MOOC stop prediction method and system based on large language model auxiliary graph node classification

Abstract

The invention discloses a MOOC stop prediction method and system based on large language model auxiliary graph node classification, and belongs to the field of MOOC stop prediction. The method comprises the following steps: firstly, acquiring course behavior information data of each learner in a target course from a MOOC platform, encoding the course behavior information data to obtain course behavior characteristic data of the learners, and constructing a final discussion graph on the basis of the course behavior characteristic data; and executing two tasks of cognitive analytic hierarchy analysis and side false label identification by the fine-tuned large language model to respectively obtain a cognitive feature matrix and a false label of each side in the final discussion graph, and finally training a MOOC stop-class prediction model and completing MOOC stop-class prediction. According to the method, the stop class risk of the student can be predicted more accurately, intervention measures are taken in time, and the course completion rate and the overall effect of online education are improved.

Description

Technical Field

[0001] The present invention belongs to the field of MOOC dropout prediction, and in particular relates to a MOOC dropout prediction method and system using a large language model to assist graph node classification. Background Art

[0002] Massive Open Online Courses (MOOC) is a large-scale online education course that integrates thousands of well-designed online courses through the network, providing learners with a rich choice of courses. The MOOC platform is favored by students for its flexible learning atmosphere and low pressure, but this free and open environment is also accompanied by a high dropout rate. Due to the large number of online education groups, manual dropout prediction is difficult to achieve. Automated dropout prediction methods based on machine learning have been widely used in the field of MOOC dropout prediction.

[0003] Current prediction methods often only focus on students' personal information and course behavior information, and use some simple machine learning methods to predict students as individuals, while the information contained in the discussion area is ignored. At present, there is no use of students' speeches and interaction records in the discussion area to predict dropouts. First, students' posts and replies in the discussion area often contain a lot of semantic information, which can be used to judge students' learning situation in the course; secondly, students' interactions in the comment area allow different students' academic evaluations to refer to each other. The difficulty of using discussion area data is that although the data in the discussion area is suitable for composition processing, the levels of participants in the discussion are often uneven, and the methods generally used to process graph data may not be effective; the massive text data in the comment area also brings difficulties to data processing. Summary of the invention

[0004] The purpose of the present invention is to solve the problems existing in the prior art and to provide a method and system for predicting dropouts of MOOCs using a large language model to assist graph node classification.

[0005] In order to achieve the above-mentioned invention object, the present invention specifically adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a method for predicting dropouts of MOOCs using a large language model-assisted graph node classification, which comprises the following steps:

[0007] S1: Obtain the course behavior information data of each learner in the target course from the MOOC platform, encode the pre-processed course behavior information data, and obtain the learner's course behavior characteristic data;

[0008] S2: Take each learner as a node in the discussion graph, the interaction relationship between learners in the discussion area as the edge connecting the nodes in the discussion graph, and the learner's course behavior feature data as the node feature to obtain the initial discussion graph, and process the initial discussion graph to obtain the final discussion graph;

[0009] S3: A randomized singular value decomposition method is used to replace the random initialization process of the LoRA parameter matrix in the O-LoRA method. The O-LoRA method is used to fine-tune the pre-trained large language model based on the LoRA parameter matrix after randomized singular value decomposition to obtain a fine-tuned large language model.

[0010] S4: inputting a set of first examples including questions and answers and a first additional question to form a first prompt word into the fine-tuned large language model, and the fine-tuned large language model outputs the cognitive level classification result corresponding to the learner under the target chapter according to the answer in the first example, and obtains a cognitive feature matrix based on the cognitive level classification result;

[0011] S5: input a set of second examples including questions and answers and a second additional question to form a second prompt word into the fine-tuned large language model, and the fine-tuned large language model determines whether the course behavior types of the two learners are the same, and outputs the course behavior type judgment result according to the answer in the second example, and encodes the course behavior type judgment result, and uses the course behavior type encoding result as the pseudo label of each edge in the final discussion graph;

[0012] S6: Take the learner's cognitive feature matrix and the final discussion graph as the input of the MOOC dropout prediction model to train the MOOC dropout prediction model;

[0013] S7: Input the course behavior information data of the learner to be predicted into the trained MOOC dropout prediction model, and output the predicted probability that the learner is a dropout.

[0014] Based on the above scheme, each step can be implemented in the following preferred specific manner.

[0015] As a preferred embodiment of the first aspect, in step S1, each course behavior information data includes N video N teaching videos for learners to watch, work Completion status of tasks, N test The test results of the test, the total number of times the learner visits the target course N visit and text data consisting of all speeches made by learners in the target course discussion area, and preprocessing the text data to form preprocessed course behavior information data;

[0016] In step S1, the specific process of encoding the pre-processed course behavior information data is as follows:

[0017] S11: For each teaching video of each learner, the ratio between the viewing time of the learner and the total viewing time of the corresponding teaching video is used as the viewing time coding value of the teaching video, and all viewing time coding values ​​of the learner constitute a corresponding viewing time coding vector;

[0018] S12: The total number of times the learner visits the target course is used as the access frequency coding vector corresponding to the learner;

[0019] S13: For each assignment completion status of each learner, if the learner completes the assignment, the ratio between the number of correct answers of the learner in each assignment and the corresponding total number of questions is used as the assignment submission status coding value; if the learner does not complete the assignment, the assignment submission status coding value is 0. Finally, all the assignment submission status coding values ​​of the learner constitute the corresponding assignment submission status coding vector;

[0020] S14: For each test case of each learner, if the learner completes the test, the ratio between the learner's test score and the corresponding test total score is used as the test score coding value; if the learner does not complete the test, the test score coding value is 0, and all the test score coding values ​​of the learner constitute the corresponding test score coding vector;

[0021] S15: for each learner's preprocessed text data, the preprocessed text data is sent to the encoding visible large language model to obtain the learner's discussion content encoding vector;

[0022] S16: Concatenate the viewing time coding vector, the access frequency coding vector, the homework submission status coding vector, the test score coding vector and the discussion content coding vector as the learner's course behavior characteristic data.

[0023] As a preferred embodiment of the above-mentioned first aspect, in step S2, the specific process of obtaining the final discussion graph is: determine whether there is an isolated node in the initial discussion graph: if so, traverse each isolated node in the initial discussion graph, calculate the cosine similarity between the node feature corresponding to an isolated node and the node features of the remaining nodes, and use the K nodes with the largest cosine similarity as the most similar nodes to the isolated node, connect the isolated node and its corresponding most similar nodes to form K edges and add them to the initial discussion graph to obtain the final discussion graph; if not, use the initial discussion graph as the final discussion graph.

[0024] As a preferred embodiment of the first aspect, in step S3, the specific process of obtaining the LoRA parameter matrix after randomized singular value decomposition by using the randomized singular value decomposition method is:

[0025] S31: Randomly sample from a normal distribution as the column vector of the random matrix to generate a complete random matrix;

[0026] S32: Use a random matrix to project the parameter matrix of the large language model into a low-dimensional space to obtain a projection matrix; perform orthogonal triangular decomposition on the projection matrix to obtain a first orthogonal matrix and an upper triangular matrix; transpose the first orthogonal matrix and multiply it with the parameter matrix of the large language model to obtain a low-dimensional matrix; perform singular value decomposition on the low-dimensional matrix to obtain a second orthogonal matrix, a third orthogonal matrix, and a diagonal matrix;

[0027] S33: The index corresponding to the largest r singular values ​​in the diagonal matrix is ​​used as the initialization index, the column vector corresponding to the initialization index in the second orthogonal matrix is ​​used as the initialized first LoRA parameter matrix, and the column vector corresponding to the initialization index in the third orthogonal matrix is ​​used as the initialized second LoRA parameter matrix. The initialized first LoRA parameter matrix and the initialized second LoRA parameter matrix constitute a set of LoRA parameter matrices after randomized singular value decomposition.

