Online education training intelligent pushing system based on AI adaptive learning

Abstract

The application relates to the technical field of intelligent education pushing, and particularly discloses an online education training intelligent pushing system based on AI adaptive learning, which comprises the following steps: collecting student historical behavior sequence data and learning achievement evaluation data, performing feature coding on the two types of data, and generating a student exclusive cognitive state vector; inputting the vector into a pre-trained adaptive learning model to output a knowledge point mastery prediction value; and searching for course resources according to the prediction value to construct a pushing candidate set. The system quantitatively calculates a teaching unit difficulty coefficient and a knowledge correlation strength, sets a pushing priority according to the calculation result, and completes the pushing. The design can accurately quantify the cognitive state of students, regularize original learning data, enrich sorting and judging dimensions, optimize course arrangement logic, and improve the matching degree of course pushing and student individualized learning features.

Description

Technical Field

[0001] This invention relates to the field of intelligent education delivery technology, specifically to an intelligent online education and training delivery system based on AI adaptive learning. Background Technology

[0002] Existing online education and training delivery systems generally employ traditional data processing methods, only collecting and storing student learning behavior sequence data and learning outcome assessment data. These systems lack data feature encoding capabilities, leaving raw learning data in a heterogeneous state and unable to integrate and process multiple types of learning data. The systems cannot convert student learning-related data into standardized data formats, making it impossible to quantitatively represent students' cognitive states. Implicit learning characteristics generated during the learning process cannot be extracted, making it difficult for the platform to accurately distinguish the cognitive differences among different students. The determination of students' cognitive states relies solely on surface-level data, resulting in limitations in data processing methods.

[0003] Existing technologies, after completing course resource retrieval and matching, mostly use fixed sorting rules to generate push sequences. The push process does not quantify the difficulty level of teaching units, nor does it assess the strength of knowledge connections between different teaching units. The push sorting relies solely on a single logic, lacking multi-dimensional quantitative judgment criteria. This fixed arrangement pattern cannot adapt to the learning adaptability of different students, the knowledge connections between teaching units lack reasonable logic, the resource push arrangement is highly uniform, and it cannot adapt to the personalized learning pace of students, resulting in low accuracy in matching push content. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent online education and training push system based on AI adaptive learning to solve the aforementioned background problems.

[0005] The objective of this invention can be achieved through the following technical solutions: An AI-based adaptive learning-based intelligent online education and training recommendation system includes: The data acquisition module obtains historical behavioral sequence data and learning outcome assessment data of students on the learning platform; The cognitive state modeling module performs feature encoding on the historical behavior sequence data and the learning outcome assessment data to generate a vector of the student's current cognitive state. The mastery prediction module inputs the student's current cognitive state vector into a pre-trained adaptive learning model and outputs mastery prediction values ​​for multiple knowledge points. The course content retrieval module retrieves matching teaching units from the course content library based on the mastery prediction value and generates a candidate set to be pushed. The decision optimization module calculates the difficulty coefficient and knowledge association strength of each teaching unit in the candidate set to be pushed, and forms a push priority ranking. The learning unit push module pushes the corresponding teaching units to the students' terminal devices according to the push priority.

[0006] As a further aspect of the present invention, the acquisition of students' historical behavioral sequence data and learning outcome evaluation data on the learning platform specifically includes: Extract the start and end times of video viewing, the time of turning pages in courseware, the time spent answering questions, and the correctness of the answers from the behavioral log database of the learning platform. The video viewing start and end times, the courseware page turning times, the time spent answering questions, and the correctness of the questions are arranged in chronological order to form a sequence of behavioral events, which is the historical behavioral sequence data. Extract the student's score for each knowledge point and the time spent answering each knowledge point from the learning platform's assessment results database. The score and time spent answering each knowledge point are arranged in order of knowledge point number to form an evaluation feature vector, which serves as the evaluation data for the learning outcome.

[0007] As a further aspect of the present invention, feature encoding is performed on the historical behavior sequence data and the learning outcome assessment data to generate a vector representing the student's current cognitive state, specifically including: For each behavioral event in the historical behavioral sequence data, calculate the corresponding action type encoding value and action intensity value; The action type encoding value and the action intensity value are combined into a behavior embedding vector; The behavior embedding vectors corresponding to all behavior events are weighted and summed according to the time decay weight to generate historical behavior aggregate features. The scores and response times for each knowledge point in the learning outcome assessment data are normalized to generate normalized scores and normalized response times. The product of the normalized score and the normalization time for each knowledge point is used as the confidence level of mastery for that knowledge point. The confidence scores of mastery of all knowledge points are concatenated into a mastery confidence vector in the order of knowledge point numbers, which serves as the current cognitive state vector of the trainee.

[0008] As a further aspect of the present invention, the student's current cognitive state vector is input into a pre-trained adaptive learning model to output predicted mastery values ​​for multiple knowledge points, specifically including: The current cognitive state vector of the learner is input into the input layer of the adaptive learning model; The hidden layer of the adaptive learning model performs a nonlinear transformation operation on the current cognitive state vector of the learner to generate a latent feature vector. The output layer of the adaptive learning model performs a fully connected mapping operation on the latent feature vector and outputs an original output vector of the same length as the total number of knowledge points. Perform a value range compression operation on each component in the original output vector so that the value range of each component falls between zero and one. Each component in the vector obtained after compression is used as the predicted mastery value for the corresponding knowledge point.

