An individualized learning path planning method and system based on LSTM fusion attention

Abstract

The application relates to the technical field of learning path planning, in particular to a personalized learning path planning method and system based on LSTM fusion Attention, which comprises the following steps: constructing a question-answering feature time sequence based on information of question-answering records, wrong question tables and knowledge point sets, and generating question-answering performance features as hidden states through a time sequence correlation trend extraction model, wherein the time sequence correlation trend extraction model is constructed based on an LSTM architecture; calculating dynamic attention weights through an Attention model by setting a time decay term, a continuous question-answering difficulty feedback term and a knowledge point correlation term, and adjusting iterations of the hidden states based on the dynamic attention weights; generating a predicted knowledge point mastery degree through a fusion mapping model, wherein the fusion mapping model is constructed based on a full connection layer and an activation function; and generating a personalized recommended learning path based on the predicted knowledge point mastery degree. The application realizes personalized learning path planning that fits the individual needs of students and improves the intelligentization of a college teaching system.

Description

Technical Field

[0001] This invention relates to the field of learning path planning technology, and in particular to a personalized learning path planning method and system based on LSTM fused with Attention. Background Technology

[0002] With the rapid development of artificial intelligence technology, higher education is gradually transforming towards intelligence and personalization. This is because if a standardized model is adopted for course learning paths, it is difficult to adapt to the differentiated needs of individual students in terms of knowledge base, learning habits, and weak areas. As a result, some students cannot keep up and their learning efficiency is reduced, which has become one of the key issues restricting the improvement of teaching quality.

[0003] In order to solve the above problems in higher education, personalized learning path planning methods have been gradually introduced into the field of educational technology in recent years, which dynamically adjust learning content by analyzing students' learning behavior data.

[0004] Currently, student answer data on online learning platforms is mostly stored in databases in the format of knowledge graphs. This is because the knowledge points learned in most subject areas are not isolated or unrelated; learners often need to master one or more knowledge points before learning others. Knowledge graphs (KG), as one of the core technologies of artificial intelligence, can reflect the hierarchical structure of knowledge in a subject area, represent the predecessors and successors between knowledge point entities, and include logical relationships such as parallel relationships.

[0005] Currently, some methods build student knowledge mastery models based on knowledge graphs, identifying weaknesses by mining the correlation between incorrect answers and knowledge points, thus more accurately tracking the evolution of learning status. However, existing methods still have the following limitations:

[0006] First, the temporal features are not fully utilized. Students' answering behavior has a significant temporal dependence; the impact of early answering behavior on current knowledge mastery decays over time, while performance on consecutive high-difficulty questions may reflect the dynamic improvement or fluctuation of students' abilities. Therefore, existing LSTM models do not fully utilize the temporal features of students' answers.

[0007] Secondly, the adaptability of attention mechanisms needs improvement. Although some methods introduce attention mechanisms to enhance focus on key learning behaviors, the calculation of attention weights is mostly based on a single dimension, such as the accuracy of answering questions, without comprehensively considering factors such as time decay. This results in insufficient adaptability of attention allocation to actual learning scenarios, ultimately affecting the accuracy of personalized path planning.

[0008] Therefore, how to accurately model students' learning status and plan personalized learning paths through a model based on LSTM and fused with Attention is a technical problem that needs to be solved. Summary of the Invention

[0009] To address this, the present invention provides a personalized learning path planning method and system based on LSTM fusion Attention. This method uses an LSTM fusion Attention model to predict the mastery of knowledge points from the knowledge graph data of an online learning platform and plans personalized learning paths to meet the individual needs of students, thereby realizing the intelligentization of the university teaching system.

[0010] To achieve the above objectives, this invention proposes a personalized learning path planning method based on LSTM fused with Attention, comprising:

[0011] The answer feature time series is constructed based on the answer records, wrong answer list and knowledge point set in the student answer knowledge graph. The time series correlation trend extraction model is constructed based on the LSTM architecture and the answer performance features are used as hidden states for iterative calculation.

[0012] In the process of generating the answer performance features, the answer performance features are used to calculate dynamic attention weights through an Attention model that sets time decay terms, continuous answer difficulty feedback terms, and knowledge point association terms, and the iteration of the hidden state of the time-series association trend extraction model is adjusted based on the dynamic attention weights.

