Learning strategy generation method and system based on student learning state

Abstract

The application discloses a learning strategy generation method and system based on a student learning state, relates to the technical field of intelligent education, and comprises the following steps: obtaining a student state vector and a teaching target vector; obtaining a candidate strategy according to the student state vector and the teaching target vector; obtaining a strategy track and a predicted state sequence according to the candidate strategy based on a state transition model; and performing multi-dimensional evaluation on the strategy track based on the predicted state sequence to obtain a to-be-recommended strategy track. Through explicit modeling of a future T-step strategy track, the candidate strategy that is suitable for the student learning state is selected through the state transition model, the learning state evolution at multiple time steps in the future can be predicted, and the technical effect that the strategy is selected according to the long-term learning state of the student is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to a method and system for generating learning strategies based on students' learning status, which extends strategy selection from short-term feedback optimization to strategy selection based on students' long-term learning status. Background Technology

[0002] In educational learning scenarios, intelligent recommendation systems, such as question recommendation, explanation content recommendation, and learning path planning, are widely used. However, existing recommendation systems suffer from a short-term feedback optimization orientation, primarily based on immediate feedback metrics. For example, question recommendation systems optimize for the accuracy of current questions, content recommendation systems optimize for click-through rates or completion rates, and learning path planning optimizes for improved scores on individual tests. This short-term optimization strategy may lead to: to improve accuracy, the system tends to recommend simple questions that students have already mastered, ignoring knowledge imbalances caused by weak areas; an excessive pursuit of completion rates, pushing a large amount of low-difficulty content, resulting in a loss of learning motivation due to a lack of challenge; and while there may be significant short-term performance improvements, gaps remain in the knowledge system, leading to long-term decline in learning effectiveness as subsequent learning encounters bottlenecks. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for generating learning strategies based on students' learning status, in order to solve the above-mentioned problems existing in the prior art.

[0004] In a first aspect, embodiments of the present invention provide a method for generating learning strategies based on student learning status, including:

[0005] Obtain the student state vector at time t and the teaching objective vector at time t+T; the student state vector represents the student's learning state information at time t; the teaching objective vector represents the student's learning state information at time t+T.

[0006] Based on the student state vector and the teaching objective vector, m candidate strategies are obtained; the m candidate strategies represent learning strategies composed of questions selected from the question bank and explanation content selected from the explanation content bank; m is a positive integer;

[0007] Based on the state transition model, m policy trajectories and m predicted state sequences are obtained according to the m candidate policies; the predicted state sequence includes the predicted state vector of the student at n time nodes within a future time period T; each policy trajectory in the m policy trajectories includes n candidate policies; the value of n is a positive integer.

[0008] Based on the m predicted state sequences, the m policy trajectories are evaluated in multiple dimensions to obtain the policy trajectory to be recommended.

[0009] Optionally, the student state vector is shown in the following formula:

[0010]

[0011] in, Let t be the student state vector at time t; Let be the knowledge point detection vector at time t, and the knowledge point detection vector is an n-dimensional vector. i in One subscript, This represents the student's evaluation value for the i-th knowledge point at time t; Let be the information evaluation value at time t, representing the number of knowledge points evaluated by the student at time t. Let be the student behavior index at time t, representing the learning intensity represented by student behavior at time t; The learning duration value represents the student's level of concentration at time point t.

[0012] Optionally, the state transition model includes a knowledge point detection structure, an information evaluation structure, a student behavior structure, and a learning duration structure; the step of obtaining m policy trajectories and m predicted state sequences based on the state transition model and the m candidate policies includes:

[0013] By using a pre-trained state transition model, the student's state vector at time t is detected after adopting the first candidate strategy, and the student's state vector at time t+1 is obtained; the first candidate strategy is any one of the m candidate strategies.

[0014] Based on the student state vector at time point t+p-1, a greedy algorithm is used to select the candidate strategy at time point t+p-1 from the m candidate strategies to obtain the p-th candidate strategy; p is a positive integer greater than 1 and less than or equal to T.

[0015] By using a pre-trained state transition model, the state of the student at time point t+p-1 after adopting the p-th candidate strategy is detected, and the student state vector at time point t+p is obtained.

[0016] The student state vectors at multiple time points are sorted to obtain the m predicted state sequences;

[0017] The candidate policies corresponding to the m predicted state sequences are sorted to obtain the m policy trajectories.

[0018] Optionally, the method for obtaining the knowledge point detection vector at time point t+1 using the knowledge point detection structure is shown in the following formula:

[0019]

[0020] Among them, the This represents the cumulative evaluation value of the i-th knowledge point in the knowledge point detection vector at time t. This represents the cumulative evaluation value of the i-th knowledge point in the knowledge point detection vector at time point t+1; the... For the learning gain term at time t, the... For the knowledge transfer term at time t, the... This is the forgetting loss term at time t;

[0021] The formula for calculating the learning gain term at time point t is as follows:

[0022]

[0023] Wherein, if the candidate strategy is the problem-recommended strategy, y represents the difficulty matching coefficient. y represents the weight corresponding to the question recommendation strategy; if the candidate strategy is the explanation recommendation strategy, y represents the explanation quality coefficient. This explains the weights corresponding to the recommendation strategy;

[0024] The formula for calculating the forgetting loss term at time point t is as follows:

[0025]

[0026] in, This represents the time length between time point t+1 and time point t. The weights of the forgetting loss terms;

[0027] The knowledge transfer term at time point t is calculated as follows:

[0028]

[0029] Among them, the Represented as in the knowledge graph and related to the above The knowledge graph represents the dependency relationships between multiple knowledge points, with the preceding knowledge point being the dependency. It represents the dependency weights extracted from the knowledge graph.

