Online educational resource recommendation method for personalized learning

Abstract

The present invention discloses an online educational resource recommendation method for personalized learning, comprising the following steps: (1) collecting learning behavior records of students watching educational videos; (2) calculating students' learning ability based on the knowledge of educational psychology, and classifying student groups according to the number of educational videos learned by students and their learning ability by using a clustering method; (3) extracting the attributes of the educational video and calculating the difficulty to form a feature vector of the educational video; (4) using the feature vector of the educational video as input, using the LinUCB algorithm to provide students with educational videos of appropriate difficulty, and forming a final recommendation list. The present invention proposes an online educational resource recommendation scheme for personalized learning based on the LinUCB algorithm, which improves the calculation speed and the utilization rate of computing resources; according to the learning ability of each student, the personalized exploration coefficient is designed in combination with the attention mechanism, and an adaptive personalized exploration mechanism is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of online education big data mining, and in particular to an online education resource recommendation method for personalized learning. Background Art

[0002] Distance education is a new educational model compared to traditional face-to-face education, and online education has become a key method of distance education. Compared to classroom learning, online education transcends the constraints of time and space, allowing students to access knowledge anytime and anywhere. It is a scientific, immediate, and interactive learning method. In recent years, a large number of online education platforms have emerged, providing students with a wide range of learning resources, such as MOOCs, Coursera, and EdX. However, faced with this vast amount of learning resources, students often struggle to easily and effectively find the right content for their needs, resulting in a lack of focus and inefficient learning. Therefore, it is crucial to use educational big data mining techniques to accurately identify and recommend appropriate learning resources to specific students based on their individual characteristics.

[0003] Recommender systems have been widely used in education, providing students with a variety of learning resources, such as articles, websites, and video courses. Recommender systems in online education are a branch of information retrieval within educational big data mining. Their mission is to provide students with relevant learning materials and help them make decisions. Traditional online educational resource recommendation methods are mostly based on collaborative filtering algorithms. They first construct a user-item matrix and then recommend learning resources based on user or item similarities. However, when recommending resources in online education scenarios, traditional recommendation methods still have significant limitations: they fail to fundamentally explore students' learning characteristics and design recommendation schemes based on these characteristics, and the recommended content is often too simple and unchallenging. Students are unable to achieve good learning outcomes and improve their learning level through the educational resources provided by the recommendation system.

[0004] Most existing online educational resource recommendation solutions leverage artificial intelligence algorithms such as neural networks and deep learning to explore students' learning patterns. However, these solutions often focus too much on the accuracy of recommendation models and overlook the most fundamental characteristics of students. In other words, they fail to design recommendation solutions based on students' individual learning characteristics. Educational psychology can be used to study students' learning behaviors, measure their learning status, and provide a theoretical basis for designing personalized recommendation solutions for different students. The Zone of Proximal Development (ZPD) theory states that teachers should provide students with learning resources that are neither too easy nor too difficult—that is, the difficulty of educational resources should be slightly beyond their current learning level to tap into their potential. Among existing recommendation solutions, the context-based bandit algorithm LinUCB can achieve a balance between exploration and utilization in news recommendation scenarios. However, when applied to learning resource recommendation in online education settings, LinUCB fails to consider students' individual characteristics, which poses a risk to the process of exploring new knowledge. Recommended educational resources may be too difficult for students, resulting in reduced recommendation performance and increased cognitive load. Summary of the Invention

[0005] Purpose of the invention: In response to the above problems, the purpose of the present invention is to provide an online education resource recommendation method for personalized learning. The online education resource recommendation plan is designed based on the students' own learning characteristics, which can tap the students' potential and improve learning effects.

[0006] Technical Solution: The present invention provides a method for recommending online educational resources for personalized learning, comprising the following steps:

[0007] (1) Collect students' learning behavior records when watching educational videos, clean the learning behavior records, and extract effective information;

[0008] (2) Integrate learning behavior records from the students' perspective, calculate students' learning ability based on the knowledge of educational psychology, and use clustering methods to classify student groups according to the number of educational videos they have learned and their learning ability;

[0009] (3) Integrate learning behavior records from the perspective of educational videos, extract the attributes of educational videos and calculate the difficulty to form the feature vector of educational videos;

[0010] (4) Taking the feature vector of the educational video as input, the LinUCB algorithm is used to provide students with educational videos of appropriate difficulty. The students' learning ability and attention mechanism are introduced to design a personalized exploration coefficient to form the final recommendation list.

[0011] Furthermore, each learning behavior record in step (1) includes four aspects of information: viewing behavior information, pause behavior information, jump behavior information, and in-class test information;

[0012] Viewing behavior information includes: the ratio of the total time students spend watching videos to the total length of the videos (fracSpent), the ratio of the time students spend playing and replaying videos to the total length of the videos (fracPlayed), and the ratio of students playing videos (fracComp);

[0013] The total time students spent watching the video included playing, pausing, and rewinding;

[0014] The proportion of students playing the video, fracComp, does not include pause and replay;

[0015] Pause behavior information includes: the number of times students paused while learning video content, numPaused, and the ratio of pause time to the total video length, fracPaused;

[0016] The jump behavior information includes: the number of times numRWs the students replayed and the number of times numFFs they fast-forwarded when learning the video content;

[0017] The quiz information refers to whether the students correctly answered the corresponding quiz after learning the video content. Each educational video corresponds to a quiz. When success = 1, it indicates a correct answer, and when success = 0, it indicates an incorrect answer.

