Intelligent teaching method, system, device and medium based on AI intelligence

Abstract

The application provides an AI intelligent-based intelligent teaching method, system, device and medium, relates to the technical field of intelligent teaching, and first collects learning track data, teaching feedback data and image data of students in a teaching process; then a dataset is constructed based on the learning track data and the teaching feedback data, meanwhile, the image data is converted into feature points and encoded to obtain feature coding data; subsequently, the K nearest neighbor algorithm is used for collaborative analysis of the dataset and the feature coding data, so as to divide sample clusters of different learning states and extract typical features therefrom; further, the long short-term memory network is used for time series analysis of the typical features, so as to obtain an evaluation result of teaching matching degree and learning efficiency; finally, the teaching resource pushing strategy of the platform is optimized based on the evaluation result. The method can realize accurate profiling and dynamic evaluation of the learning state of students, so as to drive adaptive optimization of the teaching resource pushing strategy.

Description

AI-based intelligent teaching methods, systems, equipment, and media Technical Field

[0001] This application relates to the technical field of smart teaching, and in particular to a smart teaching method, system, device and medium based on AI intelligence. Background Technology

[0002] With the popularization of online education, teaching based on artificial intelligence (AI) has become the key to improving teaching quality. It enables personalized teaching through technological means, which helps to optimize the learning experience and improve the efficiency of educational resource utilization, and has broad development prospects.

[0003] Currently, most existing smart teaching technologies rely on single-dimensional data analysis. For example, some systems recommend resources by analyzing students' answer accuracy and study time; other solutions process video surveillance data separately to identify simple states such as students' concentration.

[0004] However, the existing methods mentioned above are insufficient in integrating multi-dimensional information such as behavior, feedback, and vision when dealing with complex and continuous online learning processes. They also lack in-depth characterization of the dynamic evolution of learning status. This results in an incomplete and untimely assessment of students' actual learning needs, which in turn affects the accuracy and adaptability of teaching resource delivery. Summary of the Invention

[0005] The purpose of this application is to provide an AI-based intelligent teaching method, system, device, and medium to address the problems of low accuracy and insufficient adaptability in the delivery of teaching resources in existing technologies.

[0006] To address the aforementioned technical problems, firstly, this application provides an AI-based intelligent teaching method, comprising:

[0007] During the teaching process using the teaching resource push strategy on the AI ​​teaching platform, data on students' learning trajectory, teaching feedback, and image data are collected.

[0008] Based on the learning trajectory data and the teaching feedback data, a dataset is constructed, the image data is converted into feature point data, and the feature point data is encoded to obtain feature-encoded data.

[0009] The K-nearest neighbor algorithm is used to perform collaborative analysis on the dataset and the feature encoding data to obtain sample clusters corresponding to different learning states, and typical features are extracted from the data corresponding to the sample clusters.

[0010] Long Short-Term Memory (LSTM) networks were used to perform temporal analysis on the typical features to obtain teaching matching and efficiency evaluation results.

[0011] The teaching resource delivery strategy is optimized based on the teaching matching degree and the efficiency evaluation results.

[0012] Optionally, the step of using the K-nearest neighbor algorithm to perform collaborative analysis on the dataset and the feature-encoded data to obtain sample clusters corresponding to different learning states, and extracting typical features from the data corresponding to the sample clusters, includes:

[0013] The learning trajectory data in the dataset is aligned with the feature encoding data based on timestamps to obtain comprehensive data;

[0014] The K-nearest neighbor algorithm is used to calculate the joint distance metric by using the image feature encoding in the comprehensive data as spatial anchor points, and K neighboring data points are determined based on the joint distance metric to obtain the neighboring dataset;

[0015] Based on the feature distribution of the neighbor dataset, a mechanism combining density peak detection and K-nearest neighbor classification is used to divide the comprehensive data to obtain sample clusters corresponding to different learning states. The sample clusters and their feature distributions are used as feedback signals and input to a lightweight evaluation network to evaluate the clustering quality online and dynamically adjust the sensitivity parameters of density peak detection.

[0016] For each sample cluster, the core data region of each sample cluster is determined by calculating the density distribution and silhouette coefficient of the data points within the sample cluster in the high-dimensional feature space.

[0017] Based on the core data region, a combination of principal component analysis and canonical correlation analysis is used to extract features from the data corresponding to the sample cluster, thereby obtaining typical features.

[0018] Optionally, the K-nearest neighbor algorithm is used, with image feature encoding in the integrated data as spatial anchor points, to calculate a joint distance metric, and to determine K neighboring data points based on the joint distance metric to obtain a neighboring dataset, including:

[0019] Image feature codes of all data points are extracted from the comprehensive data, and a clustering algorithm is used to identify a preset number of cluster centers, which are then defined as spatial anchor points.

[0020] Based on the spatial anchors, the learning behavior features of each data point are weighted and calculated according to the similarity between the corresponding image feature encoding and each spatial anchor, so as to project the learning behavior features into the anchor space composed of the spatial anchors, and obtain the projected behavior feature vector.

[0021] For any two data points in the comprehensive data, calculate the first distance between the image feature codes of the two data points, and the second distance between the projected behavioral feature vectors;

[0022] A weight coefficient is introduced that is dynamically generated based on the overall correlation of the image feature encoding and the projected behavioral feature vector. The first distance and the second distance are weighted and fused using the weight coefficient to calculate the joint distance metric between data points and form the target matrix.

[0023] Based on the target matrix, select the K nearest neighbor data points with the smallest joint distance metric for each data point to form an initial neighbor list;

[0024] By combining the timestamp information of each data point in the comprehensive data, the initial neighbor list is verified to obtain a set of neighbor relationships.

[0025] For each data point in the neighborhood set, construct a structured record. The structured record contains at least the unique identifiers of all its neighboring data points, the corresponding joint distance metric, and the timestamp difference, thus obtaining the neighborhood dataset.

[0026] Optionally, the step of using a long short-term memory network to perform time-series analysis on the typical features to obtain teaching matching degree and efficiency evaluation results includes:

[0027] Based on the sequence processing capabilities of Long Short-Term Memory (LSTM) networks, the typical features are organized and processed according to the learning session to obtain a temporal feature sequence.

[0028] The time-series feature sequence is tracked and recorded, and an external memory module coupled with a long short-term memory network is introduced. The external memory module is used to store key events and corresponding teaching resource identifiers to establish a state evolution trajectory.

[0029] The forget gate, input gate, and output gate of the Long Short-Term Memory Network are used to filter, update, and control the output of information in the state evolution trajectory. The activation of the input gate is affected by key events in the external memory module to obtain state information.

[0030] By detecting gradient abrupt change points in the state information and combining them with the timestamps of events in the external memory module, the transition nodes of the learning state are obtained.

[0031] The state information of the transformation node is input into two fully connected evaluation networks with shared underlying parameters. The first fully connected evaluation network outputs a matching score between teaching resources and learning state as the teaching matching degree, and the second fully connected evaluation network outputs a learning efficiency evaluation index as the efficiency evaluation result.

[0032] Optionally, the step of tracking and recording the temporal feature sequence and introducing an external memory module coupled with a long short-term memory network, wherein the external memory module is used to store key events and corresponding teaching resource identifiers to establish a state evolution trajectory, includes:

[0033] The time-series feature sequence is analyzed using a sliding window to detect windows where feature fluctuations exceed a threshold, and these windows are marked as critical events.