[0030] Optionally, the method for obtaining the information evaluation value at time point t+1 using the information evaluation structure is shown in the following formula:

[0031]

[0032]

[0033] Wherein, L(t+1) represents the accumulated information evaluation value from time point t to time point t+1. The candidate strategy at time t is: The information evaluation value of the knowledge point detection vector; Indicates the strategy difficulty value; I represents the number of knowledge points. This is a complexity coefficient, representing the complexity of the question or explanation content; Indicates the attenuation coefficient. Indicates the coefficient of restitution;

[0034] Optionally, the method for obtaining the learning motivation index at time point t+1 using the student behavior structure is shown in the following formula:

[0035]

[0036] Among them, the This represents the cumulative learning motivation index from time point t to time point t+1. This is a positive incentive value, representing the percentage of questions answered correctly. This is a negative penalty value, representing the percentage of questions answered incorrectly. The reward factor indicates that the knowledge points in the candidate strategies depend on the knowledge points that the student has already learned. The suppression term indicates that all knowledge points in the candidate strategies are knowledge points that students have already learned.

[0037] Optionally, the method for obtaining the learning duration value at time point t+1 using the learning duration structure is shown in the following formula:

[0038]

[0039] Among them, the This represents the cumulative learning time from time t to time t+1, where ν(π) represents the decay coefficient of the i-th candidate strategy. This represents the gain coefficient of the i-th candidate strategy.

[0040] Optionally, the multi-dimensional assessment includes risk assessment, cost assessment, stability assessment, and target assessment.

[0041] Secondly, embodiments of the present invention provide a learning strategy generation system based on student learning status, comprising:

[0042] The acquisition module is used to acquire the student state vector at time t and the teaching target vector at time t+T; the student state vector represents the student's learning state information at time t; the teaching target vector represents the student's learning state information at time t+T.

[0043] The candidate strategy detection module is used to obtain m candidate strategies based on the student state vector and the teaching objective vector; the m candidate strategies represent strategies consisting of selecting questions from the question bank and selecting explanation content from the explanation content bank; m is a positive integer;

[0044] The state transition module is used to obtain m policy trajectories and m predicted state sequences based on the state transition model and the m candidate policies; the predicted state sequence includes the predicted state vectors of the student at n time nodes within a future time period T; each of the m policy trajectories includes n candidate policies; the value of n is a positive integer.

[0045] The evaluation module is used to perform multi-dimensional evaluation of the m policy trajectories based on the m predicted state sequences to obtain the policy trajectory to be recommended.

[0046] Optionally, the student state vector is shown in the following formula:

[0047]

[0048] in, Let t be the student state vector at time t; Let be the knowledge point detection vector at time t, and the knowledge point detection vector is an n-dimensional vector. i in One subscript, This represents the student's evaluation value for the i-th knowledge point at time t; Let be the information evaluation value at time t, representing the number of knowledge points evaluated by the student at time t. Let be the student behavior index at time t, representing the learning intensity represented by student behavior at time t; The learning duration value represents the student's level of concentration at time point t.

[0049] Thirdly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of any of the methods described above.

[0050] Fourthly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0051] Fifthly, embodiments of the present invention also provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of any of the methods described above.

[0052] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0053] This invention also provides a method and system for generating learning strategies based on students' learning states. By explicitly modeling the strategy trajectory for the next T steps and using a state transition model, candidate strategies adapted to students' learning states are selected. This allows for the prediction of the evolution of learning states over multiple future time steps, while simultaneously satisfying educational principles, knowledge graph structure constraints, and individual adaptability. It achieves the technical effect of selecting strategies based on students' long-term learning states.

[0054] Furthermore, it can identify strategies that pose risks to student learning before deployment, filtering and eliminating them to reduce the probability of students entering high-risk learning states. Through online updates and anomaly detection of the state transition model, it can continuously calibrate the predictive ability for student learning behavior, ensuring that the quality of generated strategies does not decline over time. Since the structural consistency of the state transition model can be predicted without interaction with real students, it supports thorough offline evaluation of candidate strategies. Before deployment to a real learning environment, new strategies can undergo hundreds of trial runs in a simulation environment to identify potential problems and achieve multi-objective balanced optimization. Through Pareto front identification and a multi-dimensional scoring system, it can achieve a balance among four dimensions: goal achievement, stability, risk control, and resource cost, avoiding the side effects of over-optimization of a single goal. It supports offline structural evaluation before deployment, making it more adaptable to the strategy selection of student learning. Attached Figure Description

[0055] Figure 1 This is a flowchart of a learning strategy generation method based on student learning status provided by an embodiment of the present invention.

[0056] Figure 2 This is a flowchart of another learning strategy generation method based on student learning status provided in an embodiment of the present invention.

[0057] Figure 3 This is a flowchart of the sub-steps of step S103 provided in the embodiment of the present invention.

