Online learning evaluation method and system based on artificial intelligence

Abstract

The invention discloses an online learning evaluation method and system based on artificial intelligence, and relates to the technical field of learning evaluation, and the method comprises the steps: collecting and preprocessing the learning data of employees, analyzing the attention of the employees to each knowledge point through a ruminant ratio algorithm, and generating a learning behavior feature matrix; and generating a preliminary learning path based on the learning behavior feature matrix, and dynamically adjusting the learning path according to the real-time learning behavior of the employee. According to the method, by using a ruminant algorithm and combining an artificial intelligence technology, not only can the attention degree of employees to each knowledge point be accurately analyzed, but also important support can be provided for subsequent path optimization and personalized recommendation by generating a learning behavior feature matrix; through combination of the graph neural network and the dynamic time warping algorithm and calculation of potential capability representation and demand variables of the nodes, deep analysis and dynamic adjustment of employee learning behaviors are realized in learning path generation, and individual differences of employees can be better adapted.

Description

Technical Field

[0001] The present invention relates to the field of learning assessment technology, and in particular to an online learning assessment method and system based on artificial intelligence. Background Art

[0002] In recent years, with the rapid development of information technology, online learning platforms have become an important part of modern education and corporate training. Especially driven by big data and artificial intelligence technologies, more and more online learning systems have begun to adopt data-driven methods to optimize learning paths and evaluate learning effects in order to improve learners' learning efficiency and learning outcomes. At present, most online learning evaluation systems on the market rely on preset fixed rules or traditional statistical methods to monitor and evaluate employees' learning behaviors. By collecting employee behavior data in scenarios such as video learning, text reading, and interactive participation, these systems can partially identify learners' learning preferences and progress and generate certain learning paths. However, when faced with learning path optimization and personalized learning content recommendation, traditional systems rely on static rule algorithms or limited behavioral analysis methods, making it difficult to dynamically adapt to employees' learning needs and behavioral changes in real time. In addition, when evaluating employees' mastery of specific knowledge points, existing learning evaluation systems are usually based only on one-time test scores or simple learning time data, lacking in-depth analysis of employees' multiple learning behaviors, which leads to limitations in personalized recommendations and accurate evaluation. Summary of the invention

[0003] In view of the problems existing in the above-mentioned existing artificial intelligence-based online learning evaluation methods and systems, the present invention is proposed.

[0004] Therefore, the problem to be solved by the present invention is that the existing technology is difficult to dynamically adapt to employees' learning needs and behavioral changes in real time, and lacks in-depth analysis of employees' multiple learning behaviors, which leads to limitations in personalized recommendations and accurate evaluation.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: an online learning evaluation method based on artificial intelligence, which includes collecting and preprocessing employee learning data, using a rumination ratio algorithm to analyze employees' attention to each knowledge point, and generating a learning behavior feature matrix; generating a preliminary learning path based on the learning behavior feature matrix, and dynamically adjusting the learning path according to the employees' real-time learning behavior; predicting the employees' learning outcomes based on the employees' learning progress data after adjusting the learning path, and generating an employee learning evaluation report based on the prediction results.

[0006] As a preferred solution of the artificial intelligence-based online learning evaluation method described in the present invention, the collection of employee learning data and preprocessing refers to the real-time collection of employees' video learning data, text reading data and interactive behavior data, time alignment of the three types of collected learning behavior data, generation of a multimodal learning data set A, division of data set A by time segments to obtain a time-segment-based sliced ​​data set, data cleaning and normalization of each type of data in data set A, and mapping of learning behavior data of different dimensions to a unified interval of [0, 1].

[0007] As a preferred solution of the online learning evaluation method based on artificial intelligence described in the present invention, wherein: the use of the rumination ratio algorithm to analyze employees' attention to each knowledge point and generate a learning behavior feature matrix means that for each knowledge point K, the learning behavior data of the knowledge point in all time segments is extracted from the data set A, and the rumination ratio R of the knowledge point K is calculated according to the video learning data:

[0008]

[0009] Where M is the total number of learning periods for knowledge point K, T is the standard learning time for knowledge point K, and t k is the length of time that employees spend studying knowledge point K in the kth learning period;

[0010] After calculating the rumination ratio, a comprehensive behavioral feature vector is generated by combining video learning, text reading and interactive behavior data. The comprehensive behavioral feature vectors of each knowledge point are merged to construct a complete learning behavior feature matrix.

[0011] As a preferred solution of the online learning evaluation method based on artificial intelligence described in the present invention, wherein: the generation of a preliminary learning path based on the learning behavior feature matrix refers to using the generated learning behavior feature matrix as the node feature matrix of the graph neural network model, and using a dynamic time warping algorithm to calculate the similarity measure between each node;

[0012] Calculate the similarity of all node pairs and set a threshold B. If the similarity measure of two nodes is less than or equal to the threshold B, establish a relationship edge between the two nodes and generate an adjacency matrix.

