Method, system, device and storage medium for predicting student learning behavior

Abstract

The application discloses a kind of fusion student learning behavior's score prediction method, system, equipment and storage medium, by analyzing the learning behavior of student, quantification each learning behavior is for the influence of student learning gain, the influence of all learning behavior for learning gain is integrated together, to measure the high-order interaction effect between learning gain under different behavior influences, and update the knowledge state of student, the application can more accurately predict the answer of student Performance, can provide better performance prediction service for online learning system, and then make the online learning system better for student provides personalized learning service, improve the experience of student when using online learning platform.

Description

Technical Field

[0001] This invention relates to the field of educational data mining technology, and in particular to a method, system, device, and storage medium for predicting academic performance by integrating student learning behaviors. Background Technology

[0002] Recently, with the rapid development of the internet and the rapid accumulation of learning resources, online learning systems have become the preferred learning method for more and more learners due to their convenience and vast amount of learning resources. Compared with traditional classroom teaching, online learning systems have their inherent advantages. Learners (e.g., students) can easily log in to online education systems anytime, anywhere via mobile phones, computers, and other devices to access diverse, open, and shared learning resources, such as watching instructional videos for course learning and completing exercises.

[0003] Currently, there are many online learning platforms on the market, such as LeetCode, China University MOOC, and Zhixue.com. Besides providing a convenient learning environment, these platforms can also assess students' abilities through their past test-taking records, thereby offering personalized learning services. These services include recommending suitable and efficient learning paths, suggesting questions of matching difficulty based on the user's level, and identifying gaps in their knowledge structure for targeted training, thus improving their learning efficiency. The key to providing personalized learning services lies in analyzing students' historical test-taking records to understand their knowledge base and accurately predict their future performance.

[0004] Existing technologies often analyze students' knowledge status through question-and-answer sequences. However, these solutions only consider the questions and the correctness of students' answers, neglecting the influence of learning behaviors. This leads to bias and makes it difficult to provide better personalized learning services. For example, when two students both get the correct answer to the same question, but student A takes 2 seconds and student B takes 20 seconds, traditional methods, which only consider the question and answer, treat these two records equally in terms of analyzing students' abilities. However, this is clearly unreasonable, because student A's short answer time suggests that they may have guessed correctly, and it cannot be considered that they obtained the correct answer through their own ability. Therefore, in addition to analyzing students' knowledge status based on their answer results, the influence of learning behaviors on students' knowledge status should also be considered.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, device, and storage medium for predicting student performance by integrating student learning behaviors. This method can accurately predict students' future answer performance, thereby providing better personalized learning services for students.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for predicting academic performance that integrates student learning behaviors includes:

[0009] Multiple answer records of students are obtained. Each answer record contains the question, answer result and several learning behaviors at the corresponding time. For time t, the question and answer result at time t are modeled by a neural network to obtain the interaction record representation vector.

[0010] Based on different learning behaviors, corresponding influencing factors are calculated, and combined with the interactive record representation vector, the learning gain factors corresponding to different learning behaviors are calculated.

[0011] By using a multimodal fusion algorithm, the learning gain factors corresponding to different learning behaviors are fused together to obtain a learning gain vector;

[0012] By setting a forgetting gate and combining it with the learning gain vector, the knowledge state at time t can be calculated;

[0013] By combining the knowledge point vector associated with the question at time t+1 with the knowledge state at time t, the student's answer can be predicted.

[0014] A performance prediction system that integrates student learning behaviors includes:

[0015] The question-answering process modeling module is used to obtain multiple question-answering records of students. Each question-answering record contains the question, answer result and several learning behaviors at the corresponding time. For time t, the question and answer result at time t are modeled by a neural network to obtain the interaction record representation vector.

[0016] The learning behavior individual effect module is used to calculate the corresponding influence factors based on different learning behaviors, and combined with the interaction record representation vector, to calculate the learning gain factors corresponding to different learning behaviors.

[0017] The learning behavior co-action module is used to fuse the learning gain factors corresponding to different learning behaviors through a multimodal fusion algorithm to obtain a learning gain vector;

[0018] The knowledge state calculation module is used to calculate the knowledge state at time t by setting a forgetting gate and combining it with the learning gain vector;

[0019] The answer prediction module is used to predict students' answer results by combining the knowledge point vector associated with the question at time t+1 and the knowledge state at time t.