[0058] Figure 4 This is a block diagram of a learning strategy generation system based on student learning status provided in an embodiment of the present invention.

[0059] Figure 5 This is a schematic diagram of the block structure of another learning strategy generation system based on student learning status provided in an embodiment of the present invention.

[0060] Figure 6 The diagram shows a block structure schematic of an electronic device provided by an embodiment of the present invention.

[0061] Icons: 410 - Acquisition module; 420 - Candidate policy detection module; 430 - State transition module; 440 - Evaluation module; 450 - Update module; 600 - Bus; 601 - Receiver; 602 - Processor; 603 - Transmitter; 604 - Memory; 605 - Bus interface. Detailed Implementation

[0062] The present invention will now be described in detail with reference to the accompanying drawings.

[0063] The embodiments of the present invention will be described in detail below with reference to examples. However, those skilled in the art will understand that the following examples are only for illustrating the present invention and should not be regarded as limiting the scope of the present invention.

[0064] Example 1

[0065] like Figure 1 As shown, this embodiment of the invention provides a method for generating learning strategies based on students' learning states, including:

[0066] S101: Obtain the student state vector at time t and the teaching objective vector at time t+T. The student state vector represents the student's learning state information at time t. The teaching objective vector represents the student's learning state information at time t+T.

[0067] S102: Based on the student state vector and the learning objective vector, obtain m candidate strategies. The m candidate strategies represent learning strategies composed of questions selected from the question bank and explanation content selected from the explanation content bank. m is a positive integer.

[0068] Where m represents the number of candidate strategies, a positive integer greater than or equal to 10 and less than or equal to 20, used to balance computational efficiency and strategy diversity. The m candidate strategies include question recommendation strategies and explanation recommendation strategies. Question recommendation strategies represent questions in a question bank. The question bank stores the associations between questions and multiple knowledge points, as well as their difficulty levels. Questions corresponding to multiple knowledge points included in the pre-teaching lesson plan are used as first pre-selected questions, resulting in multiple first pre-selected questions for each knowledge point. Among these first pre-selected questions, questions related to knowledge points that students haven't learned and those related to knowledge points where students answered pre-selected questions incorrectly are identified, resulting in multiple second pre-selected questions. Student levels are determined based on the number of incorrect answers to pre-selected questions. Among these second pre-selected questions, questions corresponding to difficulty levels matching the student level are found, resulting in multiple third pre-selected questions. Questions corresponding to knowledge points that are prerequisites of the knowledge points involved in the third pre-selected questions are found, resulting in fourth pre-selected questions. If the fourth pre-selected question achieves a correct answer rate greater than 85% when tested on students, the corresponding third pre-selected question is used as the question recommendation strategy. Explanation recommendation strategies represent explanation information for knowledge points in an explanation content library. The first pre-selected explanation includes the explanations of the three knowledge points with the lowest accuracy rates in students' answers to pre-selected questions. This first pre-selected explanation targets weak knowledge points. The first pre-selected explanation, whose explanation time is less than the distribution value of learning time, is used as the second pre-selected explanation. This second pre-selected explanation is used for time adaptation. In the second pre-selected explanation, the explanation of knowledge points that are subsequently dependent on the fourth pre-selected question is used as the explanation recommendation strategy. This strategy is used to progressively increase the difficulty of knowledge points.

[0069] S103: Based on the state transition model, obtain m policy trajectories and m predicted state sequences from m candidate policies. Each candidate policy corresponds one-to-one with a policy trajectory and a predicted state sequence. A predicted state sequence includes the predicted learning state vectors of the student at n time points within a future time period T. A policy trajectory includes n candidate policies. n is a positive integer. p represents a positive integer greater than 0 and less than n. n is the number of time steps within the time period T. The time point t+p represents a future time point.

[0070] S104: Based on m predicted state sequences, perform multi-dimensional evaluation on m policy trajectories to obtain the policy trajectory to be recommended.

[0071] Optional, such as Figure 2 As shown, the learning strategy generation method based on student learning status provided in this application also includes:

[0072] S105: Based on the trajectory of the strategy to be recommended, obtain feedback data and update the state transition model.

[0073] The feedback data consists of a student state vector obtained based on the first candidate strategy in the trajectory of the strategy to be recommended.

[0074] In this embodiment, when the distance between the feedback data at time t+1 and the student state vector predicted by the first candidate strategy in the trajectory of the recommended strategy at time t+1 is less than a first error threshold, no update is performed, and the model is considered still accurate. When the distance between the feedback data at time t+1 and the student state vector predicted by the first candidate strategy in the trajectory of the recommended strategy at time t+1 is greater than or equal to the first error threshold and less than a second error threshold, the parameters of the state transition model are fine-tuned using gradient descent, where the second error threshold is greater than the first error threshold. When the distance between the feedback data at time t+1 and the student state vector predicted by the first candidate strategy in the trajectory of the recommended strategy at time t+1 is greater than the second error threshold, a model retraining process is triggered, using the feedback data as the historical dataset to refit the parameters of the state transition model. In this embodiment, the first error threshold can be, for example, 0.05, and the second error threshold can be, for example, 0.2.

