Educational resource sequence recommendation method and related device

Abstract

The invention provides an educational resource sequence recommendation method and a related device, and relates to the technical field of resource recommendation. After student learning behavior data, learning resource attribute data and knowledge graph structure information are obtained, a regularization matrix decomposition technology is adopted to carry out potential factorization modeling on a student-resource interaction scoring matrix, and basic preference embedding features of students and resources are extracted. Aiming at diversity and complexity of student learning behaviors, a multi-view feature coding module is designed, learning sequence and short-term interest features are extracted from a behavior sequence view, static attribute features of students and courses are extracted from an attribute information view, and knowledge point pre-correction dependence and association relationship features between courses are extracted from a knowledge structure view. In order to improve the consistency and discrimination of feature representation, on the basis of multi-view fusion features, a comparative learning mechanism is introduced, and the discrimination and migration modeling ability of the model to student interests and preferences is enhanced by constructing a positive and negative sample comparison loss function.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of resource recommendation, in particular to an educational resource sequence recommendation method and related device. BACKGROUND

[0002] The educational resource recommendation system is widely used in online education platforms, smart campuses, MOOC learning systems, learning management systems and other education informationization scenarios. Its goal is to intelligently recommend personalized learning resources and dynamic learning paths according to students' historical learning behaviors, knowledge mastery states, interest preferences, etc.

[0003] The current mainstream education recommendation system technology route mainly includes: collaborative filtering (Collaborative Filtering, CF); matrix factorization (Matrix Factorization, MF); deep learning (such as neural collaborative filtering NCF, Transformer sequence model, etc.); graph neural network (GNN); attention mechanism (Attention-based Recommendation).

[0004] The matrix factorization model is widely used in education recommendation, which can embed students and courses into a latent factor space to model their interest preferences. However, the traditional matrix factorization model has limited modeling ability when facing high sparsity, high dynamics, and diversified learning behaviors, making it difficult to fully express complex learning features. At the same time, sequence modeling is gradually important in education recommendation systems. Students' learning behaviors are essentially a dynamic evolution process, with long and short term dependencies. In recent years, although sequence modeling methods have been applied in education recommendation systems, there are still problems of high sequence sparsity and serious learning bias, and the stability and generalization ability of the model need to be improved. SUMMARY

[0005] In order to solve the problems of high sequence sparsity and serious learning bias, the present application provides an educational resource sequence recommendation method and related device.

[0006] In a first aspect, the present application provides an educational resource sequence recommendation method using the following technical solution: An educational resource sequence recommendation method, comprising: Collecting interaction behavior data of students and course resources, including learning duration, completion status and learning achievement to construct a behavior sequence; Constructing a student-course rating matrix based on the behavior sequence, and extracting latent feature representations of students and courses using matrix factorization technology; Extracting features from three dimensions of behavior sequence, knowledge graph and attribute information in parallel to generate multi-dimensional extracted features; The matrix decomposition result is fused with the multi-dimensional extracted features, and the feature representation is optimized through a contrast learning mechanism; The fused features are subjected to frequency domain analysis and filtering processing to strengthen the periodic characteristics and long-term dependencies in the sequence to generate a target feature sequence; A multi-head attention mechanism is used to model the target feature sequence to generate a target model and capture long-term dependencies in learning behavior; Based on the target model, a course recommendation probability is calculated, and a knowledge graph correlation degree is combined for sorting and optimization to generate a personalized learning resource recommendation list.

[0007] Optionally, after the step of collecting the interaction behavior data of the students and the course resources, the method further comprises: Filtering abnormal learning behavior data; Using mean filling processing for missing values; Building a complete behavior sequence in chronological order.

[0008] Optionally, the step of constructing a student-course rating matrix based on the behavior sequence and extracting latent feature representation of the students and the courses using matrix decomposition technology comprises: The rating matrix considers learning completion and learning performance; Using a regularized matrix decomposition method to prevent overfitting; Using a smoothing initialization mechanism for new students or courses.

