A student-personalized time interval-aware attention mechanism knowledge tracking method

Abstract

The present invention discloses a method for tracking knowledge of personalized time interval-aware attention mechanism for students, including the following steps: 1) collecting historical interaction data of students' learning and timestamps of completed learning, designing and implementing a relationship matrix of time intervals for students' learning; 2) vector encoding of exercise sequences, learning sequences, absolute positions and personalized time intervals; 3) establishing a time-aware attention mechanism to calculate the weight of the influence of historical completed exercises on subsequent new exercises, and summarizing the initial knowledge state of students at each node based on the weight; 4) using a fully connected layer to fuse the historical knowledge state of the students that is initially summarized, and tracking the knowledge state that keeps changing during the students' learning process. Compared with other technologies, the present invention effectively utilizes the different time intervals for students to practice each question, explores the different knowledge mastery of different students under the same learning sequence, and improves the accuracy of knowledge tracking.

Description

Technical Field

[0001] The present invention specifically relates to a student-personalized time interval-aware attention mechanism knowledge tracking method. Background Art

[0002] With the further development of educational informatization, online learning systems such as massive open online course (MOOC) platforms and online referee (OJ) systems have emerged in large numbers. Generally speaking, these platforms provide students with rich learning resources (such as courses, exercises, lectures) and provide them with an open environment to learn and practice knowledge at their own pace. The advantages of autonomy and convenience have attracted public attention and attention. In response to this problem, an effective solution is to provide personalized services in online learning systems to improve students' learning experience. Among them, one of the most basic tasks is knowledge tracking.

[0003] Knowledge tracking is to model the knowledge status of students based on historical data, which represents the degree of students' mastery of knowledge. Two important theories in educational research provide basic ideas for modeling students' knowledge ability. The learning curve theory believes that students can acquire knowledge through continuous practice or practice, and the forgetting curve theory believes that students' memory of the knowledge they have learned decreases, so their knowledge proficiency shows a downward curve. Knowledge tracking research based on attention mechanism predicts students' mastery of new exercises by calculating the similarity between new exercises and historical exercises in interaction data. However, the influence of learning curve and forgetting curve on knowledge tracking is ignored. There will be some identical learning sequence fragments in the learning sequences of different students. The existing methods will assume that students have the same mastery of these exercises or knowledge points, which is contrary to reality. For example, if students have the same learning sequence and different completion time intervals, then under the influence of educational theory, their mastery of these exercises is different. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to overcome the defects of the prior art and provide a method for tracking knowledge of a personalized time interval-aware attention mechanism for students, so as to obtain the influence weight of historical exercises on students' new exercises through their personalized learning time intervals, positions and learning sequence information, so as to achieve the purpose of improving the knowledge tracking effect.

[0005] Technical solution: The present invention provides a method for tracking knowledge of a personalized time interval perception attention mechanism for students, and the implementation steps are as follows, including the following steps:

[0006] (1) Collect and organize students’ historical interactive learning data and corresponding time series during the problem-solving process, and design and implement students’ personalized time interval relationship matrix;

[0007] (2) Vector representation of exercise sequences, learning sequences, absolute positions, and personalized time intervals;

[0008] (3) Establish a time-aware attention mechanism to calculate the influence weight of previously completed exercises on subsequent new exercises, and summarize the initial knowledge state of students at each node based on this weight;

[0009] (4) Use the fully connected layer to integrate the historical knowledge state of the students that is initially summarized, and track the changing knowledge state of the students during the learning process. i .

[0010] Furthermore, the step (1) comprises the following steps:

[0011] (1.1) Student s∈S, S is the set of students in the learning system, exercises e∈E, E is the set of exercises in the student system, and the learning sequence of student s is expressed as:

[0012] s={χ1,χ2,…χ n}={(e0,a0),(e1,a1),…(e n ,a n )}

[0013] Among them, χ t =(e t ,a t ) indicates whether student s answered exercise e correctly at time t t , a t ∈{0,1}, the learning sequence s is transformed into s={s1,s2…s n |s n =e n +a n ×E}; the time series of student s’s answers is expressed as t=(t1, t2,…t n ), n is the maximum value considered for sequence processing;

[0014] (1.2) The time interval between exercises i and j practiced by student s |t i -t j |, the time interval list of student s is denoted as R s , its minimum value The personalized time intervals for student u are:

[0015]

[0016] Calculate the time interval relationship between all exercises, the time interval relationship matrix M between student s's practice exercises s for:

[0017]

[0018] Assuming that precise relative time intervals are useless beyond a certain threshold, u Perform trimming so that the maximum interval time does not exceed k, and obtain the trimmed matrix M u clipped .