[0013] The answer performance features at multiple time steps are used to generate a predicted knowledge point mastery level through a fusion mapping model, wherein the fusion mapping model is constructed based on fully connected layers and activation functions;

[0014] Based on the predicted knowledge point mastery, a personalized recommended learning path is generated in the student's answer knowledge graph.

[0015] Furthermore, the student answer knowledge graph also includes the graph relationship between the answer records and the corresponding questions, and the process of calculating the dynamic attention weights includes:

[0016] The question-answering features time sequence is used to calculate the original attention score through the multi-head attention mechanism of the Attention model;

[0017] The time decay term is calculated based on the proximity of the current time step to the set recent time step, as well as the trainable parameters of the learning rhythm.

[0018] Based on the question-answering records stored in the student's question-answering knowledge graph, the difficulty coefficient, trainable adjustable parameters, and activation function of the questions corresponding to consecutive correct answers and consecutive incorrect answers are used to calculate the consecutive answering difficulty feedback item.

[0019] Calculate the knowledge point association items based on the degree of association between the knowledge point at the current time step and the target knowledge point in the set of knowledge points;

[0020] The dynamic attention weight is determined by integrating the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term.

[0021] Furthermore, the process of calculating the original attention score includes:

[0022] The original attention score is calculated by sequentially passing the concatenation variables of the target knowledge point and the hidden state at the current time step, the trainable weighted parameters, and the trainable bias parameters through the tanh activation function and the softmax activation function.

[0023] Furthermore, the process of calculating the time decay term includes: calculating the difference between the current time step and the set recent time step, multiplying the difference by the negative of the learning rhythm trainable parameter, and then performing an exponential function operation to determine the time decay term;

[0024] The process of calculating the continuous answer difficulty feedback term includes: using the concatenated vector of the current answer result, the current continuous error label vector, and the difficulty coefficient of the answer record, and the trainable adjustable parameter to calculate the continuous answer difficulty feedback term through the sigmoid activation function.

[0025] Furthermore, the process of determining the dynamic attention weight by fusing the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term includes:

[0026] The product of the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term is used as the temporary attention score;

[0027] The temporary attention score is normalized using the softmax activation function to determine the dynamic attention weight.

[0028] Furthermore, the iterative process of adjusting the hidden state of the temporal correlation trend extraction model based on the dynamic attention weights includes:

[0029] Calculate the weighted sum vector of the hidden state of the set recent time step and the corresponding dynamic attention weights. Multiply the concatenated value of the hidden state of the current time step and the weighted sum vector, along with the gated trainable parameters, by the hidden state of the current time step after passing through the sigmoid activation function to determine the hidden state of the next time step of the iteration.

[0030] Furthermore, the training optimization process for the trainable parameters of the learning rhythm, the trainable adjustment parameters, the trainable weighted parameters, the trainable bias parameters, and the gating trainable parameters includes:

[0031] The product of the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term is used as the event importance weight. The mean square error between the true knowledge point mastery and the training knowledge point mastery of the sample dataset is calculated. The event importance weight and the mean square error are multiplied together to construct a weighted mean square error term.

[0032] Obtain the dependency relationship between the mastery of training knowledge points and historical knowledge points in the student's answer knowledge graph to construct a knowledge point relationship loss term;

[0033] The weighted mean squared error term and the knowledge point relationship loss term are weighted and summed to generate a comprehensive loss function. The Attention model and the fusion mapping model, which include the trainable parameters of the learning rhythm, trainable adjustment parameters, trainable weighted parameters, trainable bias parameters and gating trainable parameters, are then trained and optimized using the comprehensive loss function.

[0034] The above scheme achieves refined weighting of the temporal features of LSTM, which is more in line with the knowledge mastery of students' actual answering behavior. It also enables the Attention model to accurately locate key knowledge points in the learning path by combining the relationship between knowledge points in the knowledge graph. By combining the model loss function of the knowledge graph, the model training is made to better fit the knowledge progression law in actual teaching, which significantly improves the accuracy of predicting the mastery of knowledge points.

