Personalized teaching recommendation system and method based on multi-modal learning behavior analysis

Abstract

The invention discloses a personalized teaching recommendation system and method based on multi-modal learning behavior analysis, and relates to the technical field of learning behavior analysis and personalized recommendation systems, and the system comprises a cross-platform behavior representation alignment module, a behavior semantic structure modeling module, a recommendation target adaptive reconstruction module, and a pseudo tag generation and quality control module. The cross-platform behavior representation alignment module is used for performing kernel function mapping on a source platform sample vector and a target platform sample vector and calculating a distribution difference degree; the behavior semantic structure modeling module is used for performing similarity operation on source platform learning path embedding and target platform learning path embedding, and performing path regular alignment in combination with a distribution alignment result; and the recommendation target adaptive reconstruction module is used for performing nonlinear function processing on the difference value between the predicted answer accuracy and the concept understanding degree, and jointly calculating recommendation target indexes according to semantic structure results.

Description

Technical Field

[0001] This invention relates to the field of learning behavior analysis and personalized recommendation system technology, and in particular to a personalized teaching recommendation system and method based on multimodal learning behavior analysis. Background Technology

[0002] With the continuous development of educational informatization and intelligentization, personalized teaching recommendations have gradually become an important research direction for intelligent education systems. In recent years, with the help of multimodal learning analytics, researchers have been able to comprehensively utilize heterogeneous data from multiple sources, such as learners' click behavior, interaction logs, video viewing, answer records, and physiological and emotional feedback, to comprehensively model learners' learning behaviors and cognitive levels. Based on this, combined with deep learning, graph neural networks, and adaptive recommendation algorithms, intelligent education platforms can achieve recommendation services that better meet the differentiated needs of learners. Cross-platform sharing and migration of educational resources are also gradually becoming a trend. The data connection and model transfer capabilities between different educational platforms directly affect the effectiveness and application scope of personalized recommendation systems. Therefore, the combination of multimodal behavior analysis and cross-platform learning recommendations has become an emerging hot topic in the field of educational technology.

[0003] However, existing technologies still have significant shortcomings. Firstly, in cross-platform learning behavior modeling, current methods typically rely solely on single-modal data (such as answer accuracy or interaction duration) for feature alignment, ignoring the complex relationships between multimodal behavioral data. This makes it difficult to maintain the distribution consistency of the aligned representation space, thus reducing the model's transferability across target platforms. Secondly, in the behavioral semantic modeling stage, most studies only perform simple sequence matching or statistical modeling, lacking deep alignment of the semantic structure of the learning path. This results in the inability to fully capture and transfer the semantic features of learning behavior patterns, making it difficult to adapt to the differences between test-oriented and comprehension-oriented behaviors across different platforms. Thirdly, in terms of recommendation goal optimization, existing systems mostly use prediction accuracy as the primary objective function, lacking effective modeling and integration of learners' conceptual understanding. Consequently, recommendation results often tend to favor superficial score improvements rather than truly reflecting learners' knowledge mastery levels. Finally, regarding the use of pseudo-labels, existing methods generally employ fixed confidence thresholds, lacking dynamic quality control and screening mechanisms. This easily leads to the propagation of low-quality labels, further exacerbating the data distribution drift problem during the transfer process. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a personalized teaching recommendation system based on multimodal learning behavior analysis to address the problems of existing personalized teaching recommendation technologies, such as cross-platform behavior representation alignment discrepancies, low semantic structure modeling capabilities, insufficient rationality of recommendation goals, and the difficulty in achieving high-quality transfer and recommendation under insufficient pseudo-label supervision.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, this invention provides a personalized teaching recommendation system based on multimodal learning behavior analysis, including a cross-platform behavior representation alignment module, a behavior semantic structure modeling module, a recommendation target adaptive reconstruction module, and a pseudo-label generation and quality control module. The cross-platform behavior representation alignment module performs kernel function mapping on source platform sample vectors and target platform sample vectors to calculate distribution difference. The behavior semantic structure modeling module performs similarity calculations on source platform learning path embeddings and target platform learning path embeddings, and performs path regularization alignment based on the distribution alignment results. The recommendation target adaptive reconstruction module performs nonlinear function processing on the difference between predicted answer accuracy and concept comprehension, and jointly calculates the recommendation target index based on the semantic structure results. The pseudo-label generation and quality control module performs nonlinear function filtering and threshold control on pseudo-label confidence, and outputs the final transfer loss based on the recommendation target index.