[0020] A processing device includes: one or more processors; and a memory for storing one or more programs;

[0021] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0022] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0023] As can be seen from the technical solution provided by the present invention, by analyzing students' learning behaviors, quantifying the impact of each learning behavior on students' learning outcomes, and integrating the impact of all learning behaviors on learning outcomes, this invention can measure the higher-order interaction effects between learning outcomes under the influence of different behaviors and update students' knowledge status. This invention can more accurately predict students' test scores, provide better score prediction services for online learning systems, and enable online learning systems to better provide personalized learning services to students, thereby improving students' experience when using online learning platforms. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating a method for predicting academic performance by integrating student learning behaviors, provided as an embodiment of the present invention;

[0026] Figure 2 A schematic diagram of a single-function module provided for an embodiment of the present invention;

[0027] Figure 3 A schematic diagram illustrating the principle of the combined effect of learning behaviors provided in an embodiment of the present invention;

[0028] Figure 4 A schematic diagram of a performance prediction system that integrates student learning behaviors, provided as an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0031] First, the following explanations are provided for the terms that may be used in this article:

[0032] The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0033] The terms “including,” “comprising,” “containing,” “having,” or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, “including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.)” should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0034] The following provides a detailed description of the performance prediction method, system, device, and storage medium that integrates student learning behaviors provided by this invention. Contents not described in detail in the embodiments of this invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they should be performed according to conventional conditions in the art or conditions recommended by the manufacturer.

[0035] Example 1

[0036] This invention provides a method for predicting academic performance by integrating student learning behaviors, such as... Figure 1 As shown, the main steps include the following:

[0037] Step 1: Obtain multiple answer records of students. Each answer record contains the question, answer result and several learning behaviors at the corresponding time. For time t, model the question and answer result at time t using a neural network to obtain the interaction record representation vector.

[0038] In this embodiment of the invention, the step of modeling the question and answer results at time t using a neural network to obtain the interaction record representation vector includes: embedding the question and answer results at time t to obtain corresponding representation vectors; and using the representation vectors of the question and answer results at time t to model the interaction behavior through a neural network to obtain the interaction record representation vector.

[0039] Step 2: Calculate the corresponding influencing factors based on different learning behaviors, and combine them with the interaction record representation vector to calculate the learning gain factors corresponding to different learning behaviors.

[0040] In this embodiment of the invention, the learning behavior includes: the speed of answering questions, the number of attempts, and the number of requests for prompts.

[0041] To analyze the impact of students' learning behaviors on their learning outcomes, this invention first analyzes the distribution patterns of three learning behaviors: speed, trial, and prompting, and calculates the three learning behavior influencing factors based on these factors, which are then used to influence learning outcomes.

[0042] (1) The speed of answering questions refers to the time taken to answer the questions, which follows a log-normal distribution. Based on the log-normal distribution, the speed influence factor is calculated and expressed as:

[0043]

[0044] Among them, AC t μ represents the velocity influence factor at time t. t and This indicates that the information is taken from the student's answer record for question e. t The expected value and variance of the answering time obtained, a t Let ln represent the speed at which the answer is given at time t, and ln be the natural logarithm. Let P represent the normal distribution and P represent the probability distribution value.

[0045] (2) The number of attempts refers to the number of times a question is answered repeatedly, which follows a Poisson distribution. The attempt influence factor is calculated based on the Poisson distribution and is expressed as:

[0046]

[0047] Among them, PC t This represents the impact factor of the attempt at time t. This represents the average number of attempts obtained from the answer records. p represents a Poisson distribution. t Let t represent the number of attempts at time t, and P represent the probability distribution value.

[0048] (3) The number of prompts refers to the number of times a question is asked for prompts from the online learning platform, which follows a Poisson distribution. The prompt influence factor is calculated based on the Poisson distribution and is expressed as:

[0049]

[0050] Among them, NC t This indicates the influencing factor at time t. This represents the average number of request prompts obtained from the answer records. Represents a Poisson distribution, n t Let t represent the number of prompts at time t, and P represent the probability distribution value.

