Deep learning-based learning style recognition method, system, medium, and device

Abstract

The present application relates to the technical field of learning style recognition, and provides a learning style recognition method, system, medium and equipment based on deep learning, which comprises the following steps: obtaining learning behavior characteristics of a learner, and obtaining embedded features through preprocessing; gradually decomposing the embedded features into non-negative matrix factorization feature through multi-level non-negative matrix factorization operation; and fusing the embedded features and the non-negative matrix factorization feature through an adaptive attention fusion mechanism to obtain fusion features; wherein the multi-level non-negative matrix factorization operation adopts a multi-level decomposition structure, and the latent factors are increased layer by layer; and predicting the learning style through a deep classification network based on the fusion features. The accuracy of learning style recognition is improved.

Description

Technical Field

[0001] This invention belongs to the field of learning style recognition technology, and particularly relates to learning style recognition methods, systems, media and devices based on deep learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Learning style refers to a learner's habitual learning methods and tendencies during the learning process. It is a relatively stable learning style and preference that is gradually formed over a long period of learning activities, and it is an important factor reflecting individual differences among learners. For example, in the learning scenario of ideological and political courses: visual learners are good at acquiring knowledge from visual materials such as historical images, theoretical diagrams, and special courseware; auditory learners tend to deepen their understanding through auditory content such as ideological and political lectures, online audio courses, and audio recordings of theoretical recitations; kinesthetic learners efficiently absorb knowledge in hands-on scenarios such as simulated ideological and political practice activities, on-site study tours to red education bases, and interactive case discussions; balanced learners can flexibly adapt to various forms, accumulating theory through the study of ideological and political texts, and also comprehensively mastering content through videos, audio, or practical activities.

[0004] With the rapid development of online education platforms, learner behavior data is characterized by diversity, high dimensionality, and strong sparsity. For example, in ideological and political education scenarios, this data includes not only basic behaviors such as video viewing, text reading, and forum participation, but also specialized data such as the answer trajectory of ideological and political quizzes, the submission status of practical tasks, and the frequency of online group discussions. Accurately identifying students' learning styles from massive amounts of learning behavior data has become a core technical challenge for personalized education recommendation systems. Learning style identification can help education platforms provide targeted learning resources and pathways for different types of learners. For example, it can push ideological and political documentaries and theoretical visualizations to visual learners, recommend high-quality ideological and political podcasts and expert lectures to auditory learners, design interactive ideological and political practice tasks and virtual study projects for kinesthetic learners, and integrate diverse forms of ideological and political learning packages for balanced learners, thereby improving learning efficiency and outcomes.

[0005] In existing technologies, learning style recognition methods mainly fall into three categories: those based on traditional machine learning, deep learning, and hybrid methods.

[0006] Traditional machine learning methods, such as logistic regression, support vector machines, and random forests, perform statistical analysis and classification modeling on learners' basic characteristics (such as learning time, practice accuracy, and resource access frequency). While these methods have the advantages of being computationally simple and highly interpretable, they are prone to the curse of dimensionality when dealing with high-dimensional sparse data and struggle to capture the complex nonlinear relationships between learning behavior features.

[0007] In recent years, deep learning technology has been introduced into the field of learning style recognition, achieving some progress. Methods based on deep neural networks (DNNs) can automatically learn deep feature representations of learning behavior data, outperforming traditional methods in feature extraction and pattern recognition. Some studies have proposed using convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to process sequential learning behavior data, capturing learners' dynamic learning patterns through time series modeling. However, existing deep learning methods still have significant limitations when processing educational data.

[0008] First, there's the problem of identifying rare learning styles under extreme class imbalance. In real-world educational scenarios, educational data often exhibits strong imbalance, with students of different learning styles distributed extremely unevenly. For example, balanced learners might account for over 60% of the total, while auditory learners might only make up about 5%. This extreme class imbalance causes traditional deep learning models to tend to favor the majority class, resulting in a severe deficiency in identifying rare learning styles. While traditional class weight adjustment and resampling methods can alleviate this problem to some extent, they often come at the cost of sacrificing overall accuracy and are prone to introducing noise and overfitting.

[0009] Secondly, there is the problem of effective feature representation and structural modeling of high-dimensional sparse educational data. Learning behavior data typically contains 18-25 dimensions of multi-type features, such as time-based features (video viewing time, exercise completion time), count-based features (number of messages posted, number of button clicks), and proportionate features (exercise accuracy, resource skipping rate), etc. These features are not only high-dimensional and highly sparsity, but also exhibit complex potential correlations and hierarchical structures among themselves. For example, there is the causal relationship between video viewing behavior and subsequent exercise performance, and the complementarity of different learning resource usage patterns. Existing methods lack effective modeling of the intrinsic structure of features. Although traditional fully connected neural networks can theoretically learn the nonlinear mapping relationship of features, in practical applications they often fail to effectively mine this potential structured information, resulting in limited feature representation capabilities.

[0010] Furthermore, there are issues regarding the interpretability and educational applicability of the model's decision-making process. In educational applications, the results of learning style identification need to provide a basis for teachers to formulate personalized teaching strategies. This requires the model to not only have good predictive performance but also interpretability. Existing technologies have significant shortcomings in model interpretability; deep learning models are often considered "black boxes," and their decision-making processes are difficult to explain and understand. In educational applications, teachers and education administrators need to understand the basis and logic of learning style identification in order to formulate corresponding teaching strategies. Educators cannot understand which key features and behavioral patterns the model uses to determine students' learning styles. Models lacking interpretability are difficult to gain the trust of educators, limiting their application in actual teaching scenarios.

