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

Abstract

The invention relates to the technical field of learning style recognition, and provides a learning style recognition method and system based on deep learning, a medium and equipment, and the method comprises the steps: obtaining learning behavior features of a learner, obtaining embedded features through preprocessing, gradually decomposing the embedded features into non-negative matrix factorization factor features through multi-level non-negative matrix factorization operation, and obtaining a non-negative matrix factorization factor features; the embedded features and the non-negative matrix factorization factor features are fused through an adaptive attention fusion mechanism to obtain fused features; wherein the multi-layer non-negative matrix factorization operation adopts a multi-layer factorization structure, and potential factors are gradually increased layer by layer; and on the basis of the fusion features, learning styles are predicted through a deep classification network. And the learning style identification accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of learning style recognition, and particularly relates to a learning style recognition method, system, medium and device based on deep learning. BACKGROUND

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

[0003] Learning style is the habitual learning way and tendency of learners in the learning process, is the relatively stable learning way and preference formed in long-term learning activities, and is an important factor reflecting individual differences of learners. For example, in the learning scene of ideological and political courses: visual learners are good at acquiring knowledge from visual materials such as historical videos, theoretical illustrations, and special courseware of ideological and political courses; auditory learners tend to deepen their understanding through ideological and political lectures, online audio courses, and audio recitation of theories; kinesthetic learners will efficiently absorb knowledge in hands-on participation scenes such as ideological and political practice simulation, field red education base research and card punching, and interactive case study; balanced learners can adapt to various forms and can accumulate theories through ideological and political text reading, and can also master the content comprehensively through videos, audios or practical activities.

[0004] With the rapid development of online education platforms, learner behavior data presents the characteristics of diversity, high dimensionality and strong sparsity. For example, in the ideological and political learning scene, such data includes not only basic behaviors such as video watching, text reading and forum speaking, but also special data such as ideological and political special test answer tracks, practice task submission conditions and online group discussion interaction frequencies. How to accurately identify the learning style of students from massive learning behavior data has become a core technical challenge of personalized education recommendation systems. Learning style recognition can help education platforms provide targeted learning resources and paths for different types of learners, such as pushing ideological and political theme documentaries and theoretical visual illustrations for visual learners, recommending high-quality ideological and political podcasts and expert lecture audios for auditory learners, designing interactive ideological and political practice tasks and virtual research projects for kinesthetic learners, and integrating multiple forms of ideological and political learning packages for balanced learners, thereby improving learning efficiency and learning outcomes.

[0005] In the prior art, learning style recognition methods mainly include three categories based on traditional machine learning, deep learning and hybrid methods.

[0006] Traditional machine learning methods, such as logistic regression, support vector machine, random forest, etc., perform statistical analysis and classification modeling on the basic features of learners, such as learning duration, practice accuracy, resource access frequency, etc. These methods, although simple to calculate and highly interpretable, are prone to the curse of dimensionality when dealing with high-dimensional sparse data, and are difficult to capture the complex nonlinear relationships between learning behavior features.

[0007] In recent years, deep learning techniques have been introduced into the field of learning style recognition, and some progress has been made. Deep neural network (DNN) based methods can automatically learn deep feature representations of learning behavior data, and perform better than traditional methods in feature extraction and pattern recognition. Some studies have proposed using convolutional neural networks (CNN) and recurrent neural networks (RNN) to process sequential learning behavior data, capturing dynamic learning patterns of learners through time series modeling. However, existing deep learning methods still have obvious shortcomings when dealing with educational data.

[0008] First, the problem of rare learning style recognition under extreme class imbalance conditions. In actual educational scenarios, educational data often has strong imbalance characteristics, and the distribution of different learning styles is extremely uneven, for example, balanced learners may account for more than 60% of the total, while auditory learners may only account for about 5%. This extreme class imbalance causes traditional deep learning models to easily bias towards the majority class, and the recognition ability of rare learning styles is severely insufficient. Traditional class weight adjustment and resampling methods can alleviate this problem to some extent, but often at the expense of overall accuracy, and are prone to introducing noise and overfitting.

[0009] Second, the problem of effective feature representation and structural modeling of high-dimensional sparse educational data. Learning behavior data often contains 18-25 dimensional multi-type features, such as time-based features (video viewing duration, practice completion time), count-based features (message posting quantity, button click count), and proportion-based features (practice accuracy, resource skipping rate), etc. These features not only have high dimensionality and strong sparsity, but also have complex underlying relationships and hierarchical structures between learning behavior features. For example, the causal relationship between video viewing behavior and subsequent practice performance, the complementarity of different learning resource usage patterns, etc. Existing methods lack effective modeling of the intrinsic structure of features, and traditional fully connected neural networks, although theoretically capable of learning nonlinear mapping relationships between features, often fail to effectively exploit these underlying structured information in practical applications, resulting in limited feature representation capabilities.

[0010] In addition, the model decision-making process has the problems of explainability and educational applicability. In an educational application scenario, the learning style recognition result needs to provide a basis for teachers to formulate personalized teaching strategies, which requires the model not only to have good prediction performance, but also to have explainability. The existing technology has obvious defects in model explainability. Deep learning models are usually regarded as "black boxes", and their decision-making processes are difficult to explain and understand. In an educational application scenario, teachers and education managers need to understand the basis and logic of learning style recognition in order to formulate corresponding teaching strategies. Educators cannot understand which key features and behavior patterns the model is based on to judge the learning style of students. Models lacking explainability are difficult to gain the trust of educators, limiting their application in actual teaching scenarios.