[0051] In this embodiment of the invention, a control factor generated by the speed influence factor is calculated using the speed influence factor. Furthermore, the learning gain obtained through answering the questions is calculated by combining the interaction record representation vector obtained at time t with the mastery level of the corresponding knowledge points extracted from the questions at time t-1 using the interaction record representation vector. Finally, the learning gain factor under the control of the speed influence factor is calculated by combining the control factor generated by the speed influence factor with the learning gain obtained through answering the questions, expressed as:

[0052]

[0053]

[0054]

[0055] in, AC represents the velocity influence factor at time t. t The control factors, k, d, and b, are generated by a nonlinear curve and are parameters of that nonlinear curve used to control its shape; i t This represents the interaction record representation vector obtained at time t. This indicates the level of mastery of the corresponding knowledge points extracted from the questions at time t-1; Let represent the learning gain factor under the control of the velocity influence factor at time t, and σ(·) represent the sigmoid activation function.

[0056] Using the same principle, the learning gain factors under the control of the trial influence factor at time t were calculated. And the learning outcomes factors controlled by the influencing factors at time t.

[0057] Step 3: Using a multimodal fusion algorithm, the learning gain factors corresponding to different learning behaviors are fused to obtain a learning gain vector.

[0058] In this embodiment of the invention, the learning gain factor controlled by the velocity influence factor at time t is denoted as... The learning outcome factor controlled by the trial at time t is denoted as . The learning outcome factors controlled by the prompt at time t are denoted as follows: After concatenating each of the three elements with 1 and performing an outer product operation, we obtain the outer product tensor G. t ;

[0059] The learning gain vector is obtained by processing through a fully connected layer. Represented as:

[0060]

[0061] Where W3 is the weight matrix of the fully connected layer, b3 is the bias term of the fully connected layer, and ReLU is the modified linear unit activation function.

[0062] Step 4: Calculate the knowledge state at time t by setting a forget gate and combining it with the learning gain vector.

[0063] Using the problem at time t e t Related knowledge point vectors Propagate the learning gain vector to obtain the learning gain lg associated with all knowledge points. t , is represented as:

[0064]

[0065] Where T is the transpose symbol.

[0066] Set the forget gate f t , is represented as:

[0067]

[0068] Where ⊕ represents the vector concatenation operation, W4 represents the weight matrix when calculating the forget gate, and b4 represents the bias term when calculating the forget gate; AC t PC t NC t The following are the speed influence factors, in order: Try the influence factor and hint at the influence factor, i t Let represent the interaction record representation vector obtained at time t, and σ(·) represent the sigmoid activation function.

[0069] Combined with the forgetting gate f t Calculate the knowledge state h at time t. t , is represented as:

[0070] h t =f t *h t-1 +lg t

[0071] Among them, h t-1 This represents the knowledge state at time t-1.

[0072] Step 5: Combine the knowledge point vector associated with the question at time t+1 with the knowledge state at time t to predict the student's answer.

[0073] In this embodiment of the invention, the problem e at time t+1 is combined with t+1 Related knowledge point vectors The knowledge point vector associated with the question at time t+1 and the knowledge state h at time t. t The question e extracted at time t+1t+1 Mastery level of the corresponding knowledge points Represented as:

[0074]

[0075] Combined with the problem at time t+1, e t+1 The representation vector e′ t+1 The predicted student response is represented as:

[0076]

[0077] Among them, y t+1 This represents the predicted student's answer to question e at time t+1. t+1 The answer results are given, and W5 and b5 are the weight matrix and bias term used to predict the student's answer results.

[0078] The above-mentioned solution provided by the embodiments of the present invention analyzes students' learning behaviors, quantifies the impact of each learning behavior on students' learning gains, integrates the impact of all learning behaviors on learning gains, measures the higher-order interaction effects between learning gains under the influence of different behaviors, and updates students' knowledge status. The present invention can more accurately predict students' answer scores, provide better score prediction services for online learning systems, and thus enable online learning systems to better provide personalized learning services to students and improve students' experience when using online learning platforms.