[0011] Another important issue is the adaptive fusion and weight optimization of multimodal learning features. Learning behavior data contains features of various natures and magnitudes, and effectively fusing these heterogeneous features is key to improving recognition accuracy. Traditional feature fusion methods, such as simple concatenation and weighted averaging, cannot adaptively adjust the importance weights of each feature according to the characteristics of different learning styles, and cannot fully utilize the complementary information of different types of features. For example, for visual learners, image and video-related features should receive higher weights; while for kinesthetic learners, practice and interaction-related features are more important. Existing methods lack this adaptive feature importance modeling capability, resulting in poor fusion performance. Especially when processing multimodal learning behavior data, effectively integrating data of different natures, such as temporal features (learning duration), count features (number of practice sessions), and binary features (whether skipped), remains a technical challenge. Summary of the Invention

[0012] To address the technical problems mentioned above, this invention provides a learning style recognition method, system, medium, and device based on deep learning. It introduces hierarchical NMF constraints into learning style recognition, using a multi-layered decomposition structure with progressively increasing latent factors to decompose learning behavior features into interpretable semantic representations. The first layer captures basic behavioral patterns, and the second layer models high-level learning strategies. This maintains the interpretability of features while providing structured prior knowledge for deep learning, effectively avoiding the "black box" problem of traditional deep learning and improving the accuracy of learning style recognition. This enables precise delivery of red resources, theoretical learning path recommendations, and matching of practical activities for students with different ideological and political learning styles, significantly enhancing the affinity and relevance of ideological and political education in courses.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] The first aspect of the present invention provides a deep learning-based learning style recognition method, comprising:

[0015] Learners' learning behavior characteristics are obtained and preprocessed to obtain embedded features;

[0016] For the embedded features, the non-negative matrix factorization operation is used to decompose them into non-negative matrix factorization factor features step by step. The embedded features and non-negative matrix factorization factor features are then fused through an adaptive attention fusion mechanism to obtain the fused features. The multi-level non-negative matrix factorization operation adopts a multi-level decomposition structure, and the latent factors increase at each level.

[0017] Based on fusion features, the learning style is predicted through a deep classification network.

[0018] Furthermore, the learning behavior characteristics include time consumption characteristics and learning participation characteristics.

[0019] Furthermore, the preprocessing includes standardization, feature embedding transformation, and regularization.

[0020] Furthermore, the multi-level nonnegative matrix factorization operation employs a nonnegative matrix factorization constraint loss:

[0021] ;

[0022] in, It is the sparsity regularization coefficient. L1 norm, first-layer reconstruction loss , It is a reconstruction feature of the first layer. It is an embedded feature, and the second layer is a reconstruction loss. , It is a reconstruction feature of the second layer.

[0023] Furthermore, the adaptive attention fusion mechanism includes: merging embedded features and non-negative matrix factorization factor features into a unified feature representation through a concatenation operation; for the unified feature representation, calculating the importance weights of the feature dimensions through an attention network containing two fully connected layers, and multiplying the importance weights element-wise with the unified feature representation to obtain weighted features; and processing the weighted features through a feature transformation network to obtain fused features.

[0024] Furthermore, the deep classification network adopts a progressive dimensionality reduction structure, configured with a two-layer hidden layer design, and each layer includes linear transformation, layer normalization, activation function and regularization.

[0025] Furthermore, a comprehensive loss function is adopted, which adaptively weights and combines the focus loss, nonnegative matrix decomposition constraint loss, and confidence regularization loss.

[0026] A second aspect of the present invention provides a deep learning-based learning style recognition system, comprising:

[0027] The preprocessing module is configured to: acquire learners' learning behavior features and obtain embedded features through preprocessing;

[0028] The feature extraction module is configured to: for embedded features, decompose them step by step into non-negative matrix factorization features through multi-level non-negative matrix factorization operations, and fuse the embedded features and non-negative matrix factorization features through an adaptive attention fusion mechanism to obtain fused features; wherein, the multi-level non-negative matrix factorization operation adopts a multi-level decomposition structure, and the potential factors increase at each level.

[0029] The classification module is configured to predict the learned style based on fused features through a deep classification network.

[0030] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the deep learning-based learning style recognition method described above.

[0031] A fourth aspect of the present invention provides a computer device including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein the processor executes the program to implement the steps of the deep learning-based learning style recognition method described above.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] This invention introduces hierarchical NMF constraints into learning style recognition. Through a multi-layer decomposition structure with progressively increasing latent factors, the learning behavior features are decomposed into interpretable semantic representations layer by layer. The first layer captures basic behavioral patterns, and the second layer models high-level learning strategies. This maintains the interpretability of features and provides structured prior knowledge for deep learning, effectively avoiding the "black box" problem of traditional deep learning.

[0034] This invention designs a two-layer attention fusion mechanism, which includes channel attention and spatial attention. Channel attention calculates the importance of feature dimensions through global pooling and multilayer perceptron, while spatial attention assigns weights at the sample level. This mechanism can automatically adjust feature weights according to the characteristics of different learning styles, and significantly improves the accuracy and adaptability of feature representation compared with traditional fixed-weight fusion methods.

[0035] This invention addresses the problem of extreme imbalance in learning style data by innovatively combining a focus loss function with adaptive class weight adjustment. The focus loss reduces the loss contribution of easily classified samples through a modulation factor, focusing on rare samples; the adaptive weights are calculated based on the inverse square root of the class frequency, providing stronger training signals for rare classes. These two mechanisms work synergistically to significantly improve the recognition ability of rare learning styles while maintaining overall accuracy.