[0018] Furthermore, step (1) data cleaning refers to deleting learning behavior records containing empty entries and incorrect time information;

[0019] Step (1) extracting valid information includes:

[0020] In the viewing behavior information, fracComp does not include repeated or paused playback and represents the completion of the video content;

[0021] Suspension behavior information is caused by unexpected circumstances, including network freezes and students leaving temporarily;

[0022] In the jump behavior information, the number of replays numRWs reflects that the learning video has a certain degree of difficulty and needs to be watched repeatedly, which is used to calculate the difficulty of subsequent educational videos;

[0023] The success information of the in-class test reflects the student's ability and is used to calculate the difficulty of the video.

[0024] Furthermore, step (2) integrates learning behavior records from the students' perspective, calculates students' learning abilities based on the knowledge of educational psychology, and uses clustering methods to classify student groups according to the number of educational videos they have learned and their learning abilities, including:

[0025] (201) Obtain the number of educational videos each student has studied, the degree of completion of each educational video, and whether the student has correctly answered the corresponding in-class test after watching the educational video;

[0026] (202) Based on the Bloom model and the knowledge of educational psychology of classic test theory, we analyze the factors that affect students’ learning ability, including students’ participation, completion and accuracy, and use the integrated learning behavior records to estimate students’ learning ability. The expression is as follows:

[0027]

[0028] Among them, ability i represents the learning ability of student i, i = 1, 2, ..., I, I is the total number of students; the learning behavior records of student i watching educational videos constitute the set N i , n i For the set N i The number of educational videos in the , that is, the number of educational videos studied by student i; fracComp ij Represents student i for the set N i The completion degree of the jth educational video, success ij Indicates that student i watches set N i Whether the student correctly answered the class quiz after watching the jth educational video; J is the total number of educational videos;

[0029] In the above formula, n i The ratio of [number of factors] to the total number of educational videos J represents student i's participation in learning educational videos. The latter two items represent student i's average completion of educational video content and average accuracy rate of in-class quizzes, respectively. k1, k2, and k3 are weights to control the importance of the three influencing factors, where k1+k2+k3=1.

[0030] (203) According to each student's learning ability i Number of educational videos studied with each student n i ,The K-means clustering algorithm is used to divide students into three groups with different characteristics: active students, potential students, and inactive students;

[0031] The average learning ability and the average number of educational videos watched of active students, potential students, and inactive students decrease in that order.

[0032] Furthermore, step (3) integrates the learning behavior records from the perspective of educational videos, extracts the attributes of educational videos and calculates the difficulty, forming the feature vector of educational videos including:

[0033] (301) Obtain the number of students who have generated learning behavior records for each educational video, the completion degree of each educational video, the number of times it has been played back, and whether the class test corresponding to the video has been answered correctly;

[0034] (302) From the integrated learning behavior records, four attributes of educational videos are extracted: learning rate fracStudy, average degree of completion avgfracComp, playback times ratio fracnumRWs, and correctness rate fraccorrect;

[0035] (303) The difficulty of an educational video is calculated as:

[0036] dif j =l1*(1-avgfracComp j )+l2*(1-fraccorrect j )+l3*fracnumRWs j ,

[0037] In the formula, dif j represents the difficulty of educational video j; l1, l2, l3 are importance weights, l1+l2+l3=1;

[0038] (304) The five features of each educational video form a feature vector, namely:

[0039]

[0040] Furthermore, the learning rate fracStudy represents the ratio of students who generate learning behavior records with the video, and is expressed as:

[0041]

[0042] The average completion level avgfracComp represents the average completion level of all students who watched the video. The expression is:

[0043]

[0044] The playback count ratio fracnumRWs represents the ratio of the total number of playbacks of all students who watched the video to the threshold number of playbacks. The expression is:

[0045]

[0046] The correct rate fraccorrect represents the proportion of students who watched the video and answered the quiz correctly. The expression is:

[0047]

[0048] Where, subscript j represents educational video j, j = 1, 2, ..., J, J is the total number of educational videos, m j represents the number of students who have generated learning behavior records with educational video j; fracComp ijIndicates the completion degree of the i-th student for the educational video j; numRWs ij Indicates the number of times the i-th student replayed the educational video j, numRWs max The threshold for the number of times the educational video is played back; success ij Indicates whether the i-th student answered the class test correctly after watching educational video j.

[0049] Furthermore, step (4) uses the feature vector of the educational video as input, uses the LinUCB algorithm to provide students with educational videos of appropriate difficulty, introduces students' learning ability and attention mechanism to design personalized exploration ratio coefficients, and forms the final recommendation list including:

[0050] (401) Based on the results of student group clustering, all students’ learning behavior records are separated into active students, potential students, and inactive students, and are input into the LinUCB algorithm respectively to recommend educational videos to specific student groups;

[0051] (402) Based on the idea of ​​parallel matrix, the feature reward vector of each educational video is calculated. The expression is:

[0052]

[0053] Among them, D i It refers to the learning behavior record set N of student i watching educational videos in the learning behavior records of a certain type of student group separated by step (401). i The feature matrix of the corresponding educational video; Dimension is n i ×d,n i Represents the set N i The number of educational videos in , d is the feature dimension of educational videos; Represents the set N i The feature vector of the j-th educational video in is generated by step (304); It's D i The transpose of c i is the set N i The reward value vector obtained by each educational video in, Dimension is n i ×1, r j Represents the set N i The reward value obtained by the j-th educational video in the , which is defined as the completion degree of student i for the j-th educational video fracComp ij ; A i Is a diagonal matrix of dimension d×d, D i Calculated; b iIt is a d×1-dimensional vector, whose elements represent the cumulative reward value of each educational video feature; I d is the d-dimensional identity matrix;

[0054] (403) Calculate the expected value of student i’s estimated income from all videos E[r i |X], the expression is:

[0055]

[0056] Among them, r i is the feedback benefit vector of student i for each educational video in the set C of all educational videos, with a dimension of J × 1, where J is the number of educational videos in the set C, i.e., the total number of educational videos; is the feature matrix of the educational video corresponding to the set C, with dimension J×d;