[0013] Initialize each feature vector of the node feature matrix as the initial node embedding representation of the graph neural network model, use the node feature matrix and adjacency matrix to initialize the graph neural network model, and aggregate the feature information of the node and its adjacent nodes layer by layer through multi-layer graph convolution operations;

[0014] After multiple layers of graph convolution operations, the final embedding representation of each node is generated, and the node embedding representation is normalized;

[0015] The encoder network in the variational autoencoder is used to transform the node embedding representation, generate the capability mean and covariance matrix, and combine the random noise term to obtain the potential capability representation of the node, and generate the potential capability representation matrix of all nodes;

[0016] According to the potential capacity representation of the node, the node embedding representation and the contextual features of its adjacent nodes, the demand variable u of each node is generated through a weighted linear transformation model i :

[0017]

[0018] Where, d t is the required offset of the node in different learning modes, H j is the embedded feature vector representing the adjacent node j, σ() is the standard Sigmoid activation function, tanh() is the hyperbolic tangent function, ω it is the demand weight of node i under different learning modes t, β ij is the demand modification weight of neighboring node j for node i, α and γ are adjustment parameters, N(i) is the set of neighboring nodes of node i, T is the total number of learning modes, t is the index of the current learning mode, and s i is the potential capability representation vector of the i-th node in the latent space;

[0019] Generate a knowledge point demand variable matrix to represent the learning demand priority of each node in the current learning task;

[0020] Construct a weighted adjacency matrix of preliminary paths based on the required variables and node embedding representations:

[0021]

[0022] Where W i,j is the weighted adjacency matrix value between node i and node j, sim(H i ,H j ) is the similarity measure between node i and node j, ∈ is the similarity value protection term, is the demand weighting coefficient, u i is the demand variable of node i;

[0023] Determine the starting node and target node of the employee learning path, define the shortest path weight matrix based on the constructed weighted adjacency matrix to record the current shortest path weight from the starting node to each node, define the predecessor node matrix to record the predecessor node of each node in the optimal path from the initial node to the target node, and define the visited node set and the unvisited node set;

[0024] Select the node with the smallest current path weight from the unvisited node set, update the adjacent weights of all adjacent nodes of the current node, repeat the selection and update steps until the target node is added to the visited node set or the unvisited node set is empty, and start from the target node and trace back to the starting node according to the predecessor node matrix to generate a preliminary learning path.

[0025] As a preferred solution of the online learning evaluation method based on artificial intelligence described in the present invention, wherein: the dynamically adjusting the learning path according to the real-time learning behavior of employees refers to collecting the learning time, error rate and learning frequency of employees, and setting a time threshold, an error rate threshold and a learning frequency threshold;

[0026] If the error rate is greater than or equal to the error rate threshold and the learning time is greater than or equal to the learning time threshold, it means that the employee has not mastered the knowledge points well enough, so a review node is added to the learning path;

[0027] If the error rate is less than the error rate threshold and the learning time is greater than or equal to the time threshold, it means that although the employee understands the knowledge point, the learning efficiency is slow, so the current knowledge point is skipped and the next knowledge point is entered;

[0028] If the learning frequency is less than or equal to the learning frequency threshold and the error rate is greater than or equal to the error rate threshold, it means that the employee's review frequency is insufficient and the error rate is high. The current knowledge point requires more review time. In this case, the employee is prompted to strengthen the review and add a review node in the learning path.

[0029] If the learning frequency is greater than the learning frequency threshold and the error rate is lower than the error rate threshold, it means that the employee frequently reviews the knowledge point and performs well. In this case, the current knowledge point is marked as "completed" and the knowledge point is skipped on the learning path.

[0030] As a preferred solution of the online learning evaluation method based on artificial intelligence described in the present invention, wherein: predicting the learning outcomes of employees according to the employee learning progress data after adjusting the learning path refers to using the LSTM model to build a learning outcome prediction model, adding an LSTM layer, setting the input dimension to n*m, where n is the number of time steps, m is the number of features for each time step, the number of hidden layer units is 50, stacking multiple LSTM layers, adding a fully connected layer and an output layer after the last LSTM layer, setting the output layer dimension to 1, and using the mean square error as the loss function;

[0031] Collect employees’ historical learning data, pre-process it and input it into the model as a training data set. Perform iterative optimization of model parameters based on the Adam optimizer until the loss is minimized and then stop the iteration. Input the iterated model parameters into the learning outcome prediction model to obtain the learning outcome prediction model.

[0032] The employee learning data collected in real time is input into the learning outcome prediction model to obtain the employee's learning outcome prediction score.

[0033] As a preferred solution of the online learning evaluation method based on artificial intelligence described in the present invention, wherein: generating an employee's learning evaluation report based on the prediction results refers to collecting the employee's learning outcome prediction scores and the learning behavior data in the learning path to generate a learning report prediction report, displaying the learning outcome scores predicted by the LSTM model in the first position of the report, displaying the employee's learning progress and performance in the form of charts, providing the employee's future learning path planning and development direction in combination with the prediction results, and continuously updating the evaluation report according to the update of the employee's learning path.