[0009] As a further aspect of the present invention, matching teaching units are retrieved from the course content library based on the mastery prediction value to generate a candidate set for push notification, specifically including: Set a low mastery threshold and a high mastery threshold; Traverse all knowledge points and mark those with a mastery prediction value lower than the low mastery threshold as knowledge points to be strengthened. Knowledge points whose mastery prediction value is higher than the high mastery threshold are marked as mastered knowledge points; Retrieve all teaching units that are directly related to the knowledge points to be strengthened from the course content library, and add these teaching units to the initial candidate set; Retrieve all review teaching units that have prerequisite dependencies on the mastered knowledge points from the course content library, and add these teaching units to the initial candidate set; Duplicate teaching units are removed from the initial candidate set to obtain the candidate set to be pushed.

[0010] As a further aspect of the present invention, the difficulty coefficient and knowledge association strength of each teaching unit in the candidate set to be pushed are calculated to form a push priority ranking, specifically including: Retrieves the pre-marked difficulty level values ​​for each teaching unit from the course content library; The difficulty level value is directly used as the difficulty coefficient of the teaching unit; For each teaching unit in the candidate set to be pushed, calculate the size of the intersection between the set of knowledge points covered by the teaching unit and the set of knowledge points to be strengthened; Divide the size of the intersection by the total size of the set of knowledge points covered by the teaching unit to obtain the knowledge association strength of the teaching unit; The original priority score for each teaching unit is obtained by multiplying the difficulty coefficient of each teaching unit by the strength of knowledge relevance. All teaching units in the candidate set to be pushed are sorted in descending order of their original priority scores to form the push priority sort.

[0011] As a further aspect of the present invention, the corresponding teaching units are pushed to the students' terminal devices according to the push priority, specifically including: The top N teaching units ranked in the push priority sort are selected as a push batch; For each teaching unit in the push batch, generate a Uniform Resource Locator (URL) link and an estimated learning duration for that teaching unit; All teaching units' Uniform Resource Locator (URL) links and estimated learning duration values ​​are encapsulated into a single push notification. Send the push message body to the student's terminal device; On the terminal device, the Uniform Resource Locator (URL) link for each teaching unit and the corresponding estimated learning time are displayed sequentially according to the push priority.

[0012] As a further aspect of the present invention, the step of performing a weighted summation operation on the behavior embedding vectors corresponding to all behavior events according to the time decay weight to generate historical behavior aggregate features specifically includes: Obtain the timestamp corresponding to each behavioral event in the historical behavioral sequence data. The timestamp records the absolute time point when the behavioral event occurred. Calculate the difference between the current system time and the timestamp of each behavioral event to obtain the duration of each behavioral event. Based on the duration of each behavioral event, the time decay weight of each behavioral event is calculated by substituting it into a negative exponential function, where the decay coefficient of the negative exponential function is a pre-set positive constant. Multiply the behavior embedding vector corresponding to each behavior event by its time decay weight to obtain the weighted behavior embedding vector; The weighted behavioral embedding vectors of all behavioral events are subjected to vector addition to generate historical behavioral aggregate features.

[0013] As a further aspect of the present invention, the hidden layer of the adaptive learning model performs a nonlinear transformation operation on the current cognitive state vector of the learner to generate a latent feature vector, specifically including: The hidden layer of the adaptive learning model receives the learner's current cognitive state vector output from the input layer; The hidden layer has multiple neuron nodes, and each neuron node maintains a weight vector and a bias value; The student's current cognitive state vector is multiplied by the weight vector of each neuron node, and then the bias value of that neuron node is added to generate the initial activation value of each neuron node. The ReLU activation function is applied to the initial activation value of each neuron node. The ReLU activation function sets input values ​​less than zero to zero and keeps input values ​​greater than or equal to zero unchanged. The output value of each neuron node after processing by the ReLU activation function is used as the final output of that neuron node; The final outputs of all neurons are concatenated in node order to form a new vector, which is the latent feature vector.

[0014] As a further aspect of the present invention, the training steps of the adaptive learning model include: Collect historical behavior sequence data and learning outcome assessment data from multiple historical students, and obtain subsequent real knowledge point mastery assessment data from these multiple historical students; For each historical learner, a historical cognitive state vector is generated based on their historical behavior sequence data and learning outcome assessment data, using a feature encoding method. The historical cognitive state vector of each historical student is used as a training sample, and the subsequent real knowledge mastery evaluation data of that historical student is used as a training label, together forming a training sample pair. Initialize the model parameters of the adaptive learning model, including the connection weight matrix from the input layer to the hidden layer, the bias vector of the hidden layer neurons, the connection weight matrix from the hidden layer to the output layer, and the bias vector of the output layer neurons. The gradient descent optimization algorithm is adopted, with the historical cognitive state vector in the training sample pair as input and the mean square error between the mastery prediction vector output by the model and the training label as the loss function. Calculate the gradient of the loss function with respect to all parameters of the model, and update the model parameters based on the calculated gradient; Repeat the gradient calculation and parameter update process until the loss function value is less than the preset convergence threshold. At this point, stop training and obtain the trained adaptive learning model.

[0015] The beneficial effects of this invention are: Feature encoding is performed on historical behavioral sequence data and learning outcome assessment data of learners, enabling the integration of multi-dimensional heterogeneous learning data. Behavioral and assessment information within the raw data undergoes structured transformation to form a cognitive state vector with a unified format. Implicit learning features within the data are extracted, and scattered data information is fused. Subjective cognitive states are converted into quantifiable data forms, individualized learning characteristics are retained, and cognitive differences between learners can be distinguished through data dimensions. Redundant raw data is standardized, and the data structure is adapted to the computational requirements of the intelligent model, giving the raw data higher utilization value.