[0009] Optionally, the step of extracting features from three dimensions of behavior sequence, knowledge graph, and attribute information in parallel to generate multi-dimensional extracted features comprises: Extracting learning order and short-term interest features from the behavior sequence dimension as behavior sequence features; Extracting prerequisite relationships and knowledge point association features between courses from the knowledge graph dimension as knowledge graph features; Extracting student personal attributes and course metadata features from the attribute information dimension as attribute features; The behavior sequence features, the knowledge graph features, and the attribute features are multi-dimensional extracted features.

[0010] Optionally, the behavior sequence features include course ID, time information, and learning performance; The knowledge graph features are extracted through a graph convolution network; The attribute features include student grade, major, and course difficulty level.

[0011] Optionally, the step of fusing the matrix decomposition result with the multi-dimensional extracted features and optimizing the feature representation through a contrast learning mechanism further comprises: Positive samples are selected from courses with similar interests or knowledge graph associations. Negative samples are selected from courses with large differences in interests or no associations. The contrast loss function is adjusted by the temperature coefficient to optimize the features. Optionally, in the step of modeling the target feature sequence using a multi-head attention mechanism to generate a target model and capturing long-term dependencies in learning behavior, the step further includes: The number of attention heads and the number of contrast learning perspectives maintain a proportional relationship. The matrix decomposition result is introduced as a bias term for attention calculation. Residual connections and layer normalization are used to stabilize the training process.

[0012] In a second aspect, the present application provides an educational resource sequence recommendation device, comprising: A data acquisition module is configured to acquire interactive behavior data of students and course resources, including learning duration, completion status and learning performance to construct a behavior sequence. A matrix module is configured to construct a student-course rating matrix based on the behavior sequence, and extract latent feature representations of students and courses using matrix decomposition techniques. A multi-dimensional feature extraction module is configured to extract features from three dimensions of behavior sequence, knowledge graph and attribute information in parallel to generate multi-dimensional extracted features. A fusion module is configured to fuse the matrix decomposition result with the multi-dimensional extracted features, and optimize the feature representation through a contrast learning mechanism. A target feature sequence module is configured to perform frequency domain analysis and filtering processing on the fused features, and strengthen the periodic features and long-term dependencies in the sequence to generate a target feature sequence. A relationship capturing module is configured to model the target feature sequence using a multi-head attention mechanism to generate a target model and capture long-term dependencies in learning behavior. A result generation module is configured to calculate course recommendation probabilities based on the target model, sort and optimize based on knowledge graph association degrees, and generate a personalized learning resource recommendation list.

[0013] In a third aspect, the present application provides a computer device, comprising a memory and a processor, wherein the processor executes the computer instructions stored in the memory to perform the method described above.

[0014] In a fourth aspect, the present application provides a computer-readable storage medium comprising instructions, which, when executed on a computer, cause the computer to perform the method described above.

[0015] Based on the above description, after obtaining the student learning behavior data, learning resource attribute data and knowledge graph structure information, the regularization matrix decomposition technology is used to perform latent factor decomposition modeling on the student-resource interaction score matrix, and the basic preference embedding features of the student and the resource are extracted. In view of the diversity and complexity of the student learning behavior, a multi-view feature coding module is designed, and learning order and short-term interest features are extracted from the behavior sequence view, static attribute features of the student and the course are extracted from the attribute information view, and knowledge point prerequisite dependency and correlation features between courses are extracted from the knowledge structure view. In order to improve the consistency and discriminability of the feature representation, a contrast learning mechanism is introduced based on the multi-view fusion features, a positive and negative sample contrast loss function is constructed, and the discrimination and transfer modeling ability of the model to the student interest preference is strengthened. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a computer device structure schematic diagram of a hardware running environment related to an embodiment scheme of the present application. Figure 2 is a flow schematic diagram of a first embodiment of an educational resource sequence recommendation method of the present application. Figure 3 is a structural block diagram of a first embodiment of an educational resource sequence recommendation device of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below through the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0018] Referring to Figure 1 , Figure 1 is a computer device structure schematic diagram of a hardware running environment related to an embodiment scheme of the present application.