[0075] Optionally, the student state vector is shown in the following formula:

[0076]

[0077] in, Let t be the student state vector at time t, where time t represents the current time. Let be the knowledge point detection vector at time t. The knowledge point detection vector is an n-dimensional vector. i in One subscript, This represents the degree to which a student has mastered the i-th knowledge point at time t. Let be the information evaluation value at time t, representing the number of knowledge points evaluated by the student at time t. Let be the student behavior index at time t, representing the intensity of learning by students using student behavior representations at time t. The learning duration value represents the student's level of concentration at time point t.

[0078] in, , , , .

[0079] In this embodiment, Knowledge can be obtained from historical answers to pre-selected questions using a Bayesian knowledge tracing model, or, as another example, from deep knowledge tracing models.

[0080] The acquisition method and The acquisition method is the same.

[0081] The acquisition method and The acquisition method is the same.

[0082] Using an exponential decay model based on continuous learning duration: get, This represents the continuous learning duration from the start of learning to time point t. The duration of study time at the start of learning indicates the student's level of concentration.

[0083] Optionally, based on the state transition model, m policy trajectories and m predicted state sequences are obtained from m candidate policies, including sub-steps S301~S308. Please refer to the relevant documentation. Figure 3 :

[0084] S301: Obtain the pre-trained state transition model.

[0085] The state transition model includes a knowledge point detection structure, an information evaluation structure, a student behavior structure, and a learning duration structure. The pre-training method involves learning the student state vector at time t from the initial learning time point, and then training the weights in the state transition model using the labeled student state vector at time t.

[0086] S302: Obtain the first candidate strategy.

[0087] The first candidate strategy is one candidate strategy randomly selected from m candidate strategies.

[0088] S303: By using a pre-trained state transition model, detect the student's state vector at time t after adopting the first candidate strategy, and obtain the student's state vector at time t+1.

[0089] S304: Based on the student's state vector at time t+1, a greedy algorithm is used to select the candidate strategy at time t+1 from m candidate strategies to obtain the second candidate strategy.

[0090] Here, the second candidate strategy represents the strategy with the largest average sum of the elements of the student's state vector among the m candidate strategies at time point t+1.

[0091] S305: Using a pre-trained state transition model, detect the student's state vector at time t+1 after adopting the second candidate strategy, and obtain the student's state vector at time t+2.

[0092] S306: Using a pre-trained state transition model and a greedy algorithm, detect the student's state at time t+p-1 after adopting the p-th candidate strategy, and obtain the student's state vector at time t+p. p is a positive integer greater than 1 and less than or equal to T.

[0093] Where T represents the point in time when learning ends.

[0094] S307: Sort the student state vectors at time point t+p in order from early to late to obtain the predicted state sequence.

[0095] S308: Sort the candidate policies corresponding to the predicted state sequence to obtain the policy trajectory.

[0096] For each m candidate policy, m policy trajectories and m predicted state sequences are obtained.

[0097] Optionally, the method for obtaining the knowledge point detection vector at time point t+1 using the knowledge point detection structure is shown in the following formula:

[0098]

[0099] in, This represents the cumulative evaluation value of the i-th knowledge point in the knowledge point detection vector at time t. This represents the cumulative evaluation value of the i-th knowledge point in the knowledge point detection vector at time point t+1. The learning gain term at time t is given by... The i-th element of the knowledge point detection vector at time t and the learning duration value at time t are used as input. This is a candidate strategy. Let be the knowledge transfer term at time t. The knowledge transfer term is composed of the knowledge point detection vector at time t and . As input, Knowledge graph. Let be the forgetting loss term at time t. The forgetting loss term is composed of the i-th element of the knowledge point detection vector at time t and . As input, This represents the time length between time point t+1 and time point t.

[0100] in, Let i represent the candidate strategies for the problem recommendation strategy and the explanation recommendation strategy. i is a positive integer greater than 0 and less than or equal to i.

[0101] The formula for calculating the learning gain term is as follows:

[0102]

[0103] Wherein, if the candidate strategy is the problem-recommended strategy, y represents the difficulty matching coefficient. This represents the weight corresponding to the question recommendation strategy. If the candidate strategy is the explanation recommendation strategy, y represents the explanation quality coefficient. This indicates the weights corresponding to the recommendation strategy.

[0104] The difficulty matching coefficient represents the matching relationship between the difficulty of the question and the student's ability. It is the difficulty level of the question minus the student's level at time point t. Acquired through training.

[0105] The formula for calculating the forgetting loss term is as follows:

[0106]

[0107] in, This represents the time length between time point t+1 and time point t. This represents the weight of the forgetting loss term.

[0108] The calculation method for knowledge transfer items is as follows:

[0109]

[0110] Specifically, the knowledge transfer component uses a knowledge graph to detect dependencies. The knowledge graph uses a graph structure to store multiple knowledge points and the dependencies between them.

[0111] Represented as In knowledge graphs and The previous knowledge point is a dependency; a knowledge graph represents the dependency relationship between multiple knowledge points. It represents the dependency weights extracted from the knowledge graph.