[0034] For each key event, the corresponding features are extracted and compressed into an event feature vector. At the same time, the teaching resource identifier pushed within the time period of the window is obtained. The event feature vector is bound to the teaching resource identifier and a timestamp is added before being stored in the external memory module as a memory unit.

[0035] In each step of the Long Short-Term Memory network processing the temporal feature sequence, the current hidden state is used as a query vector, and the most relevant teaching resource identifier is retrieved from the external memory module through an attention mechanism to generate a context vector;

[0036] The context vector is fused with the current output of the Long Short-Term Memory network to obtain enhanced state information;

[0037] The enhanced state information of all steps is collected in chronological order, and key nodes are marked according to the timestamps in the external memory module to form the state evolution trajectory.

[0038] Optionally, the step of encoding the feature point data to obtain feature-encoded data includes:

[0039] The feature point data is categorized based on facial motion coding and human skeletal key points to obtain an expression feature set and a motion feature set.

[0040] Based on the set of facial expression features and the set of action features, facial expression units and action units are obtained;

[0041] The facial expression unit and the action unit are fused based on a cross-modal attention mechanism to generate a unified feature vector after fusion;

[0042] A self-supervised vector quantization encoder is used to map the unified feature vector into a discrete encoding index sequence, and the embedding vector corresponding to the encoding index sequence is generated to obtain the feature encoding data.

[0043] Optionally, optimizing the teaching resource recommendation strategy based on the teaching matching degree and the efficiency evaluation result includes:

[0044] A set of adjustment parameters is obtained by comprehensively analyzing the teaching matching degree and the efficiency evaluation results.

[0045] Based on the set of adjustment parameters, the actor-critic framework in reinforcement learning is used to optimize the strategy and generate resource library configuration information.

[0046] The resource library configuration information is associated with the student's learning status to update the teaching resource delivery strategy.

[0047] Secondly, this application provides an AI-based intelligent teaching system, including:

[0048] The data acquisition module is used to collect students' learning trajectory data, teaching feedback data, and image data during the teaching process of the AI ​​teaching platform using the teaching resource push strategy.

[0049] The encoding module is used to construct a dataset based on the learning trajectory data and the teaching feedback data, convert the image data into feature point data, and encode the feature point data to obtain feature-encoded data.

[0050] The extraction module is used to perform collaborative analysis of the dataset and the feature encoding data using the K-nearest neighbor algorithm to obtain sample clusters corresponding to different learning states, and to extract typical features from the data corresponding to the sample clusters.

[0051] The analysis module is used to perform time-series analysis on the typical features using a long short-term memory network to obtain teaching matching degree and efficiency evaluation results;

[0052] An optimization module is used to optimize the teaching resource recommendation strategy based on the teaching matching degree and the efficiency evaluation results.

[0053] Thirdly, this application provides an electronic device, comprising:

[0054] Memory, used to store computer programs;

[0055] A processor for executing the computer program to implement the steps of the AI-based intelligent teaching method as described in the first aspect above.

[0056] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the AI-based intelligent teaching method described in the first aspect above.

[0057] The AI-based intelligent teaching method provided in this application has the following effects:

[0058] This application first collects students' learning trajectories, feedback, and image data during teaching, providing a multi-dimensional and mutually corroborating data foundation for subsequent analysis, thus avoiding the one-sidedness of a single data source. Then, the image data is converted into coded feature data, enabling the fusion of unstructured visual information with structured learning data, creating conditions for comprehensive analysis. Next, the K-nearest neighbor algorithm is used for collaborative analysis of the fused data, comprehensively considering learning behavior and real-time status, thereby more accurately segmenting student groups reflecting different learning states and extracting key features representing their states. Then, a long short-term memory network is used to perform temporal analysis on these features, capturing the dynamic evolution of learning states, thus achieving continuous and accurate evaluation of teaching matching and learning efficiency. Finally, based on the evaluation results, the resource delivery strategy is optimized, allowing the allocation of teaching resources to dynamically adapt to students' actual learning progress and status changes, improving the accuracy and adaptability of teaching resource delivery, and ultimately enhancing the effectiveness of personalized teaching.

[0059] Furthermore, this application ensures the consistency and comparability of different modalities in measurement by using a process of fusion of multi-source data based on timestamp alignment and introducing image features as spatial anchors to calculate joint distance. Subsequently, by combining density peak detection and K-nearest neighbor classification, and using a lightweight evaluation network to dynamically adjust the clustering process, the robustness and adaptability of partitioning complex and non-uniform learning state data can be enhanced. On this basis, the core region is located by analyzing intra-cluster density and silhouette coefficient, and features are extracted by using principal component analysis and canonical correlation analysis. Finally, typical features that are more representative of the essence of different learning states and have high discriminative power are extracted, laying a more reliable foundation for subsequent time series analysis and accurate evaluation.

[0060] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 is a flowchart illustrating an AI-based intelligent teaching method provided in an embodiment of this application;

[0063] Figure 2 is a schematic diagram illustrating a specific implementation of an AI-based intelligent teaching method provided in this application embodiment;

[0064] Figure 3 is a schematic diagram of the structure of an AI-based intelligent teaching system provided in an embodiment of this application;

[0065] Figure 4 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0066] In online smart teaching practices, existing methods typically rely on single-dimensional data such as answer accuracy or learning time, making it difficult to effectively correlate and comprehensively analyze discrete learning behaviors, subjective feedback, and real-time captured visual states. Specifically, it is difficult to accurately determine whether a student is "highly focused," "confused and distracted," or "superficially browsing," because a single data source cannot fully reflect their learning engagement and true comprehension. At the same time, these methods lack the ability to depict the continuous evolution of students' learning states over time, resulting in a static and lagging assessment of the match between teaching resources and current learning needs. Therefore, existing technologies are significantly insufficient in terms of the comprehensiveness and dynamism of their assessments when dealing with the multidimensionality, continuity, and implicitness of the learning process.

[0067] In light of this, this application proposes an AI-based intelligent teaching method. The core idea of ​​this method is as follows: First, during the teaching process, students' learning trajectories, feedback questionnaires, and image data are collected simultaneously to construct a multi-dimensional data foundation characterizing the learning process. Next, by encoding the image data and fusing it with learning behavior data, and using the K-nearest neighbor algorithm for collaborative analysis, student groups in different learning states are segmented based on the comprehensive similarity of behavior and state, and their key features are extracted. Then, long short-term memory networks are used to perform temporal analysis on these features to capture the dynamic evolution of learning states, thereby achieving continuous and accurate evaluation of teaching matching and learning efficiency. Finally, resource delivery strategies are dynamically adjusted based on this evaluation result. Therefore, this solution, by constructing a complete technical closed loop of "multi-source data fusion, collaborative state segmentation, dynamic temporal evaluation, and real-time strategy optimization," can transform the complex learning process into quantifiable and traceable analytical results, effectively solving the problem of insufficient multi-dimensional dynamic evaluation capabilities of existing technologies, and ultimately achieving accurate and adaptive matching of teaching resources with students' learning needs.

[0068] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0069] The core of this application is to provide an AI-based intelligent teaching method, the flowchart of one specific implementation of which is shown in Figure 1. The method includes:

[0070] S101. During the teaching process using the teaching resource push strategy on the AI ​​teaching platform, collect students' learning trajectory data, teaching feedback data, and image data on the AI ​​teaching platform.