[0011] Another important problem is the adaptive fusion and weight optimization of multi-modal learning features. Learning behavior data contains features of different properties and magnitudes. How to effectively fuse these heterogeneous features is a key to improving recognition accuracy. Traditional feature fusion methods, such as simple concatenation and weighted average, 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 have 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 results. Especially when dealing with multi-modal learning behavior data, how to effectively integrate time-based features (learning duration), count-based features (practice frequency), and binary features (whether to skip) of different properties is still a technical problem. SUMMARY

[0012] To solve the technical problems in the background art, the present application provides a learning style recognition method, system, medium and device based on deep learning, which introduces hierarchical NMF constraints in learning style recognition. Through a multi-layer decomposition structure with incremental latent factors at each level, learning behavior features are decomposed into explainable semantic representations layer by layer. The first layer captures basic behavior patterns, and the second layer models high-level learning strategies. This not only maintains the explainability of the features, but also provides structured prior knowledge for deep learning, effectively avoiding the "black box" problem of traditional deep learning and improving the accuracy of learning style recognition. Thus, it can provide precise red resource pushing, theoretical learning path recommendation and practice activity matching for students with different ideological and political learning styles, significantly improving the affinity and pertinence of course ideological and political education.

[0013] To achieve the above purpose, the present application adopts the following technical solutions: The first aspect of the present application provides a learning style recognition method based on deep learning, which comprises: Obtaining learning behavior characteristics of a learner, and obtaining embedding features through preprocessing; For the embedding features, gradually decompose into non-negative matrix factorization features through a multi-level non-negative matrix factorization operation, and fuse the embedding features and the non-negative matrix factorization features 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. Based on the fusion features, a learning style is predicted through a deep classification network.

[0014] Further, the learning behavior characteristics include time consumption characteristics and learning participation characteristics.

[0015] Further, the preprocessing includes standardization, feature embedding transformation and regularization.

[0016] Further, the multi-level non-negative matrix factorization operation adopts a non-negative matrix factorization constraint loss: ; wherein, is a sparsity regularization coefficient, represents an L1 norm, a first layer reconstruction loss , is a reconstruction feature of the first layer, is an embedding feature, a second layer reconstruction loss , is a reconstruction feature of the second layer.

[0017] Further, the adaptive attention fusion mechanism includes: merging the embedding features and the non-negative matrix factorization feature into a unified feature representation through a splicing operation; for the unified feature representation, calculating the importance weight of the feature dimension through an attention network containing two fully connected layers, and multiplying the importance weight and the unified feature representation element by element to obtain a weighted feature; for the weighted feature, processing through a feature transformation network to obtain the fusion feature.

[0018] Further, the deep classification network adopts a progressive dimension reduction structure, which is configured as a two-layer hidden layer design, and each layer contains linear transformation, layer normalization, activation function and regularization.

[0019] Further, a comprehensive loss function is adopted, and the comprehensive loss function adaptively combines the focal loss, the non-negative matrix factorization constraint loss and the confidence regularization loss.

[0020] The second aspect of the application provides a learning style recognition system based on deep learning, which comprises: The preprocessing module is configured to: acquire learning behavior features of a learner, and obtain embedding features through preprocessing; The feature extraction module is configured to: for the embedding features, gradually decompose into non-negative matrix factorization feature through a multi-level non-negative matrix factorization operation, and fuse the embedding 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. The classification module is configured to: based on the fusion features, predict learning styles through a deep classification network.

[0021] The third aspect of the present application 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 as described above.

[0022] The fourth aspect of the present application provides 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, wherein the processor implements the steps of the deep learning-based learning style recognition method as described above when executing the program.

[0023] Compared with the prior art, the present application has the following advantages: The present application introduces hierarchical NMF constraints in learning style recognition, and through a multi-level decomposition structure with latent factors increased layer by layer, the learning behavior features are decomposed layer by layer into interpretable semantic representations, the first layer captures the basic behavior patterns, and the second layer models the high-level learning strategies, which not only maintains the interpretability of the features, but also provides structured prior knowledge for deep learning, effectively avoiding the "black box" problem of traditional deep learning.

[0024] The present application designs a double-layer attention fusion mechanism, including channel attention and spatial attention, the channel attention calculates the importance of feature dimensions through global pooling and multi-layer perceptron, and the spatial attention allocates weights at the sample level, which can automatically adjust the feature weights according to different learning style characteristics, and compared with the traditional fixed weight fusion method, the accuracy and adaptability of the feature representation are significantly improved.

[0025] The present application innovatively combines the focal loss function and the adaptive class weight adjustment to solve the problem of extreme imbalance of learning style data. The focal loss reduces the loss contribution of easy-to-classify samples through a modulation factor, and focuses on rare samples; the adaptive weight is calculated based on the square root reciprocal of the class frequency, and provides stronger training signals for rare classes. The two mechanisms work together to significantly improve the recognition ability of rare learning styles while maintaining overall accuracy.

