CN111444432A - Domain-adaptive deep knowledge tracking and personalized exercise recommendation method - Google Patents

Domain-adaptive deep knowledge tracking and personalized exercise recommendation method Download PDF

Info

Publication number
CN111444432A
CN111444432A CN202010248722.4A CN202010248722A CN111444432A CN 111444432 A CN111444432 A CN 111444432A CN 202010248722 A CN202010248722 A CN 202010248722A CN 111444432 A CN111444432 A CN 111444432A
Authority
CN
China
Prior art keywords
data
knowledge tracking
knowledge
domain
source
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202010248722.4A
Other languages
Chinese (zh)
Inventor
刘淇
陈恩红
程松
苏喻
佟威
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology of China USTC
Original Assignee
University of Science and Technology of China USTC
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Science and Technology of China USTC filed Critical University of Science and Technology of China USTC
Priority to CN202010248722.4A priority Critical patent/CN111444432A/en
Publication of CN111444432A publication Critical patent/CN111444432A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • 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—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/20—Education
    • G06Q50/205—Education administration or guidance

Abstract

The invention discloses a field self-adaptive deep knowledge tracking and personalized exercise recommendation method, wherein a related knowledge tracking method comprises the following steps: acquiring historical answer record data and corresponding exercise data of students, and dividing the historical answer record data and the corresponding exercise data into source field data and target field data; for source field data, training a knowledge tracking model of a source field by using a deep learning method; and according to the knowledge tracking model of the source field, using a transfer learning method to realize the transfer from the source field to the target field, thereby realizing the knowledge tracking of the target field data. The method can transfer the knowledge tracking model trained in the source field to the target field by using deep learning and transfer learning methods, solves the problem that a large number of models need to be trained due to a large number of fields, and also solves the problem that a reliable knowledge tracking model cannot be obtained due to the fact that the training is carried out due to a small number of field data. And by matching with a corresponding personalized exercise recommendation method, a proper exercise can be recommended to the student, and a personalized test scheme is provided for the student.