[0036] This invention employs a progressive dimensionality reduction multilayer fully connected network structure, using a two-layer hidden layer design to achieve gradual feature abstraction and compression. Each layer is equipped with a complete processing flow, including linear transformation, layer normalization, ReLU activation function, and Dropout regularization, ensuring the stability of feature propagation and the model's generalization ability. Layer normalization effectively stabilizes the training process and accelerates convergence, while Dropout regularization prevents overfitting. This progressive dimensionality reduction design avoids information bottlenecks and achieves excellent classification performance with a relatively small number of parameters, making it particularly suitable for handling high-dimensional sparse features in educational data. The network architecture is concise and efficient, ensuring both the model's expressive power and controlling computational complexity, providing a solid technical foundation for real-time learning style recognition.

[0037] This invention is particularly applicable to ideological and political theory courses and curriculum-based ideological and political education. It can accurately identify students' visual, auditory, kinesthetic, and balanced learning styles in ideological and political education, thereby realizing intelligent recommendation of ideological and political resources and planning of learning paths for each student. While ensuring the effectiveness of ideological and political education, it solves the problem of the traditional "one-size-fits-all" approach to ideological and political education, and has significant political and social value. Attached Figure Description

[0038] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0039] Figure 1 This is a flowchart of the deep learning-based learning style recognition method according to Embodiment 1 of the present invention;

[0040] Figure 2 This is an overall structural diagram of the deep learning-based learning style recognition method according to Embodiment 1 of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0043] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0044] Example 1

[0045] This embodiment provides a learning style recognition method based on deep learning.

[0046] The primary problem this embodiment needs to solve is how to significantly improve the ability to identify rare learning styles while maintaining overall classification performance, so as to achieve truly balanced classification.

[0047] This embodiment requires the design of a deep feature representation method that can simultaneously model the correlation and hierarchy between features.

[0048] This embodiment needs to provide interpretable feature decomposition and decision-making basis while ensuring classification accuracy, so that educators can understand and trust the model's judgment results.

[0049] This embodiment requires the design of a fusion mechanism that can automatically adjust feature weights based on the characteristics of different learning styles.

[0050] In summary, the goal of this embodiment is to address key issues in existing learning style recognition methods regarding class balance, feature representation, interpretability, fusion efficiency, and training stability by introducing nonnegative matrix factorization constraints, designing an adaptive feature fusion mechanism, and optimizing the loss function and training strategy. Ultimately, this aims to achieve accurate, efficient, and interpretable intelligent recognition of learning styles.

[0051] The deep learning-based learning style recognition method (hereinafter referred to as Enhanced_NMF_EDPR) provided in this embodiment is specifically designed for the automatic recognition of students' learning styles in online education platforms. It combines the interpretability advantages of nonnegative matrix factorization with the powerful feature learning capabilities of deep learning. Through hierarchical feature decomposition, adaptive fusion, and focus loss optimization, it achieves accurate classification of students' multidimensional learning behaviors, providing accurate and reliable technical support for personalized education recommendations.

[0052] The deep learning-based learning style recognition method provided in this embodiment, such as... Figure 1 and Figure 2 As shown, it includes the following steps:

[0053] Step 1: Data input and preprocessing.

[0054] In a learning style recognition system, learners' behavioral data are input into the network in the form of multidimensional feature vectors.

[0055] Let the input learner's behavioral data matrix be... ,in This represents the number of learner samples, and 18 indicates the number of dimensions for the learning behavior features. Each learner's feature vector... It contains rich information on learning behavior, covering time-consuming features such as T. image (Image content learning time), T video (Total video viewing time), T read (Text reading time) and T audio (Audio material learning duration), and learning engagement characteristics such as N msgs posted (Number of messages published), N group discussions (Number of times participating in group discussions) and N standard questions correct (Number of standard questions answered correctly), etc.

[0056] Because learning behavior data exhibits significant scale differences—time-based features typically have large values ​​measured in seconds or minutes, while proportional features like accuracy range from 0 to 1—data standardization is necessary to ensure that each feature dimension falls within the same numerical range. For each dimension of the original feature matrix... Calculate its mean and standard deviation Then, perform Z-score normalization on this dimension of the feature for all samples: ;in, Indicates the first The learner in the first Standardized values ​​for each feature dimension. The sample mean of this feature dimension. The corresponding standard deviation is... Indicates the first The learner in the first The original values ​​of each feature dimension; after standardization, the feature matrix is ​​obtained. This ensures numerical compatibility between different types of features in subsequent processing.

[0057] The standardized feature data is then subjected to dimensionality expansion and non-linear mapping through a feature embedding transformation layer. This layer employs a fully connected network structure, mapping the 18-dimensional input features to a 64-dimensional embedding space. The mapping process is defined as follows: ;in, It is the weight matrix of the embedding layer. It is a bias vector. Represents the ReLU (Modified Linear Unit) activation function. Indicates the first 18-dimensional input features of each learner Indicates the first The embedded features of each learner are used. To improve training stability and feature representation quality, a layer normalization operation, LayerNorm, is added before the activation function. Its calculation formula is as follows: ,in, and These are the input vectors. The mean and variance, It is a small constant to prevent division by zero. and These are learnable scaling and offset parameters.