[0057] (404) Introducing the attention mechanism to calculate the control exploration ratio coefficient α in the LinUCB algorithm i ;

[0058] (405) According to the LinUCB criterion, calculate the total estimated benefit p of student i for each educational video i :

[0059]

[0060] Among them, p i Refers to the total estimated benefit vector obtained by each educational video in set C for student i; Is to take the matrix The vector of diagonal elements represents the expected value of the estimated return The prediction variance of α i and The Hadamard product representation uncertainty;

[0061] (406) According to the total estimated revenue p i , select the top L programs with the highest estimated revenue to form the final recommendation list, that is,

[0062] list=top L (p ij ),p i =[p ij ],

[0063] Among them, top L (p ij ) represents selecting p from the set C of all educational videos ij The L educational videos with the largest values ​​are recommended to student i; ijis the estimated benefit value of student i for the jth educational video in set C, which constitutes p i The row elements in .

[0064] Furthermore, step (404) calculates the control exploration proportional coefficient α in the LinUCB algorithm i include:

[0065] (4041) Calculate the set N i The difference matrix with the educational videos in set C

[0066]

[0067] in, Represents the difference value matrix The p-th row and q-th column element in the set C is the difficulty of the p-th video in the set N i It is calculated by the Euclidean distance between the difficulty of the qth video in;

[0068] (4042) Calculate the attention score vector S of student i for each educational video i :

[0069]

[0070] Among them, S i is a J×1 dimensional vector, which refers to the student’s attention score, reflecting the student’s desire for unknown knowledge, represented by c i right Obtained by weighted summation of the row elements in ;

[0071] (4043) Introducing students’ learning ability and calculating the dynamic control exploration proportional coefficient α i :

[0072]

[0073] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0074] 1. Based on the LinUCB algorithm, this paper proposes an online educational resource recommendation scheme for personalized learning. The LinUCB algorithm first converts serial computations into parallel matrix computations, improving computational speed and resource utilization. Secondly, based on individual students' learning abilities, a personalized exploration coefficient is designed in conjunction with the attention mechanism, implementing an adaptive personalized exploration mechanism. This allows the recommendation scheme to not only provide students with appropriate learning resources but also, to a certain extent, tap their potential, reducing the risks associated with the exploration process.

[0075] 2. It provides a feature extraction method for student behavior records based on educational psychology, uses educational psychology to reveal the nature of learning outcomes, explain the learning process, clarify the conditions and laws of effective learning, explore students' characteristics, and provide a substantial basis for tracking students' learning status and evaluating their learning abilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 Flow chart of the method of the present invention;

[0077] Figure 2 shows the clustering results of three different student groups, namely the probability density distribution curves of the number of educational videos learned and learning ability. 2(a) is active students, 2(b) is potential students, and 2(c) is inactive students;

[0078] FIG3( a ) shows the performance of the recommendation scheme of the present invention in terms of accuracy, FIG3( b ) shows the performance of the recommendation scheme of the present invention in terms of recall, and FIG3( c ) shows the performance of the recommendation scheme of the present invention in terms of F1 value. DETAILED DESCRIPTION

[0079] The dataset used in this example comes from the online education platform Coursera. A total of 29,304 student learning behavior records were collected from the platform while watching educational videos. The dataset includes 93 educational videos, each with a corresponding in-class quiz to assess students' understanding of the video content. Students were required to complete the corresponding in-class quiz after watching each educational video.

[0080] like Figure 1 As shown in the flowchart, the method for recommending online educational resources for personalized learning described in this embodiment includes the following steps:

[0081] (1) Collect students’ learning behavior records when watching educational videos, clean the learning behavior records, and extract effective information.

[0082] Each learning behavior record includes four aspects of information: viewing behavior information, pause behavior information, jump behavior information and in-class test information.

[0083] Viewing behavior information includes: the ratio of the total time students spend watching videos to the total length of the videos (fracSpent), the ratio of the time students spend playing and replaying videos to the total length of the videos (fracPlayed), and the ratio of students playing videos (fracComp);

[0084] The total time students spent watching the video included playing, pausing, and rewinding;

[0085] The proportion of students playing the video, fracComp, does not include pause and replay;

[0086] Pause behavior information includes: the number of times students paused while learning video content, numPaused, and the ratio of pause time to the total video length, fracPaused;

[0087] The jump behavior information includes: the number of times numRWs the students replayed and the number of times numFFs they fast-forwarded when learning the video content;

[0088] The quiz information refers to whether the students correctly answered the corresponding quiz after learning the video content. Each educational video corresponds to a quiz. When success = 1, it indicates a correct answer, and when success = 0, it indicates an incorrect answer.

[0089] Data cleaning involves deleting learning behavior records containing empty entries and incorrect time information, such as excessive pauses, excessive replays or skips, and fracComp greater than 1. In this example, 549 learning behavior records containing empty entries and 1,821 incorrect learning behavior records exhibiting these conditions were deleted, ultimately retaining a total of 26,934 learning behavior records for 2,219 students.

[0090] Extracted effective information includes: in viewing behavior information, fracComp does not include repeated or paused playback, representing the completion of the video content; pause behavior information is caused by unexpected situations, including network lag and temporary departure of students; in jump behavior information, the number of playbacks numRWs reflects that the learning video is difficult and needs to be watched repeatedly, which is used to calculate the difficulty of subsequent educational videos; in-class test information success reflects student ability and is used to calculate video difficulty.

[0091] (2) Integrate learning behavior records from the students' perspective, calculate students' learning ability based on the knowledge of educational psychology, and use clustering methods to classify student groups according to the number of educational videos they have learned and their learning ability.

[0092] (201) Obtain the number of educational videos each student has studied, the degree of completion of each educational video, and whether the student has correctly answered the corresponding in-class test after watching the educational video.