Description

Domain-adaptive deep knowledge tracking and personalized exercise recommendation method
Technical Field
The invention relates to the field of transfer learning and the field of education data mining, in particular to a field-adaptive deep knowledge tracking and personalized exercise recommendation method.
Background
The current big data and data mining technology is rapidly developed, especially in the education field, online education systems are continuously appeared, and knowledge tracking is more and more important as a basic and necessary task in the education field. For example, the knowledge tracking result can be applied to personalized exercise recommendation, and a proper exercise is recommended to the student, so that the proficiency of the student on knowledge points or answering skills is enhanced, and a personalized test scheme is provided for the student.
The traditional knowledge tracking method only establishes a model for a specific school or subject and cannot be applied to other different schools and subjects. However, in reality, there are many schools, disciplines and levels, and it takes a lot of manpower and material resources to build different knowledge tracking models for different schools, disciplines and levels. Also, there are many schools that do not have enough data to train the model.
Therefore, in the prior art, due to the problems in the technical processing layer, a better knowledge tracking model cannot be obtained, the knowledge tracking effect is directly influenced, and the application of the knowledge tracking result is further influenced.
Disclosure of Invention
The invention aims to provide a field-adaptive deep knowledge tracking and personalized exercise recommendation method, which can well transfer a knowledge tracking model trained in one field (subject, school, grade) to another field (subject, school, grade), so that proper exercise can be recommended to students.
The purpose of the invention is realized by the following technical scheme:
a domain adaptive depth knowledge tracking method comprises the following steps:
acquiring a data set formed by historical answer recording data of students and corresponding exercise data, and dividing the data into source field data and target field data through a trained self-encoder;
for source field data, training a knowledge tracking model of a source field by using a deep learning method;
and according to the knowledge tracking model of the source field, migration from the source field to the target field is realized by using a migration learning method in combination with the target field data, so that the knowledge tracking of the target field data is realized.
According to the technical scheme provided by the invention, the method can transfer the knowledge tracking model trained in the source field to the target field by using deep learning and transfer learning methods, so that the problem that a large number of fields need to be trained is solved, and the problem that a reliable knowledge tracking model cannot be obtained by training due to small field data amount is also solved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on the drawings without creative efforts.
Fig. 1 is a flowchart of a domain-adaptive deep knowledge tracking and personalized problem recommendation method according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention are clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without making any creative effort, shall fall within the protection scope of the present invention.
The embodiment of the invention provides a depth knowledge tracking method of a field self-adaptation, as shown in figure 1, comprising the following steps:
step 1, acquiring a data set formed by historical answer recording data of students and corresponding exercise data, and dividing the data into source field data and target field data through a trained self-encoder.
The historical answer record data of the students comprises the scoring condition of each exercise of the students; the problem data contains text data for each problem.
In the embodiment of the invention, for the acquired data, data cleaning is firstly carried out in a conventional mode, and data with incomplete information are filtered out, so that a data set is obtained.
Then, the data is expressed by using a unified mathematical form through a standardization unit; the historical answer record data of the students are expressed as:
Figure BDA0002434741940000021
wherein
Figure BDA0002434741940000022
As a question answering record, comprises the t-th question and the corresponding score rtT is 1, 2, T is the total number of exercises; the exercise data is recorded as X ═ X1，x2，...，xT}，xtText, x, representing the tth problemt＝(w1，w2，...，wL)，wlDenotes the first word in the problem text, l 1, 2.
Then, a self-encoder is pre-trained jointly by minimizing the reconstruction error, which is defined as:
Figure BDA0002434741940000031
wherein the content of the first and second substances,
Figure BDA0002434741940000032
denotes wlThe result of the reconstruction of (a) is,
Figure BDA0002434741940000033
representing the reconstruction result of x;
in order to select a source domain sample that is favorable for a target domain, an objective function is minimized to screen out a portion of data in a data set to form a source domain data set
Figure BDA00024347419400000318
The residual data is the target domain data set
Figure BDA00024347419400000319
The objective function is defined as:
Figure BDA0002434741940000034
wherein the content of the first and second substances,
Figure BDA0002434741940000035
and
Figure BDA0002434741940000036
data representing a source domain and a target domain respectively,
Figure BDA0002434741940000037
and
Figure BDA0002434741940000038
representing corresponding reconstruction data, nS、nTRespectively representing the data quantity of the source field and the target field, and t' respectively representing the exercise serial numbers of the source field and the target field;