[0058] To prevent overfitting and enhance generalization ability, the regularization technique Dropout is applied after the feature embedding layer. Dropout uses probability... Randomly setting the output of some neurons to zero, the calculation during training is as follows: ;in, It follows a Bernoulli distribution. mask vector, This represents element-wise multiplication. Indicates the first A normalized 64-dimensional embedding representation of each learner.

[0059] After the above preprocessing steps, the original 18-dimensional learning behavior features are transformed into a normalized 64-dimensional embedding representation. This provides high-quality feature inputs for the subsequent nonnegative matrix factorization constraint module.

[0060] Step 2: Hierarchical Nonnegative Matrix Factorization (NMF) Constraints.

[0061] Preprocessed 64-dimensional embedding features The core innovation module, the hierarchical nonnegative matrix factorization constraint module, is then introduced. Through multi-level nonnegative matrix factorization operations, the input features are gradually decomposed into latent factor representations with clear semantic meanings. This maintains the interpretability advantage of traditional NMF methods while achieving end-to-end optimized training through a deep learning framework.

[0062] The hierarchical NMF module adopts a two-layer decomposition structure. The first layer uses 8 latent factors, and the second layer expands to 16 factors, forming a coarse-to-fine feature decomposition hierarchy.

[0063] For the embedded features of the input First, preliminary feature deconstruction is performed through the first-level NMF decomposition. The first-level decomposition includes a weight matrix. sum factor matrix Two learnable parameters, of which Implemented through a linear transformation layer. These are then directly optimized as network parameters. The calculation process for the first-layer factor representation is as follows: ;here Indicates the first Each sample is represented by a factor in the first layer. The ReLU activation function ensures the non-negativity constraint of the factors, meeting the basic requirements of NMF. Simultaneously, to maintain the reconstruction properties of the decomposition, the reconstruction error is calculated as a constraint term: ; ;in, and Both represent the transpose of a matrix. It is a reconstruction feature of the first layer. This is the corresponding reconstruction loss.

[0064] Building upon the first-level decomposition, the second-level NMF further refines the feature representation. It uses the factors from the first level for representation. As input, the second-level decomposition uses 16 latent factors for more refined pattern mining. The second layer includes a weight matrix. sum factor matrix The calculation process is as follows: Factor representation of each sample in the second layer ; ; ;in, To represent the transpose of a matrix, It is a reconstruction feature of the second layer. This is the reconstruction loss corresponding to the second layer; the second layer factor representation It can capture more granular learning behavior patterns, such as specific combinations of learning resource preferences and time allocation strategies.

[0065] To enhance the sparsity of the model and improve feature selection ability, a feature selection function is added to each decomposition layer. Regularization constraints. The total NMF constraint loss function is defined as:

[0066] ;

[0067] in, It is the sparsity regularization coefficient, L1 norm. This makes the factor representations sparser, which is beneficial for discovering key learning behavior patterns.

[0068] Finally, the factor representations of the two layers are merged into a complete NMF feature through a concatenation operation: .

[0069] This hierarchical decomposition strategy not only maintains the inherent partiality and interpretability of the NMF method, but also achieves stronger feature learning capabilities through end-to-end training of deep learning. The eight factors in the first layer mainly capture basic learning behavior patterns, such as time allocation preferences and selection of learning resource types; the 16 factors in the second layer further explore complex combinations of learning strategies, such as the balance between active exploration and passive acceptance, and the trade-off between deep learning and breadth learning, providing rich and meaningful feature representations for subsequent style classification.

[0070] Step 3: Adaptive attention fusion.

[0071] After processing by the hierarchical NMF constraint module, 64-dimensional original embedding features were obtained. and 24-dimensional NMF factor features To fully utilize the complementary information of these two different characteristics, an adaptive attention fusion module was designed. This module achieves intelligent integration of features through a dynamic weight allocation mechanism, ensuring that various features under different learning styles can obtain optimal fusion weights.

[0072] Step 301: The adaptive attention fusion module first merges the embedded features and NMF features into a unified feature representation through a concatenation operation. Let embed... dim =64 represents the embedding feature dimension, nmf dim =24 is the feature dimension of NMF, then the feature dimension of the fused feature is total. dim =88, the fused features are represented as follows: Where Concat represents the concatenation operation.

[0073] Step 302: Attention weight calculation adopts a multilayer perceptron structure, and learns the importance distribution between features through an attention network containing two fully connected layers.

[0074] The first layer maps the 88-dimensional input to a 64-dimensional hidden space and applies the Tanh (hyperbolic tangent) activation function to obtain the hidden representation:

[0075] ;

[0076] in, This is the first layer attention weight matrix. It is the corresponding bias vector; the choice of the Tanh activation function helps to generate bounded activation values ​​and avoid the gradient explosion problem.

[0077] The second layer maps the hidden representations back to the original dimensions and generates normalized attention weights using the Softmax (normalization exponent) function:

[0078] ;

[0079] ;

[0080] in, It is the second layer weight matrix. This represents the bias vector of the second layer. This is the final attention weight vector. Hiding indicates mapping back to the original dimension. The hidden representation maps back to the j-th dimension of the original dimension; Softmax normalization ensures that the sum of all weight components is 1, i.e. A i,j It is A i The data in the j-th dimension.

[0081] By obtaining the attention weight vector Applied to unified feature representation The weighted features obtained above : ,in, The symbol represents element-wise multiplication. It is the 88-dimensional attention weight vector obtained in step 302.