[0093] In this example, the calculation of student learning ability is guided by educational psychology, exploring the various factors that influence student learning ability. Research shows that when answering questions related to specific knowledge, students who answer correctly are more competent than those who answer incorrectly. Furthermore, students who interact more with learning resources have a greater understanding of the learning activities and stronger learning ability. Furthermore, Bloom's model indicates that the degree of completion of educational content is positively correlated with student learning ability. Classical testing theory suggests that scores can reflect a student's mastery of knowledge and measure their ability.

[0094] (202) Based on the Bloom model and the knowledge of educational psychology of classic test theory, we analyze the factors that affect students’ learning ability, including students’ participation, completion and accuracy, and use the integrated learning behavior records to estimate students’ learning ability. The expression is as follows:

[0095]

[0096] Among them, ability i represents the learning ability of student i, i = 1, 2, ..., I, I is the total number of students; the learning behavior records of student i watching educational videos constitute the set N i , n i For the set N i The number of educational videos in the , that is, the number of educational videos studied by student i; fracComp ij Represents student i for the set N i The completion degree of the jth educational video, success ij Indicates that student i watches set N i Whether the student correctly answered the class quiz after watching the jth educational video; J is the total number of educational videos;

[0097] In the above formula, n i The ratio of the number of educational videos to the total number J of educational videos represents student i's participation in educational video learning. The last two items represent student i's average completion of the educational video content and the average accuracy of the in-class test, respectively. k1, k2, and k3 are weights for controlling the importance of the three influencing factors, k1+k2+k3=1. In this embodiment, k1 is 0.3, k2 is 0.2, and k3 is 0.5.

[0098] (203) According to each student's learning ability i Number of educational videos studied with each student n i ,The K-means clustering algorithm is used to divide students into three groups with different characteristics: ,active students, potential students, and inactive students, in order to better reflect the students’ ,learning status.

[0099] The specific process of the K-means clustering algorithm is as follows:

[0100] (2031) Cluster initialization:

[0101] The number of educational videos studied and learning ability of each student are combined into a data set. This data set serves as the input for the K-means clustering algorithm. Each data point in the data set contains the number of educational videos studied and learning ability of the student, a 2×1-dimensional vector. The total number of clusters is specified as 3, i.e., K = 3. Three data points are randomly selected from the data set as the initial cluster centroids, and each cluster centroid defines a centroid cluster.

[0102] (2032)Data allocation:

[0103] Each data point in the data set is assigned to the centroid cluster to which its nearest cluster centroid belongs based on the Euclidean distance.

[0104] (2033) Cluster centroid update:

[0105] Get the average value of all data assigned to the centroid cluster, calculate the average number of educational videos learned and the average learning ability of students in the centroid cluster, and form an average vector as the cluster centroid for the next iteration.

[0106] (2034) Algorithm iteration: Repeat steps (2032) to (2033) until the stopping criteria are met, that is, no data changes its centroid cluster, and the sum of cluster distances is minimized, to obtain the final clustering result.

[0107] In this embodiment, as shown in Figures 2(a), 2(b), and 2(c), students are divided into three groups with different characteristics. The upper curve represents the probability density distribution of the number of educational videos studied by each type of student, and the lower curve represents the probability density distribution of the learning ability of each type of student. The first type of student group consists of 194 students. As shown in Figure 2(a), the number of videos studied by this type of students ranges from 29 to 92 (mean = 52.88, SD = 20.002), concentrated in the range of 30 to 60; the ability value ranges from 0.31 to 0.88 (mean = 0.65, SD = 0.100), concentrated in the range of 0.53 to 0.76. The second group of students consists of 1,265 students. As shown in Figure 2(b), the number of videos studied by this group of students ranges from 1 to 30 (mean = 5.70, SD = 6.052), with a concentration of 1 to 8. Their ability values ​​range from 0.40 to 0.74 (mean = 0.60, SD = 0.089, with a concentration of 0.58 to 0.72). The third group of students consists of 760 students. As shown in Figure 2(c), the number of videos studied by this group of students ranges from 1 to 28 (mean = 3.44, SD = 3.556), with most students studying between 0 and 4. Their ability values ​​range from 0.01 to 0.47 (mean = 0.26, SD = 0.118), with most students studying between 0.13 and 0.44. Through the above classification results and analysis, we can identify the three types of students and summarize their learning characteristics. We will subsequently analyze the performance of the recommended solution based on the characteristics of the three types of students. The characteristics of the three types of students are as follows:

[0108] The first type of students studied the largest number of educational videos and their learning ability was always high. This type of students is called active students. They have high participation in educational videos, strong learning ability, and good learning results.

[0109] The second category of students studied fewer educational videos than the first category, but their learning ability was still very high, almost the same as the first category. This category of students is called potential students. Although they have only studied a small number of videos, their learning ability is very high, which means that they have great potential to be tapped.

[0110] The third type of students have learned the least educational videos and have relatively low learning ability. Compared with the other two groups of students, they are not active and perform poorly throughout the learning process. This type of students is called inactive students. They rarely participate in the study of educational videos and have weak learning ability. Therefore, it is difficult to discover their learning patterns from their learning behavior.

[0111] (3) Integrate learning behavior records from the perspective of educational videos, extract the attributes of educational videos and calculate the difficulty to form the feature vector of educational videos;

[0112] (301) Obtain the number of students who have generated learning behavior records for each educational video, the completion degree of each educational video, the number of times it has been played back, and whether the class test corresponding to the video has been answered correctly;

[0113] (302) From the integrated learning behavior records, four attributes of educational videos are extracted: learning rate fracStudy, average completion degree avgfracComp, playback number ratio fracnumRWs, and correctness rate fraccorrect.