[0112] Optional methods for using information evaluation structures, student behavior structures, and learning duration structures include:

[0113] The method for obtaining the information evaluation value at time point t+1 using the information evaluation structure is shown in the following formula:

[0114]

[0115]

[0116] Where L(t+1) represents the accumulated information evaluation value from time point t to time point t+1. The candidate strategy at time t is: The information evaluation value of the knowledge point detection vector. This represents the strategy difficulty value. I represents the number of knowledge points. This is the complexity coefficient, representing the complexity of the problem or explanation content. Indicates the attenuation coefficient. This represents the coefficient of recovery.

[0117] The complexity of a question or explanation represents the number of knowledge points within it. The strategy difficulty value represents the difficulty level of the question or explanation; one recommended strategy for a question corresponds to one labeled question difficulty level, and one recommended strategy for a question corresponds to several labeled explanation difficulty levels. The question difficulty level or explanation difficulty level represents the ratio of related knowledge points in the question or explanation to the knowledge points already learned by the student.

[0118] The method for using student behavior structure to obtain the learning motivation index at time point t+1 is shown in the following formula:

[0119]

[0120] in, This represents the cumulative learning motivation index from time point t to time point t+1. This is a positive incentive value, representing the percentage of questions answered correctly. The negative penalty value represents the percentage of questions answered incorrectly. The reward factor indicates that the knowledge points in the candidate strategies depend on the knowledge points that the student has already learned. The suppression term indicates that all knowledge points in the candidate strategies are knowledge points that students have already learned.

[0121] If the knowledge points in the candidate strategies depend on the knowledge points that the students have already learned, that is, the fewer adjacent nodes there are between the two in the knowledge graph, then... The larger the value, the better. If the knowledge points in the candidate strategy are knowledge points that the student has already learned, that is, if the candidate strategy does not have any decay of other knowledge points besides those learned by the student, then... If the candidate strategy contains knowledge points other than those learned by students, the value is 0.1. It is 0. and The questions to be answered are pre-selected.

[0122] The method for using the attention horizontal transfer structure to obtain the learning duration value at time point t+1 is shown in the following formula:

[0123]

[0124] in, Let represent the cumulative learning time from time t to time t+1, and ν(π) represent the decay coefficient of the i-th candidate strategy. This represents the gain coefficient of the i-th candidate strategy.

[0125] Optionally, based on m predicted state sequences, the candidate policies are evaluated in multiple dimensions to obtain the predicted state sequences, including:

[0126] Based on the predicted state sequence, a risk assessment is performed to obtain a risk assessment value. The risk assessment value represents the probability that the predicted state sequence carries risk. A higher risk assessment value indicates a lower risk.

[0127] Based on the predicted state sequence, a cost assessment is performed to obtain a cost assessment value. This cost assessment value represents the estimated time and learning resources required to transition from the learning state at time t to the learning state at time t+T.

[0128] The cost assessment includes acquisition time cost, information evaluation cost, and learning resource cost. These costs are summed to obtain the overall cost evaluation value. The time cost is T, the information evaluation cost represents the sum of T information evaluation values ​​from the student's state vector at time t to the student's state vector at time t+T, and the learning resource cost represents the number of questions in the predicted state sequence's question strategy and the number of explanations in the explanation strategy.

[0129] Based on the predicted state sequence, the degree of fluctuation and convergence of the predicted state sequence are determined, and a stability evaluation value is obtained.

[0130] Among them, the stability evaluation value characterizes the degree of volatility and convergence of the predicted state sequence.

[0131] Calculate the distance between the student's state vector and the teaching objective vector at time point t+T to obtain the objective evaluation value.

[0132] The target evaluation value represents the distance between the student's state vector and the teaching target vector at time t+T.

[0133] A score set is constructed from the risk assessment value, cost assessment value, stability assessment value, and target assessment value. Each of the m predicted state sequences corresponds to one of the m score sets.

[0134] Based on m predicted state sequences and m rating sets, m policy trajectories are evaluated in multiple dimensions to obtain the policy trajectory to be recommended.

[0135] Optionally, based on the predicted state sequence, the volatility and convergence of the predicted state sequence are determined to obtain a stability evaluation value, including:

[0136] Based on the predicted state sequence, the variance of the T student state vectors between time point t and time point t+T is detected to obtain the orbital variance. The predicted state sequence represents the degree of state fluctuation of the T student state vectors.

[0137] Based on the predicted state sequence, the distance between the student state vectors corresponding to two adjacent time points is detected to obtain the first fluctuation amplitude. T time points correspond to T first fluctuation amplitudes.

[0138] The largest of the T initial fluctuation amplitudes is taken as the fluctuation amplitude. The fluctuation amplitude represents the magnitude of the change in the student's state at adjacent time points.

[0139] The convergence rate is obtained by subtracting the difference between the student's state vector and the learning target vector at time t+T from the distance between the student's state vector and the learning target vector at time t+T, and then dividing by T. The convergence rate represents how quickly the predicted state sequence approaches the target state.

[0140] Obtain variance weights, volatility weights, and convergence weights.

[0141] Among them, the variance weight, volatility weight, and convergence weight are obtained through training.

[0142] The orbital variance score is obtained by multiplying the reciprocal of the sum of the orbital variances plus 1 by the variance weight. The orbital fluctuation score is obtained by multiplying the reciprocal of the sum of the fluctuation amplitudes plus 1 by the fluctuation weight. The orbital convergence score is obtained by multiplying the convergence velocity by the fluctuation weight.