Figure BDA0002434741940000039
is an indicator of each element thereof
Figure BDA00024347419400000310
Data representing source domains
Figure BDA00024347419400000311
Whether to select or not, 1 represents select, 0 represents not select;
Figure BDA00024347419400000312
and
Figure BDA00024347419400000313
all through the preceding formula
Figure BDA00024347419400000314
Calculating to obtain; pi'e、πd' is an encoder and a decoder in a self-encoder, and the implementation mode will be described later.
Figure BDA00024347419400000320
Representing regularization terms, in order to prevent uSBecomes zero, defined as:
Figure BDA00024347419400000315
wherein λ is a regularization coefficient;
Figure BDA00024347419400000316
the value of (a) is determined according to the reconstruction error:
Figure BDA00024347419400000317
and 2, for the source field data, training a knowledge tracking model of the source field by using a deep learning method.
The knowledge tracking model of the source field mainly comprises: an auto-encoder, two neural networks, a long short term memory network, and a linear layer (output layer).
In the embodiment of the invention, a deep learning method is used for training a knowledge tracking model in the source field, so that knowledge tracking is realized. And combining the text information of the exercises and introducing two education characteristics of guessing rate and error rate to perform knowledge tracking. Because the text of the problem contains a great deal of information that adequately reflects the characteristics of the problem. Therefore, in order to obtain a better knowledge tracking effect, each problem is represented by using the coding information of the problem in the knowledge tracking process.
The self-encoder of the step is similar to the self-encoder in the step 1, and the difference is that the self-encoder of the step is obtained by training in an unsupervised mode on the exercise data, and then the self-encoder is used for representing the corresponding exercise on the encoding result of the exercise data; the self-encoder (the self-encoder of step 1 is also of the form:
encoder：q＝πe(x)
Figure BDA00024347419400000411
wherein, pie、πdRespectively representing an encoder and a decoder in a self-encoder; q represents the coding information of the problem, x represents the text of the corresponding problem,
Figure BDA0002434741940000041
representing problem text obtained by the problem text reconstructed by the self-encoder; by passing
Figure BDA0002434741940000042
And x to train the self-encoder.
After the self-encoder training is finished, the text x of the input t-th track exercise is inputtObtaining corresponding coding information qt。
In the embodiment of the invention, the coder pie(and the encoder pi mentioned in step 1)e') is implemented using a bidirectional L STM (long short term memory network) model, the bidirectional L STM model is expressed as:
Figure BDA0002434741940000043
Figure BDA0002434741940000044
wherein the content of the first and second substances,
Figure BDA0002434741940000045
and
Figure BDA0002434741940000046
is the output, w, of the L STM model in both forward and reverse directionslThe word representing the ith entry is,
Figure BDA0002434741940000047
and
Figure BDA0002434741940000048
the parameters to be learned of the positive L STM model and the negative L STM model are respectively;
the output of the L STM model in both the positive and negative directions is spliced into a vector:
Figure BDA0002434741940000049
obtaining problem representation q by pooling operationst，qtThe acquisition method of each element in (1) is expressed as:
qti＝max(η1i，η2i，...，ηLi)
in the above formula, i represents the ith dimension, and the principle is that a problem has L words, each word ηlIs an n-dimensional vector, this problem qtIs also an n-dimensional vector, and the calculation mode is as follows: q. q.stThe ith dimension is the maximum value of the values in the corresponding dimensions of all L words.
Exercise score r t1 represents a response pair, and 0 represents a wrong response; general pair problem characterization qtSplicing a zero vector:
Figure BDA00024347419400000410
wherein the content of the first and second substances,
Figure BDA00024347419400000412
is a stitching operation, stitching two vectors, where 0 ═ 0, (0, 0., 0) is a zero vector, and all elements are zero.
For decoder pid(and the decoder pi mentioned in step 1)d'), for simplicity, implemented using only the L STM model, shown as:
Figure BDA0002434741940000051
Figure BDA0002434741940000052
wherein the content of the first and second substances,
Figure BDA0002434741940000053
the output of the STM model is represented L,
Figure BDA0002434741940000054
Wdecand bdecAre parameters of the L STM model.
Using a long-short term memory network, inputting an answer record at every moment
Figure BDA0002434741940000059
Obtaining the state h of the corresponding timet：
Figure BDA00024347419400000510
Figure BDA00024347419400000511
Figure BDA00024347419400000512
Figure BDA00024347419400000513
ht＝ottanh(ct)
Wherein it，ft，ct，otRespectively representing an input gate, a forgetting gate, a memory unit and an output gate in the recurrent neural network of the long-short term memory network, W*，b*Representing the weight and bias terms in the corresponding gate, where the previous W in each equation*Show the record of answering
Figure 1
The weight of (a), the latter W*Representing the weight of the state at the last moment.
For modeling of the failure rate and the guess rate, learning is performed from the problem text. Also for simplicity, two characteristics of guess rate and error rate are introduced for knowledge tracking, which are described as:
st＝S(qt)
gt＝G(qt)
wherein S and G represent two neural networks of learning miss rate and guess rate, respectively.