[0026] The application adopts a multi-layer fully connected network structure of progressive dimension reduction, and realizes step-by-step abstraction and compression of features through two-layer hidden layer design. Each layer is equipped with a complete processing flow, including linear transformation, layer normalization, ReLU activation function and Dropout regularization, to ensure the stability of feature transmission and the generalization ability of the model. Layer normalization effectively stabilizes the training process and accelerates convergence, while Dropout regularization prevents overfitting. This progressive dimension reduction design avoids information bottlenecks and achieves excellent classification performance with relatively small parameter amount, especially suitable for processing high-dimensional sparse features in education data. The network architecture is simple and efficient, which guarantees the expression ability of the model and controls the computational complexity, providing a good technical foundation for real-time learning style recognition.

[0027] The application is particularly suitable for ideological and political theory courses and course ideological and political teaching scenes, can accurately identify the visual, auditory, dynamic, balanced and other styles of students in ideological and political learning, and then realize the intelligent recommendation of ideological and political resources and the planning of learning path of 'one person one strategy', solve the problem of 'one-handed' in traditional ideological and political teaching while ensuring the effectiveness of ideological and political education, and has great political significance and social value. BRIEF DESCRIPTION OF DRAWINGS

[0028] The drawings accompanying the specification of the application part are used to provide further understanding of the application, the schematic embodiments of the application and the description thereof are used to explain the application, and do not constitute improper limitation on the application.

[0029] Figure 1 is a flowchart of the learning style recognition method based on deep learning of the first embodiment of the application; Figure 2 is a whole structure diagram of the learning style recognition method based on deep learning of the first embodiment of the application; Figure 3 is a structural schematic diagram of a computer device of the fourth embodiment of the application. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the embodiments of the application more clear, the technical scheme in the embodiments of the application will be described clearly and completely with reference to the drawings of the embodiments of the application.

[0031] It should be pointed out that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the application belongs.

[0032] Embodiment one The embodiment provides a learning style recognition method based on deep learning.

[0033] The primary problem to be solved by the present embodiment is how to significantly improve the recognition ability of rare learning styles while maintaining overall classification performance, achieving truly balanced classification.

[0034] The present embodiment needs to design a deep feature representation method that can model the relevance and hierarchy between features at the same time.

[0035] The present embodiment needs to provide interpretable feature decomposition and decision basis while ensuring classification accuracy, so that educators can understand and trust the model's judgment results.

[0036] The present embodiment needs to design a fusion mechanism that can automatically adjust feature weights according to different learning style characteristics.

[0037] In summary, the goal of the present embodiment is to solve the key problems of existing learning style recognition methods in class balance, feature representation, interpretability, fusion efficiency, and training stability, etc. through the introduction of non-negative matrix decomposition constraints, the design of adaptive feature fusion mechanism, the optimization of loss function and training strategy, etc. Finally, accurate, efficient, and interpretable learning style intelligent recognition is achieved.

[0038] The learning style recognition method based on deep learning provided by the present embodiment (referred to as Enhanced_NMF_EDPR) is specially used for automatic recognition of students' learning styles in online education platforms. It combines the interpretability advantage of non-negative matrix decomposition with the powerful feature learning ability of deep learning, and realizes accurate classification of students' multi-dimensional learning behaviors through hierarchical feature decomposition, adaptive fusion and focal loss optimization, providing accurate and reliable technical support for personalized education recommendation.

[0039] The learning style recognition method based on deep learning provided by the present embodiment, as shown in Figure 1 and Figure 2 , includes the following steps: Step 1, data input and preprocessing.

[0040] In the learning style recognition system, the behavior data of learners is input into the network in the form of a multi-dimensional feature vector.

[0041] Let the input learner behavior data matrix be , where represents the number of learner samples, and 18 represents the number of dimensions of learning behavior features. Each learner's feature vector contains rich learning behavior information, covering time consumption features such as T image (image content learning duration), T video (total video viewing time), T read (text reading time), and Taudio (audio material learning duration), and learning engagement features like N msgs posted (message posting number), N group discussions (number of times participating in group discussion), and N standard questions correct (number of standard questions answered correctly).

[0042] Since the learning behavior data has obvious scale difference, time-related features usually have large values in seconds or minutes, while accuracy-related features are distributed between 0 and 1, so data standardization is needed to ensure that each dimension of the feature is in the same numerical range. For each dimension of the original feature matrix , calculate its mean and standard deviation , and then perform Z-score standardization transformation on the feature of all samples in this dimension: ; where represents the standardized value of the th learner in the th feature dimension, is the sample mean of the feature, is the corresponding standard deviation, represents the original value of the th learner in the th feature dimension; the standardized feature matrix is obtained, ensuring the numerical compatibility of different types of features in subsequent processing.

[0043] The standardized feature data is then processed through a feature embedding transformation layer for dimension expansion and nonlinear mapping. The feature embedding transformation layer uses a fully connected network structure to map the 18-dimensional input features to a 64-dimensional embedding space, and the mapping process is defined as: ; where is the weight matrix of the embedding layer, is the bias vector, represents the ReLU (Rectified Linear Unit) activation function, represents the 18-dimensional input feature of the th learner, represents the embedding feature of the th learner. To improve training stability and feature expression quality, a layer normalization operation LayerNorm is added before the activation function, and its calculation formula is , where and are the mean and variance of the input vector , respectively, is a small constant to prevent division by zero, and are learnable scaling and offset parameters.