[0019] Furthermore, the step (2) comprises the following steps:

[0020] (2.1) Exercise: Sequence encoding matrix E∈R E×d , learn the sequence encoding matrix X∈R 2E×d ; Find and retrieve the top n exercise sequence encodings and learning sequence encodings and stack them together:

[0021]

[0022] (2.2) Define two variable position encoding matrices By searching, we can get the code and

[0023]

[0024] (2.3) Define two variable time interval codes By searching M u clipped After that, get the code and

[0025]

[0026] Furthermore, the step (3) comprises the following steps:

[0027] (3.1) In learning interaction data, we consider both the position and time interval and calculate the similarity between the current exercise and the historical exercises:

[0028]

[0029] Among them, W Q ∈R d×d , W K ∈R d×d is the mapping between the query and the value of the attention mechanism, When the dimension is too large, it is used to avoid the inner product being too large;

[0030] (3.2) Use the softmax function for weighted calculation to obtain the weight influence factor a ij :

[0031]

[0032] (3.3) By calculating the linear weighted sum of the learning interaction and the position and time interval encoding, the weighting coefficient is obtained to obtain the final weight of the impact of the historical exercises on the new prediction exercises:

[0033]

[0034] Among them, W V ∈R d×d is the input projection of the value;

[0035] (3.4) The final weights are processed using a feedforward network, and the RuLe() activation function is used to make the linear model nonlinear, as follows:

[0036] F=FFN(z i )=ReLu(z i W (1) +b (1) )W (2) +b (2)

[0037] Among them, W (1) ,W (2) ∈R d×d , b (1) ,b (2) ∈R d , the process is unified into a function FFN(). Further, the knowledge state p i for:

[0038] p i =Sigmoid(Fw+b)

[0039] Among them, F i is the i-th row of F, w represents the weight, b represents the bias, and the activation function

[0040] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: the present invention effectively utilizes the different time intervals for students to practice each question, explores the different knowledge mastery of different students under the same learning sequence; improves the accuracy of knowledge tracking, and lays a good foundation for subsequent adaptive research; at the same time, by providing each student with the best and personalized learning experience, it brings huge educational benefits to online learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0043] The present invention provides a method for tracking students' personalized knowledge of time interval perception and attention mechanism, such as Figure 1 As shown, the following steps are included:

[0044] (1) Collect and organize the historical interactive learning data and corresponding time series of students during the problem solving process, and design and implement the students' personalized time interval relationship matrix. The specific implementation process is as follows:

[0045] (1.1) Students s∈S, S is the set of students in the learning system, exercises e∈E, E is the set of exercises in the student system, and the learning sequence of student s is s={χ1,χ2,…χ n}={(e0,a0),(e1,a1),…(e n ,a n )}, t =(e t ,a t ) indicates whether student s answered exercise e correctly at time t t , a t ∈{0,1}, the learning sequence s is transformed into s={s1,s2…s n |s n =e n +a n ×E}. The time series of student s’s answers is expressed as t=(t1, t2,…t n ), n is the maximum value considered for sequence processing.

[0046] (1.2) The time interval between exercises i and j practiced by student s |t i -t j |, the time interval list of student s is denoted as R s , its minimum value The personalized time interval of student u can be expressed as:

[0047]

[0048] Using formula (1), we can calculate the time interval relationship between all exercises. The time interval relationship matrix M between the exercises of student s is: s It can be expressed as formula (2):

[0049]

[0050] Assuming that precise relative time intervals are useless beyond a certain threshold, s Perform trimming so that the maximum interval time does not exceed k, and obtain the trimmed matrix M u clipped .

[0051] (2) Perform vector representation on the exercise sequence, learning sequence, absolute position and personalized time interval. The specific implementation is as follows:

[0052] (2.1) Exercise: Sequence encoding matrix E∈R E×d , learn the sequence encoding matrix M∈R |2E|×d .

[0053] Find and retrieve the first n exercise sequence codes and learning sequence codes and stack them together to get and Respectively expressed as:

[0054]

[0055] (2.2) Define two variable position encoding matrices By searching, we can get the code and Respectively expressed as:

[0056]

[0057] (2.3) Define two variable time interval codes By searching M u clipped After that, get the code and Respectively expressed as:

[0058]

[0059] (3) Establish a time-aware attention mechanism to calculate the weight of the impact of previously completed exercises on subsequent new exercises, and summarize the initial knowledge state of students at each node based on this weight. The specific implementation process is as follows:

[0060] (3.1) In the learning interaction data, the position and time interval are considered at the same time to calculate the similarity between the current exercise and the historical exercises. The specific calculation method is:

[0061]

[0062] Among them, W Q ∈R d×d , W K ∈R d×d is the mapping between the query and the value of the attention mechanism, When the dimension is too large, it is used to avoid the inner product being too large.