[0035] Furthermore, the process of constructing the temporal sequence of answer features includes:

[0036] The answer time, answer duration, and question difficulty contained in the answer records stored in the student answer knowledge graph are numerically normalized and used as elements of the answer time sequence feature;

[0037] The knowledge point dimensions and knowledge point IDs contained in the knowledge point set stored in the student answer knowledge graph are used as elements of the answer sequence feature through one-hot encoding.

[0038] Furthermore, the process of generating personalized recommended learning paths based on the predicted knowledge point mastery level in the student's answer knowledge graph includes:

[0039] Calculate the candidate learning path scores based on the student answer knowledge graph adapted to the predicted knowledge point mastery, cognitive load, and student preferences;

[0040] Personalized recommended learning paths are selected from the candidate learning paths based on their scores.

[0041] This invention also provides a system for applying the aforementioned personalized learning path planning method based on LSTM fusion attention, comprising:

[0042] The LSTM answer performance feature generation module is connected to the student's answer knowledge graph. It is used to construct the answer feature time series based on the answer records, wrong answer table and knowledge point set in the student's answer knowledge graph. The answer feature time series is then used to generate answer performance features through the time series correlation trend extraction model. The time series correlation trend extraction model is built based on the LSTM architecture and uses the answer performance features as hidden states for iterative calculation.

[0043] The Attention adjustment module is connected to the LSTM answer performance feature generation module. In the process of generating the answer performance features, the answer performance features are used to calculate dynamic attention weights through an Attention model that sets time decay terms, continuous answer difficulty feedback terms, and knowledge point association terms. The iteration of the hidden state of the time-series association trend extraction model is adjusted based on the dynamic attention weights.

[0044] The mapping output module is connected to the LSTM answer performance feature generation module to generate a prediction of knowledge point mastery by using the answer performance features of multiple time steps through a fusion mapping model, wherein the fusion mapping model is built based on fully connected layers and activation functions.

[0045] The learning path planning module is connected to the student answer knowledge graph and mapping output module to generate personalized recommended learning paths based on the predicted knowledge point mastery in the student answer knowledge graph.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. The knowledge graph data of the online learning platform is used to predict the mastery of knowledge points through an LSTM-integrated attention model, and personalized learning path planning is carried out to meet the individual needs of students, thus realizing the intelligentization of the university teaching system.

[0048] 2. It achieves refined weighting of the temporal features of LSTM, which is more in line with the knowledge mastery of students' actual answering behavior. It also realizes that the Attention model can accurately locate key knowledge points in the learning path by combining the relationship between knowledge points in the knowledge graph. By combining the model loss function of the knowledge graph, the model training is more in line with the knowledge progression law in actual teaching, which significantly improves the accuracy of predicting the mastery of knowledge points.

[0049] 3. Standardized construction of answer feature time sequence was achieved, improving the quality of model input. Through collaborative modeling of LSTM and dynamic Attention, the temporal dynamics of the learning state were accurately captured, enabling more accurate generation of personalized recommendation learning paths. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the personalized learning path planning method based on LSTM fusion Attention according to an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the usage structure of the personalized learning path planning system based on LSTM fusion Attention according to an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram illustrating the process of adjusting the hidden state of the Attention model in the personalized learning path planning method based on LSTM fusion Attention according to an embodiment of the present invention.

[0053] Figure 4 This is a schematic diagram of the student answer knowledge graph structure for a personalized learning path planning method based on LSTM fusion Attention, as described in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0055] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0056] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0057] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0058] like Figures 1 to 4 As shown, this invention provides a personalized learning path planning method and system based on LSTM fusion Attention. It uses an LSTM fusion Attention model to predict the mastery of knowledge points from the knowledge graph data of an online learning platform and plans personalized learning paths to meet the individual needs of students, thereby realizing the intelligentization of the university teaching system.

