석하는 방법에 있어서,복수의 문제를 포함하는 문제 데이터베이스를 구성하고, 상기 문제에 대한 사용자의 풀이 결과 데이터를 수집하고, 상기 풀이 결과 데이터를 학습하여 상기 사용자를 모델링하기 위한 데이터 분석 모델을 생성하는 A 단계;상기 데이터 분석 모델과 독립적으로 동작하고, 상기 데이터 분석 모델과 상이한 데이터를 기반으로 학습되며, 임의의 시점의 상기 데이터 분석 모델의 성능을 높이기 위하여 상기 데이터 분석 모델에서 필요한 학습 데이터를 추천하는 전문가 모델을 생성하는 B 단계;상기 전문가 모델의 추천에 따라 상기 문제 데이터베이스에서 적어도 하나 이상의 문제를 추출하고, 추출한 문제에 대한 사용자의 풀이 결과 데이터를 이용하여 상기 데이터 분석 모델을 업데이트하는 C 단계; 및상기 데이터 분석 모델의 업데이트 정보에 상기 데이터 분석 모델의 예측 정확도가 향상되는 방향으로 설정된 리워드를 적용하여 상기 전문가 모델을 업데이트하는 D 단계를 포함하며상기 B 단계는,상기 데이터 분석 모델의 제 1 상태 정보, 제 2 상태 정보 및 상기 제 1 상태가 상기 제 2 상태로 변경된 원인이 되는 데이터 정보를 학습하여 상기 전문가 모델을 생성하는 단계를 포함하는 것을 특징으로 하는 사용자 분석 방법.
- 제 1항에 있어서,상기 A 단계는, 상기 문제에 대한 사용자 각각의 특성을 설명하는 사용자 모델링 벡터를 계산하고, 상기 사용자 모델링 벡터를 이용하여 사용자의 문제에 대한 정답 확률을 추정하는 단계를 포함하며,상기 D단계는, 사용자가 문제를 실제 풀이한 결과와 상기 사용자 모델링 벡터를 이용하여 추정한 상기 문제의 정답 확률의 차이인 사용자 모델링 벡터의 예측 성능을 높이도록 설정된 리워드를 적용하여 상기 전문가 모델을 업데이트하는 단계를 포함하는 것을 특징으로 하는 사용자 분석 방법.
- 제 1항에 있어서,상기 A 단계는, 상기 문제에 대한 사용자 각각의 특성을 설명하는 사용자 모델링 벡터를 계산하고, 상기 사용자 모델링 벡터를 이용하여 상기 문제 데이터베이스를 이용하지 않고 출제된 외부 시험에 대한 사용자의 예측 점수를 추정하는 단계를 포함하며,상기 D단계는, 상기 데이터 분석 모델의 업데이트 정보에 상기 예측 점수의 표준편차가 작아지는 방향으로 설정된 리워드를 적용하여 상기 전문가 모델을 업데이트하는 단계를 포함하는 것을 특징으로 하는 사용자 분석 방법.
- 제 2항 또는 3항에 있어서,상기 C 단계는,상기 사용자 모델링 벡터의 예측 성능의 변경율이 미리 설정된 값 이내인 경우, 상기 데이터 분석 모델에 대한 추가 학습의 효과가 없는 것으로 판단하여 상기 전문가 모델의 추천을 종료하는 단계를 포함하는 것을 특징으로 하는 사용자 분석 방법.
- 제 2항 또는 3항에 있어서,상기 C 단계는,상기 사용자 모델링 벡터의 예측 성능이 미리 설정된 범위 이상인 경우, 상기 데이터 분석 모델이 추가 학습을 하지 않아도 상기 사용자 분석에 충분한 것으로 판단하여 상기 전문가 모델의 추천을 종료하는 단계를 포함하는 것을 특징으로 하는 사용자 분석 방법.