[0016] The difficulty coefficient of teaching units within the candidate set is quantified to classify them into levels of difficulty. The strength of knowledge connections between different teaching units is quantified to clarify the relationships between knowledge points. The push priority is set based on the results of these two quantifications, resulting in a sequential arrangement of teaching units. The single, fixed sorting pattern is optimized, and the reference dimensions for arrangement are expanded. The knowledge connections between teaching units remain smooth, and the knowledge arrangement maintains a reasonable hierarchical order. The arrangement logic aligns with the objective laws of human learning, the course arrangement adapts to the learning characteristics of different individuals, and the course pushes are tailored to individual learning needs. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a sequence diagram of the intelligent push system for online education and training based on AI adaptive learning as described in this invention; Figure 2 This is a flowchart for acquiring historical behavior sequence data and learning outcome evaluation data; Figure 3 This is a flowchart of feature encoding to generate the student's current cognitive state vector. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This invention is an intelligent online education and training push system based on AI adaptive learning, comprising: The specific implementation of an AI-based adaptive learning-based intelligent online education and training push system includes: a data acquisition module, a cognitive state modeling module, a mastery prediction module, a course content retrieval module, a decision optimization module, and a learning unit push module. During system operation, the data acquisition module obtains historical behavioral sequence data and learning outcome assessment data from the learning platform. The cognitive state modeling module performs feature encoding on the historical behavioral sequence data and the learning outcome assessment data to generate a current cognitive state vector for the student. The mastery prediction module inputs the student's current cognitive state vector into a pre-trained adaptive learning model and outputs predicted mastery values ​​for multiple knowledge points. The course content retrieval module retrieves matching teaching units from the course content library based on the predicted mastery values, generating a candidate set for push. The decision optimization module calculates the difficulty coefficient and knowledge association strength of each teaching unit in the candidate set for push, forming a push priority ranking. The learning unit push module pushes the corresponding teaching units to the student's terminal device according to the push priority ranking.

[0021] In one embodiment of the present invention, see [reference] Figure 2 The data acquisition module extracts the start and end times of video viewing, the time of page turning in courseware, the time spent answering questions, and the correctness of answers from the learning platform's behavior log database for each login session. The data acquisition module arranges these data in chronological order to form a behavioral event sequence, which serves as the historical behavioral sequence data. The data acquisition module also extracts the student's score for each knowledge point and the time spent answering each knowledge point from the learning platform's assessment result database. The data acquisition module arranges these scores and time spent answering each knowledge point in order of knowledge point number to form an assessment feature vector, which serves as the learning outcome assessment data.

[0022] In specific implementation, the data acquisition module extracts from the learning platform's behavior log database the start and end times of video viewing, the time of turning pages in courseware, the time spent answering questions, and the correctness indicators of the questions answered during each login session. The data acquisition module arranges these data in chronological order into a behavioral event sequence, which is then used as the historical behavioral sequence data. In some embodiments, the data acquisition module also extracts from the learning platform's assessment result database the student's score for each knowledge point and the time spent answering each knowledge point in each standardized test. The data acquisition module then sets the score for each knowledge point and the time spent answering each knowledge point... The answering time values ​​are arranged in order of knowledge point number to form an assessment feature vector as the learning outcome assessment data. It can be understood that the video viewing start and end time points extracted by the data acquisition module include the start and end time of a video playback operation. The courseware page turning operation time records the specific time point when the student switches from the current courseware page to the next courseware page. The question answering time value records the length of time from when the student sees the question to when they submit the answer. The correctness of the question answer is indicated by a value of 1 for a correct answer and a value of 0 for an incorrect answer. Optionally, when extracting the historical behavior sequence data, the data acquisition module also records the login session identifier and student device type identifier corresponding to each behavior event to distinguish the behavior patterns under different learning scenarios.

[0023] In practice, for each standardized test, the data acquisition module reads the score and time taken for each knowledge point in ascending order of knowledge point number. The data acquisition module then alternates the read score and time taken to form the assessment feature vector, which is represented as follows: ; Where: M is the total number of knowledge points, and for each knowledge point numbered j, It equals the score value of the j-th knowledge point (dimensionless). The dimensionless value is obtained by dividing the answering time of the j-th knowledge point by 1 second, thus ensuring that all components in the evaluation feature vector are dimensionless values ​​and that the dimensions on both sides of the vector are consistent. In some embodiments, the data acquisition module performs a full score verification on the score of each knowledge point before forming the evaluation feature vector. If the score of a knowledge point exceeds the preset full score value for that knowledge point, the data acquisition module truncates the score of that knowledge point to the full score value. It can be understood that the data acquisition module sorts all behavioral events generated in each login session in ascending order according to the timestamp of the event. Only when the timestamps of two behavioral events are exactly the same will they be sorted a second time according to the order of the event type encoding value. Optionally, for multiple video viewing start and end time points generated during continuous video playback, the data acquisition module only retains the first start time and the last end time of each playback session, and removes redundant time point data generated by pause and resume operations.