[0059] like Figures 1 to 4 As shown, this embodiment proposes a personalized learning path planning method based on LSTM fusion with Attention. The online learning platform integrates a student answer knowledge graph, which includes a table of incorrect answers corresponding to answer records and a graph relationship between the answer records and the sets of knowledge points. The personalized learning path planning method includes:

[0060] Based on the information of answer records, wrong answer list and knowledge point set in the student answer knowledge graph, answer feature time series is constructed, and the answer feature time series is used to generate answer performance features through time series correlation trend extraction model, wherein the time series correlation trend extraction model is built based on LSTM architecture, and the answer performance features are used as hidden states for iterative calculation;

[0061] In the process of generating the answer performance features, the answer performance features are used to calculate dynamic attention weights through an Attention model that sets time decay terms, continuous answer difficulty feedback terms, and knowledge point association terms, and the iteration of the hidden state of the time-series association trend extraction model is adjusted based on the dynamic attention weights.

[0062] The answer performance features at multiple time steps are used to generate a predicted knowledge point mastery level through a fusion mapping model, wherein the fusion mapping model is constructed based on fully connected layers and activation functions;

[0063] Based on the predicted knowledge point mastery, a personalized recommended learning path is generated in the student's answer knowledge graph.

[0064] Understandably, models that combine LSTM (Long Short-Term Memory) with attention mechanisms compensate for the shortcomings of statically mining knowledge graph information, making the prediction of knowledge mastery more accurate and in line with students' actual situation.

[0065] Specifically, the time-series correlation trend extraction model outputs a sequence of hidden states across multiple time steps, reflecting information about the student's learning up to the recent T time steps, such as recent test-taking performance and the frequency of repetition of knowledge points. An attention mechanism assigns weights to the hidden states at each time step, focusing on the learning events most relevant to the current prediction.

[0066] Specifically, the hidden layer output vector of the time-series correlation trend extraction model has a dimension of 4, the number of input nodes of the fusion mapping model is 4, the number of hidden nodes can be set to 3, the number of output nodes is 1, and the learning rate is 0.05 to avoid gradient explosion or training instability.

[0067] like Figure 3 As shown, the student answer knowledge graph further includes the graph relationship between the answer records and the corresponding questions. The process of calculating the dynamic attention weights includes:

[0068] The question-answering features time sequence is used to calculate the original attention score through the multi-head attention mechanism of the Attention model;

[0069] The time decay term is calculated based on the proximity of the current time step to the set recent time step, as well as the trainable parameters of the learning rhythm.

[0070] Based on the question-answering records stored in the student's question-answering knowledge graph, the difficulty coefficient, trainable adjustable parameters, and activation function of the questions corresponding to consecutive correct answers and consecutive incorrect answers are used to calculate the consecutive answering difficulty feedback item.

[0071] Calculate the knowledge point association items based on the degree of association between the knowledge point at the current time step and the target knowledge point in the set of knowledge points;

[0072] The dynamic attention weight is determined by integrating the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term.

[0073] Understandably, the time decay term quantifies the diminishing impact of early answering behaviors on the current state; for example, a wrong answer from one week ago has a weaker impact on current mastery than a wrong answer from yesterday. The continuous answering difficulty feedback term reflects the incentive or inhibition of learning state by changes in difficulty. The knowledge point association term combines the graph relationships of the knowledge point set to uncover weaknesses across knowledge points. Therefore, the Attention model described above can make the prediction of knowledge point mastery more accurate for LSTM-based models.

[0074] like Figure 3 As shown, the process of calculating the original attention score further includes:

[0075] The original attention score is calculated by sequentially passing the concatenation variables of the target knowledge point and the hidden state at the current time step, the trainable weighted parameters, and the trainable bias parameters through the tanh activation function and the softmax activation function.

[0076] Specifically, the calculation process for the original attention score is as follows:

[0077] score1 = tanh(W1[h] t ;e k ]+b1)

[0078] In the formula, score1 represents the original attention score, tanh represents the tanh activation function, W1 and b1 represent the trainable weighting parameters and trainable bias parameters, respectively, [h t ;e k [] represents the hidden state h at the current time step t. t Embedding with target knowledge points e k Variables that are concatenated.