[0143] The stability assessment value is obtained by adding the orbital convergence score, the orbital fluctuation score, and the orbital variance score.

[0144] Optionally, based on the predicted state sequence, a risk assessment is performed to obtain a risk assessment value, including:

[0145] Information acquisition evaluation threshold, student behavior threshold, learning duration threshold, and knowledge point detection threshold.

[0146] In this embodiment, the information evaluation threshold is 0.85; an information evaluation value greater than 0.85 indicates cognitive overload. The student behavior threshold is 0.3; a student behavior index less than 0.3 indicates a collapse in learning motivation. The learning duration threshold is 0.25; a learning duration less than 0.25 indicates attention exhaustion. The knowledge point detection threshold is 0.6; a Gini coefficient greater than 0.6 in the student's state vector indicates an imbalance in knowledge structure, meaning that some knowledge points are learned too strongly while others are learned too weakly.

[0147] The conditions that the information evaluation value at time point t+p is greater than the information evaluation threshold, or the student behavior index is less than the student behavior threshold, or the learning duration value is less than the learning duration threshold, or the Gini coefficient of the student state vector is greater than the knowledge point detection threshold are taken as risk conditions.

[0148] If the student's state vector at time t+p does not meet the risk conditions, the risk assessment value is set to 1.

[0149] If the time point t+p is different from t+p+ If the student's state vector at a given time point meets the risk condition, the risk assessment value is set as the first risk assessment value. Less than the risk time threshold.

[0150] In this embodiment, the risk time threshold is 3. The first risk assessment value is a value in the range [0.6, 0.8]. A value less than 3 indicates that the predicted state sequence briefly meets the risk condition.

[0151] If the time point t+p is different from t+p+ If the student's state vector at a given time point meets the risk condition, the risk assessment value is set as the second risk assessment value. Greater than or equal to the risk time threshold.

[0152] In this embodiment, the second risk assessment value is a value in the range [0.0, 0.4]. A value greater than or equal to 3 indicates that the predicted state sequence remains unchanged for an extended period, meeting the risk criteria. The second risk assessment value is less than the first risk assessment value, which is less than 1.

[0153] The above methods show that the smaller the risk, the higher the risk assessment value.

[0154] By employing the methods described above, a risk perception mechanism can be constructed to identify strategies that pose risks to student learning and filter them out.

[0155] This reduces the probability of students engaging in high-risk learning states and improves the long-term operational stability of the closed-loop strategy.

[0156] Optionally, based on m predicted state sequences and m rating sets, a multi-dimensional evaluation is performed on m policy trajectories to obtain the policy trajectory to be recommended, including:

[0157] Based on m predicted state sequences, hard constraint filtering is performed to select policy trajectories, resulting in... The first strategy trajectory.

[0158] Hard constraint filtering refers to filtering based on target threshold, stability threshold, risk threshold, and cost threshold.

[0159] The target evaluation value corresponding to the first strategy trajectory is greater than the target threshold, the stability evaluation value is greater than the stability threshold, the risk evaluation value is greater than the risk threshold, and the cost evaluation value is greater than the cost threshold.

[0160] In this embodiment, the target threshold is 0.6, indicating that after T time points, the student's learning state is greater than 60% of the target state of the teaching objective vector. The stability threshold is 0.5, indicating that the learning process of the candidate strategy will not abruptly or become disordered. The risk threshold is 0.7, indicating that the candidate strategy should not cause the student to meet risk conditions. The cost threshold is 0.4, representing the total cost of the candidate strategy.

[0161] according to The first strategy trajectory is obtained by using the Pareto front identification algorithm. A second strategy trajectory. A second strategy trajectory represents a first strategy trajectory that is not dominated by any other first strategy trajectory.

[0162] Among them, for The first strategy trajectory is selected using the Pareto front identification algorithm based on the Pareto optimality principle. There are two first strategy trajectories, first strategy trajectory a and first strategy trajectory b. If the student state vector at each time point corresponding to first strategy trajectory a is greater than or equal to the student state vector at each time point corresponding to first strategy trajectory b, and there exists a student state vector at one time point corresponding to first strategy trajectory a that is greater than the student state vector at one time point corresponding to first strategy trajectory b, then first strategy trajectory a is taken as the second strategy trajectory.

[0163] The comprehensive score is obtained by multiplying the target evaluation value corresponding to the second strategy trajectory by the target weight, the stability evaluation value by the stability weight, the risk evaluation value by the risk weight, and the cost evaluation value by the cost weight.

[0164] In this embodiment, the target weight is 0.30, the stability weight is 0.20, the risk weight is 0.30, and the cost weight is 0.20, representing a balanced recommendation. In another embodiment, the target weight is 0.25, the stability weight is 0.20, the risk weight is 0.45, and the cost weight is 0.10, representing a recommendation that prioritizes risk. In yet another embodiment, the target weight is 0.40, the stability weight is 0.10, the risk weight is 0.20, and the cost weight is 0.30, representing a recommendation that prioritizes efficiency.

[0165] The second-highest comprehensive score trajectory is selected as the recommended strategy trajectory. Each recommended strategy trajectory is accompanied by a detailed report on expected outcomes, including trajectory visualization, risk assessment, and stability indicators. This allows teachers and students to understand the rationale behind the recommendations, improving the system's credibility and acceptance.