[0071] Among them, learning trajectory data refers to the students' operation and interaction process on the platform, such as the changes in the progress bar of video viewing, the number of times specific knowledge points are replayed, the download records of different types of learning materials, and the distribution of the time spent in each learning stage. This data objectively reflects the students' learning path and behavior patterns. Teaching feedback data refers to the questionnaire scores filled out after the learning unit, which directly expresses the students' feelings about the teaching content, difficulty, and their own level of understanding. Image data refers to the data acquired in real time through the image acquisition module integrated on the student terminal. It continuously captures the students' facial expressions and body postures during learning, thereby transforming the real-time state of learning into visual sequence information.

[0072] In step S101, when students receive and learn from the teaching resources pushed to them on the platform, the AI ​​teaching platform automatically records all the interaction logs generated by them, such as clicks, playback, and downloads, to form learning trajectory data. At the same time, after the key learning nodes or units are completed, electronic questionnaires are pushed to students and scores are collected to form teaching feedback data. Meanwhile, the terminal image acquisition module continues to work in a non-intrusive manner in the foreground, using the recorded video stream as the raw image data. Moreover, the acquisition process of these three types of data is carried out synchronously in time and is uniformly attached with a precise timestamp, so as to ensure that subsequent correlation analysis of students' objective behavior, subjective feelings and real-time status can be performed based on the same time period.

[0073] S102. Based on the learning trajectory data and the teaching feedback data, construct a dataset, convert the image data into feature point data, encode the feature point data, and obtain feature-encoded data.

[0074] In one specific implementation, step S102 includes:

[0075] Step 1021: Perform classification processing on the feature point data based on facial action coding and human skeletal key points to obtain expression feature set and action feature set.

[0076] Among them, the expression feature set refers to a series of feature point data extracted from facial images to describe specific facial muscle movements. These feature points collectively define expressions such as "smiling" and "confused". The action feature set refers to a series of feature point data extracted from human images to describe the spatial position and angle of the body's main joints. These feature points collectively define postures such as "leaning forward" and "leaning back".

[0077] In step 1021, the facial motion encoding unit pre-integrated in the smart teaching system is invoked. This unit is a model trained on a large amount of labeled data and can automatically recognize the fine movements of facial muscles. At the same time, the human skeleton key point detection model, which is also pre-integrated and based on deep learning, is invoked. These two models perform parallel analysis on the original image data stream, thereby clearly separating the mixed image data into two structured data sets that describe facial expressions and body movements respectively.

[0078] Step 1022: Based on the set of facial expression features and the set of action features, obtain facial expression units and action units.

[0079] Among them, an expression unit refers to a semantic unit formed by combining multiple related expression feature points and capable of representing a complete expression pattern, while an action unit refers to a semantic unit formed by combining multiple related action feature points and capable of representing a coherent posture pattern.

[0080] In step 1022, feature point groups that appear continuously in time and change collaboratively in space are identified as an expression event and packaged into an expression unit; similarly, a similar analysis is performed on the action feature set to identify and package a feature point group representing a coherent posture change into an action unit.

[0081] Step 1023: The expression unit and the action unit are fused based on a cross-modal attention mechanism to generate a unified feature vector after fusion.

[0082] Among them, the cross-modal attention mechanism refers to a data fusion method that processes information from different sources. It can dynamically evaluate the importance of facial expression information and action information at a specific moment and assign different fusion weights accordingly.

[0083] In step 1023, the synchronously acquired facial expression unit sequence and action unit sequence are aligned, and then the fusion mechanism analyzes the facial expression unit and action unit at each moment. For example, when a "frowning" facial expression unit is detected, the system automatically focuses on whether it is accompanied by a "leaning back" action unit at the same moment and assigns higher weight to this potentially related information. Then, through this dynamic and focused calculation, the information from the two channels is integrated into a single, unified feature vector, which contains the joint information of facial expression and action.

[0084] Step 1024: Using a self-supervised vector quantization encoder, the unified feature vector is mapped to a discrete encoding index sequence, and the embedding vector corresponding to the encoding index sequence is generated to obtain the feature encoding data.

[0085] Among them, the self-supervised vector quantization encoder refers to a model that can learn without manual annotation. It converts continuous feature vectors into discrete numbers by matching them with a pre-set "codebook", and then uses these numbers to find the corresponding more refined representative vector in the "codebook".

[0086] In step 1024, a series of generated unified feature vectors are input into an encoder that has been pre-trained in a self-supervised manner. The encoder internally maintains a "codebook" that was built during the training phase. Then, for each input unified feature vector, the encoder searches for the most similar pattern in the "codebook" and outputs the number of the pattern, i.e., the encoding index. Then, according to the encoding index, the corresponding well-formatted and information-dense embedding vectors are extracted from the "codebook" and arranged in chronological order to form feature encoding data for subsequent complex analysis.

[0087] This application enables efficient and in-depth utilization of unstructured visual information by progressively processing, abstracting, fusing, and encoding the original image data from the physical pixel level into structured feature coding data. Subsequently, high-dimensional, redundant, and heterogeneous image information is refined into low-dimensional, compact, and computable feature coding data, enabling it to be processed collaboratively with learning behavior data within the same analytical framework. This lays a crucial data foundation for the subsequent accurate identification of the comprehensive learning state.

[0088] S103. The K-nearest neighbor algorithm is used to perform collaborative analysis on the dataset and the feature encoding data to obtain sample clusters corresponding to different learning states, and typical features are extracted from the data corresponding to the sample clusters.

[0089] In one specific implementation, as shown in FIG2, step S103 includes:

[0090] Step 1031: Align the learning trajectory data and the feature encoding data in the dataset based on timestamps to obtain comprehensive data.

[0091] Timestamp alignment refers to the process of matching data from different sources according to the precise time points of their generation or collection, ensuring that all data describing the state at the same moment can be correctly associated.

[0092] For example, during a specific learning period of an online math course, student A's video viewing progress and the accuracy rate of after-class exercises from 10:05 to 10:10 are paired with image features captured and encoded during the same period to generate a comprehensive data record. The same operation is performed on students B and C, ultimately generating three time-aligned comprehensive data records.

[0093] Step 1032: Using the K-nearest neighbor algorithm, the image feature encoding in the comprehensive data is used as the spatial anchor point to calculate the joint distance metric, and K neighboring data points are determined based on the joint distance metric to obtain the neighboring dataset.

[0094] Step 1032 may specifically include the following steps:

[0095] Step a1: Extract the image feature codes of all data points from the comprehensive data, use a clustering algorithm to identify a preset number of cluster centers, and define the cluster centers as spatial anchor points.

[0096] Among them, spatial anchor points refer to a number of representative center points summarized from all image feature codes by clustering algorithms, which are used to provide a stable reference benchmark for high-dimensional feature space; the preset number of cluster centers refers to a number pre-set based on experience or through automated methods such as the elbow rule, for example, 5, which is used to guide the algorithm to find the corresponding number of spatial anchor points.

[0097] For example, extract all image feature codes from a comprehensive data pool containing data from dozens of students; then, based on the preset number of cluster centers (3), use the K-Means algorithm for inductive analysis to obtain the coordinates of the three most representative center points, which are used as spatial anchor points and labeled as anchor point one, anchor point two, and anchor point three, respectively.

[0098] Step a2: Based on the spatial anchors, the learning behavior features of each data point are weighted and calculated according to the similarity between the corresponding image feature encoding and each spatial anchor, so as to project the learning behavior features into the anchor space composed of the spatial anchors, and obtain the projected behavior feature vector.