And then combining the error rate and the guessing rate to obtain the knowledge state of the student:
Figure BDA0002434741940000055
wherein, the symbol
Figure BDA00024347419400000515
Representing a vector by element multiplication;
based on the knowledge state of the students, the performance of the corresponding students on the exercises in the future is predicted through a linear layer:
Figure BDA0002434741940000056
yt＝sigmoid(Wout·αt+bout)
wherein, αtRepresents the output of the adaptation function, ΘadpA parameter representing an adaptation function; y istRepresenting the prediction of the sigmoid function output, Wout、boutRespectively representing the weight and the bias of the sigmoid function.
An objective function of a knowledge tracking model in the training source field is defined as:
Figure BDA0002434741940000057
wherein n represents the number of students,
Figure BDA0002434741940000058
score of student i on the t-th problem predicted for the model, function (-) representing problem
Figure BDA0002434741940000061
The conversion is to the index value t, l (x, y) is a function, the function output is 1 if x and y are the same, otherwise the output is 0.
3. And according to the knowledge tracking model of the source field, migration from the source field to the target field is realized by using a migration learning method in combination with the target field data, so that the knowledge tracking of the target field data is realized.
In the embodiment of the invention, a two-step strategy is used, wherein in the first step, based on a trained knowledge tracking model in the source field, the distance between knowledge state distributions is reduced by utilizing the Maximum Mean Difference (MMD) to update a non-output layer in the knowledge tracking model in the source field; and the second step is that on the basis of finishing the first step, the cognitive level of the student is calculated, and the output dimensionality of different fields is adapted through a fine adjustment technology, so that the output layer in the knowledge tracking model of the source field is updated.
The preferred embodiment of this step is as follows: 1) the distance between knowledge state domain distributions is narrowed according to a Maximum Mean Difference (MMD) distance metric.
In the embodiment of the invention, the target field data is input into the trained knowledge tracking model of the source field to obtain the knowledge state of the student, and then the maximum mean difference distance is calculated.
The maximum mean difference distance metric is formally defined as:
Figure BDA0002434741940000062
wherein the content of the first and second substances,
Figure BDA0002434741940000063
and
Figure BDA0002434741940000064
respectively representing knowledge states of the students calculated through the source domain data and the target domain data,
Figure BDA0002434741940000065
the method is obtained by inputting target field data into a trained knowledge tracking model of a source field, nS、nTRespectively representing the data volume of the source field and the target field; function(s)
Figure BDA0002434741940000067
Is the mapping from the original feature space to the regenerated kernel hilbert space; the difference between the two domains can be reduced by minimizing the MMD distance.
The overall objective function of the domain adaptive knowledge pursuit is then expressed as:
Figure BDA0002434741940000066
where γ is a regularization coefficient representing the degree of importance of the maximum mean difference distance metric.
This section will update the autoencoder, two neural networks and long-short term memory networks in the knowledge tracking model of the source domain.
2) And enabling the model outputs in different fields to be consistent through a fine adjustment technology.
In the embodiment of the invention, an output layer of the knowledge tracking model trained by the source domain data set is removed, all parameters in front of the output layer are fixed, the output layer matched with the target domain is replaced, and the new output layer is trained by utilizing the target domain data set.
Those skilled in the art can understand that the data in the data set are all labeled data, the label of the target field data is not considered at the training stage, and after the model training is finished, the target field data set and the corresponding label can be used for model verification and testing.
According to the technical scheme of the embodiment of the invention, the knowledge tracking model trained in the source field can be migrated into the target field by using the deep learning and migration learning methods, so that the problem that a large number of models need to be trained due to a large number of fields is solved, and the problem that a reliable knowledge tracking model cannot be obtained due to the fact that the field data amount is small is solved.
Another embodiment of the invention further provides a personalized exercise recommendation method, which utilizes the field adaptive deep knowledge tracking method to realize knowledge tracking of students, generates exercise lists according to knowledge tracking results and recommends the exercise lists to corresponding students.
The personalized exercise recommendation method provided by the embodiment of the invention can recommend proper exercise to students, so that the proficiency of the students in knowledge points or answering skills is enhanced, and a personalized test scheme is provided for the students.
Through the above description of the embodiments, it is clear to those skilled in the art that the above embodiments can be implemented by software, and can also be implemented by software plus a necessary general hardware platform. With this understanding, the technical solutions of the embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a usb disk, a removable hard disk, etc.), and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods according to the embodiments of the present invention.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. A depth knowledge tracking method of domain adaptation is characterized by comprising the following steps:
acquiring a data set formed by historical answer recording data of students and corresponding exercise data, and dividing the data into source field data and target field data through a trained self-encoder;
for source field data, training a knowledge tracking model of a source field by using a deep learning method;
and according to the knowledge tracking model of the source field, migration from the source field to the target field is realized by using a migration learning method in combination with the target field data, so that the knowledge tracking of the target field data is realized.
2. The method of claim 1, wherein the partitioning of data into source domain and target domain data by a trained auto-encoder comprises:
the data is expressed by a standardized unit in a unified mathematical form; the historical answer record data of the students are expressed as:
Figure FDA0002434741930000011
wherein
Figure FDA0002434741930000012
As a question answering record, comprises the t-th question and the corresponding score rtT is 1, 2, T is the total number of exercises; the exercise data is recorded as X ═ X1，x2，...，xT}，xtText, x, representing the tth problemt＝(w1，w2，...，wL)，，wlThe first word in the problem text is represented, and l is 1, 2.., L represents the length of the text;
an auto-encoder is jointly pre-trained by minimizing the reconstruction error, which is defined as:
Figure FDA0002434741930000013
wherein the content of the first and second substances,
Figure FDA0002434741930000014
denotes wlThe result of the reconstruction of (a) is,
Figure FDA0002434741930000015
representing the reconstruction result of x;
minimizing an objective function such that a portion of the data is filtered out of the data set to form a source domain data set
Figure FDA00024347419300000121
The residual data is the target domain data set
Figure FDA0002434741930000017
The objective function is defined as:
Figure FDA0002434741930000018
wherein the content of the first and second substances,
Figure FDA0002434741930000019
and
Figure FDA00024347419300000110
data representing a source domain and a target domain respectively,
Figure FDA00024347419300000111
and
Figure FDA00024347419300000112
representing corresponding reconstruction data, ns、nTRespectively representing the data quantity of the source field and the target field, and t' respectively representing the exercise serial numbers of the source field and the target field;
Figure FDA00024347419300000113
is an indicator of each element thereof
Figure FDA00024347419300000114
Data representing source domains
Figure FDA00024347419300000115
Whether to select or not, 1 represents select, 0 represents not select;
Figure FDA00024347419300000116
and
Figure FDA00024347419300000117
all through the preceding formula
Figure FDA00024347419300000118
Calculating to obtain; pi'e、πd' is an encoder and a decoder in the self-encoder;
Figure FDA00024347419300000119
a regularization term is represented, defined as:
Figure FDA00024347419300000120
wherein λ is a regularization coefficient;
Figure FDA0002434741930000021
value ofIs determined according to the reconstruction error:
Figure FDA0002434741930000022
3. the domain adaptive depth knowledge tracking method according to claim 1,
the knowledge tracking model of the source domain comprises: a self-encoder, two neural networks, a long-short term memory network, and a linear layer;
the self-encoder is obtained by training in an unsupervised mode on the exercise data, and the corresponding exercise is represented by using the encoding result of the self-encoder on the exercise data; after the self-encoder training is finished, the text x of the input t-th track exercise is inputtObtaining corresponding coding information qtCorresponding exercise score rt1 represents a response pair, and 0 represents a wrong response; general pair problem characterization qtSplicing a zero vector:
Figure FDA0002434741930000023
wherein the content of the first and second substances,
Figure FDA0002434741930000024
a splicing operation is performed, two vectors are spliced, 0 ═ 0, (0, 0., 0) is a zero vector, and all elements are zero;
using a long-short term memory network, inputting an answer record at every moment
Figure FDA0002434741930000025
Obtaining the state h of the corresponding timet：
Figure FDA0002434741930000026
Figure FDA0002434741930000027
Figure FDA0002434741930000028
Figure FDA0002434741930000029
ht＝ottanh(ct)
Wherein i*，f*，c*，o*Respectively representing an input gate, a forgetting gate, a memory unit and an output gate in the recurrent neural network, W*，b*Representing the weight and bias terms in the corresponding gate;
at the same time, combine qtAnd two characteristics of guess rate and error rate are introduced to carry out knowledge tracking, and the description is as follows:
st＝S(qt)
gt＝G(qt)
wherein S and G respectively represent two neural networks of a learning error rate and a guess rate;
and then combining the error rate and the guessing rate to obtain the knowledge state of the student:
Figure FDA00024347419300000210
wherein, the symbol
Figure FDA00024347419300000211
Representing a vector by element multiplication;
based on the knowledge state of the students, the performance of the corresponding students on the exercises in the future is predicted through a linear layer:
Figure FDA00024347419300000212
yt＝sigmoid(Wout·αt+bout)
wherein, αtRepresents the output of the adaptation function, ΘadpA parameter representing an adaptation function; y istRepresenting the prediction of the sigmoid function output, Wout、boutRespectively representing the weight and the bias of the sigmoid function;