[0079] To more clearly demonstrate the technical solution and its effects provided by the present invention, the method provided by the embodiments of the present invention will be described in detail below with reference to specific examples.

[0080] I. Serialization modeling and answering process.

[0081] In this embodiment of the invention, the student's answering process is recorded as a sequence, called the answer record sequence, denoted as: s = {(e1, r1, b1), (e2, r2, b2), ... (e K ,rK,b K )}, where s is the sequence of answer records, and tuple (e t r t b t ) represents a basic answer record in the sequence of learner answer records, t = 1, 2, ..., K, where K is the length of the answer record sequence, e t This represents the question that the student answers at time t, r. t This indicates the student's answer to the question, whether it is correct or incorrect (0 indicates an incorrect answer, 1 indicates a correct answer), b t This refers to the learning behaviors exhibited by students when answering questions, which are composed of (a tp t n t ) constitute, a t Indicates the student's opinion on e t The speed at which one answers a question (i.e., the time taken to answer the question), p t Let n represent the number of attempts. t Number of requests.

[0082] In this embodiment of the invention, a matrix h is used. t Represents the student's knowledge state at time t, where M represents the number of knowledge points, d h This represents the dimension of knowledge state, where R is the symbol for the set of real numbers. Each row of the knowledge state matrix corresponds to a student's cognitive vector at a specific knowledge point. The questions in the answer records are embedded using the embedding matrix E1 to obtain the question representation vector, where... Where |E| represents the total number of questions, d e It is the embedding vector dimension of the question, which represents the question e in the answer record at time t. t Embedding yields a dimension of d e The exercise representation vector e′ t The answers in the answer records (i.e., the answers to the questions) are embedded using the embedding matrix E2 to obtain the answer representation vector, where... 2 indicates two possible answers: correct or incorrect, d a It is the vector dimension of the answer embedding, and the answer representation vector in the answer record at time t is... The Q-matrix is ​​used to represent the relationship between questions and knowledge points, where Q∈R. J×M The elements in matrix Q consist of 0s and 1s, and J is the total number of questions. If Q jm =1 represents exercise e j With knowledge point k m There is a correlation between them, and vice versa. The vector of knowledge points associated with the exercises at time t is:

[0083] A fully connected neural network layer with ReLU activation function is used to deeply blend the question-answer pair interaction behavior to obtain a basic interaction record representation vector. Represented as:

[0084]

[0085] in, This represents a vector concatenation operation. Represents the weight matrix. This represents the bias term of the network layer, and ReLU is the modified linear unit activation function.

[0086] II. Learning Behavior as a Sole Role Module.

[0087] In this embodiment of the invention, the impact of learning behaviors on students' learning outcomes is quantitatively analyzed from three perspectives: speed, attempts, and prompts. Previous work has shown that answering questions quickly, with a high number of attempts and prompts, may indicate that students are not paying attention, leading to poor learning outcomes. This invention calculates the impact factors of each behavior on learning outcomes based on learning behavior characteristics.

[0088] like Figure 2 The diagram shown is a schematic of the standalone module. The standalone module is mainly responsible for calculating the corresponding influencing factors based on different learning behaviors, and combining the interaction record representation vector to calculate the learning gain factors corresponding to different learning behaviors.

[0089] 1. Calculate the impact factor.

[0090] (1) Speed.

[0091] Speed ​​refers to the student's pace of solving problems, i.e., how fast or slow they are. This can be represented by the student's answer time. It's known that shorter answer times correspond to faster answering speeds, suggesting that the student may have obtained the answer through guessing. In such cases, the student's learning gain should be minimal. Regarding question e... t Assume the time a student spends answering questions on it is 1 / 2. t Follows a log-normal distribution, i.e. Among them, the answer record is for question e t Expected value μ of the distribution of answering time t With variance This can be learned from student answer records using the maximum likelihood estimation (MLE) method. Therefore, the velocity influence factor AC... t The calculation method is as follows:

[0092]

[0093] Where P represents the probability distribution value.

[0094] The above analysis shows that the shorter the answering time, the greater the speed influence factor.

[0095] (2) Try.