[0044] To prevent model overfitting and enhance the generalization ability, a regularization technique Dropout is applied after the feature embedding layer. Dropout randomly sets some neuron outputs to zero with a probability during training, and the calculation is: ; where is a mask vector following Bernoulli distribution , and represents element-wise multiplication. represents the normalized 64-dimensional embedding representation of the th learner.

[0045] After the above preprocessing process, the original 18-dimensional learning behavior features are converted into normalized 64-dimensional embedding representation , providing high-quality feature input for the subsequent non-negative matrix factorization constraint module.

[0046] Step 2, hierarchical non-negative matrix factorization (NMF) constraint.

[0047] The preprocessed 64-dimensional embedding features then enter the core innovative module, the hierarchical non-negative matrix factorization constraint module. Through multi-level non-negative matrix factorization operation, the input features are gradually decomposed into latent factor representation with clear semantic meaning, not only maintaining the interpretability advantage of traditional NMF method, but also realizing end-to-end optimization training through deep learning framework.

[0048] The hierarchical NMF module adopts a two-level decomposition structure, with 8 latent factors in the first layer and 16 factors in the second layer, forming a feature decomposition hierarchy from coarse to fine.

[0049] For the input embedding features , first, preliminary feature deconstruction is performed through the first layer NMF decomposition. The first layer decomposition includes two learnable parameters, weight matrix and factor matrix , where is realized through a linear transformation layer, and the network parameters are directly optimized. The calculation process of the first layer factor representation is: ; here represents the factor representation of the th sample in the first layer, and the ReLU activation function ensures the non-negativity constraint of the factor, consistent with the basic requirements of NMF. At the same time, in order to maintain the reconstruction property of decomposition, the reconstruction error is calculated as a constraint term: ; ; where and both represent the transpose of the matrix, is the reconstructed feature of the first layer, is the corresponding reconstruction loss.

[0050] Based on the first layer decomposition, the second layer NMF further refines the feature representation. With the factor representation of the first layer As input, the second layer decomposition employs 16 latent factors for more fine-grained pattern mining. The second layer contains a weight matrix and a factor matrix The computation process is as follows: the factor representation of the first layer for the i-th sample at the second layer ; where, denotes the transpose of a matrix, is the reconstructed feature of the second layer, is the corresponding reconstruction loss of the second layer; the second layer factor representation can capture more fine-grained learning behavior patterns, such as specific learning resource preference combinations, time allocation strategies, etc.

[0051] To enhance the sparsity of the model and improve the feature selection ability, a regularization constraint is added in each layer decomposition. The total NMF constraint loss function is defined as: ; where, is the sparsity regularization coefficient, and the L1 norm encourages the factor representation to be more sparse, which is conducive to discovering key learning behavior patterns.

[0052] Finally, the factor representations of the two layers are merged into the complete NMF feature through a concatenation operation: .

[0053] This hierarchical decomposition strategy not only maintains the inherent partiality and interpretability of the NMF method, but also achieves stronger feature learning ability through end-to-end training of deep learning. The 8 factors of the first layer mainly capture basic learning behavior patterns, such as time allocation preference, learning resource type selection, etc.; the 16 factors of the second layer further mine complex learning strategy combinations, such as the balance between active exploration and passive acceptance, the trade-off between deep learning and broad learning, etc., providing rich and meaningful feature representation for subsequent style classification.

[0054] Step 3, adaptive attention fusion.

[0055] After processing by the hierarchical NMF constraint module, the 64-dimensional original embedding feature and the 24-dimensional NMF factor feature ​​To make full use of the complementary information of the two different feature characteristics, an adaptive attention fusion module is designed, which realizes the intelligent integration of features through a dynamic weight distribution mechanism, ensuring that each type of feature can obtain the optimal fusion weight under different learning styles.

[0056] Step 301, the adaptive attention fusion module first combines the embedding features and the NMF features into a unified feature representation through a concatenation operation, assuming that embed dim =64 is the embedding feature dimension, nmf dim =24 is the NMF feature dimension, and the dimension of the fused feature is total dim =88, and the fused feature representation is: where Concat represents the concatenation operation.

[0057] Step 302, the attention weight calculation adopts a multi-layer perceptron structure, which learns the importance distribution between features through an attention network containing two fully connected layers.

[0058] The first layer maps the 88-dimensional input to a 64-dimensional hidden space and applies a Tanh (hyperbolic tangent) activation function to obtain a hidden representation: ; where W is the first layer attention weight matrix, is the corresponding bias vector; the selection of the Tanh activation function helps to produce bounded activation values, avoiding the problem of gradient explosion.

[0059] The second layer maps the hidden representation back to the original dimension and generates normalized attention weights through a Softmax (normalized exponential) function: ; ; where W is the second layer weight matrix, is the bias vector of the second layer, is the final attention weight vector, represents the mapping of the hidden representation back to the original dimension, represents the jth dimension of the hidden representation mapped back to the original dimension; the Softmax normalization ensures that the sum of all weight components is 1, i.e. , A i,j is the jth dimension of the data in A i .