[0028] As a preferred embodiment of the first aspect, in step S4, the question in the first example and the first additional question both include all speeches of the learner in the target chapter of the target course discussion area, the name of the target course and the name of the target chapter, the cognitive level of the learner determined based on the learner's speech, the cognitive level of the online community discussion and the judgment criteria corresponding to each cognitive level; the answer in the first example is designed in the form of a thinking chain prompt, and the answer in the first example includes the classification result of the learner's cognitive level and the logical reasoning process for determining the cognitive level to which the learner belongs;

[0029] In step S4, the generation process of the cognitive feature matrix is ​​as follows: the cognitive level classification results are one-hot encoded to obtain the cognitive level encoding results, the cognitive level encoding results of all chapters constitute the learner's cognitive feature vector, the cognitive feature vector of a learner is used as a row in the cognitive feature matrix, and the cognitive feature vectors of all learners constitute a complete cognitive feature matrix.

[0030] As a preferred embodiment of the above-mentioned first aspect, in step S5, the question in the second example and the second additional question both contain course behavior information data of the two learners and determine whether the course behavior types of the two learners are the same, and the answer in the second example is designed in accordance with the thinking chain prompt, and the answer in the second example includes the judgment result of the course behavior type and the logical reasoning process for determining whether the course behavior types are the same; wherein, when the two learners are both dropouts or both are non-dropouts, the course behavior types of the two learners are deemed to be the same.

[0031] As a preferred embodiment of the first aspect, in step S6, during the training of the MOOC dropout prediction model, the course behavior information matrix is ​​formed by the course behavior information data of all learners, the course behavior information matrix is ​​concatenated with the learner's cognitive feature matrix as the learner feature matrix, two multi-layer perceptrons are used to output the same matching probability of each edge according to the edge set in the final discussion graph, the same matching probability is adjusted using the pseudo label of each edge of the final discussion graph to obtain the adjusted same matching probability, the adjusted same matching probability is used as the element in the low-pass adjacency matrix, the low-pass adjacency matrix is ​​constructed, and the learner feature matrix is ​​input into In the simplified graph convolutional network, the learner feature matrix is ​​first processed by the third multi-layer perceptron to output the initial low-pass feature matrix. The initial low-pass feature matrix is ​​processed by multiple low-pass filtering. In each low-pass filtering process, the low-pass filtering result obtained last time is multiplied by the low-pass adjacency matrix to obtain the current low-pass filtering result. The last low-pass filtering result is used as the learner low-pass feature matrix, and the result of subtracting 1 from the adjusted homonymous probability is used as the element in the high-pass adjacency matrix to construct the high-pass adjacency matrix. The unit matrix is ​​subtracted from the weighted high-pass adjacency matrix to obtain the processed result. The high-pass adjacency matrix of the learner is input into the Lap-SGC model. The learner feature matrix is ​​first processed by the fourth multi-layer perceptron to output the initial high-pass feature matrix. The initial high-pass feature matrix is ​​processed by multiple high-pass filtering. In each high-pass filtering process, the high-pass filtering result obtained in the last time is multiplied by the processed high-pass adjacency matrix to obtain the current high-pass filtering result. The high-pass filtering result obtained in the last time is used as the learner high-pass feature matrix. The learner low-pass feature matrix and the learner high-pass feature matrix are concatenated as the learner comprehensive representation matrix. After the discussion graph is embedded in position space, a nearest neighbor graph is obtained, and the nearest neighbor graph is input into the graph convolutional network to obtain the learner encoding matrix. The learner comprehensive representation matrix and the learner encoding matrix are spliced ​​and input into the fifth multi-layer perceptron, and the predicted probability of each learner being a dropout is output. The cross entropy loss between the predicted probability of each learner being a dropout and the learner's actual course behavior label is calculated, and the parameters of the MOOC dropout prediction model are updated based on minimizing the cross entropy loss until the preset iteration round threshold is reached, the MOOC dropout prediction model converges, and a trained MOOC dropout prediction model is obtained.

[0032] As a preferred embodiment of the first aspect, the specific process of using two multi-layer perceptrons to output the same matching probability of each edge is as follows: the learner feature vectors of the two nodes corresponding to each edge are respectively passed through the first multi-layer perceptron, and the corresponding outputs are the first intermediate representation vector and the second intermediate representation vector representing the two learners respectively, the first intermediate representation vector and the second intermediate representation vector are concatenated and passed through the second multi-layer perceptron to obtain the third intermediate representation vector, the second intermediate representation vector and the first intermediate representation vector are concatenated and passed through the second multi-layer perceptron to obtain the fourth intermediate representation vector, and the third intermediate representation vector and the fourth intermediate representation vector are averaged to obtain the same matching probability corresponding to each edge; wherein the learner feature vector is a row in the learner feature matrix;

[0033] The specific process of adjusting the probability of matching is as follows: when the course behavior types of two learners are different, The minimum value between and 0 is taken as the adjusted probability of matching; when the course behavior types of two learners are the same, The maximum value between 0 and 0 is taken as the adjusted homogamy probability; is the probability of homogamy before adjustment; are the preset hyperparameters.

[0034] As a preferred embodiment of the first aspect above, the specific process of obtaining the nearest neighbor graph after embedding the final discussion graph in position space is as follows: first, an embedding operation is performed on each node in the final discussion graph to map it to an implicit dimension of a position space, and the average value of the learner feature vectors of all first-order neighbor nodes of each node in the final discussion graph is used as the position encoding result of the node, and the L2 distance between the position encoding result of each node and the position encoding results of the remaining nodes in the final discussion graph is calculated, and the M nodes with the smallest L2 distance are used as the nearest neighbor nodes of each node, each node in the final discussion graph is connected to its respective nearest neighbor node, and the learner feature matrix is ​​used as the node feature to construct the nearest neighbor graph.

[0035] Furthermore, for the final discussion of the nodes in the graph , the specific form of the position encoding result is as follows:

[0036]

[0037] In the formula, Indicates the final discussion of the nodes in the graph The position encoding result of represents the degree of a node with a self-loop; Represents the first-order neighbor set of a node with a self-loop; Represents a first-order neighbor node in the first-order neighbor set; Represents the first-order neighbor node The corresponding learner feature vector.

[0038] In a second aspect, the present invention provides a MOOC dropout warning system with large language model-assisted graph node classification, comprising:

[0039] The data acquisition module is used to obtain the course behavior information data of each learner in the target course from the MOOC platform, encode the pre-processed course behavior information data, and obtain the learner's course behavior characteristic data;

[0040] The discussion graph construction module is used to take each learner as a node in the discussion graph, the interaction relationship between learners in the discussion area as the edge connecting the nodes in the discussion graph, and the course behavior feature data of the learners as the node feature to obtain the initial discussion graph, and process the initial discussion graph to obtain the final discussion graph;

[0041] A model fine-tuning module is used to replace the random initialization process of the LoRA parameter matrix in the O-LoRA method with a randomized singular value decomposition method, and the O-LoRA method is used to fine-tune the pre-trained large language model based on the LoRA parameter matrix after randomized singular value decomposition to obtain a fine-tuned large language model;

[0042] A cognitive hierarchy analysis module, for inputting a set of first examples including questions and answers and a first additional question constituting a first prompt word into a fine-tuned large language model, and the fine-tuned large language model outputs a cognitive hierarchy classification result corresponding to the learner under the target chapter according to the answer in the first example, and obtains a cognitive feature matrix based on the cognitive hierarchy classification result;

[0043] a pseudo-label acquisition module, for inputting a set of second examples including questions and answers and a second additional question constituting a second prompt word into the fine-tuned large language model, and for the fine-tuned large language model to determine whether the course behavior types of the two learners are the same, and outputting a course behavior type determination result according to the answer in the second example, and encoding the course behavior type determination result, and using the course behavior type encoding result as a pseudo-label for each edge in the final discussion graph;

[0044] A model training module is used to use the learner's cognitive feature matrix and the final discussion graph as inputs of the MOOC dropout prediction model to train the MOOC dropout prediction model;

[0045] A result acquisition module is used to input the course behavior information data of the learner to be predicted into the trained MOOC dropout prediction model, and output the predicted probability that the learner is a dropout;

[0046] The early warning module is used to send dropout warning information to teacher users or management users of the MOOC platform, or to send reminder information to learners who are dropouts to urge them to study online.