[0034] Another object of the present invention is to provide an online learning evaluation system based on artificial intelligence, which comprises:

[0035] Data collection module, used to collect employees' multimodal learning behavior data in real time and pre-process the collected data;

[0036] A matrix generation module is used to calculate the rumination ratio of video behavior data using a rumination ratio algorithm, and to generate a comprehensive behavior feature vector by combining text reading, interactive behavior, and video behavior data to construct a learning behavior feature matrix;

[0037] A preliminary path generation module is used to generate a node feature matrix and a node embedding representation based on the learning behavior feature matrix as the input of the graph neural network model, derive a potential capability representation and node demand variables based on the node embedding representation, and generate a preliminary learning path;

[0038] The learning path dynamic adjustment module is used to dynamically adjust the employee's initial learning path according to the employee's real-time learning behavior;

[0039] The outcome prediction module is used to predict the learning behavior data of employees after the path adjustment to obtain the predicted results of the employees' learning outcomes;

[0040] The report generation module is used to generate employee learning evaluation reports based on prediction results and real-time learning behavior data, and provide charts to display the reports.

[0041] A computer device comprises: a memory and a processor; the memory stores a computer program, and the processor implements the steps of an online learning evaluation method based on artificial intelligence when executing the computer program.

[0042] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of an online learning evaluation method based on artificial intelligence.

[0043] The beneficial effects of the present invention are as follows: the present invention uses the rumination ratio algorithm in combination with artificial intelligence technology, which can not only accurately analyze employees' attention to various knowledge points, but also generate a learning behavior feature matrix to provide important support for subsequent path optimization and personalized recommendations. By combining the graph neural network with the dynamic time warping algorithm and calculating the potential capability representation and demand variables of the nodes, the present invention realizes in-depth analysis and dynamic adjustment of employees' learning behaviors in learning path generation, which can better adapt to individual differences of employees. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0045] Figure 1 Schematic diagram of the process of online learning evaluation method based on artificial intelligence.

[0046] Figure 2 Schematic diagram of the structure of the online learning evaluation system based on artificial intelligence. DETAILED DESCRIPTION

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0049] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or selective embodiment that is mutually exclusive with other embodiments.

[0050] Example 1, reference Figure 1 , which is the first embodiment of the present invention, and which provides an online learning evaluation method based on artificial intelligence. The online learning evaluation method based on artificial intelligence includes:

[0051] S1. Collect and pre-process employee learning data, use the rumination ratio algorithm to analyze employees' attention to each knowledge point, and generate a learning behavior feature matrix;

[0052] Specifically, collecting employee learning data and preprocessing it refers to collecting employees' video learning data, text reading data, and interactive behavior data in real time, time-aligning the three types of collected learning behavior data to generate a multimodal learning dataset A, dividing the dataset A by time segments to obtain a time-segment-based sliced ​​dataset, cleaning and normalizing each type of data in the dataset A, and mapping the learning behavior data of different dimensions to a unified interval of [0, 1].

[0053] By collecting employees' data in video learning, text reading and interactive behavior in real time, the system can be ensured to analyze based on multi-dimensional learning behavior. By integrating video, text and interactive behavior data, the system can more comprehensively capture employees' learning patterns and behavioral characteristics. For example, the pause and double-speed playback behaviors shown by employees in video learning, combined with the number of keyword lookups in text reading and the degree of participation in interactive behavior, can more accurately reflect the employees' mastery of knowledge points. Time alignment ensures the consistency of the time dimension between video learning, text reading and interactive behavior, allowing the system to analyze different types of learning behaviors within the same time window. The generation and slicing of data set A enables the system to perform more detailed analysis based on learning behaviors within a short time window. Through normalization, the present invention maps different types of learning data to the [0,1] interval, ensuring that the dimensional differences between different dimensions can be eliminated when performing model calculations.

[0054] Furthermore, the rumination ratio algorithm is used to analyze employees’ attention to each knowledge point and generate a learning behavior feature matrix. For each knowledge point K, the learning behavior data of the knowledge point in all time segments is extracted from the data set A, and the rumination ratio R of the knowledge point K is calculated based on the video learning data:

[0055]

[0056] Where M is the total number of learning periods for knowledge point K, T is the standard learning time for knowledge point K, and t k is the length of time that employees spend studying knowledge point K in the kth learning period;

[0057] After calculating the rumination ratio, a comprehensive behavioral feature vector is generated by combining video learning, text reading and interactive behavior data. The comprehensive behavioral feature vectors of each knowledge point are merged to construct a complete learning behavior feature matrix.

[0058] In corporate training or online education systems, courses are usually pre-designed and contain multiple modules or chapters, each of which consists of multiple knowledge points. A knowledge point is the smallest unit of the course and can be a concept, skill, or task. Therefore, during the data collection process, the system will pre-divide each learning content into multiple knowledge points based on the structure of the course content to ensure that each knowledge point corresponds to a specific learning goal. For example, during the training process, a knowledge point may be a specific skill about "project management", such as "the use of time management tools". Through the input of the course designer, the system can mark this skill as an independent knowledge point and track the learning behavior of the knowledge point when the trainees learn related content.