[0060] By applying the obtained attention weight vector to the unified feature representation , the weighted feature : where, denotes element-wise multiplication, is the 88-dimensional attention weight vector obtained in step 302.

[0061] Step 303, the weighted features are then further processed by a feature transformation network, which contains a combination of linear transformation, layer normalization, and ReLU activation, to obtain the fused features: ; where, is the transformation matrix, denotes the bias vector, and layer normalization LayerNorm is calculated by the following formula: ; where, and are the mean and variance of the input vector , respectively, is a numerical stability constant, and are learnable scaling and shifting parameters, denotes the Hadamard product (element-wise multiplication); layer normalization helps stabilize the training process and accelerate model convergence.

[0062] Step 304, the fusion module simultaneously outputs the fused feature representation and the attention weight distribution, providing support for the explainability analysis of the model, and the output feature representation of the fusion module is: ; the attention weight can intuitively reflect the contribution of different feature dimensions to the current sample classification decision. By analyzing the distribution pattern of the attention weight, we can understand the key feature combination that the model relies on when identifying different learning styles. For example, for visual learners, feature dimensions related to images and videos usually have higher attention weights; for kinesthetic learners, exercise and interaction-related features are more important.

[0063] The adaptability of this fusion mechanism is reflected in the dynamic adjustment ability of the attention weight. Unlike fixed weight feature fusion methods, the adaptive attention mechanism can automatically adjust the importance of each dimension according to the feature distribution of the input sample, so that the model can allocate personalized feature weight combinations for each learner. This design not only improves the quality of feature representation, but also enhances the model's adaptability to different learning style patterns, providing high-quality fused feature input for subsequent deep classification networks.

[0064] Step 4, deep classification network.

[0065] the 88-dimensional fused features Subsequently, the deep classification network is entered to make the final learning style prediction.

[0066] The deep classification network adopts a dense connection architecture design, and realizes the nonlinear mapping from the fusion features to the learning style categories through a multi-layer fully connected network. The deep classification network adopts a progressive dimension reduction structure, and is configured as two layers of hidden layers. The deep classification network of each layer contains a complete processing flow of linear transformation, layer normalization, activation function and Dropout regularization.

[0067] Step 401, the first layer (hidden layer 1) maps the 88-dimensional fusion features to the first hidden space of 128 dimensions: ; Wherein, is the first layer weight matrix, is the corresponding bias vector, represents the first hidden space feature.

[0068] The original linear output is then standardized by layer normalization LayerNorm: ; Wherein, and are the mean and variance of the first layer output, and are the learnable normalization parameters, represents the standardization processing result.

[0069] After layer normalization, the ReLU activation function is applied to introduce nonlinearity, and the nonlinear result is obtained: ; In order to prevent overfitting and enhance the generalization ability of the model, Dropout regularization is applied after activation, the dropout rate p=0.3, and the first layer processing result is obtained: .

[0070] Step 402, the second layer network (hidden layer 2) further compresses the 128-dimensional features to 64 dimensions, and the processing flow is similar to the first layer: ; ; ; ; Wherein, is the second layer weight matrix, is the corresponding bias vector, represents the second hidden space feature, represents the second layer normalization result, represents the second layer nonlinear result, represents the second layer network output.

[0071] Step 403, the classifier layer maps the 64-dimensional deep feature to the final learning style category number n classes = 4: ; wherein, is the classifier weight matrix, is the classification bias vector, is the final category logits output.

[0072] The weight initialization of the deep classification network uses Xavier normal distribution, and different gain coefficients are set for different layers to optimize the training effect.

[0073] For the hidden layer, the standard Xavier initialization with gain coefficient gain = 1.0 is used: ; wherein, ~ represents the distribution, represents the normal distribution, is the number of input neurons, is the number of output neurons, represents the hidden layer weight.

[0074] For the output classification layer, a smaller gain coefficient gain = 0.5 is used to avoid gradient explosion in the early stage of training: ; wherein, represents the classification layer weight.

[0075] All bias vectors are initialized to zero vectors.

[0076] The final output of the deep classification network contains the following key components for subsequent loss calculation and model analysis: the classification prediction output is , representing the prediction score of the four learning style categories; the factor obtained by hierarchical NMF decomposition is , which retains the interpretable learning behavior pattern information; the reconstruction loss of the NMF module , used to constrain the quality of feature decomposition; the attention weight of the fusion module , providing the basis for visual analysis of feature importance; the original feature embedding representation , retaining the basic information of the input feature.

[0077] This multi-output design not only supports end-to-end supervised learning, but also provides rich intermediate representations for model explainability analysis and multi-task learning. Through the dense connection architecture and the carefully designed regularization strategy, the deep classification network can accurately identify and classify complex learning style patterns while maintaining good generalization ability. The progressive dimension reduction design of the network structure effectively avoids the information bottleneck problem, and the combination of layer normalization and Dropout ensures the stability of the training process and the robustness of the model.

[0078] Step 5, loss function design.

[0079] In order to effectively handle the class imbalance problem in learning style recognition and integrate multiple training objectives, this embodiment designs a comprehensive loss function that adaptively combines focal loss, NMF reconstruction loss, and confidence regularization loss, achieving collaborative optimization of classification accuracy, feature explainability, and model robustness.