[0082] Step 303: The weighted features are then further processed through a feature transformation network, which includes a combination of linear transformation, layer normalization, and ReLU activation to obtain fused features.

[0083] ;

[0084] in, It is a transformation matrix. The bias vector is represented by the LayerNorm, which is calculated using the following formula:

[0085] ;

[0086] in, and These are the input vectors. The mean and variance, It is a numerical stability constant. and These are learnable scaling and translation parameters. This represents the Hadamard product (element-wise multiplication); layer normalization helps stabilize the training process and accelerate model convergence.

[0087] Step 304: The fusion module simultaneously outputs the fused feature representation and attention weight distribution, providing support for the interpretability analysis of the model. The feature representation output by the fusion module is as follows: Attention weights It can intuitively reflect the degree of contribution of different feature dimensions to the classification decision of the current sample. By analyzing the distribution pattern of attention weights, we can understand the key feature combinations that the model relies on when recognizing different learning styles. For example, for visual learners, feature dimensions related to images and videos usually receive higher attention weights; while for kinesthetic learners, features related to practice and interaction are more important.

[0088] The adaptability of this fusion mechanism is reflected in its ability to dynamically adjust attention weights. Unlike fixed-weight feature fusion methods, the adaptive attention mechanism can automatically adjust the importance of each dimension based on the feature distribution of the input samples, enabling the model to assign personalized feature weight combinations to each learner. This design not only improves the quality of feature representations but also enhances the model's adaptability to different learning styles, providing high-quality fused feature inputs for subsequent deep classification networks.

[0089] Step 4: Deep classification network.

[0090] 88-dimensional features after fusion The system then feeds into a deep classification network for final learning style prediction.

[0091] Deep classification networks employ a densely connected architecture, using multiple fully connected layers to achieve a non-linear mapping from fused features to learned style categories. The deep classification network utilizes a progressive dimensionality reduction structure, configured as follows: The deep classification network employs a two-hidden-layer design. Each layer includes a complete processing flow of linear transformation, layer normalization, activation function, and Dropout regularization.

[0092] Step 401: The first layer (hidden layer 1) maps the 88-dimensional fused features to the 128-dimensional first hidden space:

[0093] ;

[0094] in, It is the first layer weight matrix. It is the corresponding bias vector. This represents the first hidden space feature.

[0095] The original linear output is then normalized using LayerNorm:

[0096] ;

[0097] in, and These are the mean and variance of the first layer's output, respectively. and These are learnable normalized parameters. This indicates the result of the standardization process.

[0098] After layer normalization, the ReLU activation function is applied to introduce nonlinearity, resulting in nonlinear results:

[0099] ;

[0100] To prevent overfitting and enhance the model's generalization ability, Dropout regularization is applied after activation, with a dropout rate p = 0.3, resulting in the first layer processing outcome: .

[0101] Step 402: The second layer of the network (hidden layer 2) further compresses the 128-dimensional features to 64 dimensions. The processing flow is similar to that of the first layer.

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] in, It is the second layer weight matrix. It is the corresponding bias vector. Indicates the characteristics of the second hidden space. This indicates the result of the second-level standardization process. This represents the second-level nonlinear result. This indicates the output of the second-layer network.

[0107] Step 403: The classifier layer maps the 64-dimensional deep features to the final number of learned style categories, n. classes =4:

[0108] ;

[0109] in, It is the classifier weight matrix. It is the classification bias vector. This is the final output of the category logits.

[0110] The weights of the deep classification network are initialized using a Xavier normal distribution, and different gain coefficients are set for different layers to optimize the training effect.

[0111] For hidden layers, use standard Xavier initialization with a gain coefficient of gain = 1.0:

[0112] ;

[0113] Where ~ indicates that it follows a distribution. Indicates a normal distribution. It is the number of input neurons. It is the number of output neurons. This represents the weights of the hidden layer.

[0114] For the output classification layer, a small gain coefficient of gain=0.5 is used to avoid gradient explosion in the early stages of training:

[0115] ;

[0116] in, This represents the weight of the classification layer.

[0117] All bias vectors are initialized to zero vectors.

[0118] The final output of a deep classification network includes the following key components for subsequent loss calculation and model analysis: the classification prediction output is... , representing the predicted scores for the four learning style categories; the factors obtained from the hierarchical NMF decomposition are represented as It preserves interpretable learning behavior pattern information; the reconstruction loss of the NMF module. The attention weights of the fusion module are used to constrain the quality of the feature decomposition. It provides a foundation for visual analysis of feature importance; the original feature embedding representation This preserves the basic information of the input features.

[0119] This multi-output design not only supports end-to-end supervised learning but also provides rich intermediate representations for model interpretability analysis and multi-task learning. Through a densely connected architecture and a carefully designed regularization strategy, the deep classification network can accurately identify and classify complex learning style patterns while maintaining good generalization ability. The progressive dimensionality reduction design of the network structure effectively avoids the information bottleneck problem, while the combined use of layer normalization and Dropout ensures the stability of the training process and the robustness of the model.

[0120] Step 5: Design the loss function.

[0121] To effectively address the class imbalance problem in learning style recognition and integrate multiple training objectives, this embodiment designs a comprehensive loss function. This function adaptively weights and combines the focus loss, NMF reconstruction loss, and confidence regularization loss to achieve synergistic optimization of classification accuracy, feature interpretability, and model robustness.

[0122] (1) Calculation of focal loss.

[0123] To address the problem of extreme class imbalance in learning style data, focus loss is adopted as the main classification loss function. Focus loss dynamically adjusts the loss weights of different samples by introducing a modulation factor, so that the model focuses on the rare samples that are difficult to classify.