[0024] In one embodiment of the present invention, the cognitive state modeling module calculates the action type encoding value and action intensity value corresponding to each behavioral event in the historical behavioral sequence data. The cognitive state modeling module combines the action type encoding value and the action intensity value into a behavioral embedding vector. The cognitive state modeling module performs a weighted summation operation on the behavioral embedding vectors corresponding to all behavioral events according to time decay weights to generate historical behavioral aggregate features. The cognitive state modeling module normalizes the score value and response time value of each knowledge point in the learning outcome assessment data to generate a normalized score and normalized response time. The cognitive state modeling module uses the product of the normalized score and normalized response time of each knowledge point as the mastery confidence level of that knowledge point. The cognitive state modeling module concatenates the mastery confidence levels of all knowledge points in order of knowledge point number into a mastery confidence vector as the learner's current cognitive state vector.

[0025] When the cognitive state modeling module performs a weighted summation operation on the behavior embedding vectors corresponding to all behavior events according to the time decay weight, it obtains the timestamp corresponding to each behavior event in the historical behavior sequence data. The timestamp records the absolute time point of the behavior event. The cognitive state modeling module calculates the difference between the current system time and the timestamp of each behavior event to obtain the duration of each behavior event. Based on the duration of each behavior event, the cognitive state modeling module substitutes the duration of each behavior event into a negative exponential function to calculate the time decay weight of each behavior event. The decay coefficient of the negative exponential function is a pre-set positive constant. The cognitive state modeling module multiplies the behavior embedding vector corresponding to each behavior event with its time decay weight to obtain a weighted behavior embedding vector. The cognitive state modeling module performs vector addition on all the weighted behavior embedding vectors of behavior events to generate historical behavior aggregate features.

[0026] In specific implementation, please refer to Figure 3 The cognitive state modeling module calculates the action type encoding value and action intensity value corresponding to each behavioral event in the historical behavior sequence data. The cognitive state modeling module combines the action type encoding value and the action intensity value into a behavior embedding vector. In some embodiments, the cognitive state modeling module performs a weighted summation operation on the behavior embedding vectors corresponding to all behavioral events according to the time decay weight to generate historical behavior aggregation features. The cognitive state modeling module normalizes the score value and answer time value of each knowledge point in the learning outcome assessment data to generate normalized score and normalized time. It can be understood that the cognitive state modeling module uses the product of the normalized score and normalized time of each knowledge point as the mastery confidence of the knowledge point. The cognitive state modeling module concatenates the mastery confidence of all knowledge points in the order of knowledge point number into a mastery confidence vector as the current cognitive state vector of the learner.

[0027] In practical implementation, when the cognitive state modeling module performs a weighted summation operation on the behavioral embedding vectors corresponding to all behavioral events according to the time decay weight, it obtains the timestamp corresponding to each behavioral event in the historical behavioral sequence data. The timestamp records the absolute time point of the behavioral event. The cognitive state modeling module calculates the difference between the current system time and the timestamp of each behavioral event to obtain the duration of each behavioral event. Based on the duration of each behavioral event, the cognitive state modeling module substitutes it into a negative exponential function to calculate the time decay weight of each behavioral event. The decay coefficient of the negative exponential function is a pre-set positive constant. The cognitive state modeling module multiplies the behavioral embedding vector corresponding to each behavioral event with the time decay weight to obtain a weighted behavioral embedding vector. The cognitive state modeling module performs vector addition on all the weighted behavioral embedding vectors of behavioral events to generate historical behavioral aggregate features. The time decay weight is calculated according to the following formula: ; in: Indicates the first Time decay weights for individual behavioral events Represents the natural constant. This represents the pre-set attenuation coefficient. Indicates the first The duration of the behavioral event The numerical unit is consistent with the reciprocal unit of the decay coefficient to ensure that the exponent is dimensionless. In some embodiments, the cognitive state modeling module sets the decay coefficient in the negative exponential function to 0.01, so that the time decay weight decreases monotonically as the duration of occurrence increases. Optionally, when calculating the duration of occurrence of each behavioral event, if the difference between the current system time and the timestamp is less than zero, the cognitive state modeling module sets the duration of occurrence to zero. It can be understood that after the cognitive state modeling module performs a weighted summation operation on all behavioral events, the resulting historical behavior aggregation feature is a vector with the same dimension as the dimension of the individual behavior embedding vector.

[0028] In specific implementation, when the cognitive state modeling module normalizes the score of each knowledge point in the learning outcome assessment data, it divides the score of each knowledge point by the corresponding full score to obtain the normalized score. Similarly, when normalizing the response time for each knowledge point, it divides the response time of each knowledge point by the maximum response time of all knowledge points in the standardized test to obtain the normalized time. In some embodiments, the cognitive state modeling module uses the product of the normalized score and the normalized time for each knowledge point as the mastery confidence level of that knowledge point. The knowledge point number and the confidence level range between zero and one. Optionally, before concatenating the confidence level vector, the cognitive state modeling module first checks whether there is a knowledge point whose confidence level exceeds the zero-to-one range. If it does, the excess value is truncated to the nearest interval endpoint. In other words, the cognitive state modeling module extracts the confidence level of each knowledge point in ascending order of knowledge point number and arranges these confidence levels into a one-dimensional vector. This one-dimensional vector is the student's current cognitive state vector, and the dimension of the student's current cognitive state vector is equal to the total number of knowledge points.

[0029] In one embodiment of the present invention, the mastery prediction module inputs the learner's current cognitive state vector into the input layer of the adaptive learning model. The hidden layer of the adaptive learning model performs a nonlinear transformation operation on the learner's current cognitive state vector to generate a latent feature vector. The output layer of the adaptive learning model performs a fully connected mapping operation on the latent feature vector, outputting an original output vector of the same length as the total number of knowledge points. The mastery prediction module performs a value range compression operation on each component of the original output vector, ensuring that the value of each component falls between zero and one. The mastery prediction module uses each component of the compressed vector as the predicted mastery value for the corresponding knowledge point.