[0080] (1) Focal loss calculation.

[0081] In view of the problem of extreme class imbalance in learning style data, focal loss is used as the main classification loss function. Focal loss dynamically adjusts the loss weights of different samples by introducing a modulation factor, so that the model focuses on difficult-to-classify rare samples.

[0082] For the predicted logits and the true label of the i-th sample, first calculate the softmax probability distribution: ; where represents the predicted probability of the model that sample i belongs to the true label , and C represents the total number of classes; is the logits value of sample i corresponding to the label ; represents the logits value of sample i corresponding to the true label ; for the true class , , is the predicted probability of the model for the true class.

[0083] The calculation formula of a single focal loss is: ; where is the modulation factor, which controls the loss attenuation degree of easy-to-classify samples; is the class balance weight, which sets different weight coefficients for different classes; for rare classes with fewer sample quantities, The values are larger, ensuring that these samples get enough attention in training; the total focus loss is the average of all sample losses: .

[0084] (2) NMF reconstruction loss.

[0085] To maintain the quality and interpretability of the feature decomposition, the reconstruction loss produced by the NMF module is added as a constraint term to the total loss function. The NMF reconstruction loss contains the reconstruction error of two-layer decomposition and sparsity regularization: ; where, is the reconstructed feature of the 1st layer, is the corresponding factor representation, is the non-negative factor matrix, 0.01 is the sparsity regularization coefficient. The L2 reconstruction error ensures that the decomposition result can effectively recover the original feature, while the L1 regularization promotes the sparsity of the factor representation, which is beneficial for discovering key learning behavior patterns.

[0086] (3) Confidence regularization loss.

[0087] To improve the confidence and robustness of the model's prediction, a confidence regularization term is introduced. This term avoids the model's overconfidence by maximizing the entropy of the prediction probability distribution, and the calculation formula is: ; where, represents the prediction probability of the model that sample i belongs to label , is a numerical stability constant. Confidence regularization encourages the model to produce a relatively smooth probability distribution, avoiding overfitting during the training process.

[0088] (4) Adaptive weight adjustment strategy.

[0089] To dynamically balance the influence of different loss terms during training, an adaptive weight adjustment mechanism is designed. The NMF constraint weight is linearly decayed according to the training progress: ; where, is the initial NMF weight coefficient, epoch is the current training round, and max_epochs is the total training round. This decay strategy makes the model focus on feature decomposition quality in the early training stage and gradually shift to classification performance optimization in the later stage.

[0090] (5) Comprehensive loss function.

[0091] The final total loss function combines the three loss terms by weighting: ; wherein, is the confidence regularization coefficient.

[0092] The loss function returns detailed information of each loss term for training monitoring: represents the total loss value for backpropagation optimization; represents the focal loss value reflecting the classification performance; represents the NMF reconstruction loss measuring the quality of feature decomposition; represents the confidence loss evaluating the model robustness; represents the current adaptive weight for monitoring the weight adjustment process.

[0093] Through the design of this multi-objective loss function, the model can maintain the interpretability of feature decomposition and the confidence of prediction while ensuring the classification accuracy. The adaptive weight adjustment strategy ensures the stability of the training process, avoiding conflicts between different loss terms, so that the model can receive the optimal solution with both performance and interpretability.

[0094] Step 6, training strategy and parameter optimization.

[0095] To ensure the stability of the model and achieve optimal performance, this embodiment adopts a multi-level learning rate configuration training strategy. Considering the differences in learning characteristics of different network components, the model parameters are divided into five optimization groups and set different learning rates. The NMF constraint module, due to the numerical stability of matrix decomposition, is set to a small learning rate: wherein, represents the standard learning rate.

[0096] The learning rate of the input embedding layer and the attention fusion module is the standard learning rate: .

[0097] The classifier layer is set to a larger learning rate to accelerate convergence: .

[0098] The optimization process uses the AdamW optimizer, and the weight decay coefficient is set to .

[0099] During training, a learning rate scheduler is used. When the validation loss does not improve significantly for 15 consecutive epochs (training times), the learning rate is adjusted according to the following formula: , represents the current learning rate used by the model (such as 0.001).

[0100] The trigger condition for learning rate scheduling is that the improvement amplitude of the validation loss is less than the minimum threshold: ; represents the historical lowest loss achieved by the model on the validation set so far; represents the loss of the current epoch of the model on the validation set; represents the improvement margin of the current loss compared to the historical best loss.

[0101] To prevent gradient explosion, maximum norm gradient clipping is applied to all parameters: ; represents all the learnable parameters (such as weights W and biases b) in the model, represents the "gradient" in mathematics, which is the direction and magnitude that the parameters should be updated. represents the L2 norm, which can be understood as calculating the "total length" or "total size" of this gradient vector.

[0102] The early stopping mechanism monitors the rare class F1 score (balanced F score) and triggers early stopping when there is no improvement in the performance of the rare class for 25 consecutive epochs and the improvement margin is less than 0.001. The training batch size is set to 32 and the total training round limit is set to 150 rounds. Through this carefully designed training strategy, the model can effectively improve the recognition ability of rare learning styles while maintaining overall classification performance, and finally converge to an optimal state with both performance and stability.