[0047] Compared with the traditional MOOC dropout prediction method based on machine learning, the present invention has the following beneficial effects:

[0048] The present invention uses a large language model to perform in-depth semantic analysis on students' posts and replies in the discussion area, which can more accurately identify students' understanding of course knowledge points and learning confusion, thereby more comprehensively assessing their dropout risk. The present invention combines course behavior data and discussion area interaction records by constructing a discussion graph, so that the MOOC dropout prediction model can learn the interactive behavior between students, effectively capture the implicit student learning motivation, participation and collaboration in the discussion area, and thus provide more accurate dropout predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the steps of the method of the present invention;

[0050] Figure 2 A schematic diagram of a final discussion diagram provided for an embodiment of the present invention;

[0051] Figure 3 A schematic diagram of the process of training a MOOC dropout prediction model according to the present invention;

[0052] Figure 4 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0053] In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0054] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

[0055] like Figure 1 As shown, in a preferred implementation of the present invention, the above-mentioned large language model assisted graph node classification MOOC dropout prediction method includes the following steps S1 to S7. The specific implementation process is described in detail below.

[0056] S1: Obtain the course behavior information data of each learner in the target course from the MOOC platform, encode the preprocessed course behavior information data, and obtain the learner's course behavior feature data.

[0057] It should be noted that in step S1 of the present invention, each course behavior information data includes N video N teaching videos for learners to watch, work Completion status of tasks, N test The test results of the test, the total number of times the learner visits the target course N visit And the text data consisting of all speeches made by learners in the target course discussion area, and the text data is preprocessed to form preprocessed course behavior information data.

[0058] In this embodiment, when preprocessing the text data, the text data is cleaned, specifically, irrelevant information in the text data is removed, such as advertisements, emoticons, irrelevant links, etc., to ensure the quality of the analyzed text data. Then, the text data is normalized, including conversion to a unified character encoding, removal of redundant spaces, and consistent use of punctuation marks.

[0059] It should be noted that in step S1 of the present invention, for each learner in the target course, the present invention involves encoding detailed behavioral information of the learner's activities in the target course, such as video viewing, homework submission and test participation, so as to reflect the learner's viewing of each teaching video, completion of homework and test scores, so as to facilitate the analysis of their learning behavior patterns and participation, thereby providing data support for further learning analysis and personalized feedback.

[0060] In step S1 of this embodiment, the specific process of encoding the pre-processed course behavior information data is as follows:

[0061] S11: For each teaching video of each learner, the ratio between the viewing time of the learner and the total viewing time of the corresponding teaching video is used as the viewing time encoding value of the teaching video, and all viewing time encoding values ​​of the learner constitute the corresponding viewing time encoding vector ;

[0062] S12: The total number of times a learner visits the target course is used as the access frequency encoding vector corresponding to the learner. ;

[0063] S13: For each learner's homework completion status, if the learner completes the homework, the ratio between the number of correct answers to each homework and the corresponding total number of questions is used as the homework submission status coding value. If the learner does not complete the homework, the homework submission status coding value is 0. Finally, all the homework submission status coding values ​​of the learner constitute the corresponding homework submission status coding vector ;

[0064] S14: For each test case of each learner, if the learner completes the test, the ratio between the learner's test score and the corresponding test total score is used as the test score coding value. If the learner does not complete the test, the test score coding value is 0. All the test score coding values ​​of the learner constitute the corresponding test score coding vector ;

[0065] S15: For each learner’s preprocessed text data, the preprocessed text data is fed into a large language model (such as BERT) with visible encoding to obtain the learner’s discussion content encoding vector ;

[0066] S16: Concatenate the viewing time coding vector, access frequency coding vector, homework submission status coding vector, test score coding vector, and discussion content coding vector as the learner's course behavior feature data .

[0067] S2: Take each learner as a node in the discussion graph, the interaction relationship between learners in the discussion area as the edge connecting the nodes in the discussion graph, and the learner's course behavior feature data as the node feature to obtain the initial discussion graph. Process the initial discussion graph to obtain the final discussion graph.

[0068] It should be noted that in step S2 of the present invention, in the process of analyzing the discussion area data, the method of constructing a discussion graph is particularly critical. This method can effectively record and measure the communication patterns and behavioral characteristics between learners, while flexibly responding to the evolving information in the discussion area. In this graph, learners are regarded as nodes, and the interactions between learners are reflected by the edges connecting these nodes. The construction of the discussion graph is to apply advanced analysis techniques such as graph neural networks to deeply explore and learn the complex patterns of student behavior. Therefore, the present invention constructs a final discussion graph to capture the interactions of learners in the discussion area.

[0069] In step S2 of this embodiment, the specific process of obtaining the final discussion graph is as follows: determine whether there is an isolated node in the initial discussion graph: if so, traverse each isolated node in the initial discussion graph, calculate the cosine similarity between the node feature corresponding to an isolated node and the node features of the remaining nodes, and use the K nodes with the largest cosine similarity as the most similar nodes to the isolated node, connect the isolated node and its corresponding most similar nodes to form K edges and add them to the initial discussion graph to obtain the final discussion graph; if not, use the initial discussion graph as the final discussion graph.

[0070] like Figure 2 As shown in the figure, if learner A initiates a discussion, and learners B and C respond to it, and learner D responds to learner B's response, then the edges {A, B}, {A, C} and {B, D} are constructed in the discussion graph respectively.

[0071] S3: A randomized singular value decomposition (SVD) method is used to replace the random initialization process of the LoRA parameter matrix in the O-LoRA method. Based on the LoRA parameter matrix after randomized singular value decomposition, the O-LoRA method is used to fine-tune the pre-trained large language model to obtain a fine-tuned large language model.

[0072] It should be noted that the O-LoRA method is a method for fine-tuning a large language model, and its implementation method belongs to the prior art. The large language model fine-tuned by the O-LoRA method can be used to achieve different tasks. The O-LoRA method first randomly initializes the LoRA parameter matrix. In the present invention, the randomized singular value decomposition (SVD) method is used to replace the original random initialization process of the LoRA parameter matrix in the O-LoRA method to obtain the LoRA parameter matrix after randomized singular value decomposition. This randomized singular value decomposition method is usually more computationally expensive than the complete singular value decomposition method, because it does not need to calculate the complete singular value decomposition process, but approximates it through random sampling, and initializes it by the randomized singular value decomposition method, which can promote the model to approach the effective solution space more quickly. Then, the processing is carried out according to the original steps of the O-LoRA method, and finally a fine-tuned large language model is obtained. As a result, the present invention uses a large language model to efficiently process two different tasks, while ensuring that the large language model will not forget the valuable knowledge acquired in the pre-training stage during the continuous learning process, thereby optimizing the performance of the large language model.

[0073] In step S3 of the present invention, the specific process of obtaining the LoRA parameter matrix after randomized singular value decomposition by using the randomized singular value decomposition method is:

[0074] S31: Generate a complete random matrix by randomly sampling from a normal distribution as the column vector of the random matrix.

[0075] In this embodiment, a random matrix S is first created with a dimension of k×(r+p), where k is the number of columns of the parameter matrix of the large language model; r is a positive integer much smaller than the number of columns k and the number of rows d of the parameter matrix of the large language model; p is an oversampling parameter, which is a value slightly larger than r. The column vectors of the random matrix are randomly drawn from a normal distribution to capture the key features of the column space of the parameter matrix of the large language model. This strategy can approximate the singular value decomposition of the parameter matrix of the large language model with higher computational efficiency.

[0076] S32: Use the random matrix S to project the parameter matrix W of the large language model into a low-dimensional space to obtain the projection matrix T=WS; perform orthogonal triangular (QR) decomposition on the projection matrix T to obtain the first orthogonal matrix Q and the upper triangular matrix R; transpose the first orthogonal matrix and multiply it with the parameter matrix W of the large language model to obtain a low-dimensional matrix D; perform singular value decomposition (SVD) on the low-dimensional matrix D to obtain the second orthogonal matrix , the third orthogonal matrix And the diagonal matrix ;

[0077] S33: The largest diagonal array The indices corresponding to the singular values ​​are used as initialization indices in the second orthogonal matrix The column vector corresponding to the initialization index is used as the first LoRA parameter matrix to be initialized, and the third orthogonal matrix In , the column vector corresponding to the initialization index is used as the initialized second LoRA parameter matrix, and the initialized first LoRA parameter matrix and the initialized second LoRA parameter matrix constitute a set of LoRA parameter matrices after randomized singular value decomposition.