[0079] like Figure 3 As shown, further, the process of calculating the time decay term includes: calculating the difference between the current time step and the set recent time step, calculating the difference multiplied by the negative of the learning rhythm trainable parameter, and then performing an exponential function operation to determine the time decay term;

[0080] The process of calculating the continuous answer difficulty feedback term includes: using the concatenated vector of the current answer result, the current continuous error label vector, and the difficulty coefficient of the answer record, and the trainable adjustable parameter to calculate the continuous answer difficulty feedback term through the sigmoid activation function.

[0081] Specifically, the calculation process for the time decay term is as follows:

[0082] decay t =exp(-γ·(Tt))

[0083] In the formula, decay t γ represents the time decay term, γ represents the trainable parameter of the learning rhythm, and T and t represent the recent time step and the current time step, respectively.

[0084] Specifically, the calculation process for the continuous question difficulty feedback item is as follows:

[0085] β t =σ(W b [r t ;d t c t ]+b b )

[0086] In the formula, β t This represents the feedback term for the difficulty of consecutive question answering, σ represents the sigmoid activation function, and W... b b b Represents two trainable and adjustable parameters, [r t ;d t c t [] represents the current answer result vector r at the current time step t. t The current continuous error marker vector d t and the aforementioned difficulty coefficient c t The concatenated vector, where r is the correct answer. t The difficulty coefficient is 1 for a given value and 0 for an error value. The normalized value is used to determine the difficulty level. When the number of errors in a question exceeds a set limit, d... t The set number of times is 1 if the question is difficult and 0 otherwise. The set number of times is determined according to the difficulty level. For example, a more difficult question requires 3 times and a simpler question requires 5 times.

[0087] like Figure 3 As shown, further, the process of determining the dynamic attention weight by fusing the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term includes:

[0088] The product of the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term is used as the temporary attention score;

[0089] The temporary attention score is normalized using the softmax activation function to determine the dynamic attention weight.

[0090] Specifically, the knowledge point association term is the knowledge point embedding vector e at the current time step t. t Embedding of target knowledge points k The cosine similarity.

[0091] like Figure 3As shown, further, the iterative process of adjusting the hidden state of the temporal correlation trend extraction model based on the dynamic attention weights includes:

[0092] Calculate the weighted sum vector of the hidden state of the set recent time step and the corresponding dynamic attention weights. Multiply the concatenated value of the hidden state of the current time step and the weighted sum vector, along with the gated trainable parameters, by the hidden state of the current time step after passing through the sigmoid activation function to determine the hidden state of the next time step of the iteration.

[0093] Specifically, the calculation process for the weighted summation vector is as follows:

[0094]

[0095] In the formula, c represents the weighted summation vector, T and t represent the recent time step and the current time step, respectively, and α t h represents the dynamic attention weights. t This indicates the hidden state of the current time step, which belongs to the recent time step.

[0096] Specifically, the calculation process for the hidden state in the next time step is as follows:

[0097] h t+1 =σ(W g [h t ;c]+b g )·h t

[0098] In the formula, h t+1 h t Let T represent the hidden state at the next time step and the hidden state at the current time step, respectively. Let T represent the recent time step, σ represent the sigmoid activation function, and W represent the hidden state at the current time step. g b g Represents two gating trainable parameters, [h t ;c] represents the concatenated value of the hidden state and the weighted summation vector c at the current time step, and · represents the element-wise product.

[0099] Furthermore, the training optimization process for the trainable parameters of the learning rhythm, the trainable adjustment parameters, the trainable weighted parameters, the trainable bias parameters, and the gating trainable parameters includes:

[0100] The product of the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term is used as the event importance weight. The mean square error between the true knowledge point mastery and the training knowledge point mastery of the sample dataset is calculated. The event importance weight and the mean square error are multiplied together to construct a weighted mean square error term.

[0101] Obtain the dependency relationship between the mastery of training knowledge points and historical knowledge points in the student's answer knowledge graph to construct a knowledge point relationship loss term;

[0102] The weighted mean squared error term and the knowledge point relationship loss term are weighted and summed to generate a comprehensive loss function. The Attention model and the fusion mapping model, which include the trainable parameters of the learning rhythm, trainable adjustment parameters, trainable weighted parameters, trainable bias parameters and gating trainable parameters, are then trained and optimized using the comprehensive loss function.