[0124] For the The predicted logits for each sample and real labels First, calculate the softmax probability distribution:

[0125] ;

[0126] in, This indicates that the model represents the true label for sample i. The predicted probability, where C represents the total number of categories; The label corresponding to sample i The logits value; This indicates that sample i corresponds to the true label. The logits value; for the true category , This represents the model's predicted probability for the true class.

[0127] The formula for calculating the loss of a single focus is:

[0128] ;

[0129] in, It is a modulation factor that controls the degree of loss attenuation for easily classified samples; It is a class-balanced weighting system, which sets different weight coefficients for different classes; for rare classes with a small number of samples, A larger value ensures these samples receive sufficient attention during training; the total focus loss is the average of the losses for all samples.

[0130] .

[0131] (2) NMF reconstruction loss.

[0132] To maintain the quality and interpretability of the eigenvalue decomposition, the reconstruction loss generated by the NMF module is added as a constraint term to the total loss function. The NMF reconstruction loss includes the reconstruction error of the two-layer decomposition and sparsity regularization:

[0133] ;

[0134] in, It is a reconstruction feature of layer 1. It is the corresponding factor representation. It is a nonnegative factor matrix. 0.01 is the sparsity regularization coefficient. L2 reconstruction error ensures that the decomposition results can effectively recover the original features, while L1 regularization promotes the sparsity of the factor representation, which is beneficial for discovering key learning behavior patterns.

[0135] (3) Confidence regularization loss.

[0136] To improve the confidence and robustness of the model's predictions, a confidence regularization term is introduced. This term avoids model overconfidence by maximizing the entropy of the predicted probability distribution, and its calculation formula is as follows:

[0137] ;

[0138] in, This indicates that the model represents the label for sample i. The predicted probability, It is a numerical stability constant. Confidence regularization encourages the model to produce a relatively smooth probability distribution, avoiding overfitting during training.

[0139] (4) Adaptive weight adjustment strategy.

[0140] To dynamically balance the influence of different loss terms during training, an adaptive weight adjustment mechanism was designed. The NMF-constrained weights decay linearly according to the training progress.

[0141] ;

[0142] in, Here, is the initial NMF weight coefficient, epoch is the current training epoch, and max_epochs is the total number of training epochs. This decay strategy allows the model to focus on the quality of feature decomposition in the early stages of training, and gradually shift to optimizing classification performance in the later stages.

[0143] (5) Comprehensive loss function.

[0144] The final total loss function is a weighted combination of the three loss terms:

[0145] ;

[0146] in, It is the confidence regularization coefficient.

[0147] The loss function also returns detailed information for each loss term for training monitoring: This represents the total loss value, used for backpropagation optimization. This represents the focus loss value, reflecting the classification performance. This represents the NMF reconstruction loss, which measures the quality of feature decomposition. This represents the confidence loss and is used to assess the robustness of the model. This indicates the current adaptive weights and monitors the weight adjustment process.

[0148] By designing this multi-objective loss function, the model can maintain the interpretability of feature decomposition and the confidence of prediction while ensuring classification accuracy. The adaptive weight adjustment strategy ensures the stability of the training process, avoids conflicts between different loss terms, and enables the model to obtain an optimal solution that balances performance and interpretability.

[0149] Step 6: Training strategy and parameter optimization.

[0150] To ensure model stability and achieve optimal performance, this embodiment employs a multi-level learning rate configuration training strategy. Considering the differences in learning characteristics among different network components, the model parameters are divided into five optimization groups with differentiated learning rates. The NMF constraint module, due to its involvement in the numerical stability of matrix factorization, is given a smaller learning rate. ,in, This represents the standard learning rate.

[0151] The learning rates for both the input embedding layer and the attention fusion module are standard learning rates: .

[0152] The classifier layer is set with a larger learning rate to accelerate convergence: .

[0153] The optimization process uses the AdamW optimizer, with the weight decay factor set to 1. .

[0154] During training, a learning rate scheduler is used. When the validation loss shows no significant improvement for 15 consecutive epochs (training iterations), the learning rate is adjusted according to the following formula: , This represents the learning rate (e.g., 0.001) that the model is currently using.

[0155] The learning rate scheduling is triggered when the improvement in the validation loss is less than a minimum threshold. ; This represents the lowest loss the model has ever achieved on the validation set to date. This represents the loss of the model on the validation set at the current epoch; This represents the extent to which the current loss is improved compared to the historical best loss.

[0156] To prevent gradient explosion, apply maximum norm gradient clipping to all parameters: ; This represents all the learnable parameters in the model (such as weights W and biases b). In mathematics, it represents the "gradient," which is the direction and magnitude in which parameters should be updated. Representing the L2 norm, it can be understood as calculating the "total length" or "total size" of this gradient vector.

[0157] The early stopping mechanism monitors the rare class F1 score (balanced F-score). Early stopping is triggered when the rare class performance shows no improvement and the improvement is less than 0.001 within 25 consecutive epochs. The training batch size is set to 32, and the total number of training epochs is limited to 150. This carefully designed training strategy ensures that the model can effectively improve its ability to recognize rare learning styles while maintaining overall classification performance, ultimately converging to an optimal state that balances performance and stability.

[0158] This embodiment achieves significant technical advantages in style recognition tasks by integrating hierarchical nonnegative matrix factorization constraints with a deep learning architecture. Comparative experiments with benchmark methods like Random Forest demonstrate that this technique achieves substantial improvements across multiple key performance indicators.