[0007] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the step of mapping the source platform sample vector and the target platform sample vector using a kernel function includes: the system collecting multimodal behavior data of students on the source platform and constructing corresponding sample vectors; simultaneously collecting multimodal behavior data of students on the target platform and constructing another set of sample vectors; counting the number of samples on the source platform and the target platform respectively; and mapping them to a unified embedding space through a kernel function to obtain comparable representations.

[0008] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the calculation of distribution difference includes the system calculating the mean difference of the distribution difference between the source domain and the target domain samples in the embedding space, and making a judgment based on the calculation result. When the distribution difference is lower than the first threshold, it directly proceeds to the next step. When the distribution difference is between two thresholds, lightweight alignment optimization is performed. When the distribution difference exceeds the second threshold, the distribution difference is reduced by increasing the sensitivity of the kernel function or introducing adversarial training.

[0009] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the similarity calculation includes the system converting the behavior sequence of the source platform into a learning path embedding, converting the behavior sequence of the target platform into a learning path embedding, counting the total number of paths of the source platform and the target platform, calculating the similarity between paths pair by pair, and characterizing the semantic consistency of the learning behavior pattern.

[0010] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the path regularization alignment includes the system introducing a distribution alignment index as a regularization term in the path similarity calculation, generating a semantic level alignment index, and performing logical judgment based on the distribution alignment index. When the result is close to the maximum value, the recommendation model is directly transferred. When the result is in a medium range, path clustering and local alignment are performed. When the result is lower than a set threshold, a path reconstruction mechanism is triggered and a cold start mode is entered.

[0011] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the nonlinear function processing includes: during system execution, collecting the total number of recommendation samples, obtaining the predicted answer accuracy and corresponding concept comprehension of each sample, inputting the difference into a nonlinear function for amplification or compression, averaging the results of all samples, and calculating the recommendation target adaptability index in combination with semantic modeling indicators.

[0012] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the calculation of the recommendation target index includes the system making logical judgments when calculating the recommendation target adaptability index. When the recommendation target adaptability index is close to zero, the learning path remains unchanged. When the recommendation target adaptability index is in a small range, the system distinguishes between cases where the accuracy rate is higher than the comprehension rate or the comprehension rate is higher than the accuracy rate, and recommends concept-based resources or practice-based resources respectively. When the recommendation target adaptability index exceeds a set threshold, the system triggers principle explanation-based recommendations or practical training-based recommendations according to the direction of the difference.

[0013] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the nonlinear function screening and threshold control include: during system execution, collecting a set of pseudo-labels from the target platform, obtaining the confidence values ​​of the pseudo-labels, and using a nonlinear function to convert the confidence values ​​of the pseudo-labels into probability values ​​between 0 and 1, calculating the average confidence of the pseudo-labels and comparing it with a preset threshold, and combining the recommendation target index to generate a transfer loss.

[0014] As a preferred embodiment of the personalized teaching recommendation system based on multimodal learning behavior analysis described in this invention, the final output transfer loss includes the following: when the system compares the average confidence of pseudo-labels with a threshold, it performs a logical judgment: when the difference is less than the tolerance, it maintains the pseudo-label strategy unchanged; when the average confidence is higher than the threshold, it increases the use of pseudo-labels; when the average confidence is lower than the threshold, it reduces the use of pseudo-labels and triggers a manual annotation mechanism; when all pseudo-labels fail, it stops using pseudo-labels and only retains high-confidence samples.

[0015] Secondly, this invention provides a personalized teaching recommendation method based on multimodal learning behavior analysis, including: performing kernel function mapping on source platform sample vectors and target platform sample vectors to calculate distribution difference; performing similarity calculation on source platform learning path embeddings and target platform learning path embeddings, and performing path regularization alignment based on distribution alignment results; processing the difference between predicted answer accuracy and concept comprehension using a nonlinear function, and jointly calculating the recommendation target index based on semantic structure results; performing nonlinear function screening and threshold control on pseudo-label confidence, and outputting the final transfer loss based on the recommendation target index.