an objective function of a knowledge tracking model in the training source field is defined as:
Figure FDA0002434741930000031
wherein n represents the number of students,
Figure FDA0002434741930000032
the score of student i on the tth problem predicted for the knowledge tracking model, function (-) represents the problem to be solved
Figure FDA0002434741930000033
The conversion is to the index value i, where l (x, y) is a function, where x is the same as y and the output of the function is 1, otherwise the output is 0.
4. The domain adaptive depth knowledge tracking method of claim 3, wherein the self-encoder is represented as:
encoder：q＝πe(x)
decoder：
Figure FDA0002434741930000034
wherein, pie、πdRespectively representing an encoder and a decoder in a self-encoder; q represents the coding information of the problem, x represents the text of the corresponding problem,
Figure FDA0002434741930000035
representing problem text obtained by the problem text reconstructed by the self-encoder; by passing
Figure FDA0002434741930000036
And x to train the self-encoder.
5. The domain-adaptive depth knowledge tracking method according to claim 4, wherein the coder pieThe implementation is realized by adopting a bidirectional L STM model, and the bidirectional L STM model is expressed as:
Figure FDA0002434741930000037
Figure FDA0002434741930000038
wherein the content of the first and second substances,
Figure FDA0002434741930000039
and
Figure FDA00024347419300000310
is the output, w, of the L STM model in both forward and reverse directionslThe word representing the ith entry is,
Figure FDA00024347419300000311
and
Figure FDA00024347419300000312
the parameters to be learned of the positive L STM model and the negative L STM model are respectively;
the output of the L STM model in both the positive and negative directions is spliced into a vector:
Figure FDA00024347419300000313
obtaining problem representation q by pooling operationst，qtThe acquisition method of each element in (1) is expressed as:
qti＝max(η1i，η2i，...，ηLi)
where i represents the ith dimension.
6. The domain-adaptive depth knowledge tracking method of claim 4, wherein the decoder pidThe implementation is realized by adopting L STM model, and is represented as follows:
Figure FDA00024347419300000314
Figure FDA0002434741930000041
wherein the content of the first and second substances,
Figure FDA0002434741930000042
the output of the STM model is represented L,
Figure FDA0002434741930000043
Wdecand bdecAre parameters of the L STM model.
7. The method of claim 3, wherein the using the transfer learning method to realize the transfer from the source domain to the target domain, so as to realize the knowledge tracking of the target domain data comprises:
based on a trained knowledge tracking model of the source field, reducing the distance between knowledge state distributions by using the maximum mean difference to update a non-output layer in the knowledge tracking model of the source field;
and adapting the output dimensions of different fields through a fine tuning technology so as to update the output layer in the knowledge tracking model of the source field.
8. The method of claim 7, wherein the updating the non-output layer in the knowledge tracking model of the source domain by reducing the distance between the knowledge state distributions by using the maximum mean difference based on the trained knowledge tracking model of the source domain comprises:
inputting the target field data into the trained knowledge tracking model of the source field to obtain the knowledge state of the student
Figure FDA0002434741930000044
And then calculating the maximum mean difference distance:
Figure FDA0002434741930000045
wherein the content of the first and second substances,
Figure FDA0002434741930000046
and
Figure FDA0002434741930000047
respectively representing the knowledge states of the students calculated by the source domain data and the target domain data, ns、nTRespectively representing the data volume of the source field and the target field; function(s)
Figure FDA0002434741930000048
Is the mapping from the original feature space to the regenerated kernel hilbert space;
the overall objective function of the domain adaptive knowledge pursuit is then expressed as:
Figure FDA0002434741930000049
wherein, gamma is a regularization coefficient, which represents the importance degree of the maximum mean difference distance measurement;
Figure FDA00024347419300000410
an objective function representing knowledge pursuit of the training source domain.
9. The method of claim 7, wherein the adapting the different domain output dimensions through the fine-tuning technique to update the output layers in the knowledge tracking model of the source domain comprises:
and removing a knowledge tracking output layer of the model trained by the source domain data set, fixing all parameters in front of the output layer, replacing the output layer matched with the target domain, and training the new output layer by using the target domain data set.
10. A personalized exercise recommendation method is characterized in that the field-adaptive deep knowledge tracking method of any one of claims 1 to 9 is used for realizing knowledge tracking of students, exercise lists are generated according to knowledge tracking results, and the exercise lists are recommended to the corresponding students.
CN202010248722.4A 2020-04-01 2020-04-01 Domain-adaptive deep knowledge tracking and personalized exercise recommendation method Pending CN111444432A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010248722.4A CN111444432A (en) 2020-04-01 2020-04-01 Domain-adaptive deep knowledge tracking and personalized exercise recommendation method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010248722.4A CN111444432A (en) 2020-04-01 2020-04-01 Domain-adaptive deep knowledge tracking and personalized exercise recommendation method