[0030] When the hidden layer of the adaptive learning model performs a nonlinear transformation operation on the learner's current cognitive state vector, it receives the learner's current cognitive state vector output from the input layer. The hidden layer has multiple neurons, each maintaining a weight vector and a bias value. The learner's current cognitive state vector is multiplied by the weight vector of each neuron, and the bias value of that neuron is added to generate the initial activation value for each neuron. The ReLU activation function is applied to the initial activation value of each neuron, setting input values ​​less than zero to zero and keeping input values ​​greater than or equal to zero unchanged. The output value of each neuron after processing by the ReLU activation function is taken as the final output of that neuron. The final outputs of all neurons are concatenated in node order to form a new vector, which is the hidden feature vector.

[0031] The training process of the adaptive learning model is as follows: Collect historical behavior sequence data and learning outcome assessment data from multiple historical learners, and obtain subsequent real knowledge point mastery assessment data from these learners. For each historical learner, generate a historical cognitive state vector based on their historical behavior sequence data and learning outcome assessment data using a feature encoding method. Use each learner's historical cognitive state vector as a training sample and their subsequent real knowledge point mastery assessment data as a training label, forming a training sample pair. Initialize the model parameters of the adaptive learning model, including the connection weight matrix from the input layer to the hidden layer, the bias vectors of the hidden layer neurons, the connection weight matrix from the hidden layer to the output layer, and the bias vectors of the output layer neurons. Employ the gradient descent optimization algorithm, using the historical cognitive state vectors from the training sample pair as input and the mean squared error between the model's output mastery prediction vector and the training label as the loss function. Calculate the gradient of the loss function with respect to all model parameters and update the model parameters based on the calculated gradient. Repeat the gradient calculation and parameter update process until the loss function value is less than a preset convergence threshold, at which point training stops, and the trained adaptive learning model is obtained.

[0032] In its implementation, the mastery prediction module inputs the student's current cognitive state vector into the input layer of the adaptive learning model. The hidden layer of the adaptive learning model performs a nonlinear transformation operation on the student's current cognitive state vector to generate a latent feature vector. The output layer of the adaptive learning model performs a fully connected mapping operation on the latent feature vector, outputting an original output vector of the same length as the total number of knowledge points. The mastery prediction module performs a value range compression operation on each component of the original output vector, making the value range of each component fall between zero and one. The mastery prediction module uses each component in the vector obtained after the compression operation as the mastery prediction value of the corresponding knowledge point. The value range compression operation is implemented using a logistic function, the expression of which is as follows: ; in: Represents any component in the original output vector. Represents the natural constant of Power of 1 This represents the value obtained after compression. The range of values ​​is an open interval. In some embodiments, the mastery prediction module first detects the numerical range of each component in the original output vector before performing the value range compression operation. If a component is already between zero and one, it is not compressed. Optionally, the mastery prediction module multiplies each component in the vector obtained after the compression operation by one hundred and rounds it to two decimal places, outputting the mastery prediction value in percentage form. It can be understood that the dimension of the mastery prediction value vector output by the mastery prediction module is equal to the number of all knowledge points, and each component corresponds one-to-one with the knowledge point number.

[0033] In specific implementation, when the hidden layer of the adaptive learning model performs a nonlinear transformation operation on the student's current cognitive state vector, the hidden layer receives the student's current cognitive state vector output from the input layer. The hidden layer has multiple neuron nodes, each maintaining a weight vector and a bias value. The student's current cognitive state vector is multiplied by the weight vector of each neuron node, and the bias value of the neuron node is added to generate the initial activation value of each neuron node. The ReLU activation function is applied to the initial activation value of each neuron node. The ReLU activation function sets input values ​​less than zero to zero and keeps input values ​​greater than or equal to zero unchanged. The output value of each neuron node after processing by the ReLU activation function is taken as the final output of the neuron node. The final outputs of all neuron nodes are concatenated in node order to form a new vector, which is the hidden feature vector. In some embodiments, the number of neurons in the hidden layer is set to 256, and the dimension of the weight vector of each neuron node is equal to the dimension of the student's current cognitive state vector. Optionally, the mathematical expression of the ReLU activation function is... ,in The initial activation value is represented by the hidden feature vector. It can be understood that the dimension of the hidden feature vector is equal to the number of neurons in the hidden layer, and each component in the hidden feature vector is a non-negative real number.

[0034] In its implementation, the training process of the adaptive learning model includes the following steps: collecting historical behavior sequence data and learning outcome assessment data from multiple historical learners, and obtaining subsequent real knowledge point mastery assessment data from these learners; for each learner, generating a historical cognitive state vector based on the historical behavior sequence data and learning outcome assessment data using a feature encoding method; using each learner's historical cognitive state vector as a training sample, and the learner's subsequent real knowledge point mastery assessment data as training labels, together forming a training sample pair; and initializing the model parameters of the adaptive learning model, including the connection weights from the input layer to the hidden layer. The weight matrix, bias vectors of hidden layer neurons, connection weight matrix from hidden layer to output layer, and bias vectors of output layer neurons are used. The gradient descent optimization algorithm is adopted, with the historical cognitive state vector in the training sample pair as input and the mean squared error between the mastery prediction vector output by the model and the training label as the loss function. The gradient of the loss function with respect to all model parameters is calculated, and the model parameters are updated according to the calculated gradient. The gradient calculation and parameter update process is repeated until the loss function value is less than the preset convergence threshold. At this time, the training stops and the trained adaptive learning model is obtained. See Table 1 to show the data format of some training sample pairs during the training process.