[0099] For example, the similarity between image feature A and anchor point 1 is calculated to be 0.95, the similarity with anchor point 2 is 0.05, and the similarity with anchor point 3 is 0.00. These similarities are then normalized to obtain three weights: 0.95 / (0.95+0.05+0.00)=0.95, 0.05 / (0.95+0.05+0.00)=0.05, 0.00. These weights are then used to weight and fuse behavioral feature vectors such as "answering speed" and "replay count" for image A. The specific calculation is: 0.95×[behavioral feature vector A]+0.05×[behavioral feature vector B]+0.00×[behavioral feature vector C], thus obtaining the projected behavioral feature vector V. 甲 Similarly, calculate the projected behavioral feature vectors of students B and C, i.e.: V 乙 and V 丙 .

[0100] Step a3: For any two data points in the comprehensive data, calculate the first distance between the image feature codes of the two data points and the second distance between the projected behavioral feature vectors.

[0101] For example, to calculate the feature differences between student A and student B, we first calculate the Euclidean distance between the feature codes of their original images, obtaining a first distance D1 = 2.0; then we calculate the projected behavioral feature vectors V. 甲 With V 乙 The Euclidean distance is used to obtain the second distance D2 = 1.5.

[0102] Step a4: Introduce a weight coefficient dynamically generated based on the overall correlation of the image feature encoding and the projected behavioral feature vector. Use the weight coefficient to weight and fuse the first distance and the second distance to calculate the joint distance metric between data points and form the target matrix.

[0103] The weighting coefficient is a dynamic parameter used to balance the relative importance of image feature distance and behavioral feature distance when calculating the joint distance. This coefficient is calculated based on the overall correlation of the two types of features in the entire data. The joint distance metric is a composite distance index that integrates the differences in image features and the differences in learning behavioral features, and is used to comprehensively measure the overall similarity between two data points.

[0104] For example, we analyze the global features of all current student data and assume that the overall correlation coefficient between the two sets of image features and behavioral features is 0.8. Then, we use a preset linear transformation rule to map the correlation coefficient to a weight coefficient, specifically: λ = correlation coefficient × 0.9, that is: 0.8 × 0.9 = 0.72. When rounding to one decimal place, we take λ = 0.7. Then, using this coefficient, we calculate the joint distance between A and B: D3 = (1 - 0.7) × 2.0 + 0.7 × 1.5 = 1.65. Using the same method, we calculate that the joint distance between A and C is 5.2, and the joint distance between B and C is 4.8.

[0105] Step a5: Based on the target matrix, select the K nearest neighbor data points with the smallest joint distance metric for each data point to form an initial neighbor list.

[0106] For example, based on all joint distance metrics, find the nearest neighbor for each data point and set the parameter K=2. For student A, the distances to B and C are 1.65 and 5.2 respectively. Assuming that the distance between A and D is 6.3, the two points corresponding to the smallest distances are B and C. Therefore, student A's initial neighbor list includes B and C.

[0107] Step a6: Combine the timestamp information of each data point in the comprehensive data to verify the initial neighbor list and obtain the neighbor relationship set.

[0108] For example, reading the timestamps of students A, B, and C, we get A's time as 10:08, B's time as 10:10, and C's time as 10:30. Then, we calculate the timestamp differences: |10:08 - 10:10| = 2 minutes, |10:08 - 10:30| = 22 minutes. Next, we set a reasonable time window of 15 minutes. Since the 2-minute difference between A and B is within the window, the relationship is valid; the 22-minute difference between A and C exceeds the window, so this proximity relationship is discarded. Finally, the proximity relationship set verified for student A only includes student B.

[0109] Step a7: Construct a structured record for each data point in the neighborhood relation set. The structured record should contain at least the unique identifiers of all its neighboring data points, the corresponding joint distance metric, and the timestamp difference, thus obtaining the neighborhood dataset.

[0110] Among them, structured record refers to a data archive established for each data point, which is used to systematically store key information such as its relationship with neighboring points, similarity and time difference; timestamp difference refers to the absolute time difference between the timestamps corresponding to two data points, which is used to evaluate the proximity of data points in the time dimension.

[0111] For example, generate a structured record for student A, which includes: a unique identifier for itself, unique identifiers for its neighbors, a joint distance metric of 1.65, and a timestamp difference of 2 minutes; and after performing this operation on all students, construct a structured neighborhood dataset.

[0112] Step 1033: Based on the feature distribution of the neighbor dataset, the comprehensive data is divided using a mechanism combining density peak detection and K-nearest neighbor classification to obtain sample clusters corresponding to different learning states; wherein, the sample clusters and their feature distributions are used as feedback signals and input to a lightweight evaluation network to evaluate the clustering quality online and dynamically adjust the sensitivity parameters of density peak detection.

[0113] The lightweight evaluation network refers to a small neural network used to evaluate the consistency between clustering results and teaching logic online. Its input is the "sample clusters" generated by the current clustering and their "feature distribution," specifically including the centroid features of each cluster, the dispersion of samples within a cluster, and the distance between different cluster centroids. Its output is a "clustering quality score." This network is pre-trained offline using simulation data or partially labeled data. Its training objective is to make the score reflect the rationality of the clustering results in business context, such as: whether the divided clusters correspond to interpretable learning states such as "efficient," "confused," or "inefficient"; whether there are significant differences in the mean features between different clusters; and whether the features of samples within the same cluster are relatively consistent.

[0114] When applied in the teaching scenario of this application, the network calculates scores based on real-time clustering results. If the scores are too low, such as when a cluster contains contradictory feature samples of "answering questions quickly" and "frequent frowning", the lightweight evaluation network generates a feedback signal to increase the "sensitivity parameter" of density peak detection. After the parameter is increased, the density peak detection algorithm has stricter requirements for becoming a cluster center, and may cause the original large cluster to split into multiple purer clusters, so that the clustering results are more in line with the subdivision of the actual learning state, such as further distinguishing "confusion" into "confusion about knowledge points" and "distraction of attention", thereby realizing adaptive optimization of the clustering process.

[0115] Density peak detection refers to a technique that identifies cluster centers based on the local density of data points and their distance from higher density points; K-nearest neighbor classification refers to a classification method that determines the category of a data point based on the categories of its K nearest neighbors; the sensitivity parameter is a threshold parameter used in the density peak detection algorithm to determine the local density and distance of points as cluster centers, and the adjustment of this parameter directly affects the number and boundaries of clusters; online evaluation refers to the real-time calculation and feedback of the quality of the current clustering results during the clustering process.

[0116] In step 1033, potential cluster centers are identified using density peak detection. Then, K-nearest neighbor classification is used to divide all data into different sample clusters based on these centers. At the same time, a pre-trained lightweight evaluation network is activated to receive the currently divided sample clusters and their feature distributions as input and output a clustering quality score in real time. This score is then fed back to the density peak detection module to dynamically adjust its sensitivity parameters, thereby achieving self-optimization during the division process and obtaining more stable and accurate sample clusters corresponding to different learning states.