[0059] The rumination ratio algorithm provides a fine-grained measurement standard for traditional learning behavior analysis. Compared with the evaluation of a single learning time or performance, the rumination ratio can capture the changes in employees' learning behavior in different learning periods and quantify employees' attention and learning intensity to a certain knowledge point. This not only helps to identify the learning difficulties of employees at a certain knowledge point, but also provides an accurate basis for the personalized optimization of learning paths. By integrating data from multiple learning modes such as video learning, text reading, and interactive behavior, the system can generate a feature matrix that comprehensively reflects the learning behavior of employees. This comprehensive matrix is ​​difficult to achieve in traditional learning systems because traditional methods usually rely on data from a single dimension for learning evaluation. The matrix of the present invention provides multi-dimensional behavioral features, which can provide a reliable data basis for subsequent learning path generation and dynamic optimization. The process of generating a comprehensive behavioral feature vector ensures the effective integration of multi-dimensional learning behavior data. Video learning data can reflect employees' learning behavior on dynamic knowledge points, while text reading data and interactive behavior can reflect the depth and participation of employees in the understanding process. By unifying these data into a feature vector, the system can evaluate employees' learning performance from multiple perspectives.

[0060] S2. Generate a preliminary learning path based on the learning behavior feature matrix and dynamically adjust the learning path according to the real-time learning behavior of employees;

[0061] Specifically, generating a preliminary learning path based on the learning behavior feature matrix means using the generated learning behavior feature matrix as the node feature matrix of the graph neural network model, each feature vector represents the learning behavior pattern of the employee on the knowledge point, and using the dynamic time warping algorithm to calculate the similarity measure between each node;

[0062] Calculate the similarity of all node pairs, set a threshold B based on the elbow rule, and if the similarity measure of two nodes is less than or equal to the threshold B, establish a relationship edge between the two nodes and generate an adjacency matrix;

[0063] Each feature vector of the node feature matrix is ​​initialized as the initial node embedding representation of the graph neural network model. The graph neural network model is initialized using the node feature matrix and the adjacency matrix. Through multi-layer graph convolution operations, the feature information of the nodes and their adjacent nodes is aggregated layer by layer. The feature aggregation formula of each layer of graph convolution is as follows:

[0064]

[0065] In the formula, represents the embedding representation of node i after the l-th layer of graph convolution operation, W (l) is the weight matrix of the l-th layer graph convolution operation, N(i) is the set of neighboring nodes of node i, σ is the ReLU nonlinear activation function, deg(i) is the number of connecting edges between node i and other nodes, deg(j) is the number of direct neighbors of node j, is the feature representation of the adjacent node j at the l-1 layer;

[0066] After multiple layers of graph convolution operations, the final embedding representation of each node is generated, and the node embedding representation is normalized;

[0067] The encoder network in the variational autoencoder is used to transform the node embedding representation, generate the capability mean and covariance matrix, and combine the random noise term to obtain the potential capability representation of the node:

[0068] s i =μ φ (H i )+∈·σ φ (H i ),

[0069] In the formula, s i is the potential ability representation vector of the i-th node in the latent space, which is used to describe the personalized learning ability status of the employee at this node, μ φ (H i ) is the average ability level of the ith node in the latent space, which can be understood as the “central position” or “core ability” of the node, σ φ (H i ) is the standard deviation of the learning ability distribution of the i-th node in the latent space, which is used to describe the distribution range and uncertainty of the node in the latent space, ∈ is a random noise term sampled from a standard normal distribution, which is used to introduce randomness and diversity in the representation of latent ability;

[0070] Generate the potential capacity representation matrix of all nodes;

[0071] According to the potential capacity representation of the node, the node embedding representation and the contextual features of its adjacent nodes, the demand variable u of each node is generated through a weighted linear transformation modeli :

[0072] The formula of the original weighted linear transformation model is expressed as: In the formula, w m is the weight of the mth feature, x m is the value of the mth feature, b is the bias term, and p is the total number of features;

[0073] Through the potential ability representation of the node, the node embedding representation and the contextual features of the adjacent nodes, we can get:

[0074]

[0075] In order to introduce richer interactive relationships in the demand variable generation model, nonlinear functions are used to process the potential capacity representation of nodes and the embedded representation of adjacent nodes. The Sigmoid function can map the input value to between (0, 1), which is suitable for processing probability-related demand variables. The Tanh function can map the input value to between (-1, 1), which is suitable for processing demand relationships with strong symmetry.

[0076] In the original formula, the potential capacity of a node is represented by s i and embedding representation H j The weights are added directly. This processing method ignores the overall strength differences between different features. Therefore, the norm measurement is added in the optimization process. By performing norm processing on vector features, high-dimensional vector features can be converted into a single scalar value, thereby better capturing the overall strength and change trend of the features.

[0077] There is no complex combination relationship between node features and adjacent node context features. Therefore, a fractional structure is introduced during optimization, and the dependency between different nodes is expressed through the correction term of the context features.