[0096] "Try" represents the number of times a student attempts to answer a question. Students can find the final answer by trying multiple times, and the learning gains from this approach are different from those from getting the correct answer on the first try. Suppose a student answers question e... t The number of attempts p used above t It follows a Poisson distribution, i.e. In the answer record, regarding question e tMean of the distribution of the number of attempts This can be learned through the maximum likelihood method. Then, the trial impact factor PC can be calculated. t as follows:

[0097]

[0098] Where P represents the probability distribution value.

[0099] The above analysis shows that the more attempts there are, the greater the impact factor.

[0100] (3) Hint.

[0101] The prompt represents the number of times a student requests hints from the online learning platform while answering a question. Students can obtain the correct answer by repeatedly requesting hints; excessive requests indicate less independent thought and consequently, less learning. For example, suppose a student is answering question e... t The number of prompts n used above t It follows a Poisson distribution, i.e. In the answer record, regarding question e t Mean of the distribution of request prompts This can be learned through the maximum likelihood method. Therefore, the influencing factor NC can be calculated. t as follows:

[0102]

[0103] Where P represents the probability distribution value.

[0104] The above analysis shows that the more request prompts there are, the greater the impact factor.

[0105] 2. Calculate the learning gains under the influence of each learning behavior factor.

[0106] After calculating the influencing factors for each behavior, we calculate the learning gains under the influence of each behavior. Taking the speed influencing factor as an example, the calculation process is as follows:

[0107]

[0108]

[0109]

[0110] in, Indicated by the speed influence factor AC tThe control factor generated by a nonlinear curve is used to control the students' answer gains. The larger the influence factor, the smaller the control factor, which is used to reduce the learning gains. k, d, and b are all parameters of the nonlinear curve, used to control the shape of the nonlinear curve. This refers to the answer gains obtained solely through the analysis of question-answer pairs recorded by students. These gains are determined by the knowledge state associated with the knowledge point corresponding to the current question and the interaction between the current question and the answer. This can be calculated using a multilayer perceptron (MLP). This represents the trainable parameters (weights and biases). This represents the learning gain factor after the influence of learning behavior has been applied, and σ(·) represents the sigmoid activation function.

[0111] Through the above calculations, we can obtain an estimate of student learning gains under the influence of speed behavior. Similarly, we can obtain the influence factor PC for the trial. t And the impact factor NC t Estimation of learning outcomes under control as well as Specifically, PC t NC t Substitute these values ​​into the first expression in the above calculation process, replacing AC. t Calculate the impact factor PC from the trial t The generated control factors The impact factor NC is indicated t The generated Then, the answer score was calculated using the second formula through two different multilayer perceptrons. Answering questions The difference lies in the weights and bias parameters. Substituting these into the third equation and using the σ(.) function, we can calculate... as well as

[0112] III. The combined effect of learning behaviors.

[0113] The three learning behaviors introduced above are not independent of each other; biases may arise from considering any single perspective. For example, a student's short answer time does not necessarily mean they rely on guessing; they could also be highly proficient. However, if they simultaneously exhibit repeated attempts, the likelihood of guessing increases significantly. Therefore, after analyzing the learning gains under the influence of these three behaviors, it is necessary to consider their combined effects. Here, the LMF (Low-Rank Multimodal Fusion) algorithm is used to blend the learning gains under the influence of the three behaviors, fully considering their complex interactions.

[0114] like Figure 3The diagram shown illustrates the principle of the combined effect of learning behaviors. It is mainly responsible for using a multimodal fusion algorithm to fuse the learning gain factors corresponding to different learning behaviors and obtain a learning gain vector.

[0115] First, Concatenating the vectors with 1, and then performing an outer product operation on these three vectors, yields a high-dimensional outer product tensor. This represents the higher-order interaction between learning outcomes influenced by the three behaviors:

[0116]

[0117] in, This indicates that the outer product operation is performed on vectors. For example... Figure 3 As shown, G t It retains the characteristics of each learning outcome factor and also fully considers the interactions between learning outcome factors.

[0118] Then, the outer product matrix is ​​passed through a fully connected layer to generate the final learning gain vector that considers the complex interactions between different behaviors. Represented as:

[0119]

[0120] in, These are trainable parameters, and ReLU is the modified linear unit activation function.