[0019] As Figure 1As shown, the computer device can include a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to realize the connection communication between these components. The user interface 1003 can include a display, an input unit such as a keyboard, and the optional user interface 1003 can further include a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a wireless fidelity (Wi-Fi) interface). The memory 1005 can be a high-speed random access memory (RAM), and can also be a stable non-volatile memory (NVM), such as a disk memory. The memory 1005 can also be an independent storage device from the aforementioned processor 1001.

[0020] Those skilled in the art can understand that Figure 1 The structure shown in the figure does not constitute a limitation on the computer device, and can include more or fewer components than the figure, or combine certain components, or different component arrangements.

[0021] As Figure 1 As shown, the memory 1005 as a storage medium can include an operating system, a network communication module, a user interface module, and an educational resource sequence recommendation program.

[0022] In Figure 1 In the computer device shown, the network interface 1004 is mainly used for data communication with a network server; the user interface 1003 is mainly used for data interaction with a user; the processor 1001 and the memory 1005 in the present application can be arranged in the computer device, and the computer device calls the educational resource sequence recommendation program stored in the memory 1005 through the processor 1001, and executes the educational resource sequence recommendation method provided in the present application.

[0023] The present application provides an educational resource sequence recommendation method, which refers to Figure 2 , Figure 2 The flowchart of the first embodiment of the educational resource sequence recommendation method of the present application.

[0024] In this embodiment, the educational resource sequence recommendation method includes the following steps: Step S10: Collecting the interaction behavior data of students and course resources, including learning duration, completion status and learning performance to construct a behavior sequence.

[0025] It should be noted that the existing technology mainly has the following technical problems: 1. Existing matrix factorization models are prone to interest drift and cold start issues in highly sparse educational scenarios: In educational recommendation systems, especially in the early stages of online education platforms, the interaction data between students and course resources is naturally highly sparse. Some students have just registered and have only a few interaction records. Traditional matrix factorization models struggle to fully train latent factors in this sparse data environment, which can easily lead to cold start issues. Furthermore, students' learning interests tend to change dynamically, making static matrix factorization modeling difficult to capture interest drift, resulting in reduced model stability and prediction accuracy.

[0026] 2. Existing sequence recommendation models ignore the fusion of multi-perspective features of learning behavior and are unable to model the dynamic changes of students' interests: Some existing sequence recommendation models only model based on the student behavior sequence itself, failing to effectively integrate multi-dimensional features such as student attribute characteristics, course knowledge point structure, and course content semantics. As a result, the model lacks the ability to model deep-level behaviors such as learning motivation, cognitive level, and learning path, and cannot accurately reflect the dynamic changes of students' interests and the evolution of learning rhythm.

[0027] 3. Contrastive learning has not yet been effectively integrated with matrix factorization and sequence modeling, resulting in training instability and insufficient feature representation. Using contrastive learning alone fails to fully incorporate the explicit interaction information in the student-course rating matrix. Without effective integration with matrix factorization and sequence modeling, this can lead to conflicting training objectives, unstable optimization, and an inability to consistently improve feature representation capabilities in complex and sparse data scenarios.

[0028] 4. Existing education recommendation systems lack a stable, efficient, and generalizable training framework: Recommendation systems in educational settings require long-term online service. This requires not only efficient and rapid training but also robustness and good generalization. Existing systems still have deficiencies in training architecture design, parameter update strategies, and online model fine-tuning mechanisms, limiting the effectiveness of models in practical deployments and the system's scalability.

[0029] It should be noted that after the step of collecting the interactive behavior data between students and course resources, the following steps are also included: filtering abnormal learning behavior data; using mean filling processing for missing values; and constructing a complete behavior sequence in chronological order.

[0030] In specific implementation, raw data collection will collect student-resource interaction logs: ; Course attribute metadata; Student basic information data; Knowledge graph data (prerequisite relationship, knowledge point dependency graph).

[0031] in For the students; is the course; is the timestamp; is the duration of learning; is the completion status, whose value is (0 / 1); is the course score or test score.