- 제 2항 또는 3항에 있어서,상기 C 단계는,상기 전문가 모델이 추천하는 문제의 풀이 결과 데이터가 상기 사용자 모델링 벡터에 이미 반영되어 있는 경우, 상기 전문가 모델의 추천을 종료하는 단계를 포함하는 것을 특징으로 하는 사용자 분석 방법.
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202080025985.7A CN113646788A (zh) | 2019-04-03 | 2020-03-26 | 利用主动学习技术操作机器学习框架的方法、设备和计算机程序 |
US16/964,823 US20210233191A1 (en) | 2019-04-03 | 2020-03-26 | Method, apparatus and computer program for operating a machine learning framework with active learning technique |
JP2020540290A JP7054558B2 (ja) | 2019-04-03 | 2020-03-26 | アクティブ学習手法を適用した機械学習フレームワークの運用方法、装置及びコンピュータプログラム |
CONC2021/0012810A CO2021012810A2 (es) | 2019-04-03 | 2021-09-28 | Método, aparato y programa de computadora para operar un marco de aprendizaje automático con técnica de aprendizaje activo |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
KR1020190039091A KR102096301B1 (ko) | 2019-04-03 | 2019-04-03 | 액티브 러닝 기법을 적용한 머신 러닝 프레임워크 운용 방법, 장치 및 컴퓨터 프로그램 |
KR10-2019-0039091 | 2019-04-03 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2020204468A1 true WO2020204468A1 (ko) | 2020-10-08 |
Family
ID=70281728
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/KR2020/004137 WO2020204468A1 (ko) | 2019-04-03 | 2020-03-26 | 액티브 러닝 기법을 적용한 머신 러닝 프레임워크 운용 방법, 장치 및 컴퓨터 프로그램 |
Country Status (6)
Country | Link |
---|---|
US (1) | US20210233191A1 (ko) |
JP (1) | JP7054558B2 (ko) |
KR (1) | KR102096301B1 (ko) |
CN (1) | CN113646788A (ko) |
CO (1) | CO2021012810A2 (ko) |
WO (1) | WO2020204468A1 (ko) |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11663494B2 (en) | 2019-12-05 | 2023-05-30 | Uchicago Argonne, Llc | Systems and methods for hierarchical multi-objective optimization |
US11651839B2 (en) | 2020-03-02 | 2023-05-16 | Uchicago Argonne, Llc | Systems and methods for generating phase diagrams for metastable material states |
US11710038B2 (en) * | 2020-04-13 | 2023-07-25 | Uchicago Argonne, Llc | Systems and methods for active learning from sparse training data |
KR20210148763A (ko) | 2020-06-01 | 2021-12-08 | 강릉원주대학교산학협력단 | 머신 러닝 모델링 자동화 방법 및 이를 이용한 머신 러닝 모델링 자동화 시스템 |
US11823044B2 (en) * | 2020-06-29 | 2023-11-21 | Paypal, Inc. | Query-based recommendation systems using machine learning-trained classifier |
KR102426812B1 (ko) * | 2020-08-25 | 2022-07-28 | 세종대학교산학협력단 | 강화 학습 기반의 상호작용 향상 방식 |
KR102475108B1 (ko) | 2020-11-02 | 2022-12-07 | 강릉원주대학교산학협력단 | 최적화된 하이퍼파라미터를 갖는 기계 학습 모델링 자동화 방법 및 이를 이용한 기계 학습 모델링 자동화 시스템 |
KR102274984B1 (ko) * | 2020-11-12 | 2021-07-09 | 태그하이브 주식회사 | 컨텐츠 스킬 라벨링 적합성 판단 방법 및 이를 실행하는 시스템 |
KR102398318B1 (ko) * | 2021-07-09 | 2022-05-16 | (주)뤼이드 | 학습 능력 평가 방법, 학습 능력 평가 장치, 및 학습 능력 평가 시스템 |
KR20230009816A (ko) * | 2021-07-09 | 2023-01-17 | (주)뤼이드 | 학습 능력 평가 방법, 학습 능력 평가 장치, 및 학습 능력 평가 시스템 |
CN114781194B (zh) * | 2022-06-20 | 2022-09-09 | 航天晨光股份有限公司 | 基于金属软管的数据库的构建方法 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150100530A1 (en) * | 2013-10-08 | 2015-04-09 | Google Inc. | Methods and apparatus for reinforcement learning |