[0117] For example, based on a neighboring dataset, the density peak detection algorithm is first used to calculate the local density ρ and the minimum distance δ to higher density points for each point. Assume that the calculated (ρ, δ) for student A is (8, 4), for student B is (7, 5), and for student C is (2, 10). Then, the ρ and δ of each data point are multiplied to obtain a decision value. The decision value for student A is 8×4=32, for student B is 7×5=35, and for student C is 2×10=20. Then, a core decision threshold is set, that is, the sensitivity parameter is 0.5. This threshold is used to compare the normalized decision value with the preset standard. Its initial value of 0.5 is an empirical intermediate value.

[0118] Then, according to the rule that "the decision value is greater than the threshold after normalization", A and B are identified as potential cluster centers; K-nearest neighbor classification then divides the data points into clusters represented by these centers. A and B are assigned to the same sample cluster and labeled as "efficient learning", while C is assigned to another sample cluster and labeled as "confused learning".

[0119] Simultaneously, the lightweight evaluation network is triggered to receive the data distribution of the current two cluster centers and output cluster quality scores: the score for the "efficient learning" cluster is 0.92, and the score for the "confused learning" cluster is 0.65. Since the latter score is low, the lightweight evaluation network generates a feedback signal and raises the core decision threshold in density peak detection from 0.5 to 0.6, making the clustering criteria more stringent, thereby obtaining clearer cluster boundaries in subsequent iterations.

[0120] Step 1034: For each sample cluster, the core data region of each sample cluster is determined by calculating the density distribution and contour coefficient of the data points in the high-dimensional feature space.

[0121] Among them, the density distribution in the high-dimensional feature space refers to the density of the distribution of data points in the multidimensional space composed of all features; the silhouette coefficient is a comprehensive index used to measure the cohesion of a data point with its own cluster and the separation from other clusters, and its value range is between -1 and 1.

[0122] In step 1034, the local density and contour coefficient of each data point in the feature space within each sample cluster are calculated. Based on the local density and contour coefficient, the core data region representing the most typical and purest data of the learning state can be accurately located by comprehensively analyzing high-density regions and high contour coefficient regions.

[0123] For example, the "efficient learning" sample cluster is internally analyzed, and the silhouette coefficient of all points within the cluster is calculated. The silhouette coefficient is an indicator that measures the cohesion of a data point with its own cluster and its separation from other clusters, and its value ranges from -1 to 1. Assuming that a silhouette coefficient value greater than 0.7 is a good standard for that point to be well classified into the current cluster, the silhouette coefficient of student A is calculated to be 0.75, and that of student B is 0.72. The average silhouette coefficient is (0.75+0.72) / 2=0.735. Since 0.735 is higher than the good threshold of 0.7, it indicates that the cluster has good cohesion. At the same time, the kernel density estimation shows that the feature region density of A and B is significantly higher than that of the surrounding area. Therefore, A and B together constitute the core data region of the cluster.

[0124] Step 1035: Based on the core data region, use a combination of principal component analysis and canonical correlation analysis to extract features from the data corresponding to the sample cluster to obtain typical features.

[0125] Principal component analysis is a statistical method that uses orthogonal transformation to convert potentially correlated original features into a set of linearly uncorrelated principal components, used to extract the most significant direction of change in the data; canonical correlation analysis is a multivariate statistical method used to analyze the correlation between two sets of feature vectors, aiming to find the component pairs with the strongest correlation between the two sets of features.

[0126] In step 1035, principal component analysis is first used to extract the main direction of data variation within the region and dimensionality reduction is used to retain key information. At the same time, canonical correlation analysis is used to deeply explore the deep correlation patterns between the subset of learning behavior features and the subset of image state features. The results of these two analyses are then organically combined and screened to extract a set of typical features that can most essentially and effectively distinguish different learning states.

[0127] For example, focusing on the aforementioned core data area, principal component analysis was first performed, revealing that the first principal component contributed 80% of the variance, mainly composed of "number of answers per unit time" and "inverse variance of face orientation towards the screen," with a weight of 0.6 for the number of answers per unit time and a weight of 0.4 for the inverse variance of face orientation towards the screen. Next, canonical correlation analysis was conducted, revealing a strong canonical correlation between this set of learning behavior characteristics and image state characteristics such as "frequency of smiling expressions" and "proportion of forward-leaning postures," with a hypothetical correlation coefficient of 0.85. Finally, this information was synthesized, and "number of answers per unit time," "stability of face orientation," and "coordination index between behavior and image" were extracted as typical characteristics of the "efficient learning" state.

[0128] This application constructs a dynamic joint distance metric anchored by image semantics and combines it with temporal consistency verification to achieve robust evaluation of similarity among multimodal learning data. Furthermore, it improves the adaptability and accuracy of complex learning state classification through a clustering method that integrates density peak detection and online feedback optimization mechanisms. Finally, by focusing on the core structure within clusters and integrating multivariate statistical analysis methods, it extracts typical features that can profoundly reveal the essential differences between different learning states, thus laying a reliable data foundation for subsequent temporal evolution analysis and precise teaching evaluation.

[0129] S104. Use a long short-term memory network to perform time-series analysis on the typical features to obtain the teaching matching degree and efficiency evaluation results.

[0130] In one specific implementation, step S104 includes:

[0131] Step 1041: Based on the sequence processing capability of the Long Short-Term Memory Network, the typical features are organized and processed according to the learning session to obtain the temporal feature sequence.

[0132] Among them, a learning session refers to a continuous learning activity unit with a clear start and end time for a student on an AI teaching platform, such as completing a chapter or participating in a live course; a temporal feature sequence refers to a data sequence with time dependency formed by strictly arranging the typical features belonging to the same learning session in the order of their generation.

[0133] In step 1041, all typical features generated during a complete learning session are sorted and organized according to timestamps. For example, a set of typical features extracted every minute in a 45-minute course are arranged in chronological order to form a temporal feature sequence containing 45 time steps, which serves as the input for subsequent deep temporal analysis.

[0134] For example, in a 30-minute learning session on "trigonometric functions", the typical features extracted every minute, such as "focus" and "trend value of answer accuracy", are arranged in chronological order to form a time-series feature sequence of 30 time points.

[0135] Step 1042: Track and record the time-series feature sequence, and introduce an external memory module coupled with a long short-term memory network. The external memory module is used to store key events and corresponding teaching resource identifiers to establish a state evolution trajectory.

[0136] The external memory module refers to an auxiliary storage unit independent of the main recurrent structure of the long short-term memory network. It is also a dynamically readable and writable storage array, specifically used to save key events and their contextual information identified from the learning process in chronological order. Key events refer to specific time periods in the temporal feature sequence that represent significant changes in the learning state. The feature fluctuation amplitude within these time periods exceeds a threshold derived from historical data statistics, such as a sudden and significant drop in focus or repeated playback of a certain knowledge point.

[0137] Step 1042 may specifically include the following steps:

[0138] Step b1: Analyze the time-series feature sequence using a sliding window to detect windows where feature fluctuations exceed a threshold, and mark the windows as critical events.

[0139] In step b1, a sliding window of fixed time length, such as 5 minutes, is used to traverse the time-series feature sequence and calculate the fluctuation amplitude of the features within the window in real time. When the fluctuation amplitude exceeds a preset threshold, the time window is marked as a critical event. The threshold is obtained by statistically analyzing the feature fluctuations of a large number of normal learning processes. In this embodiment, the size of the threshold is not specifically limited.

[0140] For example, a 5-minute sliding window is used to scan the sequence, and the standard deviation of the “attention” feature value suddenly rises from an average of 0.1 to 0.5 within the window from the 12th to the 17th minute, exceeding the preset fluctuation threshold of 0.3. Therefore, this window is marked as a critical event E1.