[0016] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the personalized teaching recommendation method based on multimodal learning behavior analysis as described in the first aspect of the present invention.

[0017] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the personalized teaching recommendation method based on multimodal learning behavior analysis as described in the first aspect of the present invention.

[0018] The beneficial effects of this invention are as follows: It realizes kernel function mapping and distribution difference calculation of multimodal behavioral data between source and target platforms, and triggers different alignment strategies by combining threshold judgment, thereby ensuring the comparability of heterogeneous data in a unified embedding space and effectively mitigating the distribution drift problem in cross-platform migration. Based on path embedding and similarity calculation, it introduces a distribution alignment index as a regularization term, which not only characterizes the semantic consistency of the learning path, but also links data distribution constraints. Through logical judgment, it flexibly triggers direct transfer, path clustering, or path reconstruction, enabling the model to have structural adaptability across different platforms. It processes the difference between answer accuracy and concept understanding through nonlinear functions and dynamically adjusts the recommendation target in conjunction with semantic indicators. It can output different types of learning resources in different contexts when accuracy and understanding are inconsistent, thereby improving the personalization and educational fit of recommendations. It dynamically filters pseudo-labels through nonlinear function transformation and threshold control, and generates transfer loss in combination with recommendation target indicators, ensuring the effectiveness and robustness of pseudo-labels in semi-supervised learning, thereby improving the stability and generalization ability of the overall system. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is an overall module diagram of a personalized teaching recommendation system based on multimodal learning behavior analysis.

[0021] Figure 2 This is a flowchart of a personalized teaching recommendation method based on multimodal learning behavior analysis. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] 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 phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figure 1 This embodiment provides a personalized teaching recommendation system based on multimodal learning behavior analysis, including: The module includes a cross-platform behavior representation alignment module, a behavior semantic structure modeling module, a recommendation target adaptive reconstruction module, and a pseudo-label generation and quality control module.

[0026] The cross-platform behavior representation alignment module is used to perform kernel function mapping on the source platform sample vector and the target platform sample vector and calculate the distribution difference.

[0027] It should also be noted that the kernel function mapping of the source platform sample vector and the target platform sample vector includes the system collecting multimodal behavior data of students on the source platform and constructing corresponding sample vectors, while simultaneously collecting multimodal behavior data of students on the target platform and constructing another set of sample vectors, counting the number of samples on the source platform and the target platform respectively, and mapping them to a unified embedding space through a kernel function to obtain comparable representations.

[0028] It should also be noted that the calculation of distribution difference includes the system calculating the mean difference of the distribution difference between the source domain and the target domain samples in the embedding space, and making a judgment based on the calculation result. When the distribution difference is lower than the first threshold, it directly proceeds to the next step. When the distribution difference is between the two thresholds, lightweight alignment optimization is performed. When the distribution difference exceeds the second threshold, the distribution difference is reduced by increasing the sensitivity of the kernel function or introducing adversarial training.

[0029] It should also be noted that a preferred approach to calculating the distribution dissimilarity specifically includes calculating the cross-platform behavioral representation alignment index. , represented as: ; in, This represents the total number of samples from the source platform, used to normalize source domain samples. This represents the total number of samples on the target platform, used to normalize samples in the target domain. Indicates the first in the source platform Index of a student sample Indicates the first in the target platform Index of a student sample Indicates the source platform number The multimodal behavior vectors of each student, including clicks and viewing duration. Indicates the target platform number The multimodal behavior vectors of each student include interaction logs and answering habits. This represents the Gaussian kernel scale parameter, which determines the distribution sensitivity between different samples.

[0030] when This indicates that the behavioral distributions of the source platform and the target platform are highly similar, and the behavioral characteristics of students cannot be distinguished under a unified embedding space. This suggests that no additional complex distribution alignment training is required, and semantic modeling can be directly entered, reducing redundant computation.