[0032] It should be noted that the course attribute metadata belongs to the course side static feature parameter. Typical examples include course difficulty level, course major category, and course credit / hour operation processing mode: first, the discrete attribute (such as major category) is embedded and coded, and the continuous attribute (such as difficulty level) is standardized; then it is mapped into a course metadata vector is spliced with the course embedding vector, which is used for multi-perspective feature fusion.

[0033] The student basic information data belongs to the user side static feature parameter. Typical examples include grade, major, and historical average score level. Operation processing mode: student category attribute (such as grade, major) → embedding; student continuous attribute (such as average score) → standardization; get student attribute vector .

[0034] The knowledge graph is represented as a graph , where the nodes are knowledge points / courses, and the edges are prerequisite or dependency relationships; the node representation is calculated through a graph convolution network (GCN) or graph embedding method to obtain knowledge graph features In the fusion stage, it is spliced with the behavior sequence feature and the attribute feature to generate multi-dimensional extraction features. It is fused into the student latent factor representation obtained by matrix decomposition as a supplementary input feature.

[0035] The three types of parameter data (course metadata, student basic information, and knowledge graph structure) are input into the fusion module together with the behavior sequence feature and the matrix decomposition latent factor; Fusion method: splicing + projection (such as MLP), and enhance the consistency and discriminability under the optimization of contrastive learning; in the subsequent sequence modeling stage, these fusion features will enter the Fourier frequency domain enhancement and multi-head attention mechanism to participate in the generation of the recommendation result.

[0036] The data is de-duplicated, and abnormal behaviors are filtered (such as ), and missing values are filled: where is the mean fill value, represents the value of the th sample.

[0037] The behavior sequence is constructed and sorted in chronological order to construct the behavior sequence:​ in is the sequence length.

[0038] Step S20: constructing a student-course rating matrix based on the behavior sequence, and using matrix decomposition technology to extract potential feature representations of students and courses.

[0039] It should be noted that the steps of constructing a student-course rating matrix based on the behavior sequence and using matrix decomposition technology to extract potential feature representations of students and courses include: the rating matrix comprehensively considers learning completion and academic performance; using a regularized matrix decomposition method to prevent overfitting; and using a smooth initialization mechanism for newly added students or courses.

[0040] In the specific implementation, the student-course rating matrix is ​​constructed and the rating calculation formula is as follows: ; The score normalization formula is: in, For completion status; To score; is the weighting coefficient; For rating; The upper and lower bounds of the scoring interval.

[0041] It should be noted that continuous feature standardization and Z-score standardization are performed: in is the original eigenvalue; is the characteristic mean; is the characteristic standard deviation; The results are standardized.

[0042] In the specific implementation, the matrix decomposition and loss function formula are: The rating prediction formula is: in is the global score mean; , bias for students and courses; , Embedding vectors for students and courses, represents the set of observed student-course interactions, Typically represents a latent factor matrix of students / users. No. Row, that is, The latent factor vector of students, where It is represented by transposing the matrix; is the regularization coefficient.

[0043] Step S30: Extract features from three dimensions, namely, behavior sequence, knowledge graph, and attribute information, in parallel to generate multi-dimensional extracted features.

[0044] It can be understood that the steps of extracting features in parallel from the three dimensions of behavior sequence, knowledge graph and attribute information to generate multi-dimensional extracted features include: extracting learning order and short-term interest features from the behavior sequence dimension as behavior sequence features; extracting prerequisite relationships and knowledge point association features between courses from the knowledge graph dimension as knowledge graph features; extracting student personal attributes and course metadata features from the attribute information dimension as attribute features; and combining the behavior sequence features, the knowledge graph features and the attribute features into multi-dimensional extracted features.

[0045] It should be noted that the behavior sequence features include course ID, time information and learning performance; the knowledge graph features are extracted through a graph convolutional network; and the attribute features include student grade, major and course difficulty level.

[0046] In the specific implementation, the cold start smoothing mechanism and the initialization embedding of new users / new courses are as follows: in, Expressing with students A collection of interactive courses with graded content. Representation and Courses a collection of students with graded interactions; Training optimization, using SGD update: in is the learning rate.