[0103] Specifically, the calculation process for the weighted mean square error term is as follows:

[0104]

[0105] In the formula, L WMSE y represents the weighted mean square error term. t ′、y t These represent the mastery of training knowledge points and the mastery of real knowledge points, respectively, and η represents the weight of the importance of the event.

[0106] Specifically, the calculation process for the knowledge point relationship loss term is as follows:

[0107]

[0108] In the formula, L prior The term y′ represents the knowledge point relationship loss term, where k1→k2∈ξ indicates that in the student's answer knowledge graph ξ, the current knowledge point k1 is a predecessor of the historical knowledge point k2. k1 y k1 These represent the training knowledge point mastery and the actual knowledge point mastery at the current time step, respectively, with ε representing a small error value, such as 0.1, to avoid absolute coercion. Therefore, the knowledge point relationship loss term utilizes the fact that the current knowledge point mastery is less than the previous knowledge point mastery to ensure it conforms to the learning process relationship between knowledge points.

[0109] The calculation process of the comprehensive loss function is as follows:

[0110] Loss=λL WMSE +(1-λ)L prior

[0111] In the formula, Loss represents the comprehensive loss function, and L WMSE L represents the weighted mean square error term. prior λ represents the knowledge point relationship loss term, and λ represents the weighted value.

[0112] The above scheme achieves refined weighting of the temporal features of LSTM, which is more in line with the knowledge mastery of students' actual answering behavior. It also enables the Attention model to accurately locate key knowledge points in the learning path by combining the relationship between knowledge points in the knowledge graph. By combining the model loss function of the knowledge graph, the model training is made to better fit the knowledge progression law in actual teaching, which significantly improves the accuracy of predicting the mastery of knowledge points.

[0113] like Figure 4 As shown, further, the information in the answer record in the student answer knowledge graph includes answer time, answer duration, and question difficulty; the information in the knowledge point set includes knowledge point dimension and knowledge point ID; and the process of constructing the answer feature time sequence includes:

[0114] The answer time, answer duration, and question difficulty contained in the answer records stored in the student answer knowledge graph are numerically normalized and used as elements of the answer time sequence feature;

[0115] The knowledge point dimensions and knowledge point IDs contained in the knowledge point set stored in the student answer knowledge graph are used as elements of the answer sequence feature through one-hot encoding.

[0116] Specifically, the information in the incorrect question table includes the answer results and the score rate, which are directly used as elements of the answer sequence feature.

[0117] Furthermore, the process of generating personalized recommended learning paths based on the predicted knowledge point mastery level in the student's answer knowledge graph includes:

[0118] Calculate the candidate learning path scores based on the student answer knowledge graph adapted to the predicted knowledge point mastery, cognitive load, and student preferences;

[0119] Personalized recommended learning paths are selected from the candidate learning paths based on their scores.

[0120] Specifically, the cognitive load, used to avoid focusing on highly difficult content, is the ratio of the current difficulty level to the total difficulty level of the content. The student preference adaptation refers to students' preferred learning time for video resources, practice resources, etc., and is the ratio of each type of resource to the total learning time.

[0121] like Figure 2 As shown, this embodiment also provides a system applying the personalized learning path planning method based on LSTM fusion Attention. The online learning platform integrates a student answer knowledge graph, which includes a table of incorrect answers corresponding to answer records and a graph relationship between the answer records and sets of knowledge points. The system applied to the online learning platform includes:

[0122] The LSTM answer performance feature generation module is connected to the student answer knowledge graph. It is used to construct the answer feature time series based on the answer records, wrong answer table and knowledge point set in the student answer knowledge graph. The answer feature time series is then used to generate answer performance features through the time series correlation trend extraction model. The time series correlation trend extraction model is built based on the LSTM architecture and uses the answer performance features as hidden states for iterative calculation.

[0123] The Attention adjustment module is connected to the LSTM answer performance feature generation module. In the process of generating the answer performance features, the answer performance features are used to calculate dynamic attention weights through an Attention model that sets time decay terms, continuous answer difficulty feedback terms, and knowledge point association terms. The iteration of the hidden state of the time-series association trend extraction model is adjusted based on the dynamic attention weights.