[0035] Table 1: Format of Training Sample Data for Adaptive Learning Models

[0036] In some embodiments, the preset convergence threshold is set to 0.001. Training stops when the decrease in the loss function value is less than the preset convergence threshold in ten consecutive iterations. Optionally, the gradient descent optimization algorithm uses the Adam optimizer, and the learning rate is initially set to 0.001. It can be understood that the connection weight matrix from the input layer to the hidden layer and the connection weight matrix from the hidden layer to the output layer in the trained adaptive learning model are both fixed and used for subsequent mastery prediction of new learner cognitive state vectors.

[0037] In one embodiment of the present invention, the course content retrieval module sets a low mastery threshold and a high mastery threshold. The course content retrieval module iterates through all knowledge points, marking knowledge points with predicted mastery values ​​below the low mastery threshold as knowledge points to be strengthened. The course content retrieval module marks knowledge points with predicted mastery values ​​above the high mastery threshold as mastered knowledge points. The course content retrieval module retrieves all teaching units directly related to the knowledge points to be strengthened from the course content library and adds these teaching units to an initial candidate set. The course content retrieval module retrieves all review teaching units with prerequisite dependencies on the mastered knowledge points from the course content library and adds these teaching units to the initial candidate set. The course content retrieval module removes duplicate teaching units from the initial candidate set to obtain the candidate set to be pushed.

[0038] In practical implementation, the course content retrieval module sets a low mastery threshold and a high mastery threshold. The module iterates through all knowledge points, marking those with predicted mastery values ​​below the low threshold as knowledge points to be strengthened, and those with predicted mastery values ​​above the high threshold as already mastered knowledge points. The module then retrieves all teaching units directly related to the knowledge points to be strengthened from the course content library and adds them to an initial candidate set. Next, it retrieves all review teaching units with prerequisite dependencies on the already mastered knowledge points and adds them to the initial candidate set. Finally, the module removes duplicate teaching units from the initial candidate set to obtain the candidate set to be pushed. For each knowledge point, the condition for determining whether it belongs to the knowledge points to be strengthened can be expressed by the following formula: ; in: Indicates the first Predicted value of mastery of each knowledge point This indicates the low mastery threshold. and All are dimensionless real numbers, and the inequality holds when the first number is... A knowledge point is marked as a knowledge point to be strengthened. In some embodiments, the course content retrieval module also marks knowledge points with a mastery prediction value equal to the low mastery threshold as knowledge points to be strengthened. Optionally, the course content retrieval module also sets an intermediate threshold. Knowledge points with a mastery prediction value between the low mastery threshold and the high mastery threshold are not marked. It can be understood that the high mastery threshold is greater than the low mastery threshold, and both are pre-configured constants.

[0039] In practice, when the course content retrieval module searches the course content library for teaching units that are directly related to the knowledge points to be reinforced, it reads the pre-stored knowledge point coverage list for each teaching unit in the course content library. If the knowledge point coverage list of a teaching unit contains at least one knowledge point to be reinforced, the teaching unit is determined to have a direct relationship. When the course content retrieval module searches the course content library for review teaching units that have a prerequisite dependency relationship with mastered knowledge points, it reads the pre-stored prerequisite knowledge point list for each review teaching unit in the course content library. If all the knowledge points in the prerequisite knowledge point list of a review teaching unit are mastered knowledge points, the review teaching unit is determined to meet the retrieval conditions. See Table 2 for some knowledge points and their corresponding tagging results.

[0040] Table 2: Knowledge Point Marking Results

[0041] In some embodiments, before performing the operation of removing duplicate teaching units, the course content retrieval module first generates a unique identifier for each teaching unit in the initial candidate set, and determines whether it is a duplicate teaching unit by comparing the unique identifiers. Optionally, the low mastery threshold is set to 0.40 and the high mastery threshold is set to 0.85. It can be understood that when there are two teaching units with the same unique identifier in the initial candidate set, the course content retrieval module only retains one of the teaching units, and removes the remaining duplicate teaching units from the initial candidate set to obtain the candidate set to be pushed.

[0042] In one embodiment of the present invention, the decision optimization module calculates the difficulty coefficient and knowledge association strength for each teaching unit in the candidate set to be pushed. The decision optimization module reads the pre-labeled difficulty level value of each teaching unit from the course content library and directly uses the difficulty level value as the difficulty coefficient of that teaching unit. For each teaching unit in the candidate set to be pushed, the decision optimization module calculates the intersection between the set of knowledge points covered by the teaching unit and the set of knowledge points to be strengthened, and divides the intersection by the total size of the set of knowledge points covered by the teaching unit to obtain the knowledge association strength of that teaching unit. The decision optimization module multiplies the difficulty coefficient and knowledge association strength of each teaching unit to obtain the original priority score of that teaching unit. The decision optimization module sorts all teaching units in the candidate set to be pushed in descending order of their original priority scores to form the push priority ranking.

[0043] The learning unit push module pushes corresponding teaching units to students' terminal devices according to the push priority. The learning unit push module selects the top N teaching units from the push priority ranking as a push batch. For each teaching unit in the push batch, the learning unit push module generates a Uniform Resource Locator (URL) link and an estimated learning duration for that teaching unit. The learning unit push module encapsulates all the URLs and estimated learning durations of all teaching units into a push message body. The learning unit push module sends the push message body to the student's terminal device. On the terminal device, the URL and corresponding estimated learning duration of each teaching unit are displayed sequentially according to the push priority ranking.