[0063] (3.2) Use the softmax function for weighted calculation to obtain the weight influence factor a ij :

[0064]

[0065] (3.3) By calculating the linear weighted sum of the learned interaction and the position and time interval encoding, the weighting coefficient a ij The final weight of the impact of historical exercises on new predicted exercises is obtained by formula (7):

[0066]

[0067] Among them, W V ∈R d×d is the input projection of values.

[0068] (3.4) The final weights are processed using a feedforward network, and the RuLe() activation function is used to make the linear model nonlinear, as follows:

[0069] F=FFN(z i )=ReLu(z i W (1) +b (1) )W (2) +b (2) (9)

[0070] Among them, W (1) ,W (2) ∈R d×d , b (1) ,b (2) ∈R d , the process is unified into a function FFN().

[0071] (4) Use the fully connected layer to integrate the historical knowledge state of the students that is initially summarized, and track the changing knowledge state of the students during the learning process. i :

[0072] p i =Sigmoid(F i w+b) (10)

[0073] Among them, F i is the i-th row of F, w represents the weight, b represents the bias, and the activation function

[0074] The present invention effectively utilizes the different time intervals for students to practice each question, and explores the different knowledge mastery of different students under the same learning sequence; the present invention is conducive to improving the accuracy of knowledge tracking and laying a good foundation for subsequent personalized research. At the same time, by providing each student with the best and personalized learning experience, it brings huge educational benefits to online learning.

Claims

1. A student-personalized time interval-aware attention mechanism knowledge tracking method, characterized in that: The following steps are involved: (1) Collect and organize students’ historical interactive learning data and corresponding time series during the problem-solving process, and design and implement students’ personalized time interval relationship matrix; (2) Vector representation of exercise sequences, learning sequences, absolute positions, and personalized time intervals; (3) Establish a time-aware attention mechanism to calculate the influence weight of previously completed exercises on subsequent new exercises, and summarize the initial knowledge state of students at each node based on this weight; (4) Use the fully connected layer to integrate the historical knowledge state of the students that is initially summarized, and track the changing knowledge state of the students during the learning process. i ; The step (1) comprises the following steps: (1.1) Student s∈S, S is the set of students in the learning system, exercises e∈E, E is the set of exercises in the student system, and the learning sequence of student s is expressed as: s={χ1,χ2,…χ n }={(e0,a0),(e1,a1),…(e n ,a n )} Among them, χ t =(e t ,a t ) indicates whether student s answered exercise e correctly at time t t , a t ∈{0,1}, the learning sequence s is transformed into s={s1,s2…s n |s n =e n +a n ×E}; the time series of student s’s answers is expressed as t=(t1, t2,…t n ), n is the maximum value considered for sequence processing; (1.2) The time interval between exercises i and j practiced by student s |t i -t j |, the time interval list of student s is denoted as R s , its minimum value The personalized time intervals for student u are: Calculate the time interval relationship between all exercises, the time interval relationship matrix M between student s's practice exercises s for: Assuming that precise relative time intervals are useless beyond a certain threshold, u Perform trimming so that the maximum interval time does not exceed k, and obtain the trimmed matrix M u clipped ; The step (2) comprises the following steps: (2.1) Exercise: Sequence encoding matrix E∈R |E|×d , learn the sequence encoding matrix X∈R |2E|×d ; Find and retrieve the top n exercise sequence encodings and learning sequence encodings and stack them together: (2.2) Define two variable position encoding matrices By searching, we can get the code and (2.3) Define two variable time interval codes By searching M u clipped After that, get the code and 2. The method for tracking knowledge of a personalized time interval-aware attention mechanism for students according to claim 1, characterized in that: The step (3) comprises the following steps: (3.1) In learning interaction data, we consider both the position and time interval and calculate the similarity between the current exercise and the historical exercises: Among them, W Q ∈R d×d , W K ∈R d×d is the mapping between the query and the value of the attention mechanism, When the dimension is too large, it is used to avoid the inner product being too large; (3.2) Use the softmax function for weighted calculation to obtain the weight influence factor a ij : (3.3) By calculating the linear weighted sum of the learning interaction and the position and time interval encoding, the weighting coefficient is obtained to obtain the final weight of the impact of the historical exercises on the new predicted exercises: Among them, W V ∈R d×d is the input projection of the value; (3.4) The final weights are processed using a feedforward network, and the RuLe() activation function is used to make the linear model nonlinear, as follows: F=FFN(z i )=ReLu(from i IN (1) +b (1) )IN (2) +b (2) Among them, W (1) ,W (2) ∈R d×d , b (1) ,b (2) ∈R d , the process is unified into a function FFN().

3. The method for tracking knowledge of personalized time interval-aware attention mechanism for students according to claim 1, characterized in that: The knowledge state p i for: p i =Sigmoid(Fw+b) Among them, F i is the i-th row of F, w represents the weight, b represents the bias, and the activation function

Images