Publications (1)

Publication Number Publication Date
CN111444432A true CN111444432A (en) 2020-07-24

Family

ID=71652708

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010248722.4A Pending CN111444432A (en) 2020-04-01 2020-04-01 Domain-adaptive deep knowledge tracking and personalized exercise recommendation method

Country Status (1)

Country Link
CN (1) CN111444432A (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112015780A (en) * 2020-08-24 2020-12-01 上海松鼠课堂人工智能科技有限公司 Intelligent proposition analysis processing method and system based on deep learning
CN112182308A (en) * 2020-09-29 2021-01-05 华中师范大学 Multi-feature fusion depth knowledge tracking method and system based on multi-thermal coding
CN112232430A (en) * 2020-10-23 2021-01-15 浙江大华技术股份有限公司 Neural network model testing method and device, storage medium and electronic device
CN112949929A (en) * 2021-03-15 2021-06-11 华中师范大学 Knowledge tracking method and system based on collaborative embedded enhanced topic representation
CN112990464A (en) * 2021-03-12 2021-06-18 东北师范大学 Knowledge tracking method and system
CN113094351A (en) * 2021-03-26 2021-07-09 鹏城实验室 Production process knowledge mining and migrating method, intelligent terminal and storage medium
CN117131022A (en) * 2023-09-25 2023-11-28 广州东方电力有限公司 Heterogeneous data migration method of electric power information system
CN117743699A (en) * 2024-02-20 2024-03-22 山东省计算中心（国家超级计算济南中心） Problem recommendation method and system based on DKT and Topson sampling algorithm
CN117743699B (en) * 2024-02-20 2024-05-14 山东省计算中心（国家超级计算济南中心） Problem recommendation method and system based on DKT and Topson sampling algorithm

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180357538A1 (en) * 2017-06-09 2018-12-13 Korea Advanced Institute Of Science And Technology Electronic apparatus and method for generating trained model
CN109902298A (en) * 2019-02-13 2019-06-18 东北师范大学 Domain Modeling and know-how estimating and measuring method in a kind of adaptive and learning system
CN110428010A (en) * 2019-08-05 2019-11-08 中国科学技术大学 Knowledge method for tracing

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180357538A1 (en) * 2017-06-09 2018-12-13 Korea Advanced Institute Of Science And Technology Electronic apparatus and method for generating trained model
CN109902298A (en) * 2019-02-13 2019-06-18 东北师范大学 Domain Modeling and know-how estimating and measuring method in a kind of adaptive and learning system
CN110428010A (en) * 2019-08-05 2019-11-08 中国科学技术大学 Knowledge method for tracing