[0078] In step S3, for the first cognitive hierarchy analysis task, the present invention first generates a random matrix corresponding to the first task, and executes the above-mentioned S31-S33 process to obtain a set of randomized singular value decomposition LoRA parameter matrices corresponding to the first task. ,in, The first LoRA parameter matrix and the second LoRA parameter matrix initialized for the first task are respectively. For the second edge pseudo-label identification, the present invention also first generates a random matrix corresponding to the second task, and executes the above S31~S33 process to obtain a set of randomized singular value decomposition LoRA parameter matrices corresponding to the second task ,in, The first LoRA parameter matrix and the second LoRA parameter matrix initialized corresponding to the second task respectively.

[0079] S4: A set of first examples including questions and answers and a first additional question constituting a first prompt word is input into the fine-tuned large language model, and the fine-tuned large language model outputs the cognitive level classification result corresponding to the learner under the target chapter according to the answer in the first example, and obtains a cognitive feature matrix based on the cognitive level classification result.

[0080] It should be noted that in step S4 of the present invention, the question in the first example and the first additional question both include all speeches of the learner under the target chapter in the target course discussion area, the name of the target course and the name of the target chapter, the cognitive level to which the learner belongs based on the learner's speech, the cognitive level of online community discussion and the judgment criteria corresponding to each cognitive level; the answer in the first example is designed in the form of a thinking chain prompt, and the answer in the first example includes the learner's cognitive level classification result and the logical reasoning process for judging the learner's cognitive level.

[0081] In this embodiment, it is considered that the large language model has a series of significant advantages when performing cognitive level analysis tasks of online discussions, such as its excellent automation ability, which can quickly process and analyze large-scale discussion data and significantly improve work efficiency. Through deep learning technology training, the large language model can understand the complexity of language and capture the nuances and deep meanings in the discussion content. In addition, the large language model has strong generalization ability and can adapt to different topics and diverse expression styles to ensure the accuracy and universality of the analysis results. The large language model can also reduce the deviation caused by human factors and provide more objective and consistent analysis results. Therefore, the method of thinking chain prompts in the present invention designs the large language model prompt word prompt, and the logical reasoning process is as step-by-step as possible to teach the large language model to think step by step. Even in the case of only a small number of examples, the large language model can also show complex reasoning ability, and finally help complete the cognitive level analysis of the discussion content. In addition to the method of thinking chain prompts, this embodiment also designs the online community discussion cognitive level and the judgment criteria corresponding to each cognitive level into the large language model prompt word. Specifically, the cognitive level evaluation table for online community discussions proposed by Henri in 1992 was used to add the five cognitive levels of online discussion content of discussants and the judgment criteria of each cognitive level into the large language model prompt words. The five cognitive levels are basic identification, deep identification, reasoning, judgment and strategy. Among them, the judgment criteria of basic identification are: in the cognitive process, first observe and study the problem, identify its basic elements, and explore the connection between these elements, so as to build a basic understanding of the problem; the judgment criteria of deep identification are: on the basis of basic understanding, further analyze the problem and deeply understand the values, beliefs and assumptions behind it. This level requires a deeper description of the problem to form a more comprehensive understanding; the judgment criteria of reasoning are: adopting the strategy of logical reasoning, adopting existing views or proposing new views. In this process, the recognized propositions serve as the basis for building new connections and arguments, reflecting the logical reasoning ability in the cognitive process; the judgment criteria of judgment are: judgment involves multiple forms such as decision-making, description, appreciation and criticism. The learner's judgment is based on certain values ​​or rules, reflecting the individual's ability to evaluate and select information based on specific criteria; the judgment criteria for strategy are: based on understanding and judgment, propose a coordinated action strategy for implementation, or make a choice among multiple options and make a decision. This level reflects the learner's ability to develop effective strategies to achieve goals when facing problems.

[0082] In this embodiment, the first prompt word designed is as follows: "Question: The following are the online discussion cognitive level judgment standards, which are divided into five levels. Basic identification: <Basic identification judgment standard>; Deep identification: <Deep identification judgment standard>; Reasoning: <Reasoning judgment standard>; Judgment: <Judgment judgment standard>; Strategy: <Strategy judgment standard>; Learner A posted a discussion content under the "Learning Theory" chapter of the "Introduction to Modern Psychology" course: "I believe that reinforcement learning theory can explain how rewards promote the frequent occurrence of specific behavior patterns. "Please analyze A's cognitive level for the knowledge points contained in this chapter. Answer: Learner A's speech directly quoted the concept of "reinforcement learning theory" and applied it to explain the changes in behavior frequency. This shows that A is able to recognize and use the basic terms and concepts in the course. A's speech did not show an exploration of the deeper meaning behind reinforcement learning theory, nor did it provide additional analysis or critical thinking. Since the speech mainly focused on the basic recognition of concepts without further reasoning or in-depth analysis, it was judged to be at the basic recognition level. Question: The following are the judgment standards for the cognitive level of online discussions, which are divided into five levels. Basic recognition: <basic recognition judgment standard>; deep recognition: <deep recognition judgment standard>; reasoning: <reasoning judgment standard>; judgment: <judgment judgment standard>; strategy: <strategy judgment standard>; please analyze the cognitive level of the discussion content I entered based on this standard and the course and chapter names I will enter next, and tell me the cognitive level it belongs to and your reasons for your judgment. Course name: <course name> Chapter name: <chapter name> Discussion content: <specific discussion content>."

[0083] It should be noted that in step S4 of the present invention, the generation process of the cognitive feature matrix is ​​specifically as follows: the cognitive level classification results are one-hot encoded to obtain the cognitive level encoding results, the cognitive level encoding results of all chapters constitute the learner's cognitive feature vector, the cognitive feature vector of a learner is used as a row in the cognitive feature matrix, and the cognitive feature vectors of all learners constitute a complete cognitive feature matrix.

[0084] S5: Input a set of second examples including questions and answers and a second additional question to form a second prompt word into the fine-tuned large language model, and the fine-tuned large language model determines whether the course behavior types of the two learners are the same, and outputs the course behavior type judgment result according to the answer in the second example, and encodes the course behavior type judgment result, and uses the course behavior type encoding result as the pseudo label of each edge in the final discussion graph.

[0085] It should be noted that in step S5 of the present invention, the question in the second example and the second additional question both contain the course behavior information data of the two learners and determine whether the course behavior types of the two learners are the same. The answer in the second example is designed in accordance with the thinking chain prompt, and the answer in the second example includes the judgment result of the course behavior type and the logical reasoning process for determining whether the course behavior types are the same; wherein, when the two learners are both dropouts or both are non-dropouts, the course behavior types of the two learners are deemed to be the same.