[0031] when When the value is in the middle range, it indicates that there is a certain difference in the behavioral distribution between the source platform and the target platform, but the difference is within a controllable range. If the system detects... , A moderate threshold (e.g., 0.5) indicates that lightweight distribution alignment optimization should continue, such as mini-batch stochastic gradient descent (SGD) adjustments, or the use of a small kernel function regularization to ensure that subsequent semantic modeling is not amplified by distribution differences.

[0032] when A larger value indicates a significant difference in behavior between the source and target platforms, suggesting that students' behavioral patterns are completely different across different platforms; for example, one platform may favor "practice questions," while the other may favor "interaction." This indicates that it is necessary to strengthen the distribution alignment mechanism and increase the sensitivity parameter of the kernel function. To enhance the perception of sample differences, adversarial training is introduced, which forces two distributions to be closer through the discriminator, increases sample sampling coverage, and expands the training set to reduce bias.

[0033] It should also be noted that the system first collects and vectorizes multimodal behavioral data from the source and target platforms, and then projects sample data from different sources into a unified embedding space through kernel function mapping. Further, it quantifies the distribution difference by calculating the mean difference and triggers different alignment strategies based on the magnitude of the difference. When the distribution difference is low, it can directly proceed to the next step; when the difference is in the intermediate range, it corrects it through lightweight optimization; and when the difference is too large, it achieves distribution approximation by increasing the sensitivity of the kernel function or introducing adversarial training. This achieves unified modeling of behavioral data from heterogeneous platforms, making data from different platforms comparable at the feature level, breaking down the barriers of cross-platform data differences, reducing model performance degradation caused by data distribution drift, and improving the cross-platform transferability and stability of the recommendation model.

[0034] The behavioral semantic structure modeling module is used to perform similarity calculations between the learning path embeddings of the source platform and the learning path embeddings of the target platform, and to perform path regularization alignment based on the distribution alignment results.

[0035] Furthermore, the similarity calculation includes the system converting the behavior sequence of the source platform into a learning path embedding, converting the behavior sequence of the target platform into a learning path embedding, counting the total number of paths of the source platform and the target platform, calculating the similarity between paths pairwise, and characterizing the semantic consistency of the learning behavior pattern.

[0036] It should be noted that path regularization alignment includes the system introducing a distribution alignment index as a regularization term in the path similarity calculation, generating a semantic level alignment index, and performing logical judgment based on the distribution alignment index. When the result is close to the maximum value, the recommendation model is directly transferred. When the result is in a medium range, path clustering and local alignment are performed. When the result is below a set threshold, the path reconstruction mechanism is triggered and the system enters the cold start mode.

[0037] It should also be noted that a preferred scheme for path regularization alignment based on distribution alignment results specifically includes calculating the behavioral semantic structure alignment index. , represented as: ; in, Indicates the source platform number Learning path index Indicates the target platform number Learning path index This indicates the total number of learning paths on the source platform. This indicates the total number of learning paths on the target platform. Indicates the source platform number Embedded vectors of learning paths Indicates the target platform number Embedded vectors of learning paths This represents the semantic similarity adjustment coefficient, used to control the magnitude of similarity calculation between paths. This represents the regularization coefficient, used to incorporate the results of the first step. This ensures that both distribution consistency and semantic consistency are optimized simultaneously.

[0038] when Furthermore, when the value remains stable at a high level, it indicates that the behavioral path structures of the source and target platforms are highly similar, and the students' learning trajectories do not differ significantly between the two platforms. This indicates that most path differences are close to 0, meaning the recommendation model can be directly transferred without additional adjustments, and the recommendation path alignment is sufficient to support subsequent recommendation target adaptation steps.

[0039] when This indicates that while there is some consistency in the learning paths between the source and target platforms, significant differences still exist. For example, students on the source platform tend to prefer a "question-answering + review" approach, while students on the target platform prefer a "video" approach. Discussion path, if part If the results are high, some are low, and the average result is in the middle, then "soft alignment" is needed for inconsistent paths. This can be achieved by using path clustering or local alignment methods, focusing on adjusting the paths with the greatest differences, and introducing additional weights in subsequent steps (reconstruction of recommendation goals) to avoid the inconsistent paths having an excessive impact on the recommendation results.