[0121] Preferably, to save memory overhead, the high-order weight matrix of the fully connected layer is... Perform a low-rank decomposition, assuming the weight matrix W3 is composed of d h matrix Stacked together, each matrix With G t The final learning gain vector is obtained by multiplying the results. The one-dimensional matrix can be further divided into three-dimensional matrices. Decomposing it into the outer product of three one-dimensional vectors, we have the following decomposition form:

[0122]

[0123] in, It is through The three one-dimensional vectors obtained from the decomposition, R, is the smallest value that makes the decomposition meaningful, and is called the rank of the matrix. In the following calculations, a fixed rank r is used to represent R.

[0124] Will To stack, that is to Therefore, the original calculation can be performed High-dimensional operation formula W3·G t Replace it with the following format:

[0125]

[0126] This means that the learning outcomes that need to be aggregated can be combined with the decomposed low-dimensional weight matrix. By performing matrix multiplication to transform the vectors and then multiplying the transformed eigenvectors element-wise, the storage of high-dimensional inner product matrices and weight matrices is eliminated, greatly reducing memory overhead.

[0127] IV. Update Knowledge Status

[0128] After obtaining the students' learning gain vector Afterwards, their knowledge status needs to be updated accordingly. Considering that students' knowledge will be forgotten over time, a forgetting gate is used to control student forgetting. The forgetting gate f... t The student's knowledge level at the previous moment, current answer result, and learning behavior are all jointly influenced by this, as represented by:

[0129]

[0130] Where ⊕ represents the vector concatenation operation, Represents the weight matrix. σ represents the bias term, and σ(·) represents the sigmoid activation function.

[0131] Due to the learning gain vector This calculation only considers the knowledge points relevant to the currently answered question. To calculate the overall learning gain across all knowledge point dimensions, the learning gains associated with a specific knowledge point need to be propagated to all other knowledge point dimensions. This paper relies on the knowledge point vector associated with the current question. ( (This indicates that its transpose) propagates the learning gains, thereby obtaining the learning gains (lg) associated with all knowledge points. t :

[0132]

[0133] Finally, combined with the forgetting gate f t And learning gains lg t The student's knowledge status h t The update is as follows:

[0134] h t =f t *h t-1 +lg t

[0135] Among them, h t-1 This represents the knowledge state at time t-1.

[0136] V. Prediction of student performance.

[0137] Based on the student's knowledge status h at time t t It can be predicted that at time t+1, the student's understanding of question e will be correct. t+1 The answer result y t+1 Due to the knowledge state h t It records the students' mastery of each knowledge point, so it can predict their understanding of e. t+1 When performing this, you need to first use the method with e. t+1 Related knowledge point vector The level of mastery of the knowledge points corresponding to this question Extracted, and then according to And the title e t+1 The representation vector e′ t+1 Predict the answer result y using a fully connected layer with a sigmoid activation function. t+1 :

[0138]

[0139]

[0140] Where ⊕ represents the vector concatenation operation. Represents the weight matrix. The term y represents the bias term of the network layer, and σ(·) represents the sigmoid activation function. The output y... t+1 ∈(0,1) indicates that the learner's prediction of the problem e at time t+1 is correct. t+1 The probability of a correct answer can be used to determine whether a learner can answer correctly based on a set threshold. t+1 If the value is greater than the threshold, the predicted answer is considered correct; otherwise, the predicted answer is considered incorrect. The threshold is generally set to 0.5.

[0141] VI. Model Training.

[0142] The entirety of the network model provided by this invention, including individual action modules, joint action modules, knowledge state update modules, and score prediction modules, is considered as a single network model. The parameters of the network model are updated through training, and the objective loss function set during training is:

[0143]

[0144] Among them, y t r represents the predicted response at time t (i.e., the predicted value). tθ represents the answer result (i.e., the true value) at time t in the answer record, K is the length of the answer record during training, and θ represents all the parameters that need to be learned in the model. During training, the Adam optimizer is used to learn the parameter values ​​of the model by minimizing the objective function.