[0103] This embodiment has achieved significant technical advantages in the learning style recognition task by combining hierarchical non-negative matrix factorization constraints with deep learning architecture. Through comparative experiments with the random forest benchmark method, this technology has achieved substantial improvements in multiple key performance indicators.

[0104] Table 1, Algorithm Effect Table in this Embodiment

[0105] As shown in Table 1, in terms of overall accuracy, the Enhanced_NMF_EDPR method of this embodiment achieves 0.864, which is 1.4% higher than the 0.852 of the random forest, indicating that the method of this embodiment has better recognition accuracy while maintaining high overall classification performance. More importantly, in the key technical difficulty of rare class recognition, this embodiment has made a breakthrough. The rare class F1 score reaches 0.633, while the random forest method cannot recognize the rare class at all (F1=0.000), which fully proves the effectiveness of the method of this embodiment in solving the problem of extreme class imbalance. The significant improvement in rare class recognition ability is of great significance for personalized education applications, ensuring that students with learning styles such as auditory and kinesthetic can also receive accurate style recognition and personalized recommendations.

[0106] On the macro-averaged F1 indicator, the embodiment achieves 0.756, which is much higher than the 0.435 of the random forest, with an increase of 73.8%. The macro-averaged F1 can fairly evaluate the recognition ability of the model to all classes. The significant improvement of the indicator shows that the method of the embodiment not only performs excellently on the majority classes, but more importantly, effectively identifies the rare classes, embodying the balanced classification ability of the model. The weighted F1 score reaches 0.855, which is 3.0% higher than the 0.830 of the random forest, further verifying the superiority of the method of the embodiment in comprehensive performance.

[0107] The technical advantages of the embodiment mainly reflect in three aspects: first, the hierarchical NMF constraint mechanism decomposes the learning behavior features into two layers through 8+16 factors, which not only maintains the interpretability of the features, but also provides rich semantic representation, effectively solving the "black box" problem of traditional deep learning; second, the adaptive attention fusion mechanism can dynamically adjust the feature weights according to different learning styles, significantly improving the adaptability of the model to complex learning patterns; finally, the focal loss and adaptive class weight optimization strategy are designed specifically for the class imbalance problem, ensuring the effective identification of rare learning styles.

[0108] The experimental results show that the embodiment not only maintains a competitive advantage in the traditional overall accuracy indicator, but also makes a qualitative breakthrough in the key technical challenge of rare class identification, providing an accurate, reliable, and balanced learning style recognition technology solution for personalized education recommendation systems.

[0109] The embodiment is specifically applied to the automatic identification and classification of students' learning styles in online education platforms. Through innovative modules such as hierarchical non-negative matrix factorization constraint, adaptive feature fusion, focal loss optimization, and multi-scale attention mechanism, the embodiment realizes the deep mining of students' multi-dimensional learning behavior features and the accurate identification of learning styles, improving the accuracy and effectiveness of personalized education recommendation.

[0110] The embodiment first introduces hierarchical NMF constraint in learning style recognition, decomposing learning behavior features into two layers through 8 and 16 factors, and decomposing them into interpretable semantic representations layer by layer. The first layer captures basic behavior patterns (learning time preference, resource usage habit), and the second layer models high-level learning strategies (active exploration vs. passive acceptance). This design not only maintains the interpretability of the features, but also provides structured prior knowledge for deep learning, effectively avoiding the "black box" problem of traditional deep learning.

[0111] The embodiment designs a double-layer attention fusion mechanism, including channel attention and spatial attention. The channel attention calculates the importance of feature dimensions through global pooling and multi-layer perceptron, and the spatial attention allocates weights at the sample level. The mechanism can automatically adjust the feature weight according to different learning style characteristics, and significantly improves the accuracy and adaptability of feature representation compared with the traditional fixed weight fusion method.

[0112] The embodiment innovatively combines the focal loss function and adaptive class weight adjustment to solve the problem of extreme imbalance of learning style data. The focal loss reduces the loss contribution of easy-to-classify samples through a modulation factor , and focuses on rare samples; the adaptive weight is calculated based on the square root reciprocal of the class frequency, providing stronger training signals for rare classes. The two mechanisms work together to significantly improve the recognition ability of rare learning styles while maintaining overall accuracy.

[0113] The embodiment adopts a progressive dimension reduction multi-layer fully connected network structure, which realizes step-by-step abstraction and compression of features through a two-layer hidden layer design of 128→64. Each layer is equipped with a complete processing flow, including linear transformation, layer normalization, ReLU activation function and Dropout regularization, to ensure the stability of feature transmission and the generalization ability of the model. Layer normalization effectively stabilizes the training process and accelerates convergence, while Dropout regularization prevents overfitting. This progressive dimension reduction design avoids information bottlenecks and achieves excellent classification performance with relatively small parameter amount, especially suitable for processing high-dimensional sparse features in educational data. The network architecture is simple and efficient, ensuring the expression ability of the model and controlling the computational complexity, providing a good technical foundation for real-time learning style recognition.

[0114] The embodiment proposes a three-stage training strategy of preheating, balancing and fine-tuning. The preheating stage trains the NMF module to establish a stable feature decomposition foundation; the balancing stage introduces a classification loss and dynamically adjusts the constraint weight; the fine-tuning stage is optimized end-to-end. This phased strategy avoids conflicts between different loss terms, ensures stable convergence of the model, and has better performance and convergence than traditional end-to-end training.