[0040] when (Setting a threshold) This indicates that the learning paths of the source and target platforms differ significantly and lack transferability. For example, the source platform is highly exam-driven, while the target platform is entirely inquiry-based learning, and most... Path re-mapping is required, which involves retraining the local path embeddings on the target platform to avoid direct transfer. This can be combined with step one. If both the distribution alignment index and the distribution alignment index are low, the system should be deemed "not suitable for direct migration" and a cold start recommendation should be adopted instead.

[0041] It should also be noted that the system converts the behavior sequences of cross-platform learners into path embedding vectors and calculates the similarity between the paths of the source platform and the target platform. Furthermore, a distribution alignment index is introduced as a regularization term in the similarity calculation, thereby generating a semantic alignment index. Different processing logics are triggered according to the numerical range. When the index is close to the maximum value, the recommendation model is directly transferred. When the index is in a medium range, path clustering and local alignment are performed. When the index is below the threshold, path reconstruction is triggered and a cold start mode is entered. This achieves consistent modeling of the semantic patterns of the learning path. It not only considers the matching at the sequence level, but also links the distribution level alignment with the semantic level alignment through the regularization mechanism. This captures the potential semantic patterns in the behavior patterns of learners on different platforms, thereby avoiding recommendation distortion caused by simple sequence differences and improving the model's adaptability to the transfer of learning paths.

[0042] The recommended target adaptive reconstruction module is used to perform nonlinear function processing on the difference between the predicted answer accuracy and the concept comprehension, and jointly calculate the recommended target index based on the semantic structure results.

[0043] Furthermore, the nonlinear function processing includes collecting the total number of recommended samples during system execution, obtaining the predicted answer accuracy and corresponding concept comprehension of each sample, inputting the difference into a nonlinear function for amplification or compression, averaging the results of all samples, and combining semantic modeling indicators to calculate the recommendation target adaptability index.

[0044] It should be noted that the calculation of the recommended target metric includes the system making logical judgments when calculating the recommended target adaptability metric. When the recommended target adaptability metric is close to zero, the learning path remains unchanged. When the recommended target adaptability metric is in a small range, the system distinguishes between cases where the accuracy is higher than the comprehension or the comprehension is higher than the accuracy, and recommends concept-based resources or practice-based resources respectively. When the recommended target adaptability metric exceeds a set threshold, the system triggers recommendations of principle explanation-based resources or practical training-based resources according to the direction of the difference.

[0045] It should also be noted that a preferred method for calculating the recommendation target index specifically includes calculating the adaptive index of the recommendation target. , represented as: ; in, This represents the adaptive metric for recommendation targets, used to measure the fit of recommendation targets across different platforms. This indicates the index of the target sample to be recommended. This represents the total number of target samples for recommendation. Indicates the first The accuracy rate of predicted answers for each sample measures the degree of knowledge mastery. Indicates the first A concept comprehension index for each sample measures the level of understanding of knowledge points. This represents the sensitivity parameter of the Tanh function, used to control whether the difference between accuracy and comprehension is amplified or compressed. Representing the hyperbolic tangent function enhances the ability to characterize nonlinearity. This represents the adjustment coefficient, used to introduce the results of the second step. .

[0046] when When the recommended goals are highly consistent and the differences in answer accuracy and comprehension are small, the student not only answers the questions but also truly understands the knowledge points. In this case, the recommendation system does not need to make significant adjustments to the recommendation strategy, maintains the existing learning path, and does not require additional intervention. It prioritizes recommending higher-level learning tasks or extended resources.

[0047] when hour, A positive difference indicates that although students have a high accuracy rate in the questions on the source platform, their understanding of the concepts on the target platform is insufficient. Students obtain correct answers through memorization or rote practice, but their understanding of the knowledge points is inadequate. The recommendation system should appropriately insert concept explanation videos and interactive examples, reduce purely rote practice resources, increase explanatory materials, and improve students' comprehension.

[0048] when When this occurs, it indicates a significant difference between students' accuracy and comprehension levels, resulting in an imbalance in learning outcomes. This indicates that students can solve the problems but do not understand the underlying principles; the recommendation system should recommend tasks that explain the principles plus application examples. Students understand the concepts but frequently make mistakes in answering specific questions. The recommendation system should recommend "practical training tasks + answering skills guidance".