[0166] By employing Pareto front identification and a multidimensional scoring system, strategies that excel in one particular indicator but are severely imbalanced in other dimensions, especially risk control and stability, are accurately identified and eliminated. This mechanism eliminates systemic risks caused by neglecting any single dimension and proactively avoids the phenomenon of "dimensional overfitting." It achieves a balance among the four dimensions of goal achievement, stability, risk control, and resource cost, avoiding the side effects of over-optimizing a single objective.

[0167] Example 2

[0168] Based on the above-described learning strategy generation method based on student learning status, this embodiment of the invention also provides an execution subject for performing the above steps. This execution subject can be... Figure 4 The learning strategy generation system 200 based on student learning status is shown below. Please refer to [reference needed]. Figure 4 The system includes:

[0169] The acquisition module 410 is used to acquire the student state vector at time t and the teaching target vector at time t+T. The student state vector represents the learning state information of the student detected at time t. The teaching target vector represents the learning state information of the student expected to be detected at time t+T.

[0170] The candidate strategy detection module 420 is used to obtain m candidate strategies based on the student state vector and the teaching objective vector. The candidate strategy represents a strategy consisting of selecting questions from the question bank and selecting explanation content from the explanation content library. m is a positive integer.

[0171] The state transition module 430 is used to obtain m policy trajectories and m predicted state sequences based on the state transition model and the m candidate policies. Each candidate policy corresponds one-to-one with a policy trajectory and a predicted state sequence. The predicted state sequence includes the predicted state vectors of the student at n time points within a future time period T. The policy trajectory includes n candidate policies. The value of n is a positive integer.

[0172] Evaluation module 440 is used to perform multi-dimensional evaluation of the m policy trajectories based on the m predicted state sequences to obtain the policy trajectory to be recommended.

[0173] As an optional implementation, the learning strategy generation system 200 based on student learning status also includes, for example, Figure 5 The update module 450 shown:

[0174] The update module 450 is used to obtain feedback data based on the trajectory of the strategy to be recommended and update the state transition model; the feedback data is the student state vector obtained based on the first candidate strategy in the trajectory of the strategy to be recommended.

[0175] Optionally, the student state vector is shown in the following formula:

[0176]

[0177] in, Let t be the student state vector at time t. Let t be the knowledge point detection vector, and let t be an n-dimensional vector. i in One subscript, This represents the evaluation value of the i-th knowledge point by the student at time t. Let be the information evaluation value at time t, and represent the number of knowledge points evaluated by the student at time t. Let be the student behavior index at time t, representing the learning intensity represented by student behavior at time t. The learning duration value represents the student's level of concentration at time point t.

[0178] Regarding the systems and apparatus in the above embodiments, the specific methods by which each module performs its operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0179] This invention also provides an electronic device, such as... Figure 6 As shown, it includes a memory 604, a processor 602, and a computer program stored in the memory 604 and executable on the processor 602. When the processor 602 executes the program, it implements the steps of the learning strategy generation method based on student learning status described above.

[0180] Among them, Figure 6 In this document, a bus architecture (represented by bus 600) is used. Bus 600 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 602 and memory represented by memory 604. Bus 600 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 605 provides an interface between bus 600 and receiver 601 and transmitter 603. Receiver 601 and transmitter 603 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 602 is responsible for managing bus 600 and general processing, while memory 604 can be used to store data used by processor 602 during operation.

[0181] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the learning strategy generation method based on student learning status described above.

[0182] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the learning strategy generation method based on student learning status described above.

[0183] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0184] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0185] The various component embodiments of the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components in the apparatus according to embodiments of the present invention. The present invention can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such programs implementing the present invention can be stored on a computer-readable medium or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

Claims

1. A method for generating learning strategies based on students' learning states, characterized in that, include: Obtain the student state vector at time t and the teaching objective vector at time t+T; the student state vector represents the student's learning state information at time t; the teaching objective vector represents the student's learning state information at time t+T. Based on the student state vector and the teaching objective vector, m candidate strategies are obtained; the m candidate strategies represent learning strategies composed of questions selected from the question bank and explanation content selected from the explanation content bank. m is a positive integer; Based on the state transition model, m policy trajectories and m predicted state sequences are obtained according to the m candidate policies; the predicted state sequence includes the predicted state vector of the student at n time nodes within a future time period T; each policy trajectory in the m policy trajectories includes n candidate policies; the value of n is a positive integer. Based on the m predicted state sequences, the m policy trajectories are evaluated in multiple dimensions to obtain the policy trajectory to be recommended.

2. The learning strategy generation method based on student learning status according to claim 1, characterized in that, The student state vector is shown in the following formula: in, Let t be the student state vector at time t; Let be the knowledge point detection vector at time t, and the knowledge point detection vector is an n-dimensional vector. i in One subscript, This represents the student's evaluation value for the i-th knowledge point at time t; Let be the information evaluation value at time t, representing the number of knowledge points evaluated by the student at time t. Let be the student behavior index at time t, representing the learning intensity represented by student behavior at time t; The learning duration value represents the student's level of concentration at time point t.