[0047] The multi-view module is divided into the following parts: The behavior sequence perspective and the characteristics of the behavior sequence perspective are expressed as follows: in Embed for the course ID, is the time position code, For the completed state, is the last learning interval, For academic performance.

[0048] Knowledge graph perspective, knowledge graph Using GCN encoding: in is the normalized adjacency matrix; For the Layer node representation; No. layer parameter matrix; is the activation function.

[0049] The attribute semantic perspective is , student attributes are , the course attributes are .

[0050] Step S40: Fusing the matrix decomposition result with the multi-dimensional extracted features, and optimizing the feature representation through a contrastive learning mechanism.

[0051] It should be noted that the step of fusing the matrix decomposition results with the multi-dimensional extracted features and optimizing the feature representation through the contrastive learning mechanism also includes: positive samples select courses with similar interests or knowledge graph associations; negative samples select courses with large interest differences or no associations; and using a contrastive loss function adjusted by a temperature coefficient to optimize features.

[0052] In the specific implementation, multi-view fusion is performed and the view features are finally integrated: in is the vector concatenation operation, Represents the embedded feature representation extracted from the knowledge graph.

[0053] Contrastive learning enhancement module, sample pair construction, positive samples are data within the interest cluster or similar to the knowledge graph, and negative samples are data outside the interest cluster or unrelated courses.

[0054] Calculate InfoNCE loss, the loss function is in is the cosine similarity; is the temperature coefficient, Indicates the The representation vector of the samples Represents The matching positive sample representation is usually another view / intent vector of the same sequence, Indicates the number of The representation of candidate samples is used for negative sample comparison.

[0055] Step S50: Perform frequency domain analysis and filtering on the fused features to enhance the periodic features and long-term dependencies in the sequence to generate a target feature sequence. Step S60: modeling the target feature sequence using a multi-head attention mechanism to generate a target model and capture long-term dependencies in learning behavior.

[0056] It can be understood that the step of modeling the target feature sequence using a multi-head attention mechanism to generate a target model and capture long-term dependencies in learning behavior further includes: the number of attention heads and the number of contrastive learning perspectives maintain a proportional relationship; introducing matrix decomposition results as bias terms for attention calculation; using residual connection and layer normalization to stabilize the training process.

[0057] In a specific implementation, the Fourier frequency domain enhancement module and the fused feature input are: First, do a fast Fourier transform: After Fourier transform, complex weighted filtering is performed Then inverse transform back to time domain After that, position encoding is added: The calculation of the position encoding is: Multi-head attention mechanism structure: Single head attention: Multi-head attention integration: Each attention head is: Where is the query, key, and value matrix; is the attention dimension; , , , is a trainable projection matrix.

[0058] Add residual and normalization: The feedforward neural network feedforward layer is: The steps of generating a recommended Top-N resource sequence include: predicting the next step learning resource distribution: Generating a recommendation result: Joint training total loss: wherein is a sequence prediction error; is a matrix decomposition error; is a contrastive learning error; is a full model parameter set; is a loss weighting coefficient.

[0059] Step S70: Calculate the course recommendation probability based on the target model, sort and optimize based on the knowledge graph correlation degree, and generate a personalized learning resource recommendation list.

[0060] It should be noted that the technical problem to be solved by the present embodiment is: Deeply integrate matrix decomposition and contrastive learning to alleviate data sparsity and cold start: The present scheme introduces traditional matrix decomposition into the overall framework of contrastive learning, fully utilizes the explicit student-course interaction information in the rating matrix as the prior embedding of contrastive learning, and improves the stability and modeling accuracy of the model in the extremely sparse cold start scenario. The matrix decomposition embedding provides a good initialization semantic space for contrastive learning, effectively alleviating the feature deviation problem of sparse samples.

[0061] Multi-view feature fusion design to improve dynamic understanding of student learning behavior: The present scheme comprehensively introduces multi-view feature spaces such as student attributes, course knowledge graph structure, behavior sequence context, and course metadata, to model the long-term interest changes, stage cognitive level and knowledge point mastery process of student learning behavior, and improve the accurate description ability of the recommendation model for individualized learning path of students.