[0086] In this embodiment, the designed second prompt is as follows: "Question: The following is the behavior information of two learners in the MOOC course. Please help me judge whether they belong to the same type of learners, that is, whether they are both dropouts or non-dropouts, and give your reasons for judgment. Learner A: Total course video viewing time: 300 minutes Total course video duration: 500 minutes; Course access frequency: 47 times; Homework submission: 1, 1, 1, 1; Test status: 87, 95, 90; Learner B: Total course video viewing time: 200 minutes; Course video duration: 500 minutes; Course access frequency: 16 times; Homework submission: 0, 1, 0, 1; Test status: 54, 0, 63. Answer: Video viewing: Learner A watched 60% of the total course time (300 / 500), while learner B only watched 40% (200 / 500). This reflects that A has better learning participation. Course access frequency: Learner A visits 3 times a week, while learner B visits once a week. More frequent access behavior may indicate stronger learning enthusiasm. Homework submission: Learner A submitted all homework, while learner B's homework submission was unstable and there were records of non-submission. Submitting homework is an important indicator of participation. Test scores: Learner A's test scores are generally excellent, which suggests that his learning effect is good. Learner B's test scores fluctuate greatly, and there are records of not taking exams, which may be a sign of insufficient learning participation. Conclusion: According to the above analysis, learner A shows high learning participation and learning effect, and is unlikely to be a dropout. Learner B's performance in video viewing, course access frequency, homework submission and test scores is unstable, showing low learning participation. They may be dropouts. Therefore, they are not the same type of learners. Question: The following is the behavior information of two learners in the MOOC course. Please help me judge whether they are the same type of learners, that is, whether they are both dropouts or non-dropouts, and give your reasons for your judgment. Learner A: Total viewing time of course videos: <total viewing time>, total course video duration: <total duration>, course access frequency: <access frequency>, homework submission status: 0 means not submitted, 1 means submitted, homework submission sequence is <1 , 1...1>, <Test status>: 0 means that the student did not take the test, and the test score sequence is <87, 95...90>; Learner B: Total course video viewing time: <Total viewing time>, Total course video duration: <Total duration>, Course access frequency: <Access frequency>, Homework submission status: 0 means not submitted, 1 means submitted, and the homework submission sequence is <0, 1...0>, <Test status>: 0 means that the student did not take the test, and the test score sequence is <54, 0...63>. "

[0087] It should be noted that in step S5 of the present invention, when encoding the course behavior type judgment result, when the course behavior types of two learners are the same, they are encoded as 1, and when the course behavior types of the two learners are different, they are encoded as 0, thereby obtaining the course behavior type encoding result.

[0088] S6: Take the learner’s cognitive feature matrix and the final discussion graph as the input of the MOOC dropout prediction model to train the MOOC dropout prediction model.

[0089] In step S6, during the training of the MOOC dropout prediction model, the course behavior information matrix is ​​constructed from the course behavior information data of all learners, and the course behavior information matrix is ​​concatenated with the learner's cognitive feature matrix as the learner feature matrix. Two multi-layer perceptrons are used to output the same matching probability of each edge according to the edge set in the final discussion graph, and the same matching probability is adjusted using the pseudo label of each edge in the final discussion graph to obtain the adjusted same matching probability. The adjusted same matching probability is used as an element in the low-pass adjacency matrix to construct the low-pass adjacency matrix, and the learner feature matrix is ​​input into the simplified graph convolutional network. First, the learner feature matrix is ​​processed by the third multi-layer perceptron to output the initial low-pass feature matrix. The initial low-pass feature matrix is ​​processed by multiple low-pass filtering. In each low-pass filtering process, the low-pass filtering result obtained last time is multiplied by the low-pass adjacency matrix to obtain the current low-pass filtering result. The last low-pass filtering result is used as the learner low-pass feature matrix. The result after subtracting 1 from the adjusted homonymy probability is used as the element in the high-pass adjacency matrix to construct the high-pass adjacency matrix. The unit matrix is ​​subtracted from the weighted high-pass adjacency matrix to obtain the processed high-pass adjacency matrix. Array, input the learner feature matrix into the Lap-SGC model, first process the learner feature matrix through the fourth multi-layer perceptron, output the initial high-pass feature matrix, and process the initial high-pass feature matrix through multiple high-pass filtering. In each high-pass filtering process, the high-pass filtering result obtained in the last time is multiplied by the processed high-pass adjacency matrix to obtain the current high-pass filtering result. The high-pass filtering result obtained in the last time is used as the learner high-pass feature matrix, and the learner low-pass feature matrix and the learner high-pass feature matrix are concatenated as the learner comprehensive representation matrix. The final discussion diagram After position space embedding, a nearest neighbor graph is obtained, and the nearest neighbor graph is input into a graph convolutional network to obtain a learner encoding matrix. The learner comprehensive representation matrix and the learner encoding matrix are concatenated and input into a fifth multi-layer perceptron, and the predicted probability of each learner being a dropout is output. The cross entropy loss between the predicted probability of each learner being a dropout and the learner's actual course behavior label is calculated, and the parameters of the MOOC dropout prediction model are updated based on minimizing the cross entropy loss until a preset iteration round threshold is reached, the MOOC dropout prediction model converges, and a trained MOOC dropout prediction model is obtained.

[0090] It should be noted that, in the present invention, the specific process of using two multilayer perceptrons to output the same matching probability of each edge is: the learner feature vectors of the two nodes corresponding to each edge are respectively passed through the first multilayer perceptron (Multilayer Perceptron, MLP), and the corresponding outputs represent the first intermediate representation vector and the second intermediate representation vector of the two learners respectively, the first intermediate representation vector and the second intermediate representation vector are concatenated and passed through the second multilayer perceptron to obtain the third intermediate representation vector, the second intermediate representation vector and the first intermediate representation vector are concatenated and passed through the second multilayer perceptron to obtain the fourth intermediate representation vector, the third intermediate representation vector and the fourth intermediate representation vector are averaged to obtain the same matching probability corresponding to each edge; wherein, the learner feature vector is a row in the learner feature matrix.

[0091] In this embodiment, for the edge connecting learner a and learner b in the final discussion graph , the following formula is used to estimate the edge The probability of assortativeness is:

[0092]

[0093]

[0094]

[0095] in, Represents a matrix concatenation operation; and They represent the feature vector of learner a and the feature vector of learner b, respectively, and both are derived from a row in the learner feature matrix; and They represent the intermediate representation vectors obtained after embedding the feature vectors of learner a and learner b respectively; and Represents two multi-layer perceptrons. The present invention transforms and The feature embedding method is used together to discuss the undirected nature of the edges in the graph data. The resulting edges The probability of homogeneity is the final discussion graph edge The probability that the two learners being connected are of the same type.

[0096] It should be noted that in the present invention, after obtaining the probability of matching each edge, the probability of matching is further adjusted. When the course behavior types of two learners are different, The minimum value between and 0 is taken as the adjusted matching probability; when the course behavior types of two learners are the same, then The maximum value between 0 and 0 is taken as the adjusted homogamy probability; is the probability of homogeneity before adjustment. The above specific process is expressed by formula:

[0097]

[0098] in, is a hyperparameter used to adjust the probability of edge matching. In this embodiment, The value of is between 0 and 0.5; is the adjusted probability of homogamy; represents the pseudo-label of the edge corresponding to node (learner) a and node (learner) b, when When , it means that the course behavior types of the two learners are the same. When , it indicates that the course behavior types of the two learners are different; and They represent the maximum value function and the minimum value function respectively.

[0099] In this embodiment, when the large language model identifies an edge connecting learners of the same course behavior type, the edge should be increased accordingly. In order to strengthen the model's understanding of the consistency of learners' course behavior types. On the contrary, if the edge connects learners with different course behavior types, the edge will be reduced. The probability value of the same match is used to reflect the model's recognition of the differences in learners' course behavior types.

[0100] It should be noted that, in the present invention, according to the adjusted edge matching probability, the present invention performs adaptive encoding on the final discussion graph. Specifically, a basic low-pass graph neural network (GNN), namely a simplified graph convolution network (SGC), is used to encode the learner feature matrix Perform low-pass filtering. This process aims to capture and retain the similarity feature information between adjacent nodes. The simplified low-pass filtering operation of the graph convolutional network can be expressed as:

[0101]

[0102]

[0103] in, is a low-pass adjacency matrix, and the element in the a-th row and b-th column of the low-pass adjacency matrix is ​​the adjusted assortative probability of the edge connecting learner a and learner b, that is, ; represents the initial low-pass feature matrix; represents the third multi-layer perceptron; and Respectively represent Second and The low-pass filtering result obtained by this step; is the filtering times index; The low-pass filtering result is That is the learner low-pass feature matrix.

[0104] After completing the low-pass filtering, the present invention uses a Lap-SGC model to perform high-pass filtering on the learner feature matrix, thereby capturing the difference feature information between learners. First, construct a high-pass adjacency matrix The element in the ath row and bth column of the high-pass adjacency matrix is ​​1 minus the adjusted assortative probability of the edge connecting learner a and learner b, that is, Then start executing Second high pass filtering process.