[0114] The learning rate fracStudy represents the ratio of students who generate learning behavior records with the video, and is expressed as:

[0115]

[0116] The average completion level avgfracComp represents the average completion level of all students who watched the video. The expression is:

[0117]

[0118] The playback count ratio fracnumRWs represents the ratio of the total number of playbacks of all students who watched the video to the threshold number of playbacks. The expression is:

[0119]

[0120] The correct rate fraccorrect represents the proportion of students who watched the video and answered the quiz correctly. The expression is:

[0121]

[0122] Where, subscript j represents educational video j, j = 1, 2, ..., J, J is the total number of educational videos, m j represents the number of students who have generated learning behavior records with educational video j; fracComp ij Indicates the completion degree of the i-th student for the educational video j; numRWs ij Indicates the number of times the i-th student replayed the educational video j, numRWs max The threshold for the number of times the educational video is played back; success ij Indicates whether the i-th student answered the class test correctly after watching educational video j.

[0123] (303) The difficulty of an educational video is calculated as:

[0124] dif j =l1*(1-avgfracComp j )+l2*(1-fraccorrect j )+l3*fracnumRWs j , where dif j represents the difficulty of educational video j; l1, l2, l3 are importance weights, l1+l2+l3=1; in this embodiment, l1 is 0.2, l2 is 0.6, and l3 is 0.2.

[0125] (304) The five features of each educational video form a feature vector, namely:

[0126]

[0127] (4) Taking the feature vector of the educational video as input, the LinUCB algorithm is used to provide students with educational videos of appropriate difficulty. The students' learning ability and attention mechanism are introduced to design a personalized exploration coefficient to form the final recommendation list.

[0128] (401) Based on the results of student group clustering, all students’ learning behavior records are separated into active students, potential students, and inactive students, and are input into the LinUCB algorithm respectively to recommend educational videos to specific student groups, which facilitates the subsequent statistical analysis of the performance of the recommendation method for different student groups;

[0129] (402) Based on the idea of ​​parallel matrix, parallel matrix calculation is used instead of the traditional LinUCB serial vector to improve the computing efficiency and computer resource utilization, and the feature reward vector of each educational video is calculated. is a parameter to be learned, with a dimension of d×1, Parameters representing the k-th educational video feature. The calculation is as follows:

[0130]

[0131] Among them, D i It refers to the learning behavior record set N of student i watching educational videos in the learning behavior records of a certain type of student group separated by step (401). i The feature matrix of the corresponding educational video; Dimension is n i ×d,n i Represents the set N i The number of educational videos in the video, d is the feature dimension of educational videos; x j T Represents the set N i The feature vector of the j-th educational video in is generated by step (304); It's D i The transpose of c i is the set N i The reward value vector obtained by each educational video in, Dimension is n i ×1, r j Represents the set N i The reward value obtained by the j-th educational video in the , which is defined as the completion degree of student i for the j-th educational video fracComp ij ; A i Is a diagonal matrix of dimension d×d, D i Calculated; b i It is a d×1-dimensional vector, whose elements represent the cumulative reward value of each educational video feature; I d is the d-dimensional identity matrix;

[0132] (403) Calculate the expected value of student i’s estimated income from all videos E[r i |X], the expression is:

[0133]

[0134] Among them, r i is the feedback benefit vector of student i for each educational video in the set C of all educational videos, with a dimension of J × 1, where J is the number of educational videos in the set C, i.e., the total number of educational videos; is the feature matrix of the educational video corresponding to the set C, with dimension J×d;

[0135] (404) Introducing the attention mechanism to calculate the control exploration ratio coefficient α in the LinUCB algorithm i ;

[0136] (4041) Calculate the set N i The difference matrix with the educational videos in set C

[0137]

[0138] in, Represents the difference value matrix The p-th row and q-th column element in the set C is the difficulty of the p-th video in the set N i It is calculated by the Euclidean distance between the difficulty of the qth video in;

[0139] (4042) Calculate the attention score vector S of student i for each educational video i :

[0140]

[0141] Among them, S i is a J×1 dimensional vector, which refers to the student’s attention score, reflecting the student’s desire for unknown knowledge, represented by c i right The weighted sum of the row elements in is obtained; therefore, S i Students' attention to educational videos they have already watched is reduced, while their attention to educational videos they have not yet watched is increased. The greater the difference in difficulty between the educational videos they have already watched and those they have not yet watched, the easier it is to attract students' attention and the stronger their desire to explore unknown knowledge.

[0142] (4043) Introducing students’ learning ability and calculating the dynamic control exploration proportional coefficient α i :

[0143]

[0144] In this embodiment, each student's learning ability and attention are different. Therefore, the personalized exploration coefficient designed by the method of this embodiment can make full use of the student's learning state and adaptively adjust the exploration ratio according to the student's learning ability and attention. On the other hand, the personalized α i You can ensure that the difficulty of the educational videos explored is within the students' ability range, thereby reducing the risk of exploration.

[0145] (405) According to the LinUCB criterion, calculate the total estimated benefit p of student i for each educational video i :

[0146]

[0147] Among them, p iRefers to the total estimated benefit vector obtained by each educational video in set C for student i; Is to take the matrix The vector of diagonal elements represents the expected value of the estimated return The prediction variance of ; its dimension is J×1. Combined with the personalized exploration coefficient α calculated in step (404) i , α i and The Hadamard product can be interpreted as Therefore, in the above formula, the first half represents the use of educational videos that students have watched, and the second half realizes the guaranteed exploration of new educational videos.

[0148] (406) According to the total estimated revenue p i , select the top L programs with the highest estimated revenue to form the final recommendation list, that is,

[0149] list=top L (p ij ),p i =[p ij ],

[0150] Among them, top L (p ij ) represents selecting p from the set C of all educational videos ij The L educational videos with the largest values ​​are recommended to student i; ij is the estimated benefit value of student i for the jth educational video in set C, which constitutes p i The row elements in .