[0062] Sequence-dependent modeling and contrastive learning collaborative training to improve long-term learning path prediction accuracy: In the sequence modeling stage, the present scheme uses a long sequence Transformer structure to cooperate with multi-view features for deep sequence modeling, enhances the long-term dependence capturing ability and sequence feature robustness through a contrastive learning module, realizes the improvement of complex learning path prediction accuracy, and supports dynamic learning route planning and intelligent course recommendation.

[0063] Construct a complete and efficient educational resource sequence recommendation system framework to improve the actual application deployment ability: The present scheme designs an integrated system architecture of training-optimization-online deployment, supports parallel training, dynamic fine-tuning, and rolling update, improves the self-adaptive expansion ability and long-term online learning effect of the model in actual application, and guarantees the stable operation of large-scale education platforms and the effect of intelligent recommendation services.

[0064] In a specific implementation, a matrix decomposition enhancement, multi-view feature fusion, Fourier frequency domain enhancement, and multi-view contrast learning joint optimization education resource sequence recommendation method is proposed. After obtaining student learning behavior data, learning resource attribute data, and knowledge graph structure information, the method first uses regularized matrix decomposition technology to model the latent factor decomposition of the student-resource interaction score matrix, and extracts the basic preference embedding features of students and resources. In view of the diversity and complexity of student learning behavior, a multi-view feature coding module is designed to extract learning order and short-term interest features from the behavior sequence perspective, student and course static attribute features from the attribute information perspective, and knowledge point prerequisite dependency and correlation features between courses from the knowledge structure perspective. To improve the consistency and discriminability of feature representation, a contrast learning mechanism is introduced based on multi-view fusion features, and a positive and negative sample contrast loss function is constructed to enhance the model's ability to discriminate and transfer student interest preferences.

[0065] In the sequence modeling stage, to enhance the model's ability to depict the periodicity, rhythm, and long-term dependency features in students' long-term learning behavior, the system introduces a Fourier frequency domain enhancement module in the Transformer time series model. The fused feature sequence is mapped to the frequency domain space through the Fast Fourier Transform, and the trainable complex weighting mechanism is used to weight and adjust different frequency components, filter out high-frequency noise and redundant fluctuations, and strengthen the overall trend features of the sequence. Then, through inverse Fourier transform, the sequence modeling input is restored to the time domain to form the Fourier enhanced sequence modeling input. The Fourier enhanced feature sequence is input into the Transformer encoder composed of multi-head attention mechanism and feedforward neural network to realize dynamic modeling of sequence dependency in long-term learning paths.

[0066] Finally, the method organically fuses the results of matrix decomposition embedding, multi-view feature representation, contrast learning enhanced features, and Fourier frequency domain enhanced time series modeling to complete the intelligent recommendation of students' personalized learning resources. The invention fully combines the high sparsity, dynamic behavior, multi-modal feature correlation, and periodic rhythm features of learning data in the education scene, effectively improves the recommendation accuracy, representation robustness, and model generalization ability of the recommendation system in complex educational data environments, and is suitable for online education platforms, smart campus systems, and multi-scene personalized education resource recommendation applications.

[0067] After obtaining student learning behavior data, learning resource attribute data, and knowledge graph structure information, this embodiment uses regularized matrix decomposition technology to perform latent factor decomposition modeling on the student-resource interaction rating matrix, and extracts the basic preference embedding features of students and resources. In view of the diversity and complexity of student learning behavior, a multi-perspective feature encoding module is designed to extract learning sequence and short-term interest features from the perspective of behavior sequence, extract student and course static attribute features from the perspective of attribute information, and extract knowledge point prerequisite dependency and association relationship features between courses from the perspective of knowledge structure. In order to improve the consistency and discriminability of feature representation, a comparative learning mechanism is introduced on the basis of multi-perspective fusion features. By constructing a positive and negative sample comparison loss function, the model's ability to discriminate and transfer student interest preferences is enhanced.