[0141] Step b2: For each key event, extract the corresponding features and compress and encode the features into an event feature vector. At the same time, obtain the teaching resource identifier pushed within the time period of the window, bind the event feature vector with the teaching resource identifier, attach a timestamp, and store it in the external memory module as a memory unit.

[0142] Among them, the teaching resource identifier refers to the unique identifier of a specific teaching resource on the platform, such as the number or code of a certain instructional video or a set of practice questions.

[0143] For example, the mean value of features within the E1 window is extracted and encoded, and it is found that “Auxiliary Explanation Video V123” was pushed to students during that time period; then the encoded event vector, the “V123” identifier and the 14.5-minute timestamp are packaged and stored in the external memory module.

[0144] Step b3: In each step of processing the temporal feature sequence in the Long Short-Term Memory Network, the current hidden state is used as the query vector, and the most relevant teaching resource identifier is retrieved from the external memory module through the attention mechanism to generate a context vector.

[0145] In this context, the hidden state refers to a vector generated and maintained at each time step during the internal cyclic processing of the Long Short-Term Memory (LSTM) network. The hidden state condenses all the historical feature information processed by the network from the beginning of the sequence to the current moment, representing the LSM network's internal memory and summary of the observed learning state evolution. In the teaching scenario of this application, the hidden state essentially dynamically represents a comprehensive vector of the student's accumulated understanding of the currently learned knowledge points, attention level, and emotional tendency.

[0146] Attention mechanisms are information processing techniques that simulate selective human attention. By calculating the correlation between a "query" vector and a set of "key" vectors, information is aggregated from the corresponding "value" vectors in a targeted manner, thereby focusing on the most critical parts.

[0147] In step b3, when the Long Short-Term Memory Network processes each moment of the temporal feature sequence in sequence, the current hidden state is used as the query vector. Then, the similarity between the query vector and the key vectors of all memory units in the external memory is calculated using the attention mechanism to obtain a set of attention weights. The teaching resource identifiers and other information associated with the memory unit with the highest weight are retrieved and aggregated to generate a context vector. The key vector is usually converted from the event feature vector.

[0148] For example, when a Long Short-Term Memory network processes features up to the 20th minute, its current hidden state... It is used as a query vector; then the key vector corresponding to the E1 event in external memory is calculated. The similarity can be calculated using a dot product formula: Here, sim represents the similarity score. Assuming the calculated sim value is 8.5, since there is only one event E1 in the current external memory, its attention weight is directly determined to be 1.0. Based on this weight, the embedding vector corresponding to "auxiliary explanation video V123" in unit E1 is retrieved and used as the context vector. .

[0149] Step b4: Fuse the context vector with the current output of the Long Short-Term Memory network to obtain enhanced state information.

[0150] The context vector refers to the vector composed of historical key event information most relevant to the current processing moment, which is retrieved and aggregated from the external memory module through the attention mechanism.

[0151] For example, the original output vector of the Long Short-Term Memory network at the 20th minute. With the retrieved context vector The fusion is performed using a weighted summation method: ,in, This is the enhanced state information at the 20th minute. λ is a preset fusion coefficient, which is set to 0.5 here. The enhanced state representation can be calculated using the formula above.

[0152] Step b5: Gather the enhanced state information of all steps in chronological order, and mark key nodes according to the timestamps in the external memory module to form the state evolution trajectory.

[0153] Among them, the state evolution trajectory refers to an enhanced record that reflects the continuous and dynamic changes in the learning state, formed by fusing the real-time processing results of the temporal feature sequence with relevant historical event information retrieved from external memory.

[0154] For example, steps b3 and b4 are repeated for each time point in the time series to generate a series of enhanced state information [h1, h2, ..., h] arranged in chronological order. 30 Simultaneously, based on the timestamp "14" of the key event E1 recorded in the external memory module, the enhanced state information h corresponding to the 14th position... 14 Key nodes are marked on the graph to form the final state evolution trajectory.

[0155] Step 1043: Using the forget gate, input gate, and output gate of the Long Short-Term Memory Network, the information in the state evolution trajectory is filtered, updated, and output controlled. The activation of the input gate is affected by key events in the external memory module to obtain state information.

[0156] In this application, the Long Short-Term Memory Network is specifically configured to simulate and predict the continuous evolution of a student's learning state. Its input is a sequence of "typical features" organized in chronological order, and these features already contain fusion information of learning behavior and visual state. Subsequently, the Long Short-Term Memory Network learns to capture temporal patterns of state changes such as "from confusion to understanding" and "from focus to fatigue".

[0157] The forget gate, input gate, and output gate refer to the three core control units within the Long Short-Term Memory (LSTM) network. The forget gate is responsible for determining which past, irrelevant learning state information can be weakened, such as weakening the learning state information of previously resolved confusions. The input gate, based on the new features of the current input and the context provided by external memory, decides which important new state information to store in the long-term memory unit, such as storing newly generated confusion learning states based on a recently watched explanatory video. The output gate, based on the state of the current memory unit, outputs a "state information" that comprehensively reflects the student's current learning depth. This entire gating mechanism together ensures the network's dependence on long-term learning, such as the effective modeling of the impact of prior knowledge on current understanding.

[0158] In step 1043, the state evolution trajectory is input into the gating unit of the Long Short-Term Memory network for fine processing. Then, the information in the trajectory is analyzed through the forgetting gate, and it is decided which outdated and irrelevant state details are faded out. Then, new state information is received through the input gate. The degree of opening of the input gate is affected by the context vector. That is, when the current state is highly related to a certain historical key event, the input gate will more actively integrate the information corresponding to the event. Finally, the output gate integrates the current memory and input, and outputs the filtered and refined more representative state information.

[0159] For example, the state evolution trajectory is fed into the gating unit of the Long Short-Term Memory network for deep processing. When the enhanced information is processed around the 20th minute, the input gate is activated to enhance the information by incorporating contextual information, which enables the network to more effectively integrate the background of "having watched the detailed video" into memory, thereby outputting more accurate state information.

[0160] Step 1044: By detecting gradient abrupt change points in the state information and combining them with the timestamps of events in the external memory module, the transition nodes of the learning state are obtained.

[0161] Among them, the gradient mutation point refers to the position in the curve of state information changing over time where the rate of change suddenly increases or decreases significantly, indicating that the state has undergone a qualitative change or turning point.

[0162] In step 1044, the state information sequence after gating is analyzed and the gradient of the state information at each time point is calculated. When a sudden change in the gradient value is detected, the time point is marked as a potential learning state transition node. Then, the timestamps of these nodes are compared with the timestamps of key events in the external memory module. If they match or are close in time, the node is confirmed as a valid learning state transition node triggered by a specific teaching event.

[0163] For example, analyzing the gradient of all state information changes, it was found that between the 14th and 15th minutes, the gradient value of the state change jumped from 0.2 to 1.8, exceeding the preset mutation threshold of 1.0. This time point highly coincides with the timestamp "14" of the key event E1, and is therefore identified as a learning state transition node triggered by watching V123.

[0164] Step 1045: Input the state information of the transformation node into two fully connected evaluation networks with shared underlying parameters. The first fully connected evaluation network outputs the matching score between teaching resources and learning state as the teaching matching degree, and the second fully connected evaluation network outputs the learning efficiency evaluation index as the efficiency evaluation result.