US20170364831A1 (en) * | 2016-06-21 | 2017-12-21 | Sri International | Systems and methods for machine learning using a trusted model |
KR20180061999A (ko) * | 2016-11-30 | 2018-06-08 | 한국전자통신연구원 | 개인 맞춤형 학습 제공 장치 및 그 방법 |
KR101895961B1 (ko) * | 2017-10-31 | 2018-09-06 | (주)뤼이드 | 점수 추정 방법, 장치 및 컴퓨터 프로그램 |
KR20180127188A (ko) * | 2018-04-23 | 2018-11-28 | (주)뤼이드 | 개인 맞춤형 교육 컨텐츠를 제공하기 위한 기계학습 프레임워크 운용 방법, 장치 및 컴퓨터 프로그램 |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10290221B2 (en) * | 2012-04-27 | 2019-05-14 | Aptima, Inc. | Systems and methods to customize student instruction |
US20140272914A1 (en) * | 2013-03-15 | 2014-09-18 | William Marsh Rice University | Sparse Factor Analysis for Learning Analytics and Content Analytics |
US10706734B2 (en) * | 2013-12-06 | 2020-07-07 | Act, Inc. | Methods for improving test efficiency and accuracy in a computer adaptive test (CAT) |
JP6364037B2 (ja) | 2016-03-16 | 2018-07-25 | セコム株式会社 | 学習データ選択装置 |
KR101853091B1 (ko) | 2017-05-19 | 2018-04-27 | (주)뤼이드 | 기계학습이 적용된 사용자 답변 예측 프레임워크를 통한 개인 맞춤형 교육 컨텐츠 제공 방법, 장치 및 컴퓨터 프로그램 |
US11651702B2 (en) * | 2018-08-31 | 2023-05-16 | Pearson Education, Inc. | Systems and methods for prediction of student outcomes and proactive intervention |
KR102213479B1 (ko) * | 2019-02-28 | 2021-02-08 | (주)뤼이드 | 교육 컨텐츠를 제공하는 방법, 장치 및 컴퓨터 프로그램 |
-
2019
- 2019-04-03 KR KR1020190039091A patent/KR102096301B1/ko active IP Right Grant
-
2020
- 2020-03-26 WO PCT/KR2020/004137 patent/WO2020204468A1/ko active Application Filing
- 2020-03-26 CN CN202080025985.7A patent/CN113646788A/zh not_active Withdrawn
- 2020-03-26 US US16/964,823 patent/US20210233191A1/en not_active Abandoned
- 2020-03-26 JP JP2020540290A patent/JP7054558B2/ja active Active
-
2021
- 2021-09-28 CO CONC2021/0012810A patent/CO2021012810A2/es unknown
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150100530A1 (en) * | 2013-10-08 | 2015-04-09 | Google Inc. | Methods and apparatus for reinforcement learning |
US20170364831A1 (en) * | 2016-06-21 | 2017-12-21 | Sri International | Systems and methods for machine learning using a trusted model |
KR20180061999A (ko) * | 2016-11-30 | 2018-06-08 | 한국전자통신연구원 | 개인 맞춤형 학습 제공 장치 및 그 방법 |
KR101895961B1 (ko) * | 2017-10-31 | 2018-09-06 | (주)뤼이드 | 점수 추정 방법, 장치 및 컴퓨터 프로그램 |
KR20180127188A (ko) * | 2018-04-23 | 2018-11-28 | (주)뤼이드 | 개인 맞춤형 교육 컨텐츠를 제공하기 위한 기계학습 프레임워크 운용 방법, 장치 및 컴퓨터 프로그램 |
Also Published As
Publication number | Publication date |
---|---|
CN113646788A (zh) | 2021-11-12 |
JP7054558B2 (ja) | 2022-04-14 |
US20210233191A1 (en) | 2021-07-29 |
JP2021520529A (ja) | 2021-08-19 |
CO2021012810A2 (es) | 2021-11-19 |
KR102096301B1 (ko) | 2020-04-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2020204468A1 (ko) | 액티브 러닝 기법을 적용한 머신 러닝 프레임워크 운용 방법, 장치 및 컴퓨터 프로그램 | |