[0078]

[0079] Where, d t It is the demand offset of the node under different learning modes (such as "beginner mode" or "review mode"). By analyzing the historical learning data, the changes in employees' demand characteristics under different learning modes are determined, which is used to adjust the potential ability representation. j is the embedded feature vector representing the adjacent node j, reflecting the feature representation of the node in the learning path graph model, σ() is the standard Sigmoid activation function, which is used to map the input value to (0,1), tanh() is the hyperbolic tangent function, which is used to map the input value to (-1,1), ∥s i +d t∥ represents the overall strength of the potential ability to represent the learning mode offset, measuring the comprehensive performance of node features, ∥H j -s i ∥ is the difference between the embedding representation of the adjacent node j and the potential ability representation of the current node i, ω it is the demand weight of node i under different learning modes t, which is fitted by historical learning behavior data and optimized by back propagation algorithm. ij is the demand correction weight of the adjacent node j for the node i, which is fitted by analyzing the contextual relationship between nodes and the historical learning behavior data. α and γ are adjustment parameters used to control the nonlinearity of the Sigmoid and Tanh activation functions. They are optimized by cross-validation or grid search, and the value range is usually (0,1). N(i) is the set of adjacent nodes of node i, T is the total number of learning modes, and t is the index of the current learning mode.

[0080] Generate a knowledge point demand variable matrix to represent the learning demand priority of each node in the current learning task;

[0081] Construct a weighted adjacency matrix of preliminary paths based on the required variables and node embedding representations:

[0082]

[0083] Where W i,j is the weighted adjacency matrix value between node i and node j, indicating the connectivity weight between the two nodes and used to describe the smoothness of the path from node i to node j. sim(H i ,H j ) is the similarity measure between node i and node j, which measures the similarity between two nodes (knowledge points) in the feature space and is calculated by cosine similarity. ∈ is the similarity value protection term, which is used to prevent the denominator calculation problem when the similarity is zero. is the demand weighting coefficient, used to balance the demand variable u i The influence degree in the weight matrix usually ranges from 0.1 to 0.5, depending on the actual scenario requirements, and can be dynamically adjusted according to the distribution characteristics of the demand variables. i is the demand variable of node i, indicating the priority learning demand of this node (knowledge point) in the current learning task;

[0084] Determine the starting node (employee's learning status) and target node (the final target knowledge point that the employee needs to reach) of the employee's learning path, define the shortest path weight matrix based on the constructed weighted adjacency matrix to record the current shortest path weight from the starting node to each node, define the predecessor node matrix to record the predecessor node of each node in the optimal path from the initial node to the target node, and define the visited node set and the unvisited node set;

[0085] Select the node with the smallest current path weight from the unvisited node set, update the adjacent weights of all adjacent nodes of the current node, repeat the selection and update steps until the target node is added to the visited node set or the unvisited node set is empty, and start from the target node and trace back to the starting node according to the predecessor node matrix to generate a preliminary learning path.

[0086] By aggregating the feature information of nodes and their adjacent nodes layer by layer through GNN, the system can capture complex node relationships, making the generated learning path more flexible and intelligent. The traditional system only relies on preset rules, while the present invention can dynamically adjust the learning path through feature aggregation and similarity measurement to ensure that the learning path is continuously optimized according to the actual learning behavior of employees. When processing the time series data of employees at different learning stages, the DTW algorithm can dynamically match the learning behavior of employees at different time periods, improve the flexibility and adaptability of path generation, and make up for the shortcomings of the traditional system in the time dimension. The potential ability representation can accurately model the personalized learning ability of employees at different knowledge points, and introduce randomness and diversity in combination with random noise terms. This design enhances the adaptability and elasticity of the learning path, especially when dealing with uncertain employee learning performance, it can provide a more reasonable learning plan for each knowledge point. In the generation of learning paths, the potential ability representation not only provides the core learning ability of employees, but also takes into account the range of variation of ability distribution, which enables the learning path to be dynamically adjusted according to the learning ability status of employees to avoid over-learning or under-learning. The use of nonlinear functions to introduce more complex interactive relationships enables the demand variable generation model to better express the dependencies between nodes. Compared with the traditional linear weight model, this design can more accurately capture the learning needs of employees under different learning modes, and thus generate a more personalized learning path. By optimizing the node features in the weighted matrix through norm measurement, the system can better capture the overall strength and change trend of high-dimensional features. This optimization method greatly improves the robustness and expressiveness of the model when processing complex learning behavior data. The weighted adjacency matrix not only takes into account the similarity between nodes, but also balances the personalized learning needs of employees through the demand weighting coefficient, ensuring that the generated path can reasonably match the actual learning level of the employees. Through this optimization, the present invention realizes the efficient generation of personalized learning paths and effectively improves learning efficiency. During the path generation process, the system can adjust the path in real time by dynamically updating the node weights, thereby ensuring that employees always move along the optimal learning path. This design greatly improves the adaptability and intelligence level of the online learning system.

[0087] Furthermore, dynamically adjusting the learning path according to the real-time learning behavior of employees means collecting the learning time, error rate and learning frequency of employees, and setting a time threshold, an error rate threshold and a learning frequency threshold;

[0088] The duration threshold is set based on the standard learning duration of the knowledge point;

[0089] The error rate threshold can be flexibly adjusted according to the characteristics of different knowledge points. For example, for relatively basic knowledge points, the error rate threshold may be set to 10%. That is, when the employee's error rate exceeds 10%, the system will consider that the employee has insufficient grasp of the knowledge point, and thus increase the review node. For advanced knowledge points, the error rate threshold can be appropriately increased to avoid triggering a review too early on complex knowledge points.