[0068] In addition, an embodiment of the present application also proposes a computer-readable storage medium, on which a program for recommending educational resource sequences is stored. When the program for recommending educational resource sequences is executed by a processor, the steps of the method for recommending educational resource sequences as described above are implemented.

[0069] Reference Figure 3 , Figure 3 This is a structural block diagram of the first embodiment of the educational resource sequence recommendation device of this application.

[0070] like Figure 3 As shown, the educational resource sequence recommendation device proposed in the embodiment of the present application includes: The data collection module 10 is used to collect the interaction behavior data between students and course resources, including learning time, completion status and learning results to construct a behavior sequence; A matrix module 20 is used to construct a student-course rating matrix based on the behavior sequence and extract potential feature representations of students and courses using matrix decomposition technology; A multi-dimensional feature extraction module 30 is used to extract features from three dimensions, namely, behavior sequence, knowledge graph, and attribute information, in parallel to generate multi-dimensional extracted features; A fusion module 40 is used to fuse the matrix decomposition result with the multi-dimensional extracted features and optimize the feature representation through a contrastive learning mechanism; The target feature sequence module 50 is used to perform frequency domain analysis and filtering on the fused features, and strengthen the periodic features and long-term dependencies in the sequence to generate a target feature sequence; A relationship capture module 60 is configured to model the target feature sequence using a multi-head attention mechanism to generate a target model, thereby capturing long-term dependencies in the learning behavior; The result generation module 70 is used to calculate the course recommendation probability based on the target model, perform sorting optimization based on the knowledge graph relevance, and generate a personalized learning resource recommendation list.

[0071] It should be understood that the above is only illustrative, and does not constitute any limitation on the technical solutions of the present application. In specific applications, those skilled in the art can set up as needed, and the present application does not limit this.

[0072] After obtaining the student learning behavior data, learning resource attribute data and knowledge graph structure information, the embodiment adopts the regularization matrix decomposition technology to perform latent factor decomposition modeling on the student-resource interaction score matrix, and extracts the basic preference embedding features of the students and the resources. In view of the diversity and complexity of the student learning behavior, a multi-view feature coding module is designed to extract learning order and short-term interest features from the behavior sequence perspective, to extract student and course static attribute features from the attribute information perspective, and to extract knowledge point prerequisite dependency and correlation relationship features between courses from the knowledge structure perspective. In order to improve the consistency and discriminability of the feature representation, a contrast learning mechanism is introduced based on the multi-view fusion features, a positive and negative sample contrast loss function is constructed, and the discrimination and transfer modeling ability of the model to the student interest preference is strengthened.

[0073] It should be noted that the above workflow is only illustrative and does not limit the scope of protection of the present application. In actual applications, those skilled in the art can select part or all of them to achieve the purpose of the embodiment scheme according to actual needs, which is not limited here.

[0074] In addition, technical details not described in detail in the embodiment can be referred to the method for recommending educational resource sequences provided by any embodiment of the present application, which will not be described here.

[0075] In addition, it should be noted that in this paper, the term "includes" "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or system. Without more limitations, the element defined by the sentence "includes a" does not exclude the presence of another identical element in the process, method, article or system including the element.

[0076] The above sequence number of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments.

[0077] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, also can be through hardware, but in many cases the former is the better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the contribution to the prior art can be embodied in the form of software products, the computer software product is stored in a storage medium (such as read-only memory (ReadOnly Memory, ROM) / RAM, disk, optical disc), including a number of instructions to make a terminal device (may be a mobile phone, computer, server, or network equipment, etc.) executes the method of each embodiment of the present application. The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, any equivalent structure or equivalent process transformation made by using the contents of the present application specification and drawings, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for recommending educational resource sequences, characterized in that: include: Collect students' interaction behavior data with course resources, including learning time, completion status, and learning results to construct behavior sequences; Constructing a student-course rating matrix based on the behavior sequence, and using matrix decomposition technology to extract the latent feature representations of students and courses; Extract features from three dimensions: behavior sequence, knowledge graph, and attribute information in parallel to generate multi-dimensional extracted features; Fusing the matrix decomposition results with the multi-dimensional extracted features, and optimizing the feature representation through a contrastive learning mechanism; Perform frequency domain analysis and filtering on the fused features to enhance the periodic features and long-term dependencies in the sequence to generate the target feature sequence; A multi-head attention mechanism is used to model the target feature sequence to generate a target model, capturing the long-term dependencies in the learning behavior; The course recommendation probability is calculated based on the target model, and the ranking is optimized based on the knowledge graph relevance to generate a personalized learning resource recommendation list.