[0105]

[0106]

[0107]

[0108] in, is the processed high-pass adjacency matrix; is a hyperparameter that controls the filtering strength; represents the initial high-pass feature matrix; Represents the fourth multi-layer perceptron; and Respectively represent Second and The high-pass filtering result obtained in the first The high-pass filtering result is obtained That is the learner’s high-pass feature matrix.

[0109] It should be noted that in the present invention, the learner's comprehensive representation is supplemented by searching for the same matching nodes of the learner node in the entire graph. This process can overcome the heterogeneity of the discussion graph data to a certain extent. Therefore, the present invention first embeds the final discussion graph in the position space to obtain the nearest neighbor graph. The specific process is: first, an embedding operation is performed on each node in the final discussion graph, and it is mapped to an implicit dimension of a position space. The average value of the learner feature vectors of all first-order neighbor nodes of each node in the final discussion graph is used as the position encoding result of the node, and the L2 distance between the position encoding result of each node and the position encoding results of the remaining nodes in the final discussion graph is calculated. The M nodes with the smallest L2 distance are used as the nearest neighbor nodes of each node, and each node in the final discussion graph is connected to its respective nearest neighbor node, and the learner feature matrix is ​​used as the node feature to construct the nearest neighbor graph.

[0110] In this embodiment, for the final discussion of the nodes in the graph , the specific form of the position encoding result is as follows:

[0111]

[0112] In the formula, Indicates the final discussion of the nodes in the graph The position encoding result of represents the degree of a node with a self-loop; Represents the first-order neighbor set of a node with a self-loop; Represents a first-order neighbor node in the first-order neighbor set; Represents the first-order neighbor node The corresponding learner feature vector.

[0113] In this embodiment, based on the final discussion graph, the nearest neighbor graph is constructed by the position encoding result of each node, that is, each node is connected to M other nodes with the closest L2 distance. Each node in the nearest neighbor graph is only connected to the node whose position encoding is closer to itself, which overcomes the heterogeneity of the final discussion graph to a certain extent.

[0114] In general, in the present invention, if Figure 3As shown in the figure, since the final discussion graph shows heterogeneity in structure and content, the general graph data processing method may not perform well. Therefore, the present invention first estimates the degree of homogeneity of the edges in the final discussion graph data, and then performs adaptive dual-channel encoding on the nodes in the final discussion graph according to the homogeneity probability of the edges. The cognitive feature matrix obtained by S4 is used as a supplement to the learner node features in the final discussion graph, and the pseudo-label obtained by S5 is used to adjust the node adaptive encoding. The course behavior information matrix and the learner's cognitive feature matrix are spliced ​​as the learner feature matrix. After the above process, the learner comprehensive representation matrix and the learner encoding matrix are obtained respectively. The two matrices are spliced ​​and input into the multi-layer perceptron to perform predictive analysis of learner dropouts, and finally output the predicted probability of each learner being a dropout. Then, the present invention measures the deviation between the predicted probability and the learner's true course behavior label through cross entropy loss. For each learner, it corresponds to a learner's true course behavior label, and its value is 1, indicating that the learner has dropped out of class, and its value is 0, indicating that the learner has not dropped out of class, that is, the learner has successfully completed the course. Then the functional form of the above cross entropy loss is:

[0115]

[0116] in, represents the number of learners; Indicates The predicted probability that a learner is a dropout; Indicates The actual course behavior labels of each learner.

[0117] S7: Input the course behavior information data of the learner to be predicted into the trained MOOC dropout prediction model, and output the predicted probability that the learner is a dropout.

[0118] The present invention will now use a specific example to demonstrate the application effect of the MOOC dropout prediction method based on large language model-assisted graph node classification described in S1 to S7 in the above embodiments on a specific data set, so as to facilitate understanding of the essence of the present invention.

[0119] Example

[0120] The specific implementation process of the MOOC dropout prediction method using a large language model to assist graph node classification used in this embodiment is as described above and will not be repeated here.

[0121] In order to demonstrate the technical effect of the method proposed in the present invention, the method of the present invention is verified in an actual MOOC scenario. The data selected by the present invention is an educational data set, in which the data is collected from a real online course learning area. Students learn courses through videos in this course area, and complete homework after each course and unit tests after each unit, and students need to obtain regular course scores by posting and replying in the course discussion area. Based on this data, this embodiment encodes the student's course behavior information and uses it as a feature of the learner node in the graph. The learner is used as a node, and the edge connecting the learner is constructed according to the student's posting and reply records in the discussion area, that is, each edge represents the communication between two students under the same post, thereby constructing a learner discussion graph. Accuracy is selected as an indicator to evaluate the effect of the model. The value range of the accuracy is 0 to 1. The larger the accuracy value, the better the model effect. The calculation formula for the above accuracy is as follows:

[0122]

[0123] Among them, True Positive (TP) represents the number of positive classes predicted as positive classes, True Negative (TN) represents the number of negative classes predicted as negative classes, False Positive (FP) represents the number of negative classes predicted as positive classes, and False Negative (FN) represents the number of positive classes predicted as negative classes.

[0124] The comparison between the method proposed in the present invention and the benchmark machine learning model is shown in Table 1. For the sake of fairness, each method was tested 5 times, and the average of the results was taken as the final dropout prediction result. In Table 1, the implementation processes of the three methods, Attn, which introduces the attention mechanism and conditional random field, EMIF, which is based on multi-view interaction, and HMM, which is based on the hidden Markov model, are all prior art and will not be described in detail. The experimental results in Table 1 show that the method proposed in the present invention achieves the best results compared with the benchmark method. One of the main reasons is that the method proposed in the present invention utilizes the rich information in the course discussion area.

[0125] Table 1. Dropout prediction results of different methods on education datasets method ACC (%) Attn 90.83 EMIF 89.55 HMM 91.34 The present invention 93.32

[0126] In addition, this embodiment also conducts an ablation study to better demonstrate the contribution of the innovation proposed by the present invention, and the experimental results are shown in Table 2. In Table 2, W / O RA represents the process of obtaining the cognitive feature matrix in the ablation step S4, and only the course behavior information matrix composed of the learner's course behavior feature data is used as the learner feature matrix; W / O ED represents the process of obtaining the pseudo label of each edge in the final discussion graph in the ablation step S5, and in the training process of the MOOC dropout prediction model, the same matching probability of each edge in the final discussion graph is not adjusted, and only the unadjusted same matching probability is used for low-pass and high-pass filtering; W / O HW represents the process of ablation embedding the position space of the final discussion graph to obtain the nearest neighbor graph, and only the classic graph neural network method GCN is used for graph representation learning and node classification.

[0127] Table 2. Ablation test results of the innovative module of the present invention on the education dataset method ACC (%) W / O RA 88.23 W / O ED 91.83 W / O HW 85.46 The present invention 93.32

[0128] The results of ablation experiments show that the three innovative parts proposed in the present invention all have significant effects. The main reason is that the cognitive hierarchy analysis and pseudo-label identification processes can use the excellent text reading and analysis capabilities of the large language model to supplement the learner node features and the structural information in the final discussion graph; the process of obtaining the nearest neighbor graph can effectively deal with the mismatch problem caused by the uneven levels of learners communicating with each other in the discussion area. In summary, the method of the present invention can provide an efficient and accurate student dropout detection solution for online education scenarios.