[0151] The experimental method of this embodiment is further described below.

[0152] In this embodiment, the performance indicators used to evaluate the recommendation scheme proposed by the present invention are divided into two categories: precision indicators and non-precision indicators. Precision indicators include accuracy, recall rate, and F1 value, and their purpose is to measure whether the recommended educational videos accurately fit the students' learning characteristics; non-precision indicators include adaptability and personalization, which focus on measuring the characteristics of the recommendation results. Adaptability can reflect whether the recommendation results are suitable for students, and personalization can reflect the overall diversity of the recommendation scheme. The specific meanings of these five indicators are as follows, where N represents the number of educational videos in the recommendation list:

[0153] Precision @ N: refers to the proportion of successfully recommended educational videos in the recommended list. In this embodiment, a successful recommendation not only means that the student watched the educational video, but also correctly completed the class quiz corresponding to the educational video.

[0154] Recall (recall@N): This refers to the proportion of successfully recommended educational videos to the number of educational videos actually watched by students. This metric is calculated using the same numerator as precision, while the denominator is the number of videos actually watched by students.

[0155] F1 value (F1@N): refers to the harmonic mean of precision and recall, which comprehensively evaluates precision and recall.

[0156] Adaptivity@N: This refers to the difference between the average difficulty of the educational videos in the recommended list and the average difficulty of the educational videos the student has already watched. This evaluation metric is used to measure whether the recommended educational videos are consistent with the student's learning ability level and reflects the adaptability of the recommendation results to the student.

[0157] Personalization@N: This refers to the average difference between the recommendation lists received by different students. It represents the overall diversity of the recommendation options. The higher the personalization level, the more personalized the recommendation options are, and the more personalized the learning experience students can obtain.

[0158] The experimental results are analyzed below with reference to the accompanying figures and tables.

[0159] In this example, 2,582 learning behavior logs with almost zero participation were deleted from the cleaned dataset. Ultimately, the number of active students, potential students, and inactive students was 194, 890, and 517, respectively; and the corresponding number of learning behavior logs for these three groups of students was 10,258, 10,932, and 3,162, respectively.

[0160] First, combining the characteristics of the three types of student groups, the accuracy, recall rate and F1 value of the solution proposed in this embodiment are evaluated and analyzed. As can be seen from Figure 3(a), the accuracy corresponding to active students is always the highest, followed by potential students and inactive students. Since active students provide relatively rich learning behavior records, the solution of the present invention can more accurately describe their learning status. In addition, according to Figure 3(b), the recall rate of active students is generally lower, while that of inactive students is higher, which is consistent with the characteristics of the three types of student groups. Active students watch the most videos, while the number of videos watched by potential students and inactive students is far less than that of active students. Therefore, when calculating the recall rate, if the denominator is smaller, that is, the number of educational videos watched by this type of students is smaller, the resulting recall rate will be larger. Although the number of successfully recommended videos for active students is greater than that for the other two types of students, this value is still small compared to the dozens of videos they have watched.

[0161] The results in Figure 3(c) show that potential students have the highest F1 value. In this example, the F1 values ​​are compared horizontally based on the characteristics of three student groups:

[0162] 1) The F1 value of active students increases continuously. These students watch the most videos, so a larger N value leads to better performance.

[0163] 2) For potential students, both F1@8 and F1@10 are relatively high, because most potential students have watched 1 to 10 educational videos. When the number of recommended videos is too large for them, such as N ≥ 10, F1@10 begins to drop slightly.

[0164] 3) The F1 value of inactive students is decreasing. They watch very few videos (0-4) and their abilities are relatively weak. Based on the above analysis, when the length of the recommendation list is constant, the accuracy rate can more objectively and reasonably evaluate the accuracy of the recommendation algorithm. This is because it refers to the proportion of educational videos in the recommendation list that are successfully recommended and is not affected by the classification results of the student group.

[0165] As shown in Table 1 below, this example evaluates the adaptability of the recommended solution of the present invention. According to the ZPD theory, the content to be learned in the next stage should be challenging for students, which can stimulate their enthusiasm, tap their potential, and achieve higher cognitive levels. In conjunction with the ZPD theory, this example specifically analyzes the adaptability of the recommended solution's personalized exploration strategy to three types of student groups:

[0166] (1) For active students, the average difficulty of the educational videos recommended to them is about 0.08 to 0.14 higher than the average difficulty of the educational videos watched by students. This is reasonable for active students with strong abilities. The present invention provides slightly more difficult content through personalized exploration strategies to stimulate the potential of these students;

[0167] (2) For potential students, due to their higher abilities, the difficulty of the recommended educational videos is slightly higher than that of the educational videos they have watched. However, the difficulty gap is smaller than that of active students, ranging from 0.06 to 0.09. This is because potential students watch fewer educational videos, so a slight increase in difficulty can effectively stimulate their enthusiasm and encourage them to participate more in learning.

[0168] (3) For inactive students, some of the recommended educational videos are slightly more difficult, while others are still within their ability range. The difference ranges from -0.01 to 0.02, which is consistent with their weaker abilities. The difficulty of learning resources should not be too high for their abilities, otherwise it may cause cognitive overload.

[0169] Table 1 Adaptability performance of three types of student groups under personalized exploration strategy

[0170] adaptive@5 adaptive@8 adaptive@10 Active students 0.1396 0.1109 0.0898 potential students 0.0883 0.0893 0.0617 Inactive students -0.0140 -0.0130 0.0136

[0171] In order to verify whether the personalized coefficient designed by the present invention improves the personalization level of the recommendation algorithm, this example selects the most active students with the highest accuracy and further analyzes the impact of different exploration coefficients on the personalization level. This example sets three different exploration coefficients:

[0172] The first type has no specific value, α = 1, which means that the ratio of exploration and exploitation is the same and has not been adjusted.