[0049] when When the error occurs, it indicates a significant difference in students' answer accuracy and comprehension, suggesting insufficient semantic path alignment. This indicates a significant gap in the learning behavior patterns between the two platforms, with students' learning paths differing considerably between the two platforms (e.g., the source platform is a fast-paced question-and-answer type, while the target platform is a slow-paced deep learning type). The system needs to make "dual adjustments": on the one hand, re-planning the learning path (e.g., transitioning from a fast-paced question bank to concept-guided learning); on the other hand, regenerating the weighted loss function of the recommendation target to ensure that the transfer still conforms to the learning logic of the target platform.

[0050] It should also be noted that the system collects and compares the predicted answer accuracy and conceptual understanding of the recommended samples, and inputs the difference between the two into a nonlinear function for amplification or compression. Then, the results of all samples are averaged, and combined with semantic modeling indicators, a recommendation target adaptability index is calculated. The system makes logical judgments based on the recommendation target adaptability index. When the recommendation target adaptability index is close to zero, the existing learning path is maintained. When the index is in a small range, different situations are distinguished as "accuracy is higher than understanding" or "understanding is higher than accuracy", and concept reinforcement or practice training resources are recommended respectively. When the index exceeds a set threshold, principle explanation or practical training resources are triggered. In this way, adaptive reconstruction of learning objectives is achieved. Even if there are differences between the optimization objectives of the source platform and the target platform, the system can still dynamically adjust the recommendation strategy. By balancing the differences between test performance and conceptual understanding, the recommendation results are made more in line with the educational objectives of the target platform, improving the accuracy and personalization of model recommendations.

[0051] The pseudo-label generation and quality control module is used to perform non-linear function filtering and threshold control on the confidence of pseudo-labels, and outputs the final migration loss in combination with the recommended target index.

[0052] Furthermore, the nonlinear function screening and threshold control include the system collecting a set of pseudo-labels on the target platform during execution, obtaining the confidence values ​​of the pseudo-labels, and using a nonlinear function to convert the confidence values ​​of the pseudo-labels into probability values ​​between 0 and 1. The average confidence of the pseudo-labels is then calculated and compared with a preset threshold, and the migration loss is generated in conjunction with the recommended target indicators.

[0053] It should be noted that the final output migration loss includes the logical judgment performed by the system when comparing the average confidence of pseudo-labels with the threshold. When the difference is less than the tolerance, the pseudo-label strategy remains unchanged. When the average confidence is higher than the threshold, the use of pseudo-labels is increased. When the average confidence is lower than the threshold, the use of pseudo-labels is reduced and the manual annotation mechanism is triggered. When the pseudo-labels are completely ineffective, the use of pseudo-labels is stopped and only high-confidence samples are retained.

[0054] It should also be noted that a preferred approach for outputting the final migration loss based on the recommended target metric specifically includes calculating the final migration loss function. , represented as: ; in, Indicates the pseudo-label sample index. Indicates the total number of pseudo-tags. Indicates the first The confidence value of each pseudo-label. This represents the kurtosis adjustment parameter of the Sigmoid function, which controls the sensitivity of pseudo-label filtering. It is an exponential function used to construct the Sigmoid activation mechanism. This represents the confidence threshold for pseudo-labels, distinguishing between high-quality and low-quality labels. This represents the adjustment coefficient, used to balance the pseudo-label control mechanism and the recommendation objective.

[0055] When the output value When the overall confidence level of the pseudo-labels is very close to the threshold, at the critical level of 1.3, the denominator is observed to be... If the value is close to 0, the existing pseudo-label training strategy can be maintained, but monitoring needs to be increased to avoid fluctuations in pseudo-label quality.

[0056] When the output value When the overall confidence level of the pseudo-labels is significantly higher than the threshold, it indicates that most pseudo-labels are too trustworthy and more pseudo-labels should be introduced to expand the training. When the overall confidence level of the pseudo-labels is significantly lower than the threshold, it indicates that most pseudo-labels are too trustworthy and the use of pseudo-labels should be reduced. If necessary, expert annotation should be introduced to improve data quality.