[0129] It should also be noted that the large language model-assisted graph node classification MOOC dropout prediction method in the above embodiment can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a large language model-assisted graph node classification MOOC dropout warning system corresponding to the large language model-assisted graph node classification MOOC dropout prediction method provided in the above embodiment, such as Figure 4 As shown, it includes:

[0130] The data acquisition module is used to obtain the course behavior information data of each learner in the target course from the MOOC platform, encode the pre-processed course behavior information data, and obtain the learner's course behavior characteristic data;

[0131] The discussion graph construction module is used to take each learner as a node in the discussion graph, the interaction relationship between learners in the discussion area as the edge connecting the nodes in the discussion graph, and the course behavior feature data of the learners as the node feature to obtain the initial discussion graph, and process the initial discussion graph to obtain the final discussion graph;

[0132] A model fine-tuning module is used to replace the random initialization process of the LoRA parameter matrix in the O-LoRA method with a randomized singular value decomposition method, and the O-LoRA method is used to fine-tune the pre-trained large language model based on the LoRA parameter matrix after randomized singular value decomposition to obtain a fine-tuned large language model;

[0133] A cognitive hierarchy analysis module, for inputting a set of first examples including questions and answers and a first additional question constituting a first prompt word into a fine-tuned large language model, and the fine-tuned large language model outputs a cognitive hierarchy classification result corresponding to the learner under the target chapter according to the answer in the first example, and obtains a cognitive feature matrix based on the cognitive hierarchy classification result;

[0134] a pseudo-label acquisition module, for inputting a set of second examples including questions and answers and a second additional question constituting a second prompt word into the fine-tuned large language model, and for the fine-tuned large language model to determine whether the course behavior types of the two learners are the same, and outputting a course behavior type determination result according to the answer in the second example, and encoding the course behavior type determination result, and using the course behavior type encoding result as a pseudo-label for each edge in the final discussion graph;

[0135] A model training module is used to use the learner's cognitive feature matrix and the final discussion graph as inputs of the MOOC dropout prediction model to train the MOOC dropout prediction model;

[0136] A result acquisition module is used to input the course behavior information data of the learner to be predicted into the trained MOOC dropout prediction model, and output the predicted probability that the learner is a dropout;

[0137] The early warning module is used to send dropout warning information to teacher users or management users of the MOOC platform, or to send reminder information to learners who are dropouts to urge them to study online.

[0138] It should be noted that in the early warning prompt module, the predicted probability corresponding to the learner can be used to determine whether each learner is a dropout. When the predicted probability is greater than the preset probability threshold, the learner is considered to be a dropout or may be a dropout. When the predicted probability is less than or equal to the preset probability threshold, the learner is considered not to be a dropout. Therefore, an early warning prompt can be given based on the predicted probability output by the result acquisition module. Specifically, a dropout warning message can be sent to the teacher user or the management user to inform the teacher user or the management user that a learner may drop out of class. Reminder messages can also be sent to learners who are dropouts to urge them to study online, so as to reduce the number of dropouts, improve the course completion rate and the overall effect of online education.

[0139] It should also be noted that those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the various embodiments provided in this application, the division of steps or modules in the system and method is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0140] It should also be noted that the learners' course behavior information data involved in the present invention is obtained with full consent and authorization, and the collection, use and processing of relevant information must comply with relevant laws, regulations and standards of relevant countries and regions.

[0141] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A large language model-assisted graph node classification method for predicting MOOC dropouts, characterized in that: The following steps are involved: S1: Obtain the course behavior information data of each learner in the target course from the MOOC platform, encode the pre-processed course behavior information data, and obtain the learner's course behavior characteristic data; S2: Take each learner as a node in the discussion graph, the interaction relationship between learners in the discussion area as the edge connecting the nodes in the discussion graph, and the learner's course behavior feature data as the node feature to obtain the initial discussion graph, and process the initial discussion graph to obtain the final discussion graph; S3: A randomized singular value decomposition method is used to replace the random initialization process of the LoRA parameter matrix in the O-LoRA method. The O-LoRA method is used to fine-tune the pre-trained large language model based on the LoRA parameter matrix after randomized singular value decomposition to obtain a fine-tuned large language model. S4: inputting a set of first examples including questions and answers and a first additional question to form a first prompt word into the fine-tuned large language model, and the fine-tuned large language model outputs the cognitive level classification result corresponding to the learner under the target chapter according to the answer in the first example, and obtains a cognitive feature matrix based on the cognitive level classification result; S5: input a set of second examples including questions and answers and a second additional question to form a second prompt word into the fine-tuned large language model, and the fine-tuned large language model determines whether the course behavior types of the two learners are the same, and outputs the course behavior type judgment result according to the answer in the second example, and encodes the course behavior type judgment result, and uses the course behavior type encoding result as the pseudo label of each edge in the final discussion graph; S6: Take the learner's cognitive feature matrix and the final discussion graph as the input of the MOOC dropout prediction model to train the MOOC dropout prediction model; S7: Input the course behavior information data of the learner to be predicted into the trained MOOC dropout prediction model, and output the predicted probability that the learner is a dropout.

2. The method for predicting dropouts from a MOOC using a large language model-assisted graph node classification as claimed in claim 1, characterized in that: In step S1, each course behavior information data contains N video N teaching videos for learners to watch, work Completion status of tasks, N test The test results of the test, the total number of times the learner visits the target course N visit and text data consisting of all speeches made by learners in the target course discussion area, and preprocessing the text data to form preprocessed course behavior information data; In step S1, the specific process of encoding the pre-processed course behavior information data is as follows: S11: For each teaching video of each learner, the ratio between the viewing time of the learner and the total viewing time of the corresponding teaching video is used as the viewing time coding value of the teaching video, and all viewing time coding values ​​of the learner constitute a corresponding viewing time coding vector; S12: The total number of times the learner visits the target course is used as the access frequency coding vector corresponding to the learner; S13: For each assignment completion status of each learner, if the learner completes the assignment, the ratio between the number of correct answers of the learner in each assignment and the corresponding total number of questions is used as the assignment submission status coding value; if the learner does not complete the assignment, the assignment submission status coding value is 0. Finally, all the assignment submission status coding values ​​of the learner constitute the corresponding assignment submission status coding vector; S14: For each test case of each learner, if the learner completes the test, the ratio between the learner's test score and the corresponding test total score is used as the test score coding value; if the learner does not complete the test, the test score coding value is 0, and all the test score coding values ​​of the learner constitute the corresponding test score coding vector; S15: for each learner's preprocessed text data, the preprocessed text data is sent to the encoding visible large language model to obtain the learner's discussion content encoding vector; S16: Concatenate the viewing time coding vector, the access frequency coding vector, the homework submission status coding vector, the test score coding vector and the discussion content coding vector as the learner's course behavior characteristic data.

3. The method for predicting dropouts from MOOCs using large language model-assisted graph node classification as claimed in claim 1, characterized in that: In step S2, the specific process of obtaining the final discussion graph is as follows: determine whether there is an isolated node in the initial discussion graph: if there is, traverse each isolated node in the initial discussion graph, calculate the cosine similarity between the node feature corresponding to an isolated node and the node features of the remaining nodes, take the K nodes with the largest cosine similarity as the most similar nodes of the isolated node, connect the isolated node and its corresponding most similar nodes, form K edges and add them to the initial discussion graph, and obtain the final discussion graph; If it does not exist, the initial discussion graph will be used as the final discussion graph.

4. The method for predicting dropouts from MOOCs using large language model-assisted graph node classification as claimed in claim 1, characterized in that: In step S3, the specific process of using the randomized singular value decomposition method to obtain the LoRA parameter matrix after randomized singular value decomposition is: S31: Randomly sample from a normal distribution as the column vector of the random matrix to generate a complete random matrix; S32: Use a random matrix to project the parameter matrix of the large language model into a low-dimensional space to obtain a projection matrix; perform orthogonal triangular decomposition on the projection matrix to obtain a first orthogonal matrix and an upper triangular matrix; transpose the first orthogonal matrix and multiply it by the parameter matrix of the large language model to obtain a low-dimensional matrix; Perform singular value decomposition on the low-dimensional matrix to obtain a second orthogonal matrix, a third orthogonal matrix and a diagonal matrix; S33: The index corresponding to the largest r singular values ​​in the diagonal matrix is ​​used as the initialization index, the column vector corresponding to the initialization index in the second orthogonal matrix is ​​used as the initialized first LoRA parameter matrix, and the column vector corresponding to the initialization index in the third orthogonal matrix is ​​used as the initialized second LoRA parameter matrix. The initialized first LoRA parameter matrix and the initialized second LoRA parameter matrix constitute a set of LoRA parameter matrices after randomized singular value decomposition.