[0165] Among them, the two fully connected evaluation networks refer to two parallel neural network evaluators that share the same underlying parameters. This means that the two fully connected evaluation networks first learn general deep feature representations related to learning quality from the input state information, such as "cognitive engagement" and "knowledge mastery trend".

[0166] In step 1045, the state information sequence containing the confirmed transition nodes is fed into two fully connected evaluation networks with shared underlying parameters. The intrinsic relationship between the state information and the pushed teaching resources is analyzed in the first fully connected evaluation network. Its input is the state information sequence containing the transition nodes, and its output is a scalar score. After being normalized by functions such as sigmoid, the score can reflect the effectiveness of a specific teaching resource in triggering or adapting to the learning state transition, i.e., the "teaching matching degree".

[0167] Meanwhile, the overall learning process reflected by the state information sequence is analyzed in the second fully connected evaluation network. Its input is also a state information sequence, and its output is a "learning efficiency evaluation index". This index comprehensively considers factors such as the proportion of effective learning time, the frequency and magnitude of positive state transitions, and can be used to quantitatively evaluate the overall rate and effect of knowledge acquisition and internalization during the learning session.

[0168] This application enhances the understanding of causal relationships in state changes by introducing an external memory module and an attention mechanism, enabling the Long Short-Term Memory Network to dynamically associate and utilize historical key teaching event information when analyzing the temporal evolution of learning states. Then, through a gating mechanism and gradient mutation point detection, it can accurately locate learning state transition nodes triggered by teaching interventions. Finally, through a shared underlying dual-path evaluation network, it can achieve synergistic and accurate evaluation of teaching resource matching degree and learning process efficiency, thereby providing a decision-making basis with both temporal depth and event correlation for optimizing teaching strategies.

[0169] S105. Optimize the teaching resource recommendation strategy based on the teaching matching degree and the efficiency evaluation results.

[0170] In one specific implementation, step S105 includes:

[0171] Step 1051: Perform a comprehensive analysis of the teaching matching degree and the efficiency evaluation results to obtain a set of adjustment parameters.

[0172] The set of adjustment parameters refers to a combination of quantitative instructions and rules used to specifically guide the modification of teaching resource delivery strategies. It is not a single value, but includes adjustment suggestions for multiple dimensions such as resource type, delivery timing, content difficulty, and presentation order.

[0173] In step 1051, the teaching matching score and the learning efficiency evaluation index are jointly analyzed, and a series of analysis rules and mapping relationships are preset. For example, when the matching degree is low but the efficiency is high, it may indicate that the resource type is not suitable but the student has strong self-learning ability, and the adjustment direction is to replace the resource type; when the matching degree is high but the efficiency is low, it may indicate that the resource is suitable but the student is confused, and the adjustment direction is to insert auxiliary explanatory resources or adjust the push rhythm. Then, by traversing these preset rules and matching and calculating with the current evaluation results, a specific and executable set of adjustment parameters is output, such as "after knowledge point K1, switch the resource type from video to interactive practice questions and extend the push interval by 20%".

[0174] Step 1052: Based on the set of adjustment parameters, the actor-critic framework in reinforcement learning is used to optimize the strategy and generate resource library configuration information.

[0175] In step 1052, in order to apply the actor-critic framework to the smart teaching scenario of this application, its core elements need to be mapped to the specific data generated in the aforementioned steps. That is, the input "state" is defined as the student's current learning state information obtained from the aforementioned time-series analysis steps. This information is the state information output after processing by the Long Short-Term Memory Network. The output "action" is defined as a set of specific resource push operations that the AI ​​teaching platform can execute. Its content corresponds to pushing a resource represented by a specific teaching resource identifier to the student, or executing control commands such as "pause push". The "reward" for instantly evaluating the teaching effect produced after executing a certain "action" in a specific "state" is defined as a scalar value calculated by comprehensively calculating the optimization goal implied by the obtained set of adjustment parameters and the teaching matching degree and efficiency evaluation results.

[0176] Then, in this framework, the "actor" is a policy network whose function is to learn a mapping from "state" to "action" probability distribution, that is, to learn which teaching resources should be prioritized in which learning state. The "critic" is a value network whose function is to evaluate the expected cumulative "reward" that can be obtained by performing a certain "action" in a specific "state". Subsequently, the two networks work together: the "critic" evaluates the "actor's" action selection and provides gradient information; the "actor" adjusts its decision-making strategy according to the evaluation to select actions that can obtain higher evaluations. Through a large number of interactive iterations with simulated or historical teaching environments, the framework can eventually learn an optimized policy model. This model can output the optimal resource push action probability distribution for any input student learning state. The set of these actions and their triggering conditions are solidified into new "resource pool configuration information".

[0177] Step 1053: Associate the resource library configuration information with the student's learning status to update the teaching resource push strategy.

[0178] In step 1053, the latest resource library configuration information is loaded into the strategy engine. The core of this engine is a real-time matcher. When it detects that a student has entered a specific learning state, it immediately queries the new resource library configuration information and finds the resource or resource sequence that best matches the state. At the same time, it dynamically determines the timing and method of push according to the rules in the configuration information. Ultimately, this new logic, which integrates optimized resource configuration and dynamic state matching rules, will completely replace the original, fixed or simple push rules, thereby realizing the overall update and upgrade of the teaching resource push strategy.

[0179] This application achieves a fundamental shift in push strategies from static and general to dynamic and personalized by transforming evaluation results into specific adjustment instructions and utilizing a reinforcement learning framework to autonomously explore the optimal push strategy through simulation and iteration. This enables the push of teaching resources to accurately adapt to students' complex and changing learning states, forming an adaptive closed loop of "evaluation, optimization, execution, and re-evaluation," thereby continuously improving the accuracy of teaching interventions and overall learning efficiency.

[0180] Figure 3 is a schematic diagram of a specific implementation of an AI-based intelligent teaching system provided in this application. Referring to Figure 3, the system may include:

[0181] The data acquisition module 31 is used to collect students' learning trajectory data, teaching feedback data, and image data during the teaching process of the AI ​​teaching platform using the teaching resource push strategy.

[0182] The encoding module 32 is used to construct a dataset based on the learning trajectory data and the teaching feedback data, convert the image data into feature point data, and encode the feature point data to obtain feature encoded data.

[0183] Extraction module 33 is used to perform collaborative analysis on the dataset and the feature encoding data using the K-nearest neighbor algorithm to obtain sample clusters corresponding to different learning states, and extract typical features from the data corresponding to the sample clusters.

[0184] Analysis module 34 is used to perform time-series analysis on the typical features using a long short-term memory network to obtain teaching matching degree and efficiency evaluation results.

[0185] The optimization module 35 is used to optimize the teaching resource push strategy based on the teaching matching degree and the efficiency evaluation result.

[0186] The AI-based intelligent teaching system of this application is used to implement the aforementioned AI-based intelligent teaching method. Therefore, the specific implementation of the AI-based intelligent teaching system can be found in the embodiment section of the AI-based intelligent teaching method above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0187] As shown in Figure 4, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the AI-based intelligent teaching methods described above.

[0188] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the AI-based intelligent teaching methods described above.

[0189] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0190] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the AI-based intelligent teaching method.