[0090] The learning frequency threshold directly sets the recommended number of reviews based on the course requirements. For example, if the review recommendation for a certain knowledge point is 3 times, the learning frequency threshold is set to 3 times. If the employee's review frequency is less than 3 times and the error rate of the knowledge point is high, the system will prompt the employee to review more. If the review frequency reaches or exceeds 3 times and the error rate is low, the system can assume that the knowledge point has been well mastered.

[0091] If the error rate is greater than or equal to the error rate threshold and the learning time is greater than or equal to the learning time threshold, it means that the employee has not mastered the knowledge points well enough, so a review node is added to the learning path;

[0092] If the error rate is less than the error rate threshold and the learning time is greater than or equal to the time threshold, it means that although the employee understands the knowledge point, the learning efficiency is slow, so the current knowledge point is skipped and the next knowledge point is entered;

[0093] If the learning frequency is less than or equal to the learning frequency threshold and the error rate is greater than or equal to the error rate threshold, it means that the employee's review frequency is insufficient and the error rate is high. The current knowledge point requires more review time. In this case, the employee is prompted to strengthen the review and add a review node in the learning path.

[0094] If the learning frequency is greater than the learning frequency threshold and the error rate is lower than the error rate threshold, it means that the employee frequently reviews the knowledge point and performs well. In this case, the current knowledge point is marked as "completed" and skipped on the learning path to avoid unnecessary repeated learning.

[0095] By collecting multi-dimensional data on learning time, error rate and learning frequency, the present invention can comprehensively analyze employees' learning behavior on different knowledge points, thereby providing a reliable basis for subsequent path adjustment. Setting the time, error rate and learning frequency thresholds can make it clear when the path adjustment should be triggered. For example, the learning time threshold can prevent employees from wasting too much time on a certain knowledge point, and the error rate threshold helps the system identify the knowledge points that employees need to review, greatly enhancing the personalization and flexibility of the learning path. In the prior art, many learning path adjustments are based only on simple preset rules, lacking dynamic adjustments for employees' personal learning status. The present invention can adjust the learning path in real time according to the specific performance of employees by setting multiple thresholds, thereby improving learning efficiency. Adjusting the path according to the learning time and error rate ensures the balance between the flexibility and efficiency of the learning path, and by combining the learning time and error rate, the real-time optimization of the learning path is achieved. When an employee shows that he has insufficient mastery of certain knowledge points, the system will add review nodes, and when an employee shows that he has a good understanding but low learning efficiency, the system can skip unnecessary repeated learning. By combining the learning frequency, the system can more accurately determine whether an employee needs to review a certain knowledge point. Compared with the traditional approach that path adjustment only relies on test results or learning time, the present invention uses the indicator of learning frequency to effectively improve the optimization accuracy of the learning path.

[0096] S3. Predict the learning outcomes of employees based on their learning progress data after adjusting their learning paths, and generate a learning evaluation report for them based on the prediction results;

[0097] Specifically, predicting the learning outcomes of employees based on the learning progress data of employees after adjusting the learning path refers to using the LSTM model to build a learning outcome prediction model, adding an LSTM layer, setting the input dimension to n*m, where n is the number of time steps, m is the number of features for each time step, the number of hidden layer units is 50, stacking multiple LSTM layers, adding a fully connected layer and an output layer after the last LSTM layer, setting the output layer dimension to 1, and using the mean square error as the loss function;

[0098] Collect employees’ historical learning data, pre-process it and input it into the model as a training data set. Perform iterative optimization of model parameters based on the Adam optimizer until the loss is minimized and then stop the iteration. Input the iterated model parameters into the learning outcome prediction model to obtain the learning outcome prediction model.

[0099] The employee learning data collected in real time is input into the learning outcome prediction model to obtain the employee's learning outcome prediction score.

[0100] The LSTM model is very suitable for time series data analysis of employee learning behavior because of its ability to handle long-term dependencies. In online learning scenarios, employees' learning behaviors often contain rich time information, such as the number of times the same knowledge point is learned in different time periods, the length of time it takes to learn, and the learning results. Through its memory gate mechanism, LSTM can capture long-term behavior patterns in the learning process of employees, which is crucial for predicting employees' future learning outcomes. Traditional linear regression or simple neural network models are difficult to capture such complex time dependencies, while LSTM can more accurately predict employees' future learning outcomes by modeling dynamic changes in time series. Traditional learning outcome assessment usually relies on periodic exams or tests, while the present invention can continuously evaluate employees during the learning process by inputting employees' learning behavior data in real time, timely discover employees' learning weaknesses, and provide a basis for adjusting subsequent learning paths. This real-time nature provides companies with timely employee training feedback, which can better help companies develop personalized training plans.

[0101] Furthermore, generating an employee's learning evaluation report based on the prediction results refers to collecting the employee's learning outcome prediction scores and the learning behavior data in the learning path to generate a learning report prediction report, displaying the learning outcome scores predicted by the LSTM model in the first position of the report, providing an intuitive evaluation of the employee's learning outcomes, and displaying the employee's learning progress and performance in the form of charts. In combination with the prediction results, the employee's future learning path planning and development direction are provided, for example: "The system predicts that you have strong learning potential in advanced knowledge point D. It is recommended that you continue to study advanced courses to further improve your professional ability." The evaluation report is continuously updated according to the update of the employee's learning path.