2. The method according to claim 1, characterized in that After the step of collecting the interactive behavior data between students and course resources, the method further includes: Filter abnormal learning behavior data; Missing values ​​are filled with mean values; Construct a complete sequence of actions in chronological order.

3. The method according to claim 1, characterized in that The step of constructing a student-course rating matrix based on the behavior sequence and extracting potential feature representations of students and courses using matrix decomposition technology includes: The grading matrix takes into account both learning completion and academic performance; Regularized matrix factorization method is used to prevent overfitting; A smooth initialization mechanism is adopted for new students or courses.

4. The method according to claim 1, wherein The steps of extracting features from three dimensions, namely behavior sequence, knowledge graph, and attribute information, in parallel to generate multi-dimensional extracted features include: Extract learning sequence and short-term interest features from the behavior sequence dimension as behavior sequence features; Extract prerequisite relationships between courses and knowledge point association features from the knowledge graph dimension as knowledge graph features; Extract student personal attributes and course metadata features from the attribute information dimension as attribute features; The behavior sequence features, the knowledge graph features and the attribute features are converted into multi-dimensional features.

5. The method according to claim 4, characterized in that The behavior sequence features include course ID, time information and learning performance; The knowledge graph features are extracted through a graph convolutional network; The attribute characteristics include student grade, major and course difficulty level.

6. The method according to claim 1, characterized in that The step of fusing the matrix decomposition result with the multi-dimensional extracted features and optimizing the feature representation through a contrastive learning mechanism further includes: Positive samples select courses with similar interests or knowledge graph associations; Negative samples select courses with large interest differences or no correlation; The contrast loss function with temperature coefficient adjustment is used to optimize the features.

7. The method according to claim 1, characterized in that The step of using a multi-head attention mechanism to model the target feature sequence to generate a target model and capture the long-term dependencies in the learning behavior also includes: The number of attention heads is proportional to the number of contrastive learning views; Introduce the matrix decomposition result as the bias term for attention calculation; Residual connections and layer normalization are used to stabilize the training process.

8. An educational resource sequence recommendation device, characterized in that: include: The data collection module is used to collect students' interactive behavior data with course resources, including learning time, completion status and learning results to construct behavior sequences; A matrix module is used to construct a student-course rating matrix based on the behavior sequence and extract potential feature representations of students and courses using matrix decomposition technology; Multi-dimensional feature extraction module, which is used to extract features from three dimensions, namely behavior sequence, knowledge graph and attribute information, in parallel to generate multi-dimensional extracted features; A fusion module is used to fuse the matrix decomposition results with the multi-dimensional extracted features and optimize the feature representation through a contrastive learning mechanism; The target feature sequence module is used to perform frequency domain analysis and filtering on the fused features, strengthen the periodic features and long-term dependencies in the sequence to generate the target feature sequence; A relationship capture module is used to model the target feature sequence using a multi-head attention mechanism to generate a target model and capture long-term dependencies in learning behaviors; The result generation module is used to calculate the course recommendation probability based on the target model, perform sorting optimization based on the knowledge graph relevance, and generate a personalized learning resource recommendation list.

9. A computer device, characterized in that: The device comprises: a memory and a processor, wherein the processor executes the method according to any one of claims 1 to 7 when running computer instructions stored in the memory.

10. A computer-readable storage medium, characterized in that The method comprises instructions which, when executed on a computer, cause the computer to execute the method according to any one of claims 1 to 7.

Images