3. The learning strategy generation method based on student learning status according to claim 1, characterized in that, The state transition model includes a knowledge point detection structure, an information evaluation structure, a student behavior structure, and a learning duration structure. The state transition model, which obtains m policy trajectories and m predicted state sequences based on the m candidate policies, includes: By using a pre-trained state transition model, the student's state vector at time t is detected after adopting the first candidate strategy, and the student's state vector at time t+1 is obtained; the first candidate strategy is any one of the m candidate strategies. Based on the student state vector at time point t+p-1, a greedy algorithm is used to select the candidate strategy at time point t+p-1 from the m candidate strategies to obtain the p-th candidate strategy; p is a positive integer greater than 1 and less than or equal to T; By using a pre-trained state transition model, the state of the student at time point t+p-1 after adopting the p-th candidate strategy is detected, and the student state vector at time point t+p is obtained. The student state vectors at multiple time points are sorted to obtain the m predicted state sequences; The candidate policies corresponding to the m predicted state sequences are sorted to obtain the m policy trajectories.

4. The learning strategy generation method based on student learning status according to claim 3, characterized in that, The method for obtaining the knowledge point detection vector at time point t+1 using the knowledge point detection structure is shown in the following formula: Among them, the This represents the cumulative evaluation value of the i-th knowledge point in the knowledge point detection vector at time t. This represents the cumulative evaluation value of the i-th knowledge point in the knowledge point detection vector at time point t+1; the... For the learning gain term at time t, the... For the knowledge transfer term at time t, the... This is the forgetting loss term at time t; The formula for calculating the learning gain term at time point t is as follows: Wherein, if the candidate strategy is the problem-recommended strategy, y represents the difficulty matching coefficient. y represents the weight corresponding to the question recommendation strategy; if the candidate strategy is the explanation recommendation strategy, y represents the explanation quality coefficient. This explains the weights corresponding to the recommendation strategy; The formula for calculating the forgetting loss term at time point t is as follows: in, This represents the time length between time point t+1 and time point t. The weights of the forgetting loss terms; The knowledge transfer term at time point t is calculated as follows: Among them, the Represented as in the knowledge graph and related to the above The knowledge graph represents the dependency relationships between multiple knowledge points, with the preceding knowledge point being the dependency. It represents the dependency weights extracted from the knowledge graph.

5. The learning strategy generation method based on student learning status according to claim 3, characterized in that, The method for obtaining the information evaluation value at time point t+1 using the information evaluation structure is shown in the following formula: Wherein, L(t+1) represents the accumulated information evaluation value from time point t to time point t+1. The candidate strategy at time t is: The information evaluation value of the knowledge point detection vector; Indicates the strategy difficulty value; I represents the number of knowledge points. This is a complexity coefficient, representing the complexity of the question or explanation content; Indicates the attenuation coefficient. This represents the coefficient of recovery.

6. The learning strategy generation method based on student learning status according to claim 3, characterized in that, The method for obtaining the learning motivation index at time point t+1 using the student behavior structure is shown in the following formula: Among them, the This represents the cumulative learning motivation index from time point t to time point t+1. This is a positive incentive value, representing the percentage of questions answered correctly. This is a negative penalty value, representing the percentage of questions answered incorrectly. The reward factor indicates that the knowledge points in the candidate strategies depend on the knowledge points that the student has already learned. The suppression term indicates that all knowledge points in the candidate strategies are knowledge points that students have already learned.

7. The learning strategy generation method based on student learning status according to claim 3, characterized in that, The method for obtaining the learning duration value at time point t+1 using the learning duration structure is shown in the following formula: Among them, the This represents the cumulative learning time from time t to time t+1, where ν(π) represents the decay coefficient of the i-th candidate strategy. This represents the gain coefficient of the i-th candidate strategy.

8. The learning strategy generation method based on student learning status according to any one of claims 1-7, characterized in that, The multi-dimensional assessment includes risk assessment, cost assessment, stability assessment, and target assessment.

9. A learning strategy generation system based on student learning status, characterized in that, include: The acquisition module is used to acquire the student state vector at time t and the teaching target vector at time t+T; the student state vector represents the student's learning state information at time t; the teaching target vector represents the student's learning state information at time t+T. The candidate strategy detection module is used to obtain m candidate strategies based on the student state vector and the teaching objective vector; the m candidate strategies represent strategies consisting of selecting questions from the question bank and selecting explanation content from the explanation content library. m is a positive integer; The state transition module is used to obtain m policy trajectories and m predicted state sequences based on the state transition model and the m candidate policies; the predicted state sequence includes the predicted state vectors of the student at n time nodes within a future time period T; each of the m policy trajectories includes n candidate policies; the value of n is a positive integer. The evaluation module is used to perform multi-dimensional evaluation of the m policy trajectories based on the m predicted state sequences to obtain the policy trajectory to be recommended.

10. The learning strategy generation system based on student learning status according to claim 9, characterized in that, The student state vector is shown in the following formula: in, Let t be the student state vector at time t; Let be the knowledge point detection vector at time t, and the knowledge point detection vector is an n-dimensional vector. i in One subscript, This represents the student's evaluation value for the i-th knowledge point at time t; Let be the information evaluation value at time t, representing the number of knowledge points evaluated by the student at time t. Let be the student behavior index at time t, representing the learning intensity represented by student behavior at time t; The learning duration value represents the student's level of concentration at time point t.

Images