[0102] The learning assessment report not only displays the predicted scores of employees' learning outcomes, but also intuitively displays employees' learning progress and performance in the form of charts. This visual learning assessment report, combined with multi-dimensional data analysis, can clearly present the learning change trajectory of employees at different time points, and reflect the dynamic changes of learning outcomes through trend charts or bar charts. At the same time, as the learning path of employees is dynamically adjusted, the assessment report will also be updated in real time to ensure that employees and managers can keep abreast of the latest learning progress and potential problems, help employees identify problems in the learning process, and make adjustments to weak links. Through the visual display form, employees can not only intuitively see their learning progress, but also get systematic learning suggestions to ensure the continuous optimization of the learning path.

[0103] Example 2, reference Figure 2, which is the second embodiment of the present invention, and which is different from the previous embodiment, provides an online learning evaluation system based on artificial intelligence, which includes a data acquisition module for collecting employees' multimodal learning behavior data in real time and preprocessing the collected data;

[0104] A matrix generation module is used to calculate the rumination ratio of video behavior data using a rumination ratio algorithm, and to generate a comprehensive behavior feature vector by combining text reading, interactive behavior, and video behavior data to construct a learning behavior feature matrix;

[0105] A preliminary path generation module is used to generate a node feature matrix and a node embedding representation based on the learning behavior feature matrix as the input of the graph neural network model, derive a potential capability representation and node demand variables based on the node embedding representation, and generate a preliminary learning path;

[0106] The learning path dynamic adjustment module is used to dynamically adjust the employee's initial learning path according to the employee's real-time learning behavior;

[0107] The outcome prediction module is used to predict the learning behavior data of employees after the path adjustment to obtain the predicted results of the employees' learning outcomes;

[0108] The report generation module is used to generate employee learning evaluation reports based on prediction results and real-time learning behavior data, and provide charts to display the reports.

[0109] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0110] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0111] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0112] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

Claims

1. An online learning evaluation method based on artificial intelligence, characterized by: include, Collect and pre-process employee learning data, use the rumination ratio algorithm to analyze employees' attention to each knowledge point, and generate a learning behavior feature matrix; Generate preliminary learning paths based on the learning behavior feature matrix, and dynamically adjust the learning paths according to employees' real-time learning behaviors; The employee's learning outcomes are predicted based on the employee's learning progress data after the learning path is adjusted, and an employee learning evaluation report is generated based on the prediction results.

2. The online learning evaluation method based on artificial intelligence as claimed in claim 1, characterized in that: The collecting of employee learning data and preprocessing thereof refers to collecting employees' video learning data, text reading data and interactive behavior data in real time, time-aligning the three types of collected learning behavior data to generate a multimodal learning data set A, dividing the data set A by time segments to obtain a time-segment-based segmented data set, cleaning and normalizing each type of data in the data set A, and mapping learning behavior data of different dimensions to a unified interval of [0, 1].

3. The online learning evaluation method based on artificial intelligence as claimed in claim 2, characterized in that: The use of the rumination ratio algorithm to analyze employees' attention to each knowledge point and generate a learning behavior feature matrix means that for each knowledge point K, the learning behavior data of the knowledge point in all time segments is extracted from the data set A, and the rumination ratio R of the knowledge point K is calculated based on the video learning data: Where M is the total number of learning periods for knowledge point K, T is the standard learning time for knowledge point K, and t k is the length of time that employees spend studying knowledge point K in the kth learning period; After calculating the rumination ratio, a comprehensive behavioral feature vector is generated by combining video learning, text reading and interactive behavior data. The comprehensive behavioral feature vectors of each knowledge point are merged to construct a complete learning behavior feature matrix.

4. The online learning evaluation method based on artificial intelligence as claimed in claim 3, characterized in that: Generating a preliminary learning path based on the learning behavior feature matrix refers to using the generated learning behavior feature matrix as a node feature matrix of the graph neural network model, and using a dynamic time warping algorithm to calculate the similarity measure between each node; Calculate the similarity of all node pairs and set a threshold B. If the similarity measure of two nodes is less than or equal to the threshold B, establish a relationship edge between the two nodes and generate an adjacency matrix. Initialize each feature vector of the node feature matrix as the initial node embedding representation of the graph neural network model, use the node feature matrix and adjacency matrix to initialize the graph neural network model, and aggregate the feature information of the node and its adjacent nodes layer by layer through multi-layer graph convolution operations; After multiple layers of graph convolution operations, the final embedding representation of each node is generated, and the node embedding representation is normalized; The encoder network in the variational autoencoder is used to transform the node embedding representation, generate the capability mean and covariance matrix, and combine the random noise term to obtain the potential capability representation of the node, and generate the potential capability representation matrix of all nodes; According to the potential capacity representation of the node, the node embedding representation and the contextual features of its adjacent nodes, the demand variable u of each node is generated through a weighted linear transformation model i : Where, d t is the required offset of the node in different learning modes, H j is the embedded feature vector representing the adjacent node j, σ() is the standard Sigmoid activation function, tanh() is the hyperbolic tangent function, ω it is the demand weight of node i under different learning modes t, β ij is the demand modification weight of neighboring node j for node i, α and γ are adjustment parameters, N(i) is the set of neighboring nodes of node i, T is the total number of learning modes, t is the index of the current learning mode, and s i is the potential capability representation vector of the i-th node in the latent space; Generate a knowledge point demand variable matrix to represent the learning demand priority of each node in the current learning task; Construct a weighted adjacency matrix of preliminary paths based on the required variables and node embedding representations: Where W i,j is the weighted adjacency matrix value between node i and node j, sim(H i ,H j ) is the similarity measure between node i and node j, ∈ is the similarity value protection term, is the demand weighting coefficient, u i is the demand variable of node i; Determine the starting node and target node of the employee learning path, define the shortest path weight matrix based on the constructed weighted adjacency matrix to record the current shortest path weight from the starting node to each node, define the predecessor node matrix to record the predecessor node of each node in the optimal path from the initial node to the target node, and define the visited node set and the unvisited node set; Select the node with the smallest current path weight from the unvisited node set, update the adjacent weights of all adjacent nodes of the current node, repeat the selection and update steps until the target node is added to the visited node set or the unvisited node set is empty, and start from the target node and trace back to the starting node according to the predecessor node matrix to generate a preliminary learning path.