WO2018212394A1 (ko) | 기계 학습 프레임워크를 운용하는 방법, 장치 및 컴퓨터 프로그램 | |
WO2017146344A1 (ko) | 개인 맞춤형 교육 컨텐츠를 제공하는 방법, 장치 및 컴퓨터 프로그램 | |
Jonassen | Operationalizing mental models: strategies for assessing mental models to support meaningful learning and design-supportive learning environments. | |
WO2021251690A1 (ko) | 인공 지능 학습 기반의 학습 컨텐츠 추천 시스템 및 그것의 동작 방법 | |
US20210133598A1 (en) | Machine learning method, apparatus, and computer program for providing personalized educational content based on learning efficiency | |
Verdú et al. | A genetic fuzzy expert system for automatic question classification in a competitive learning environment | |
KR20220136952A (ko) | 복수의 시험 도메인에 적용되는 전이요소로 학습된 인공지능 모델을 통해 사용자의 실력을 평가하는 장치, 시스템 및 그것의 동작방법 | |
WO2018212396A1 (ko) | 데이터를 분석하는 방법, 장치 및 컴퓨터 프로그램 | |
CN110651294A (zh) | 估计考试分数的方法、设备及计算机程序 | |
WO2022059921A1 (ko) | 사용자의 학습효과를 반영하여 추천 문제를 결정하는 학습 컨텐츠 추천 장치, 시스템 및 그것의 동작 방법 | |
Giabbanelli et al. | Overcoming the PBL assessment challenge: Design and development of the incremental thesaurus for assessing causal maps (ITACM) | |
Samarakou et al. | An e-learning system for extracting text comprehension and learning style characteristics | |
Andaroon et al. | The effect of individual counseling on attitudes and decisional conflict in the choice of delivery among nulliparous women | |
Sedelmaier et al. | How can we find out what makes a good requirements engineer in the age of digitalization? | |
WO2022145829A1 (ko) | 잠재인자에 기반한 협업 필터링을 사용하여 사용자의 정답확률을 예측하는 학습 컨텐츠 추천 시스템 및 그것의 동작방법 | |
WO2018079968A1 (ko) | 개인 맞춤형 교육 컨텐츠를 제공하는 방법, 장치 및 컴퓨터 프로그램 | |
Mavrikis | Data-driven modelling of students' interactions in an ILE | |
KR102567043B1 (ko) | 인공지능 모델을 활용한 독서능력진단 기반 책 큐레이션 서비스 제공 방법, 장치 및 시스템 | |
WO2022245009A1 (ko) | 메타인지능력 평가방법 및 이를 위한 평가시스템 | |
Mahnane et al. | Automatic detection of learning styles based on dynamic Bayesian network in adaptive e-learning system | |
KR20190025871A (ko) | 사용자 맞춤형 컨텐츠를 제공하기 위한 방법, 장치 및 컴퓨터 프로그램 | |
Liu et al. | Student performance assessment using Bayesian network and web portfolios | |
Mueller et al. | Authoring guide for cognitive tutorials for artificial intelligence: Purposes and methods | |
Shen et al. | Constructing Behavioral Representation of Computational Thinking based on Event Graph: A new approach for learning analytics |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
ENP | Entry into the national phase |
Ref document number: 2020540290 Country of ref document: JP Kind code of ref document: A |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 20784793 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: P6001778/2021 Country of ref document: AE |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20784793 Country of ref document: EP Kind code of ref document: A1 |