5. The method for predicting dropouts from a MOOC using a large language model-assisted graph node classification as claimed in claim 1, characterized in that: In step S4, the question in the first example and the first additional question both include all speeches of the learner in the target chapter of the target course discussion area, the name of the target course and the name of the target chapter, the cognitive level of the learner determined based on the learner's speech, the cognitive level of the online community discussion and the judgment criteria corresponding to each cognitive level; the answer in the first example is designed in the form of a thinking chain prompt, and the answer in the first example includes the classification result of the learner's cognitive level and the logical reasoning process for determining the cognitive level to which the learner belongs; In step S4, the generation process of the cognitive feature matrix is ​​as follows: the cognitive level classification results are one-hot encoded to obtain the cognitive level encoding results, the cognitive level encoding results of all chapters constitute the learner's cognitive feature vector, the cognitive feature vector of a learner is used as a row in the cognitive feature matrix, and the cognitive feature vectors of all learners constitute a complete cognitive feature matrix.

6. The method for predicting dropouts from a MOOC using a large language model-assisted graph node classification as claimed in claim 1, characterized in that: In step S5, the question in the second example and the second additional question both contain the course behavior information data of the two learners and determine whether the course behavior types of the two learners are the same. The answer in the second example is designed in accordance with the thinking chain prompt, and the answer in the second example includes the judgment result of the course behavior type and the logical reasoning process for determining whether the course behavior types are the same; wherein, when the two learners are both dropouts or both are non-dropouts, the course behavior types of the two learners are deemed to be the same.

7. The method for predicting dropouts from a MOOC using a large language model-assisted graph node classification as claimed in claim 1, characterized in that: In step S6, during the training of the MOOC dropout prediction model, the course behavior information matrix is ​​constructed from the course behavior information data of all learners, and the course behavior information matrix is ​​concatenated with the learner's cognitive feature matrix as the learner feature matrix. Two multi-layer perceptrons are used to output the same matching probability of each edge according to the edge set in the final discussion graph, and the same matching probability is adjusted using the pseudo label of each edge in the final discussion graph to obtain the adjusted same matching probability. The adjusted same matching probability is used as an element in the low-pass adjacency matrix to construct the low-pass adjacency matrix, and the learner feature matrix is ​​input into the simplified graph convolutional network. First, the learner feature matrix is ​​processed by the third multi-layer perceptron to output the initial low-pass feature matrix. The initial low-pass feature matrix is ​​processed by multiple low-pass filtering. In each low-pass filtering process, the low-pass filtering result obtained last time is multiplied by the low-pass adjacency matrix to obtain the current low-pass filtering result. The last low-pass filtering result is used as the learner low-pass feature matrix. The result after subtracting 1 from the adjusted homonymy probability is used as the element in the high-pass adjacency matrix to construct the high-pass adjacency matrix. The unit matrix is ​​subtracted from the weighted high-pass adjacency matrix to obtain the processed high-pass adjacency matrix. Array, input the learner feature matrix into the Lap-SGC model, first process the learner feature matrix through the fourth multi-layer perceptron, output the initial high-pass feature matrix, and process the initial high-pass feature matrix through multiple high-pass filtering. In each high-pass filtering process, the high-pass filtering result obtained in the last time is multiplied by the processed high-pass adjacency matrix to obtain the current high-pass filtering result. The high-pass filtering result obtained in the last time is used as the learner high-pass feature matrix, and the learner low-pass feature matrix and the learner high-pass feature matrix are concatenated as the learner comprehensive representation matrix. The final discussion diagram After position space embedding, a nearest neighbor graph is obtained, and the nearest neighbor graph is input into a graph convolutional network to obtain a learner encoding matrix. The learner comprehensive representation matrix and the learner encoding matrix are concatenated and input into a fifth multi-layer perceptron, and the predicted probability of each learner being a dropout is output. The cross entropy loss between the predicted probability of each learner being a dropout and the learner's actual course behavior label is calculated, and the parameters of the MOOC dropout prediction model are updated based on minimizing the cross entropy loss until a preset iteration round threshold is reached, the MOOC dropout prediction model converges, and a trained MOOC dropout prediction model is obtained.

8. The method for predicting dropouts from MOOCs using large language model-assisted graph node classification as claimed in claim 7, characterized in that: The specific process of using two multi-layer perceptrons to output the probability of matching each edge is as follows: the learner feature vectors of the two nodes corresponding to each edge are respectively passed through the first multi-layer perceptron, and the corresponding outputs are the first intermediate representation vector and the second intermediate representation vector representing the two learners respectively, the first intermediate representation vector and the second intermediate representation vector are concatenated and passed through the second multi-layer perceptron to obtain the third intermediate representation vector, the second intermediate representation vector and the first intermediate representation vector are concatenated and passed through the second multi-layer perceptron to obtain the fourth intermediate representation vector, the third intermediate representation vector and the fourth intermediate representation vector are averaged to obtain the probability of matching each edge; wherein the learner feature vector is a row in the learner feature matrix; The specific process of adjusting the probability of matching is as follows: when the course behavior types of two learners are different, The minimum value between and 0 is taken as the adjusted probability of matching; when the course behavior types of two learners are the same, The maximum value between 0 and 0 is taken as the adjusted homogamy probability; is the probability of homogamy before adjustment; are the preset hyperparameters.

9. The method for predicting dropouts from MOOCs using large language model-assisted graph node classification as claimed in claim 8, characterized in that: The specific process of obtaining the nearest neighbor graph after embedding the final discussion graph in position space is as follows: first, perform embedding operation on each node in the final discussion graph to map it to an implicit dimension of position space, take the average of the learner feature vectors of all first-order neighbor nodes of each node in the final discussion graph as the position encoding result of the node, calculate the L2 distance between the position encoding result of each node and the position encoding results of the remaining nodes in the final discussion graph, take the M nodes with the smallest L2 distance as the nearest neighbor nodes of each node, connect each node in the final discussion graph with its respective nearest neighbor node, and use the learner feature matrix as the node feature to construct the nearest neighbor graph.

10. A MOOC dropout warning system with large language model-assisted graph node classification, characterized in that: include: The data acquisition module is used to obtain the course behavior information data of each learner in the target course from the MOOC platform, encode the pre-processed course behavior information data, and obtain the learner's course behavior characteristic data; The discussion graph construction module is used to take each learner as a node in the discussion graph, the interaction relationship between learners in the discussion area as the edge connecting the nodes in the discussion graph, and the course behavior feature data of the learners as the node feature to obtain the initial discussion graph, and process the initial discussion graph to obtain the final discussion graph; A model fine-tuning module is used to replace the random initialization process of the LoRA parameter matrix in the O-LoRA method with a randomized singular value decomposition method, and the O-LoRA method is used to fine-tune the pre-trained large language model based on the LoRA parameter matrix after randomized singular value decomposition to obtain a fine-tuned large language model; A cognitive hierarchy analysis module, for inputting a set of first examples including questions and answers and a first additional question constituting a first prompt word into a fine-tuned large language model, and the fine-tuned large language model outputs a cognitive hierarchy classification result corresponding to the learner under the target chapter according to the answer in the first example, and obtains a cognitive feature matrix based on the cognitive hierarchy classification result; a pseudo-label acquisition module, for inputting a set of second examples including questions and answers and a second additional question constituting a second prompt word into the fine-tuned large language model, and for the fine-tuned large language model to determine whether the course behavior types of the two learners are the same, and outputting a course behavior type determination result according to the answer in the second example, and encoding the course behavior type determination result, and using the course behavior type encoding result as a pseudo-label for each edge in the final discussion graph; A model training module is used to use the learner's cognitive feature matrix and the final discussion graph as inputs of the MOOC dropout prediction model to train the MOOC dropout prediction model; A result acquisition module is used to input the course behavior information data of the learner to be predicted into the trained MOOC dropout prediction model, and output the predicted probability that the learner is a dropout; The early warning module is used to send dropout warning information to teacher users or management users of the MOOC platform, or to send reminder information to learners who are dropouts to urge them to study online.

Images