5. The online learning evaluation method based on artificial intelligence as claimed in claim 4, characterized in that: The dynamically adjusting the learning path according to the real-time learning behavior of employees refers to collecting the learning time, error rate and learning frequency of employees, and setting a time threshold, an error rate threshold and a learning frequency threshold; If the error rate is greater than or equal to the error rate threshold and the learning time is greater than or equal to the learning time threshold, it means that the employee has not mastered the knowledge points well enough, so a review node is added to the learning path; If the error rate is less than the error rate threshold and the learning time is greater than or equal to the time threshold, it means that although the employee understands the knowledge point, the learning efficiency is slow, so the current knowledge point is skipped and the next knowledge point is entered; If the learning frequency is less than or equal to the learning frequency threshold and the error rate is greater than or equal to the error rate threshold, it means that the employee's review frequency is insufficient and the error rate is high. The current knowledge point requires more review time. In this case, the employee is prompted to strengthen the review and add a review node in the learning path. If the learning frequency is greater than the learning frequency threshold and the error rate is lower than the error rate threshold, it means that the employee frequently reviews the knowledge point and performs well. In this case, the current knowledge point is marked as "completed" and skipped on the learning path.

6. The online learning evaluation method based on artificial intelligence as claimed in claim 5, characterized in that: The predicting of the learning outcomes of employees according to the employee learning progress data after the learning path is adjusted refers to building a learning outcome prediction model using an LSTM model, adding an LSTM layer, setting the input dimension to n*m, where n is the number of time steps, m is the number of features for each time step, the number of hidden layer units is 50, stacking multiple LSTM layers, adding a fully connected layer and an output layer after the last LSTM layer, setting the output layer dimension to 1, and using the mean square error as the loss function; Collect employees’ historical learning data, pre-process it and input it into the model as a training data set. Perform iterative optimization of model parameters based on the Adam optimizer until the loss is minimized and then stop the iteration. Input the iterated model parameters into the learning outcome prediction model to obtain the learning outcome prediction model. The employee learning data collected in real time is input into the learning outcome prediction model to obtain the employee's learning outcome prediction score.

7. The online learning evaluation method based on artificial intelligence as claimed in claim 6, characterized in that: Generating an employee's learning evaluation report based on the prediction results refers to collecting the employee's learning outcome prediction scores and the learning behavior data in the learning path to generate a learning report prediction report, displaying the learning outcome scores predicted by the LSTM model in the first position of the report, displaying the employee's learning progress and performance in the form of charts, providing the employee's future learning path planning and development direction in combination with the prediction results, and continuously updating the evaluation report according to the update of the employee's learning path.

8. An artificial intelligence-based online learning evaluation system based on the artificial intelligence-based online learning evaluation method according to any one of claims 1 to 7, characterized in that: include, Data collection module, used to collect employees' multimodal learning behavior data in real time and pre-process the collected data; A matrix generation module is used to calculate the rumination ratio of video behavior data using a rumination ratio algorithm, and to generate a comprehensive behavior feature vector by combining text reading, interactive behavior, and video behavior data to construct a learning behavior feature matrix; A preliminary path generation module is used to generate a node feature matrix and a node embedding representation based on the learning behavior feature matrix as the input of the graph neural network model, derive a potential capability representation and node demand variables based on the node embedding representation, and generate a preliminary learning path; The learning path dynamic adjustment module is used to dynamically adjust the employee's initial learning path according to the employee's real-time learning behavior; The outcome prediction module is used to predict the learning behavior data of employees after the path adjustment to obtain the predicted results of the employees' learning outcomes; The report generation module is used to generate employee learning evaluation reports based on prediction results and real-time learning behavior data, and provide charts to display the reports.

9. A computer device comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of the online learning evaluation method based on artificial intelligence described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the online learning evaluation method based on artificial intelligence described in any one of claims 1 to 7 are implemented.

Images