[0044] In specific implementation, the decision optimization module calculates the difficulty coefficient and knowledge association strength for each teaching unit in the candidate set to be pushed. The module reads the pre-labeled difficulty level value of each teaching unit from the course content library and uses this value directly as the difficulty coefficient. For each teaching unit in the candidate set, the module calculates the intersection between the set of knowledge points covered by the teaching unit and the set of knowledge points to be strengthened. The intersection is divided by the total size of the set of knowledge points covered by the teaching unit to obtain the knowledge association strength. The module multiplies the difficulty coefficient and knowledge association strength of each teaching unit to obtain its original priority score. The module then sorts all teaching units in the candidate set according to their original priority scores from highest to lowest to form the push priority ranking. The formula for calculating the original priority score is as follows: ; in: Indicates the first The original priority scores for each teaching unit. Indicates the first The difficulty level of each teaching unit. Indicates the first The set of knowledge points covered by each teaching unit This represents the set of knowledge points to be strengthened. Represents a set With sets The size of the intersection, Represents a set The total size, the result of the division operation is the knowledge association strength, and the value ranges from zero to one. For dimensionless real numbers, in some embodiments, the decision optimization module first determines whether the knowledge association strength is zero before performing the multiplication operation. If the knowledge association strength is zero, the original priority score is directly set to zero. Optionally, the decision optimization module assigns integers with difficulty coefficients from 1 to 5, corresponding to five levels from easy to difficult. It can be understood that the push priority sorting is a list, and each element in the list contains a teaching unit identifier and its corresponding original priority score. The list is arranged in descending order of the original priority score.

[0045] In practice, the learning unit push module pushes corresponding teaching units to students' terminal devices according to the push priority. The module selects the top N teaching units from the priority ranking as a push batch. For each teaching unit in the push batch, the module generates a Uniform Resource Locator (URL) link and an estimated learning duration. The module then encapsulates all the URLs and estimated learning durations of all teaching units into a push message body, which is sent to the student's terminal device. On the terminal device, the URLs of each teaching unit are displayed sequentially according to the push priority ranking. In some embodiments, the N value in the top N ranked resources locator links and corresponding estimated learning time values ​​are pre-set to three. The learning unit push module sends only one push batch in each push cycle. Optionally, the learning unit push module adds a timeliness parameter when generating the unified resource locator links, so that the unified resource locator links are only valid within a specified time period. It can be understood that the estimated learning time values ​​are read from the course content library. Each teaching unit pre-stores the average learning time value required to complete the teaching unit. After receiving the push message body, the terminal device parses out the list of unified resource locator links and the corresponding list of estimated learning time values, and renders and displays them sequentially according to the order of receipt.

[0046] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A smart online education and training recommendation system based on AI adaptive learning is disclosed, characterized in that: include: The data acquisition module obtains historical behavioral sequence data and learning outcome assessment data of students on the learning platform; The cognitive state modeling module performs feature encoding on the historical behavior sequence data and the learning outcome assessment data to generate a vector of the student's current cognitive state. The mastery prediction module inputs the student's current cognitive state vector into a pre-trained adaptive learning model and outputs mastery prediction values ​​for multiple knowledge points. The course content retrieval module retrieves matching teaching units from the course content library based on the mastery prediction value and generates a candidate set to be pushed. The decision optimization module calculates the difficulty coefficient and knowledge association strength of each teaching unit in the candidate set to be pushed, and forms a push priority ranking. The learning unit push module pushes the corresponding teaching units to the students' terminal devices according to the push priority.

2. The intelligent online education and training push system based on AI adaptive learning disclosed in claim 1 is characterized in that, The acquisition of students' historical behavior sequence data and learning outcome evaluation data on the learning platform specifically includes: Extract the start and end times of video viewing, the time of turning pages in courseware, the time spent answering questions, and the correctness of the answers from the behavioral log database of the learning platform. The video viewing start and end times, the courseware page turning times, the time spent answering questions, and the correctness of the questions are arranged in chronological order to form a sequence of behavioral events, which is the historical behavioral sequence data. Extract the student's score for each knowledge point and the time spent answering each knowledge point from the learning platform's assessment results database. The score and time spent answering each knowledge point are arranged in order of knowledge point number to form an evaluation feature vector, which serves as the evaluation data for the learning outcome.

3. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 1 is characterized in that, Feature encoding is performed on the historical behavior sequence data and the learning outcome assessment data to generate a vector of the learner's current cognitive state, specifically including: For each behavioral event in the historical behavioral sequence data, calculate the corresponding action type encoding value and action intensity value; The action type encoding value and the action intensity value are combined into a behavior embedding vector; The behavior embedding vectors corresponding to all behavior events are weighted and summed according to the time decay weight to generate historical behavior aggregate features. The scores and response times for each knowledge point in the learning outcome assessment data are normalized to generate normalized scores and normalized response times. The product of the normalized score and the normalization time for each knowledge point is used as the confidence level of mastery for that knowledge point. The confidence scores of mastery of all knowledge points are concatenated into a mastery confidence vector in the order of knowledge point numbers, which serves as the current cognitive state vector of the trainee.

4. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 1, characterized in that, The current cognitive state vector of the learner is input into a pre-trained adaptive learning model, which outputs predicted mastery values ​​for multiple knowledge points, specifically including: The current cognitive state vector of the learner is input into the input layer of the adaptive learning model; The hidden layer of the adaptive learning model performs a nonlinear transformation operation on the current cognitive state vector of the learner to generate a latent feature vector. The output layer of the adaptive learning model performs a fully connected mapping operation on the latent feature vector and outputs an original output vector of the same length as the total number of knowledge points. Perform a value range compression operation on each component in the original output vector so that the value range of each component falls between zero and one. Each component in the vector obtained after compression is used as the predicted mastery value for the corresponding knowledge point.

5. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 1, characterized in that, Based on the predicted mastery level, matching teaching units are retrieved from the course content library to generate a candidate set for push notifications, specifically including: Set a low mastery threshold and a high mastery threshold; Traverse all knowledge points and mark those with a mastery prediction value lower than the low mastery threshold as knowledge points to be strengthened. Knowledge points whose mastery prediction value is higher than the high mastery threshold are marked as mastered knowledge points; Retrieve all teaching units that are directly related to the knowledge points to be strengthened from the course content library, and add these teaching units to the initial candidate set; Retrieve all review teaching units from the course content library that have prerequisite dependencies on the mastered knowledge points, and add these teaching units to the initial candidate set; Duplicate teaching units are removed from the initial candidate set to obtain the candidate set to be pushed.

6. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 5, characterized in that, For each teaching unit in the candidate set to be pushed, the difficulty coefficient and knowledge relevance strength are calculated to form a push priority ranking, specifically including: Retrieves the pre-marked difficulty level values ​​for each teaching unit from the course content library; The difficulty level value is directly used as the difficulty coefficient of the teaching unit; For each teaching unit in the candidate set to be pushed, calculate the size of the intersection between the set of knowledge points covered by the teaching unit and the set of knowledge points to be strengthened; Divide the size of the intersection by the total size of the set of knowledge points covered by the teaching unit to obtain the knowledge association strength of the teaching unit; The original priority score for each teaching unit is obtained by multiplying the difficulty coefficient of each teaching unit by the strength of knowledge relevance. All teaching units in the candidate set to be pushed are sorted in descending order of their original priority scores to form the push priority sort.

7. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 1, characterized in that, The corresponding teaching units are pushed to the students' terminal devices according to the push priority, specifically including: The top N teaching units ranked in the push priority sort are selected as a push batch; For each teaching unit in the push batch, generate a Uniform Resource Locator (URL) link and an estimated learning duration for that teaching unit; All teaching units' Uniform Resource Locator (URL) links and estimated learning duration values ​​are encapsulated into a single push notification. Send the push message body to the student's terminal device; On the terminal device, the Uniform Resource Locator (URL) link for each teaching unit and the corresponding estimated learning time are displayed sequentially according to the push priority.

8. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 3 is characterized in that, The step of performing a weighted summation operation on the behavior embedding vectors corresponding to all behavior events according to the time decay weight to generate historical behavior aggregate features specifically includes: Obtain the timestamp corresponding to each behavioral event in the historical behavioral sequence data. The timestamp records the absolute time point when the behavioral event occurred. Calculate the difference between the current system time and the timestamp of each behavioral event to obtain the duration of each behavioral event. Based on the duration of each behavioral event, the time decay weight of each behavioral event is calculated by substituting it into a negative exponential function, where the decay coefficient of the negative exponential function is a pre-set positive constant. Multiply the behavior embedding vector corresponding to each behavior event by its time decay weight to obtain the weighted behavior embedding vector; The weighted behavioral embedding vectors of all behavioral events are subjected to vector addition to generate historical behavioral aggregate features.

9. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 4, characterized in that, The hidden layer of the adaptive learning model performs a nonlinear transformation operation on the learner's current cognitive state vector to generate a latent feature vector, specifically including: The hidden layer of the adaptive learning model receives the learner's current cognitive state vector output from the input layer; The hidden layer has multiple neuron nodes, and each neuron node maintains a weight vector and a bias value; The student's current cognitive state vector is multiplied by the weight vector of each neuron node, and then the bias value of that neuron node is added to generate the initial activation value of each neuron node. The ReLU activation function is applied to the initial activation value of each neuron node. The ReLU activation function sets input values ​​less than zero to zero and keeps input values ​​greater than or equal to zero unchanged. The output value of each neuron node after processing by the ReLU activation function is used as the final output of that neuron node; The final outputs of all neurons are concatenated in node order to form a new vector, which is the latent feature vector.

10. The online education and training intelligent push system based on AI adaptive learning disclosed in claim 4, characterized in that, The training steps of the adaptive learning model include: Collect historical behavior sequence data and learning outcome assessment data from multiple historical students, and obtain subsequent real knowledge point mastery assessment data from these multiple historical students; For each historical learner, a historical cognitive state vector is generated based on their historical behavior sequence data and learning outcome assessment data, using a feature encoding method. The historical cognitive state vector of each historical student is used as a training sample, and the subsequent real knowledge mastery evaluation data of that historical student is used as a training label, together forming a training sample pair. Initialize the model parameters of the adaptive learning model, including the connection weight matrix from the input layer to the hidden layer, the bias vector of the hidden layer neurons, the connection weight matrix from the hidden layer to the output layer, and the bias vector of the output layer neurons. The gradient descent optimization algorithm is adopted, with the historical cognitive state vector in the training sample pair as input and the mean square error between the mastery prediction vector output by the model and the training label as the loss function. Calculate the gradient of the loss function with respect to all parameters of the model, and update the model parameters based on the calculated gradient; Repeat the gradient calculation and parameter update process until the loss function value is less than the preset convergence threshold. At this point, stop training and obtain the trained adaptive learning model.

Images