[0145] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0146] Example 2

[0147] This invention also provides a performance prediction system that integrates student learning behaviors, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 4 As shown, the system mainly includes:

[0148] The question-answering process modeling module is used to obtain multiple question-answering records of students. Each question-answering record contains the question, answer result and several learning behaviors at the corresponding time. For time t, the question and answer result at time t are modeled by a neural network to obtain the interaction record representation vector.

[0149] The learning behavior individual effect module is used to calculate the corresponding influence factors based on different learning behaviors, and combined with the interaction record representation vector, to calculate the learning gain factors corresponding to different learning behaviors.

[0150] The learning behavior co-action module is used to fuse the learning gain factors corresponding to different learning behaviors through a multimodal fusion algorithm to obtain a learning gain vector;

[0151] The knowledge state calculation module is used to calculate the knowledge state at time t by setting a forgetting gate and combining it with the learning gain vector;

[0152] The answer prediction module is used to predict students' answer results by combining the knowledge point vector associated with the question at time t+1 and the knowledge state at time t.

[0153] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0154] Example 3

[0155] The present invention also provides a processing device, such as Figure 5 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0156] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0157] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example:

[0158] Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc.

[0159] The output device can be a display terminal;

[0160] The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0161] Example 4

[0162] The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0163] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0164] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting academic performance that integrates student learning behaviors, characterized in that, The method comprises the following steps: obtaining a plurality of answer records of a student, each answer record comprising a question, an answer result and a plurality of learning behaviors at a corresponding time; for t time, modeling the question and the answer result at t time through a neural network to obtain an interaction record representation vector; calculating corresponding influence factors according to different learning behaviors, and combining the interaction record representation vector to calculate learning gain factors corresponding to different learning behaviors; fusing the learning gain factors corresponding to different learning behaviors through a multi-modal fusion algorithm to obtain a learning gain vector; calculating a knowledge state at t time by setting a forgetting gate and combining the learning gain vector; combining a knowledge point vector associated with a question at t+1 time and the knowledge state at t time to predict the answer result of the student; wherein the calculation of the corresponding influence factors according to different learning behaviors comprises: the learning behaviors include speed at answering, number of attempts and number of requests for hints, and the corresponding influence factors are referred to as speed influence factors, attempt influence factors and hint influence factors; wherein: the speed at answering refers to the answering time, which is subject to a lognormal distribution, and the speed influence factor is calculated in combination with the lognormal distribution and is represented as: where AC t represents the speed influence factor at time t, μ t represents the expected value of the answer time of question e t represents the expected value and variance obtained from the answer record of question e t represents the speed at time t, ln is the natural logarithm, represents the normal distribution, P represents the probability distribution value;​ the number of attempts refers to the number of repeated answering times, which is subject to a Poisson distribution, and the attempt influence factor is calculated in combination with the Poisson distribution and is represented as: where PC t denotes the attempt influence factor at time t, denotes the average number of attempts obtained from the answer records, denotes the Poisson distribution, p t denotes the number of attempts at time t, P denotes the probability distribution value; the number of hints refers to the number of times of requesting hints from an online learning platform during answering, which is subject to a Poisson distribution, and the hint influence factor is calculated in combination with the Poisson distribution and is represented as: wherein NC t denotes the prompting influence factor at time t, denotes the average number of times of requesting prompting obtained from the answer record, denotes a Poisson distribution, n t denotes the number of prompts at time t, P denotes a probability distribution value; a control factor generated by the speed influence factor is calculated using the speed influence factor, and a learning gain obtained through the answer record is calculated using the interaction record representation vector obtained at t time, the mastery degree of the corresponding knowledge point extracted from the question at t-1 time, the speed influence factor generated control factor and the learning gain obtained through the answer record to calculate the learning gain factor under the control of the speed influence factor, which is represented as: wherein, denotes the speed impact factor at time t t a control factor generated through a non-linear curve, k, d and b are parameters of the non-linear curve, used to control the shape of the non-linear curve; i t denotes the interaction record representation vector obtained at time t, denotes the mastery degree of the corresponding knowledge point extracted from the question at time t-1; denotes the learning gain factor under the speed impact factor at time t, σ(·) denotes a sigmoid activation function; Using the same principle, the learning harvest factor under the control of the attempt influence factor at time t is calculated and the learning harvest factor under the control of the hint influence factor at time t 2. The method of claim 1, wherein the learning behavior of the student is fused with the performance of the student. the modeling of the question and the answer result at t time through the neural network to obtain the interaction record representation vector comprises: embedding the question and the answer result at t time respectively to obtain corresponding representation vectors; modeling the interaction behavior through the neural network using the representation vectors of the question and the answer result at t time to obtain the interaction record representation vector.