[0173] The second is a fixed value of α = 0.5, which means that the ratio of exploration to exploitation is the same for each student without considering the student's characteristics.

[0174] The third is the personalized exploration coefficient α proposed in step (404) i .

[0175] As shown in Table 2, compared with α = 1, when α = 0.5, the personalization of the recommendation scheme has been slightly improved, which proves that the exploration coefficient can improve the personalization of the recommendation results. Compared with the fixed exploration coefficient α = 0.5, the personalized exploration coefficient α designed by the present invention is i The personalized ability of the recommendation scheme is significantly improved, which proves that incorporating students' learning ability into the exploration coefficient can enhance the personalized ability of the recommendation algorithm.

[0176] Table 2 Personalization performance of recommended solutions under different exploration coefficients

[0177] personalized@5 personalized@8 personalized@10 No specific value (α=1) 38.15％ 36.78％ 36.49％ Fixed value (α = 0.5) 44.70％ 39.83％ 38.71％ Personalization coefficient 55.23％ 53.75％ 52.22％

[0178] Finally, to verify the effectiveness of the recommendation scheme in the present invention, this embodiment still uses active students as the target group and compares the performance with some existing recommendation schemes:

[0179] Popularity: This is the most common recommendation scheme, which recommends some educational videos with the most viewers to students. This is similar to traditional offline teaching, where teachers teach the most in-demand courses based on the needs of most students.

[0180] User-based collaborative filtering: Recommends educational videos based on videos watched by students similar to the target student. The similarity between students is calculated by analyzing their learning behavior records.

[0181] Item-based collaborative filtering: recommends educational videos similar to those watched by the target student. The similarity between educational videos is calculated by analyzing the student's learning behavior records.

[0182] As shown in Table 3, this embodiment uses accuracy, recall, and F1 value as evaluation indicators. The recommendation method proposed in the present invention shows better performance. Among the three comparison schemes, the user-based collaborative filtering method has the best performance. In comparison, the method proposed in the present invention has improvements of 16.06%, 1.08%, and 0.23% in precision@8, recall@8, and F1@8, respectively, and improvements of 19.22%, 0.48%, and 2.75% in precision@10, recall@10, and F1@10, respectively. The results show that the recommendation scheme proposed in the present invention can provide students with the most suitable online educational resources and reduce the risks brought by exploration.

[0183] Table 3 Accuracy, recall and F1 value performance of different recommendation schemes

[0184]

Claims

1. An online educational resource recommendation method for personalized learning, characterized in that: The steps include: (1) Collect students' learning behavior records when watching educational videos, clean the learning behavior records, and extract effective information; (2) Integrate learning behavior records from the students' perspective, calculate students' learning abilities based on knowledge of educational psychology, and use clustering methods to classify student groups according to the number of educational videos they have watched and their learning abilities; (3) Integrate learning behavior records from the perspective of educational videos, extract the attributes of educational videos and calculate the difficulty to form the feature vector of educational videos; (4) Taking the feature vector of the educational video as input, the LinUCB algorithm is used to provide students with educational videos of appropriate difficulty. The personalized exploration coefficient is designed by introducing the students' learning ability and attention mechanism to form the final recommendation list. Step (2) integrates learning behavior records from the students' perspective, calculates students' learning abilities based on knowledge of educational psychology, and uses clustering methods to classify student groups according to the number of educational videos they have learned and their learning abilities, including: (201) obtaining the number of educational videos studied by each student, the degree of completion of each educational video, and whether the student correctly answered the corresponding in-class test after watching the educational video; (202) Based on Bloom’s model and the knowledge of educational psychology of classic test theory, we analyze the factors that affect students’ learning ability, including students’ participation, completion and accuracy, and use integrated learning behavior records to estimate students’ learning ability. The expression is as follows: Among them, ability i represents the learning ability of student i, i = 1, 2, ..., I, I is the total number of students; the learning behavior records of student i watching educational videos constitute a set N i , n i For the set N i The number of educational videos in the text box, that is, the number of educational videos studied by student i; fracComp ij Represents student i for set N i The completion degree of the jth educational video in ij Indicates that student i watches set N i Whether the student correctly answered the in-class quiz after the jth educational video; J is the total number of educational videos; In the above formula, n i The ratio of the total number of educational videos J represents the participation of student i in educational video learning, and the last two items represent the average completion of the educational video content and the average correct rate of the in-class test by student i, respectively; k1, k2, k3 are the weights for controlling the importance of the three influencing factors, k1+k2+k3=1; (203) According to each student's learning ability i Number of educational videos studied with each student n i ,The K-means clustering algorithm is used to divide students into three groups with different characteristics: active students, potential students, and inactive students; The average learning ability and the average number of educational videos watched by active students, potential students, and inactive students decrease in that order.

2. The online educational resource recommendation method according to claim 1, characterized in that: Step (3) integrates the learning behavior records from the perspective of educational videos, extracts the attributes of educational videos and calculates the difficulty, and forms a feature vector of educational videos including: (301) Obtaining the number of students who have generated learning behavior records for each educational video, the completion degree of each educational video, the number of times it has been played back, and whether the in-class test corresponding to the video has been answered correctly; (302) Extract four attributes of educational videos from the integrated learning behavior records: learning rate fracStudy, average degree of completion avgfracComp, playback times ratio fracnumRWs, and correctness rate fraccorrect; (303) Calculate the difficulty of the educational video, the expression is: dif j =l1*(1-avgfracComp j )+l2*(1-fraccorrect j )+l3*fracnumRWs j , In the formula, dif j represents the difficulty of educational video j; l1, l2, l3 are importance weights, l1+l2+l3=1; (304) The five features of each educational video form a feature vector, namely:

3. The online educational resource recommendation method according to claim 2, characterized in that: The learning rate fracStudy represents the ratio of students who produce learning behavior records with the video, and the expression is: The average completion level avgfracComp represents the average completion level of all students who watched the video. The expression is: The playback frequency ratio fracnumRWs represents the ratio of the total playback frequency of all students who watched the video to the threshold playback frequency. The expression is: The correct rate fraccorrect represents the proportion of students who watched the video and answered the quiz correctly. The expression is: Wherein, subscript j represents educational video j, j = 1, 2, ..., J, J is the total number of educational videos, m j Represents the number of students who have learning behavior records with educational video j; fracComp ij represents the completion degree of the i-th student for educational video j; numRWs ij Indicates the number of times the i-th student replayed the educational video j, numRWs max is the threshold for the number of times the educational video is played back; success ij Indicates whether the i-th student correctly answered the class quiz after watching educational video j.

4. The online educational resource recommendation method according to claim 3, characterized in that: Step (4) takes the feature vector of the educational video as input, uses the LinUCB algorithm to provide students with educational videos of appropriate difficulty, introduces students' learning ability and attention mechanism to design personalized exploration ratio coefficients, and forms the final recommendation list including: (401) According to the results of student group clustering, all students’ learning behavior records are separated into active students, potential students and inactive students, and are respectively input into the LinUCB algorithm to recommend educational videos for specific student groups; (402) Calculate the feature reward vector for each educational video The expression is: Among them, D i It refers to the learning behavior record set N of student i watching educational videos in the learning behavior records of a certain type of student group separated by step (401). i The feature matrix of the corresponding educational video; The dimension is n i ×d,n i Represents the set N i The number of educational videos in , d is the feature dimension of educational videos; Represents the set N i The feature vector of the j-th educational video in is generated by step (304); Yes D i The transpose of i is the set N i The reward value vector obtained by each educational video in, The dimension is n i ×1, r j Represents the set N i The reward value obtained by the j-th educational video in is defined as the completion degree of student i for the j-th educational video fracComp ij ; A i is a diagonal matrix of dimension d×d, represented by D i Calculated; b i is a d×1-dimensional vector whose elements represent the cumulative reward value of each educational video feature; I d is the identity matrix of d dimensions; (403) Calculate the expected value of student i’s estimated benefit from all videos E[r i |X], the expression is: Among them, r i is the feedback benefit vector of student i for each educational video in the set C of all educational videos, with a dimension of J×1, where J is the number of educational videos in the set C, i.e., the total number of educational videos; is the feature matrix of the educational video corresponding to the set C, with a dimension of J×d; (404) Introduce the attention mechanism and calculate the control exploration ratio coefficient α in the LinUCB algorithm i ; (405) According to the LinUCB criterion, calculate the total estimated benefit p of student i for each educational video i : Among them, p i Refers to the total estimated benefit vector obtained by each educational video in set C for student i; Is to take the matrix The vector of diagonal elements of represents the expected value of the estimated return The prediction variance of i and The Hadamard product representation The uncertainty of (406) Based on the total estimated revenue p i , select the top L programs with the highest estimated revenue to form the final recommendation list, that is, list=top L (p ij ),p i =[p ij ], Among them, top L (p ij ) represents selecting p from the set C of all educational videos ij The L educational videos with the largest values ​​are recommended to student i; ij is the estimated benefit value of student i for the jth educational video in set C, forming p i The row elements in .

5. The online educational resource recommendation method according to claim 4, characterized in that: Step (404) calculates the control exploration proportional coefficient α in the LinUCB algorithm i include: (4041) Calculate the set N i The difference matrix with the educational videos in set C in, Represents the difference value matrix The p-th row and q-th column element in the set C is the sum of the difficulty of the p-th video in the set N i It is calculated by the Euclidean distance between the difficulty of the qth video in; (4042) Calculate the attention score vector S of student i for each educational video i : Among them, S i is a J×1-dimensional vector, which refers to the student's attention score, composed of c i right It is obtained by weighted summing of the row elements in; (4043) Introducing students’ learning ability, calculating dynamic control exploration proportional coefficient α i :

6. The online educational resource recommendation method according to claim 1, characterized in that: Step (1) Each learning behavior record includes four aspects of information: viewing behavior information, pause behavior information, jump behavior information and in-class test information; The viewing behavior information includes: the ratio of the total time students spend watching videos to the total length of the videos (fracSpent), the ratio of the time students spend playing and replaying videos to the total length of the videos (fracPlayed), and the ratio of students playing videos (fracComp); The total time students spent watching the video included playing, pausing, and rewinding; The proportion of students playing the video, fracComp, does not include pause and replay; The pause behavior information includes: the number of times students paused when learning video content, numPaused, and the ratio of the pause time to the total video duration, fracPaused; The jump behavior information includes: the number of times numRWs the students replay and the number of times numFFs the students fast forward when learning the video content; The quiz information refers to whether the students correctly answered the corresponding quiz after learning the video content. Each educational video corresponds to a quiz. Success = 1 indicates a correct answer, and success = 0 indicates an error.

7. The online educational resource recommendation method according to claim 1, characterized in that: Step (1) data cleaning refers to deleting learning behavior records containing empty entries and incorrect time information; Step (1) extracting valid information includes: In the viewing behavior information, fracComp does not include repeated or paused playback, which represents the completion of the video content; The suspension behavior information is caused by unexpected circumstances, including network freeze and students leaving temporarily; In the jump behavior information, the number of replays numRWs reflects that the learning video has a certain degree of difficulty and needs to be watched repeatedly, which is used for the difficulty calculation of subsequent educational videos; The success information of the in-class test reflects the student's ability and is used to calculate the difficulty of the video.

Images