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
SONG CHENG, QI LIU, ENHONG CHEN: ""Domain Adaption for Knowledge Tracing"", 《ARXIV》 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112015780B (en) * 2020-08-24 2021-06-01 上海松鼠课堂人工智能科技有限公司 Intelligent proposition analysis processing method and system based on deep learning
CN112015780A (en) * 2020-08-24 2020-12-01 上海松鼠课堂人工智能科技有限公司 Intelligent proposition analysis processing method and system based on deep learning
CN112182308A (en) * 2020-09-29 2021-01-05 华中师范大学 Multi-feature fusion depth knowledge tracking method and system based on multi-thermal coding
CN112232430A (en) * 2020-10-23 2021-01-15 浙江大华技术股份有限公司 Neural network model testing method and device, storage medium and electronic device
CN112990464A (en) * 2021-03-12 2021-06-18 东北师范大学 Knowledge tracking method and system
CN112949929B (en) * 2021-03-15 2022-06-07 华中师范大学 Knowledge tracking method and system based on collaborative embedded enhanced topic representation
CN112949929A (en) * 2021-03-15 2021-06-11 华中师范大学 Knowledge tracking method and system based on collaborative embedded enhanced topic representation
CN113094351A (en) * 2021-03-26 2021-07-09 鹏城实验室 Production process knowledge mining and migrating method, intelligent terminal and storage medium
CN113094351B (en) * 2021-03-26 2024-02-06 鹏城实验室 Knowledge mining and migration method for industrial automatic production process, intelligent terminal and storage medium
CN117131022A (en) * 2023-09-25 2023-11-28 广州东方电力有限公司 Heterogeneous data migration method of electric power information system
CN117131022B (en) * 2023-09-25 2024-03-29 广州东方电力有限公司 Heterogeneous data migration method of electric power information system
CN117743699A (en) * 2024-02-20 2024-03-22 山东省计算中心（国家超级计算济南中心） Problem recommendation method and system based on DKT and Topson sampling algorithm
CN117743699B (en) * 2024-02-20 2024-05-14 山东省计算中心（国家超级计算济南中心） Problem recommendation method and system based on DKT and Topson sampling algorithm

Similar Documents

Publication Publication Date Title
CN111444432A (en) Domain-adaptive deep knowledge tracking and personalized exercise recommendation method
CN110264091B (en) Student Cognitive Diagnosis Method
Liu et al. Ekt: Exercise-aware knowledge tracing for student performance prediction
US20210256354A1 (en) Artificial intelligence learning-based user knowledge tracing system and operating method thereof
Zhao et al. A memory-augmented neural model for automated grading
CN111724083A (en) Training method and device for financial risk recognition model, computer equipment and medium
CN112508334B (en) Personalized paper grouping method and system integrating cognition characteristics and test question text information
CN111538868B (en) Knowledge tracking method and problem recommendation method
US20140272914A1 (en) Sparse Factor Analysis for Learning Analytics and Content Analytics
CN112257966B (en) Model processing method and device, electronic equipment and storage medium
Yeung et al. Incorporating features learned by an enhanced deep knowledge tracing model for stem/non-stem job prediction
CN111047482B (en) Knowledge tracking system and method based on hierarchical memory network
CN114254208A (en) Identification method of weak knowledge points and planning method and device of learning path
CN113793239B (en) Personalized knowledge tracking method and system integrating learning behavior characteristics
CN114299349B (en) Crowdsourcing image learning method based on multi-expert system and knowledge distillation
CN112184089B (en) Training method, device and equipment of test question difficulty prediction model and storage medium
CN114254127A (en) Student ability portrayal method and learning resource recommendation method and device
CN114429212A (en) Intelligent learning knowledge ability tracking method, electronic device and storage medium
CN114722182A (en) Knowledge graph-based online class recommendation method and system
CN111382619B (en) Picture recommendation model generation method, picture recommendation method, device, equipment and medium
CN113283488B (en) Learning behavior-based cognitive diagnosis method and system
Hagedoorn et al. Massive open online courses temporal profiling for dropout prediction
CN114298299A (en) Model training method, device, equipment and storage medium based on course learning
Peeperkorn et al. Can recurrent neural networks learn process model structure?
CN115510932A (en) Model training method and device, electronic equipment and storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20200724

RJ01 Rejection of invention patent application after publication