3. The method of claim 1, wherein the learning behavior of the student is fused with the performance of the student. the fusing of the learning gain factors corresponding to different learning behaviors through the multi-modal fusion algorithm to obtain the learning gain vector comprises: the learning behaviors include speed at answering, number of attempts and number of requests for hints, and the corresponding influence factors are referred to as speed influence factors, attempt influence factors and hint influence factors; The learning harvest factor under the control of the speed influence factor at time t is denoted as The learning harvest factor under the control of the attempt influence factor at time t is denoted as The learning harvest factor under the control of the hint influence factor at time t is denoted as The three are spliced with 1 respectively, and then the outer product operation is performed to obtain the outer product tensor G t ; processing by a fully connected layer to obtain a learning harvest vector is represented as: wherein W3 is a weight matrix of a full connection layer, b3 is a bias term of the full connection layer, and ReLU is a rectified linear unit activation function.

4. The method of claim 1, wherein the learning behavior of the student is fused with the performance of the student. the calculation of the knowledge state at t time by setting the forgetting gate and combining the learning gain vector comprises: The topic e at time t is used t The associated knowledge point vector The learning gain vector is propagated to obtain the learning gain lg associated with all knowledge points t , which is expressed as: wherein T is a transpose symbol; Setting the forget gate f t is represented as: wherein, ⊕ represents a vector splicing operation, W4 represents a weight matrix when calculating the forgetting gate, b4 represents a bias term when calculating the forgetting gate; AC t , PC t , NC t denote the speed influence factor, the attempt influence factor and the prompt influence factor, respectively, i t denotes the interaction record representation vector obtained at time t, and σ(·) denotes a sigmoid activation function. In conjunction with the forget gate f t , the knowledge state h t at time t is calculated as: h t = f t *h t-1 + lg t where h t-1 represents the knowledge state at time t-1.

5. The method of claim 1, wherein, the prediction of the answer result of the student by combining the knowledge point vector associated with the question at t+1 time and the knowledge state at t time comprises: The question e at time t+1 t+1 The relevant knowledge point vector The knowledge point vector associated with the question at time t+1 and the knowledge state h at time t t The extracted question e at time t+1 t+1 The degree of mastery under the corresponding knowledge point Is expressed as: Again, the representation vector e' of the question e at time t+1 t+1 t+1 The predicted answer of the student is represented as:​ where y t+1 represents the predicted answer of the student to the question e t+1 at time t+1, W5 and b5 are the weight matrix and bias term for predicting the answer of the student, and σ(·) represents the sigmoid activation function.

6. A system for predicting performance of a student based on fusion of learning behavior, characterized in that, a device for implementing the method of any one of claims 1-5, comprising: The answer process modeling module is configured to obtain a plurality of answer records of a student, each answer record including a question, an answer result and a plurality of learning behaviors at a corresponding time; for a time t, a neural network is used to model the question and the answer result at the time t to obtain an interaction record representation vector; The learning behavior individual action module is configured to calculate corresponding influence factors according to different learning behaviors, and combine the interaction record representation vector to calculate learning gain factors corresponding to the different learning behaviors; The learning behavior joint action module is configured to fuse the learning gain factors corresponding to the different learning behaviors by using a multi-modal fusion algorithm to obtain a learning gain vector; The knowledge state calculation module is configured to set a forgetting gate and combine the learning gain vector to calculate a knowledge state at the time t; The answer result prediction module is configured to combine a knowledge point vector associated with a question at a time t+1 and the knowledge state at the time t to predict an answer result of the student.

7. A processing device, characterized by The computer program product comprises: one or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-5.

8. A readable storage medium, storing a computer program, characterized in that, The computer program product comprises: one or more processors; a memory for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-5.

Images