[0115] Through the organic combination of the above innovative technologies, the embodiment is significantly better than existing methods in rare class recognition, overall classification accuracy and model interpretability, providing a practical and reliable intelligent solution for educational technology applications.

[0116] The embodiment can balance feature expression ability, model interpretability and class balance processing, effectively handle high-dimensional sparse educational data, and provide reliable technical support for educational practice.

[0117] The embodiment is mainly applied to a personalized recommendation system of an online education platform, an intelligent learning management system, an adaptive learning software, and an education data analysis tool. In these products, the embodiment can perform real-time analysis on multi-dimensional behavior data of learners, automatically identify the learning style types (visual type, auditory type, kinesthetic type, and balanced type) of the learners, provide accurate learner portraits for the platform, and thus realize personalized course recommendation of the online learning platform, learning path planning of online education, content matching of enterprise training systems, exercise pushing of intelligent question banks, and learning strategy suggestions of teaching auxiliary software, and the like. In addition, the embodiment also has important application value in a learning behavior analysis tool of an education research institution, a teacher teaching auxiliary system, and an education evaluation platform, and can help education workers better understand the characteristics of students and develop targeted teaching plans.

[0118] Embodiment Two The embodiment provides a learning style recognition system based on deep learning, which specifically comprises: a preprocessing module configured to obtain learning behavior characteristics of learners and obtain embedding features through preprocessing; a feature extraction module configured to gradually decompose the embedding features into non-negative matrix factorization feature through a multi-level non-negative matrix decomposition operation, and fuse the embedding 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 decomposition operation adopts a multi-level decomposition structure, and the latent factors are increased layer by layer; a classification module configured to predict learning styles based on the fusion features through a deep classification network.

[0119] It should be noted that each module in the embodiment corresponds to each step in Embodiment One, and the specific implementation process is the same, which will not be repeated here.

[0120] Embodiment Three The embodiment provides a computer readable storage medium having a computer program stored thereon, and the program is executed by a processor to realize the steps in the learning style recognition method based on deep learning as described in Embodiment One.

[0121] Embodiment Four The embodiment provides a computer device, which comprises a processor and a memory. Figure 3As shown, it comprises a computer readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored in the computer readable storage medium 1003 and capable of running on the processor 1001, wherein the processor 1001, the communication interface 1002 and the computer readable storage medium 1003 are connected through a bus or other means. The communication interface 1002 is configured to receive and send data, and the processor 1001 implements the steps of the deep learning-based learning style recognition method according to the program.

[0122] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. The present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A learning style recognition method based on deep learning, characterized by, The method comprises the following steps: obtaining learning behavior characteristics of a learner, and obtaining embedding features through preprocessing; for the embedding features, gradually decomposing into non-negative matrix factorization feature through multi-level non-negative matrix factorization operation, and fusing the embedding 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; based on the fusion features, predicting learning styles through a deep classification network. 2.The deep learning-based learning style recognition method of claim 1, wherein, The learning behavior characteristics include time consumption characteristics and learning participation characteristics. 3.The deep learning-based learning style recognition method of claim 1, wherein, The preprocessing includes standardization, feature embedding transformation and regularization. 4.The deep learning-based learning style recognition method of claim 1, wherein, The multi-level non-negative matrix factorization operation adopts a non-negative matrix factorization constraint loss: ； wherein, is a sparsity regularization coefficient, denotes the L1 norm, a first layer reconstruction loss , is a reconstructed feature of the first layer, is an embedding feature, a second layer reconstruction loss , is a reconstructed feature of the second layer. 5.The deep learning-based learning style recognition method of claim 1, wherein, The adaptive attention fusion mechanism includes: merging the embedding features and the non-negative matrix factorization feature into a unified feature representation through a splicing operation; for the unified feature representation, calculating the importance weight of the feature dimension through an attention network containing two fully connected layers, and multiplying the importance weight and the unified feature representation element by element to obtain a weighted feature; for the weighted feature, processing through a feature transformation network to obtain the fusion feature. 6.The deep learning-based learning style recognition method of claim 1, wherein, The deep classification network adopts a progressive dimension reduction structure, is configured as a two-layer hidden layer design, and each layer contains linear transformation, layer normalization, activation function and regularization. 7.The deep learning-based learning style recognition method of claim 1, wherein, A comprehensive loss function is adopted, and the comprehensive loss function adaptively combines the focal loss, the non-negative matrix factorization constraint loss and the confidence regularization loss.

8. A learning style recognition system based on deep learning, characterized by, The method comprises the following steps: a preprocessing module configured to obtain learning behavior characteristics of a learner, and obtain embedding features through preprocessing; a feature extraction module configured to, for the embedding features, gradually decompose into non-negative matrix factorization feature through multi-level non-negative matrix factorization operation, and fuse the embedding 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; a classification module configured to, based on the fusion features, predict learning styles through a deep classification network.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the steps in the learning style recognition method based on deep learning in any one of claims 1-7.

10. 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, wherein, The processor executes the program to realize the steps in the learning style recognition method based on deep learning in any one of claims 1-7.

Images