[0057] When the output value If the overall confidence level of the pseudo-labels is significantly mismatched with the threshold, the use of large-scale pseudo-labels should be stopped immediately, and only high-confidence samples should be retained, or the system should be reverted to using only manually labeled data.

[0058] When the output value Meet the conditions ,and When the output value is close to This indicates that the pseudo-label mechanism is working well, and the overall credibility of the pseudo-labels is highly consistent with the threshold. The current training strategy should be maintained, and the number of training iterations should be increased appropriately to accelerate model convergence.

[0059] It should also be noted that when the supervision signal from the target platform is insufficient, the system collects a set of pseudo-labels and calculates the confidence level of each pseudo-label. It then uses a non-linear function to map these pseudo-labels to a probability interval, calculates the average confidence level, compares it with a threshold, and generates a transfer loss based on the recommended target metric. When the difference between the average confidence level of the pseudo-label and the threshold is less than the tolerance, the usage strategy remains unchanged. When the average confidence level is significantly higher than the threshold, the use of pseudo-labels is increased. When the average confidence level is significantly lower than the threshold, the use of pseudo-labels is reduced, and a manual annotation mechanism is triggered. When all pseudo-labels fail, the use of pseudo-labels is stopped, and only high-confidence samples are retained. Through this mechanism, dynamic screening and adaptive control of pseudo-label quality are achieved, ensuring the reliability of the semi-supervised transfer process, avoiding error propagation caused by low-quality pseudo-labels, and fully utilizing limited supervision signals to support model training, thereby improving the system's robustness and stability.

[0060] Reference Figure 2 As one embodiment of the present invention, this embodiment also provides a personalized teaching recommendation method based on multimodal learning behavior analysis, including the following steps: S1: Perform kernel function mapping on the source platform sample vector and the target platform sample vector, and calculate the distribution difference.

[0061] S2: Perform similarity calculations between the learning path embeddings of the source platform and the learning path embeddings of the target platform, and combine the distribution alignment results to perform path regularization alignment.

[0062] S3: Perform nonlinear function processing on the difference between the predicted answer accuracy and the concept comprehension, and jointly calculate the recommended target index based on the semantic structure results.

[0063] S4: Perform non-linear function filtering and threshold control on the confidence of pseudo-labels, and output the final migration loss in combination with the recommended target index.

[0064] This embodiment also provides a computer device applicable to the personalized teaching recommendation method based on multimodal learning behavior analysis, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the personalized teaching recommendation method based on multimodal learning behavior analysis as proposed in the above embodiment.

[0065] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0066] This embodiment also provides a storage medium on which a computer program is stored. When executed by a processor, the program implements the personalized teaching recommendation method based on multimodal learning behavior analysis as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0067] In summary, this invention achieves more accurate, dynamic, intelligent, and context-adaptive personalized clothing design recommendations than traditional recommendation systems by: deeply integrating multimodal user data, dynamically capturing changes in interests, introducing reinforcement learning to optimize long-term value, and considering real-time context. This results in a more effective improvement in user satisfaction and lifetime value.

[0068] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A personalized teaching recommendation system based on multimodal learning behavior analysis, characterized in that: include, Cross-platform behavior representation alignment module, behavior semantic structure modeling module, recommendation target adaptive reconstruction module, pseudo-label generation and quality control module; The cross-platform behavior representation alignment module is used to perform kernel function mapping on the source platform sample vector and the target platform sample vector, and calculate the distribution difference degree. The behavioral semantic structure modeling module is used to perform similarity calculations between the learning path embeddings of the source platform and the learning path embeddings of the target platform, and to perform path regularization alignment in combination with the distribution alignment results. The adaptive reconstruction module for the recommendation target is used to perform nonlinear function processing on the difference between the predicted answer accuracy and the concept comprehension, and jointly calculate the recommendation target index based on the semantic structure result. The pseudo-label generation and quality control module is used to perform nonlinear function filtering and threshold control on the confidence of pseudo-labels, and outputs the final migration loss in combination with the recommended target index.

2. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 1, characterized in that: The kernel function mapping of the source platform sample vector and the target platform sample vector includes the system collecting multimodal behavior data of students on the source platform and constructing corresponding sample vectors, while simultaneously collecting multimodal behavior data of students on the target platform and constructing another set of sample vectors. The number of samples on the source platform and the target platform are counted respectively, and then mapped to a unified embedding space through a kernel function to obtain comparable representations.

3. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 2, characterized in that: The calculation of distribution difference includes the system calculating the mean difference of the distribution difference between the source domain and the target domain samples in the embedding space, and making a judgment based on the calculation result. When the distribution difference is lower than the first threshold, it directly proceeds to the next step. When the distribution difference is between the two thresholds, lightweight alignment optimization is performed. When the distribution difference exceeds the second threshold, the distribution difference is reduced by increasing the sensitivity of the kernel function or introducing adversarial training.

4. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 3, characterized in that: The similarity calculation includes the system converting the behavior sequence of the source platform into a learning path embedding, converting the behavior sequence of the target platform into a learning path embedding, counting the total number of paths of the source platform and the target platform, calculating the similarity between paths pairwise, and characterizing the semantic consistency of the learning behavior pattern.

5. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 4, characterized in that: The path regularization alignment includes the system introducing a distribution alignment index as a regularization term in the path similarity calculation, generating a semantic level alignment index, and performing logical judgment based on the distribution alignment index. When the result is close to the maximum value, the recommendation model is directly transferred. When the result is in a medium range, path clustering and local alignment are performed. When the result is below a set threshold, a path reconstruction mechanism is triggered and a cold start mode is entered.

6. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 5, characterized in that: The nonlinear function processing includes the system collecting the total number of recommended samples during execution, obtaining the predicted answer accuracy and corresponding concept comprehension of each sample, inputting the difference into a nonlinear function for amplification or compression, averaging the results of all samples, and combining semantic modeling indicators to calculate the recommendation target adaptability index.

7. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 6, characterized in that: The calculation of the recommended target index includes the system making logical judgments when calculating the recommended target adaptability index. When the recommended target adaptability index is close to zero, the learning path remains unchanged. When the recommended target adaptability index is in a small range, the system distinguishes between cases where the accuracy is higher than the comprehension or the comprehension is higher than the accuracy, and recommends concept-based resources or practice-based resources respectively. When the recommended target adaptability index exceeds a set threshold, the system triggers principle explanation-based recommendations or practical training-based recommendations according to the direction of the difference.

8. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 7, characterized in that: The nonlinear function screening and threshold control process includes the system collecting a set of pseudo-labels from the target platform during execution, obtaining the confidence values ​​of the pseudo-labels, and using a nonlinear function to convert the confidence values ​​of the pseudo-labels into probability values ​​between 0 and 1. The average confidence of the pseudo-labels is then calculated and compared with a preset threshold, and the migration loss is generated in conjunction with the recommended target indicators.

9. The personalized teaching recommendation system based on multimodal learning behavior analysis as described in claim 8, characterized in that: The final migration loss output includes the logical judgment performed by the system when comparing the average confidence of pseudo-labels with a threshold. When the difference is less than the tolerance, the pseudo-label strategy remains unchanged. When the average confidence is higher than the threshold, the use of pseudo-labels is increased. When the average confidence is lower than the threshold, the use of pseudo-labels is reduced and a manual annotation mechanism is triggered. When the pseudo-labels are completely ineffective, the use of pseudo-labels is stopped, and only high-confidence samples are retained.

10. A personalized teaching recommendation method based on multimodal learning behavior analysis, based on the personalized teaching recommendation system based on multimodal learning behavior analysis as described in any one of claims 1 to 9, characterized in that: include, Perform kernel function mapping on the source platform sample vector and the target platform sample vector, and calculate the distribution difference. The similarity calculation is performed between the learning path embedding of the source platform and the learning path embedding of the target platform, and the path regularization alignment is performed in combination with the distribution alignment result. The difference between the predicted answer accuracy and the concept comprehension is processed by a nonlinear function, and the recommended target index is jointly calculated based on the semantic structure results. The confidence scores of pseudo-labels are filtered using a non-linear function and thresholded, and the final migration loss is output in combination with the recommended target index.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the personalized teaching recommendation method based on multimodal learning behavior analysis as described in claim 8.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the personalized teaching recommendation method based on multimodal learning behavior analysis as described in claim 8.

Images