[0124] The mapping output module is connected to the LSTM answer performance feature generation module to generate a prediction of knowledge point mastery by using the answer performance features of multiple time steps through a fusion mapping model, wherein the fusion mapping model is built based on fully connected layers and activation functions.

[0125] The learning path planning module is connected to the student answer knowledge graph and mapping output module to generate personalized recommended learning paths based on the predicted knowledge point mastery in the student answer knowledge graph.

[0126] The above scheme achieves standardized construction of the temporal sequence of answer features, improves the quality of model input, and accurately captures the temporal dynamics of the learning state through collaborative modeling of LSTM and dynamic attention, thus enabling more accurate generation of personalized recommendation learning paths.

[0127] In this embodiment, the knowledge graph data of the online learning platform is used to predict the mastery of knowledge points through an LSTM-Attention model, and personalized learning path planning is performed to meet the individual needs of students, thus realizing the intelligentization of the university teaching system. It achieves refined weighting of the temporal features of LSTM, which better reflects the knowledge mastery contained in students' actual answering behavior. It also enables the Attention model to accurately locate key knowledge points in the learning path by combining the relationships between knowledge points in the knowledge graph. By combining the model loss function of the knowledge graph, the model training is made more in line with the knowledge progression pattern in actual teaching, significantly improving the accuracy of predicting the mastery of knowledge points. It achieves standardized construction of the temporal sequence of answer features, improving the quality of model input. Through collaborative modeling of LSTM and dynamic Attention, it accurately captures the temporal dynamics of the learning state, achieving more accurate generation of personalized recommended learning paths.

[0128] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A personalized learning path planning method based on LSTM fused with Attention, characterized in that, include: Based on the information of answer records, wrong answer list and knowledge point set in the student answer knowledge graph, answer feature time series is constructed, and the answer feature time series is used to generate answer performance features through time series correlation trend extraction model, wherein the time series correlation trend extraction model is built based on LSTM architecture, and the answer performance features are used as hidden states for iterative calculation; In the process of generating the answer performance features, the answer performance features are used to calculate dynamic attention weights through an Attention model that sets time decay terms, continuous answer difficulty feedback terms, and knowledge point association terms, and the iteration of the hidden state of the time-series association trend extraction model is adjusted based on the dynamic attention weights. The answer performance features at multiple time steps are used to generate a predicted knowledge point mastery level through a fusion mapping model, wherein the fusion mapping model is constructed based on fully connected layers and activation functions; Based on the predicted knowledge point mastery, a personalized recommended learning path is generated in the student's answer knowledge graph; The process of calculating the dynamic attention weights includes: The question-answering features time sequence is used to calculate the original attention score through the multi-head attention mechanism of the Attention model; The time decay term is calculated based on the proximity of the current time step to the set recent time step, as well as the trainable parameters of the learning rhythm. Based on the question-answering records stored in the student's question-answering knowledge graph, the difficulty coefficient, trainable adjustable parameters, and activation function of the questions corresponding to consecutive correct answers and consecutive incorrect answers are used to calculate the consecutive answering difficulty feedback item. The process of calculating the difficulty feedback item for consecutive questions is as follows: ， In the formula, This indicates the difficulty level of consecutive question-answering. This represents the sigmoid activation function. This represents two trainable and adjustable parameters. This represents the current answer result vector at the current time step t. Current continuous error marker vector and the aforementioned difficulty level The concatenated vector; Calculate the knowledge point association items based on the degree of association between the knowledge point at the current time step and the target knowledge point in the set of knowledge points; The dynamic attention weight is determined by integrating the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term. The iterative process of adjusting the hidden state of the time-series correlation trend extraction model based on the dynamic attention weights includes: ， ， In the formula, This represents a weighted summation vector, where T and t represent the recent time step and the current time step, respectively. Represents dynamic attention weights. This indicates the hidden state of the current time step, which belongs to the recent time step. These represent the hidden states at the next time step. This represents the sigmoid activation function. This represents two gating trainable parameters. This represents element-wise product.