[0159] Table 1. Algorithm Performance in this Embodiment

[0160]

[0161] As shown in Table 1, in terms of overall accuracy, the Enhanced_NMF_EDPR method in this embodiment achieves 0.864, a 1.4% improvement compared to the Random Forest's 0.852. This indicates that the method in this embodiment maintains high overall classification performance while exhibiting superior recognition accuracy. More importantly, this embodiment has achieved a breakthrough in addressing the key technical challenge of rare category recognition. The rare category F1 score reaches 0.633, while the Random Forest method fails to recognize rare categories at all (F1=0.000). This comparison fully demonstrates the effectiveness of the method in this embodiment in solving the problem of extreme class imbalance. The significant improvement in rare category recognition capability is of great significance for personalized education applications, ensuring that students with auditory, kinesthetic, or other learning styles can also receive accurate style recognition and personalized recommendations.

[0162] In terms of the macro-average F1 score, this embodiment achieves 0.756, significantly surpassing the random forest's 0.435, representing an improvement of 73.8%. The macro-average F1 score fairly evaluates the model's ability to identify all classes. This significant improvement indicates that the method in this embodiment not only performs well in most classes but, more importantly, achieves effective identification in rare classes, demonstrating the model's balanced classification ability. The weighted F1 score reaches 0.855, a 3.0% improvement compared to the random forest's 0.830, further validating the superior overall performance of the method in this embodiment.

[0163] The technical advantages of this embodiment are mainly reflected in three aspects: First, the hierarchical NMF constraint mechanism, through a two-layer decomposition structure of 8+16 factors, maintains the interpretability of features and provides rich semantic representations, effectively solving the "black box" problem of traditional deep learning; Second, the adaptive attention fusion mechanism can dynamically adjust feature weights according to different learning styles, significantly improving the model's adaptability to complex learning patterns; Finally, the collaborative optimization strategy of focus loss and adaptive class weights is specifically designed for the class imbalance problem, ensuring the effective identification of rare learning styles.

[0164] The comprehensive experimental results show that this embodiment not only maintains a competitive advantage in traditional overall accuracy metrics, but also achieves a qualitative breakthrough in the key technical challenge of rare category identification, providing an accurate, reliable, and balanced learning style identification technology solution for personalized education recommendation systems.

[0165] This embodiment is specifically applied to the automatic identification and classification of students' learning styles in online education platforms. Through innovative modules such as hierarchical nonnegative matrix factorization constraints, adaptive feature fusion, focus loss optimization, and multi-scale attention mechanisms, it achieves in-depth mining of students' multi-dimensional learning behavior characteristics and accurate identification of learning styles, thereby improving the accuracy and effectiveness of personalized education recommendations.

[0166] This embodiment introduces hierarchical NMF constraints into learning style recognition for the first time. Through a two-layer decomposition structure with 8 and 16 factors, learning behavior features are progressively decomposed into interpretable semantic representations. The first layer captures basic behavioral patterns (learning time preferences, resource usage habits), while the second layer models high-level learning strategies (active exploration vs. passive acceptance). This design maintains feature interpretability while providing structured prior knowledge for deep learning, effectively avoiding the "black box" problem of traditional deep learning.

[0167] This embodiment designs a two-layer attention fusion mechanism, including channel attention and spatial attention. Channel attention calculates the importance of feature dimensions through global pooling and a multilayer perceptron, while spatial attention assigns weights at the sample level. This mechanism can automatically adjust feature weights according to different learning styles, significantly improving the accuracy and adaptability of feature representation compared to traditional fixed-weight fusion methods.

[0168] This embodiment addresses the problem of extreme imbalance in learning style data by innovatively combining a focus loss function with adaptive class weight adjustment. The focus loss is achieved through a modulation factor. The loss contribution of easily classified samples is reduced, with a focus on rare samples; adaptive weights are calculated based on the inverse square root of the class frequency, providing stronger training signals for rare classes. These two mechanisms work synergistically to significantly improve the recognition ability of rare learning styles while maintaining overall accuracy.

[0169] This embodiment employs a progressive dimensionality reduction multilayer fully connected network structure, using a two-layer hidden layer design (128→64) to achieve gradual feature abstraction and compression. Each layer is equipped with a complete processing flow, including linear transformation, layer normalization, ReLU activation function, and Dropout regularization, ensuring the stability of feature propagation and the model's generalization ability. Layer normalization effectively stabilizes the training process and accelerates convergence, while Dropout regularization prevents overfitting. This progressive dimensionality reduction design avoids information bottlenecks and achieves excellent classification performance with a relatively small number of parameters, making it particularly suitable for handling high-dimensional sparse features in educational data. The network architecture is concise and efficient, ensuring both the model's expressive power and controlling computational complexity, providing a solid technical foundation for real-time learning style recognition.

[0170] This embodiment proposes a three-stage training strategy: warm-up, balancing, and fine-tuning. The warm-up stage trains the NMF module to establish a stable feature decomposition foundation; the balancing stage introduces classification loss and dynamically adjusts constraint weights; and the fine-tuning stage performs end-to-end optimization. This phased strategy avoids conflicts between different loss terms, ensuring stable model convergence and exhibiting better performance and convergence compared to traditional end-to-end training.

[0171] By organically combining the above-mentioned innovative technologies, this embodiment significantly outperforms existing methods in terms of rare category identification, overall classification accuracy, and model interpretability, providing a practical and reliable intelligent solution for educational technology applications.