[0191] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0192] The above provides a detailed description of an AI-based intelligent teaching method, system, device, and medium provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A smart teaching method based on AI intelligence, characterized in that, include: During the teaching process using a teaching resource recommendation strategy on an AI teaching platform, learning trajectory data, teaching feedback data, and image data of students learning on the platform are collected. Based on the learning trajectory data and the teaching feedback data, a dataset is constructed. The image data is converted into feature point data, and the feature point data is encoded to obtain feature-encoded data. The K-nearest neighbor algorithm is used to perform collaborative analysis on the dataset and the feature-encoded data to obtain sample clusters corresponding to different learning states, and typical features are extracted from the data corresponding to the sample clusters. A long short-term memory network is used to perform time-series analysis on the typical features to obtain teaching matching degree and efficiency evaluation results. The teaching resource recommendation strategy is optimized based on the teaching matching degree and the efficiency evaluation results. The process involves using the K-nearest neighbor algorithm to collaboratively analyze the dataset and the feature-encoded data to obtain sample clusters corresponding to different learning states, and extracting typical features from the data corresponding to the sample clusters. This includes: aligning the learning trajectory data in the dataset with the feature-encoded data based on timestamps to obtain comprehensive data; using the K-nearest neighbor algorithm, calculating a joint distance metric using image feature encoding in the comprehensive data as spatial anchors, and determining K neighboring data points based on the joint distance metric to obtain a neighboring dataset; based on the feature distribution of the neighboring dataset, using a mechanism combining density peak detection and K-nearest neighbor classification to partition the comprehensive data, obtaining sample clusters corresponding to different learning states; wherein the sample clusters and their feature distributions serve as feedback signals, input to a lightweight evaluation network to evaluate the clustering quality online and dynamically adjust the sensitivity parameters of density peak detection; for each sample cluster, calculating the density distribution and silhouette coefficient of data points within the cluster in a high-dimensional feature space to determine the core data region of each sample cluster; and based on the core data region, using a method combining principal component analysis and canonical correlation analysis to extract features from the data corresponding to the sample clusters to obtain typical features.

2. The method according to claim 1, characterized in that, The method employs the K-nearest neighbor algorithm, using image feature codes in the comprehensive data as spatial anchors, to calculate a joint distance metric and determine K neighboring data points based on the joint distance metric to obtain a neighboring dataset. This includes: extracting image feature codes from all data points in the comprehensive data; identifying a predetermined number of cluster centers using a clustering algorithm, defining the cluster centers as spatial anchors; based on the spatial anchors, weighting the learning behavior features of each data point according to the similarity between the corresponding image feature codes and each spatial anchor, projecting the learning behavior features onto the anchor space formed by the spatial anchors to obtain a projected behavior feature vector; and for any two data points in the comprehensive data, calculating a first distance between the image feature codes of the two data points and the projected distance. The second distance between behavioral feature vectors is calculated. A weighting coefficient, dynamically generated based on the overall correlation of the image feature encoding and projected behavioral feature vectors, is introduced. The first distance and the second distance are weighted and fused using the weighting coefficient to calculate the joint distance metric between data points, forming a target matrix. Based on the target matrix, the K nearest neighbor data points with the smallest joint distance metric are selected for each data point to form an initial neighbor list. The initial neighbor list is verified by combining the timestamp information of each data point in the comprehensive data to obtain a neighbor relationship set. A structured record is constructed for each data point in the neighbor relationship set. The structured record contains at least the unique identifiers of all its neighbor data points, the corresponding joint distance metric, and the timestamp difference, resulting in a neighbor dataset.

3. The method according to claim 1, characterized in that, The step of using a Long Short-Term Memory (LSTM) network to perform temporal analysis on the typical features to obtain teaching matching degree and efficiency evaluation results includes: organizing and processing the typical features according to learning sessions based on the sequence processing capability of the LSTM network to obtain a temporal feature sequence; tracking and recording the temporal feature sequence and introducing an external memory module coupled to the LSTM network, which is used to store key events and corresponding teaching resource identifiers to establish a state evolution trajectory; using the forget gate, input gate, and output gate of the LSTM network to filter, update, and control the output of information in the state evolution trajectory, wherein the activation of the input gate is affected by key events in the external memory module to obtain state information; obtaining the learning state transition node by detecting gradient mutation points in the state information and combining them with the timestamps of events in the external memory module; and inputting the state information containing the transition node into two fully connected evaluation networks with shared underlying parameters. The first fully connected evaluation network outputs a matching degree score between teaching resources and learning states as the teaching matching degree, and the second fully connected evaluation network outputs a learning efficiency evaluation index as the efficiency evaluation result.

4. The method according to claim 3, characterized in that, The process of tracking and recording the temporal feature sequence and introducing an external memory module coupled with a long short-term memory network (LSTM) to store key events and corresponding teaching resource identifiers to establish a state evolution trajectory includes: analyzing the temporal feature sequence using a sliding window to detect windows where feature fluctuations exceed a threshold and marking these windows as key events; for each key event, extracting corresponding features and compressing and encoding these features into an event feature vector, while simultaneously obtaining the teaching resource identifiers pushed during the window's time period, binding the event feature vector with the teaching resource identifiers, and adding a timestamp before storing it as a memory unit in the external memory module; at each step of the LTM processing the temporal feature sequence, using the current hidden state as a query vector, retrieving the most relevant teaching resource identifier from the external memory module through an attention mechanism to generate a context vector; fusing the context vector with the current output of the LTM to obtain enhanced state information; and collecting the enhanced state information of all steps in chronological order and marking key nodes according to the timestamps in the external memory module to form the state evolution trajectory.

5. The method according to claim 1, characterized in that, The process of encoding the feature point data to obtain feature-encoded data includes: classifying the feature point data based on facial action coding and human skeletal key points to obtain an expression feature set and an action feature set; obtaining expression units and action units based on the expression feature set and the action feature set; fusing the expression units and the action units based on a cross-modal attention mechanism to generate a fused unified feature vector; and using a self-supervised vector quantization encoder to map the unified feature vector into a discrete encoding index sequence and generate an embedding vector corresponding to the encoding index sequence to obtain feature-encoded data.

6. The method according to claim 1, characterized in that, The step of optimizing the teaching resource delivery strategy based on the teaching matching degree and the efficiency evaluation result includes: comprehensively analyzing the teaching matching degree and the efficiency evaluation result to obtain a set of adjustment parameters; optimizing the strategy using the actor-critic framework in reinforcement learning based on the set of adjustment parameters to generate resource library configuration information; and associating the resource library configuration information with the student's learning status to update the teaching resource delivery strategy.

7. A smart teaching system based on AI, characterized in that, A smart teaching method based on AI as described in claim 1, comprising: a data acquisition module for collecting learning trajectory data, teaching feedback data, and image data of students learning on the AI ​​teaching platform during the teaching process using a teaching resource push strategy; an encoding module for constructing a dataset based on the learning trajectory data and the teaching feedback data, converting the image data into feature point data, and encoding the feature point data to obtain feature-encoded data; an extraction module for performing collaborative analysis of the dataset and the feature-encoded data using the K-nearest neighbor algorithm to obtain sample clusters corresponding to different learning states, and extracting typical features from the data corresponding to the sample clusters; an analysis module for performing time-series analysis of the typical features using a long short-term memory network to obtain teaching matching degree and efficiency evaluation results; and an optimization module for optimizing the teaching resource push strategy based on the teaching matching degree and the efficiency evaluation results.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the AI-based intelligent teaching method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the AI-based intelligent teaching method as described in any one of claims 1 to 6.

Images