2. The personalized learning path planning method based on LSTM fusion Attention as described in claim 1, characterized in that, The process of calculating the original attention score includes: The original attention score is calculated by sequentially passing the concatenation variables of the target knowledge point and the hidden state at the current time step, the trainable weighted parameters, and the trainable bias parameters through the tanh activation function and the softmax activation function.

3. The personalized learning path planning method based on LSTM fusion Attention as described in claim 1, characterized in that, The process of calculating the time decay term includes: calculating the difference between the current time step and the set recent time step, multiplying the difference by the negative of the learning rhythm trainable parameter, and then performing an exponential function operation to determine the time decay term.

4. The personalized learning path planning method based on LSTM fusion Attention as described in claim 1, characterized in that, The process of determining the dynamic attention weight based on the fusion calculation of the original attention score, the time decay term, the continuous answering difficulty feedback term, and the knowledge point association term includes: The product of the original attention score, the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term is used as the temporary attention score; The temporary attention score is normalized using the softmax activation function to determine the dynamic attention weight.

5. The personalized learning path planning method based on LSTM fusion Attention as described in claim 1, characterized in that, The training optimization process for the trainable parameters of the learning rhythm, trainable adjustment parameters, trainable weighted parameters, trainable bias parameters, and gating trainable parameters includes: The product of the time decay term, the continuous answer difficulty feedback term, and the knowledge point association term is used as the event importance weight. The mean square error between the true knowledge point mastery and the training knowledge point mastery of the sample dataset is calculated. The event importance weight and the mean square error are multiplied together to construct a weighted mean square error term. Obtain the dependency relationship between the mastery of training knowledge points and historical knowledge points in the student's answer knowledge graph to construct a knowledge point relationship loss term; The weighted mean squared error term and the knowledge point relationship loss term are weighted and summed to generate a comprehensive loss function. The Attention model and the fusion mapping model, which include the trainable parameters of the learning rhythm, trainable adjustment parameters, trainable weighted parameters, trainable bias parameters and gating trainable parameters, are then trained and optimized using the comprehensive loss function.

6. The personalized learning path planning method based on LSTM fusion Attention according to any one of claims 1 to 5, characterized in that, The process of constructing the time series of answer features includes: The answer time, answer duration, and question difficulty contained in the answer records stored in the student answer knowledge graph are numerically normalized and used as elements of the answer feature time sequence; The knowledge point dimensions and knowledge point IDs contained in the knowledge point set stored in the student answer knowledge graph are used as elements of the answer feature time sequence through one-hot encoding.

7. The personalized learning path planning method based on LSTM fusion Attention according to any one of claims 1 to 5, characterized in that, The process of generating a personalized recommended learning path based on the predicted knowledge point mastery in the student's answer knowledge graph includes: Calculate the candidate learning path scores based on the student answer knowledge graph adapted to the predicted knowledge point mastery, cognitive load, and student preferences; Personalized recommended learning paths are selected from the candidate learning paths based on their scores.

8. A system applying the personalized learning path planning method based on LSTM fusion Attention as described in any one of claims 1 to 7, characterized in that, include: The LSTM answer performance feature generation module is connected to the student's answer knowledge graph. It is used to construct the answer feature time series based on the answer records, wrong answer table and knowledge point set in the student's answer knowledge graph. The answer feature time series is then used to generate answer performance features through the time series correlation trend extraction model. The time series correlation trend extraction model is built based on the LSTM architecture and uses the answer performance features as hidden states for iterative calculation. The Attention adjustment module is connected to the LSTM answer performance feature generation module. In the process of generating the answer performance features, the answer performance features are used to calculate dynamic attention weights through an Attention model that sets time decay terms, continuous answer difficulty feedback terms, and knowledge point association terms. The iteration of the hidden state of the time-series association trend extraction model is adjusted based on the dynamic attention weights. The mapping output module is connected to the LSTM answer performance feature generation module to generate a prediction of knowledge point mastery by using the answer performance features of multiple time steps through a fusion mapping model, wherein the fusion mapping model is built based on fully connected layers and activation functions. The learning path planning module is connected to the student answer knowledge graph and mapping output module to generate personalized recommended learning paths based on the predicted knowledge point mastery in the student answer knowledge graph.

Images