[0172] This embodiment can balance feature representation ability, model interpretability, and class balance processing, effectively handling high-dimensional sparse educational data and providing reliable technical support for educational practice.

[0173] This embodiment is primarily applied to personalized recommendation systems, intelligent learning management systems, adaptive learning software, and educational data analysis tools in online education platforms. In these products, this embodiment can perform real-time analysis of learners' multidimensional behavioral data, automatically identifying their learning style types (visual, auditory, kinesthetic, and balance-based), providing the platform with accurate learner profiles. This enables personalized course recommendations on online learning platforms, learning path planning in online education, content matching in corporate training systems, exercise delivery from intelligent question banks, and learning strategy suggestions from teaching aid software. Furthermore, this embodiment also has significant application value in learning behavior analysis tools, teacher instructional support systems, and educational evaluation platforms in educational research institutions, helping educators better understand student characteristics and develop targeted teaching plans.

[0174] Example 2

[0175] This embodiment provides a deep learning-based learning style recognition system, which specifically includes:

[0176] The preprocessing module is configured to: acquire learners' learning behavior features and obtain embedded features through preprocessing;

[0177] The feature extraction module is configured to: for embedded features, decompose them step by step into non-negative matrix factorization features through multi-level non-negative matrix factorization operations, and fuse the embedded features and non-negative matrix factorization features through an adaptive attention fusion mechanism to obtain fused features; wherein, the multi-level non-negative matrix factorization operation adopts a multi-level decomposition structure, and the potential factors increase at each level.

[0178] The classification module is configured to predict the learned style based on fused features through a deep classification network.

[0179] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0180] Example 3

[0181] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the deep learning-based learning style recognition method described in Embodiment 1 above.

[0182] Example 4

[0183] This embodiment provides a computer device, such as... Figure 3As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data. When the processor 1001 executes the program, it implements the steps in the deep learning-based learning style recognition method described in Embodiment 1 above.

[0184] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A learning style recognition method based on deep learning, characterized in that, include: Learners' learning behavior characteristics are obtained and preprocessed to obtain embedded features; The learning behavior characteristics include time consumption characteristics and learning participation characteristics. Time consumption characteristics include the learning time of image content, the total video viewing time, the text reading time, and the learning time of audio materials. Learning participation characteristics include the number of messages posted, the number of times participating in group discussions, and the number of standard questions answered correctly. For the embedded features, the non-negative matrix factorization operation is used to decompose them into non-negative matrix factorization factor features step by step. The embedded features and non-negative matrix factorization factor features are then fused through an adaptive attention fusion mechanism to obtain fused features. The multi-level non-negative matrix factorization operation adopts a multi-level decomposition structure with progressively increasing potential factors. The first layer captures basic behavioral patterns, and the second layer models high-level learning strategies. Based on fusion features, the learning style is predicted through a deep classification network.

2. The deep learning-based learning style recognition method as described in claim 1, characterized in that, The preprocessing includes standardization, feature embedding transformation, and regularization.

3. The deep learning-based learning style recognition method as described in claim 1, characterized in that, The multi-level nonnegative matrix factorization operation employs a nonnegative matrix factorization constraint loss: ； in, It is the sparsity regularization coefficient. L1 norm, first-layer reconstruction loss , It is a reconstruction feature of the first layer. It is an embedded feature, and the second layer is a reconstruction loss. , It is a reconstruction feature of the second layer. Represented as a factor, It is a factor matrix.

4. The deep learning-based learning style recognition method as described in claim 1, characterized in that, The adaptive attention fusion mechanism includes: merging embedded features and non-negative matrix factorization features into a unified feature representation through a concatenation operation; for the unified feature representation, calculating the importance weights of the feature dimensions through an attention network containing two fully connected layers, and multiplying the importance weights element-wise with the unified feature representation to obtain weighted features; and processing the weighted features through a feature transformation network to obtain fused features.

5. The deep learning-based learning style recognition method as described in claim 1, characterized in that, The deep classification network adopts a progressive dimensionality reduction structure, configured with two hidden layers, and each layer includes linear transformation, layer normalization, activation function and regularization.

6. The deep learning-based learning style recognition method as described in claim 1, characterized in that, A comprehensive loss function is adopted, which adaptively weights and combines the focus loss, nonnegative matrix decomposition constraint loss, and confidence regularization loss.

7. A deep learning-based learning style recognition system, characterized in that, include: The preprocessing module is configured to: acquire learners' learning behavior features and obtain embedded features through preprocessing; The learning behavior characteristics include time consumption characteristics and learning participation characteristics. Time consumption characteristics include the learning time of image content, the total video viewing time, the text reading time, and the learning time of audio materials. Learning participation characteristics include the number of messages posted, the number of times participating in group discussions, and the number of standard questions answered correctly. The feature extraction module is configured to: for embedded features, decompose them into non-negative matrix factorization features through multi-level non-negative matrix factorization operations, and fuse the embedded features and non-negative matrix factorization features through an adaptive attention fusion mechanism to obtain fused features; wherein, the multi-level non-negative matrix factorization operation adopts a multi-level decomposition structure, and the potential factors increase at each level, the first layer captures basic behavioral patterns, and the second layer models high-level learning strategies. The classification module is configured to predict the learned style based on fused features through a deep classification network.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the deep learning-based learning style recognition method as described in any one of claims 1-6.

9. A computer device comprising a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the deep learning-based learning style recognition method as described in any one of claims 1-6.

Images