Education resource sequence recommendation method and device based on three-view structure distillation, equipment and storage medium

Abstract

The application discloses a kind of education resource sequence recommendation method, device and equipment based on three-view structure distillation and storage medium, the method includes: based on student-course interaction log, the learning sequence of time sequence course index is constructed, three types of view hypergraph are constructed according to semantic, path, concept structure multidimensional feature, three-view course representation matrix is generated by structure coding fusion, and the recommended list is output by inputting education resource recommendation model;Education resource recommendation model generates student state vector by the coding of Transformer, and the total course score is obtained by fusing three-view branch score and predicted;The present application predicts the education resource recommendation by three-view collaborative scoring from multiple perspectives, accurately captures course representation, effectively depicts learning path high-order law, improves the accuracy, explainability and generalization ability of education resource recommendation in sparse interaction and cold start scene.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to a method, apparatus, device, and storage medium for recommending educational resource sequences based on three-view structural distillation. Background Technology

[0002] With the continued deepening of "Internet + Education," various online learning management systems have accumulated a wealth of course resources and student interaction logs. Providing course path planning and resource recommendations tailored to individual learners has become one of the key capabilities in intelligent education.

[0003] Existing technical approaches for educational recommendation systems mainly include collaborative filtering, deep sequence recommendation (such as RNN / GRU / LSTM, Transformer sequence models), graph neural network-based recommendation methods, attention-enhanced recommendation, and multi-source information fusion recommendation. In recent years, research focus has further shifted towards "deep integration of graph structure representation and sequence modeling" and "joint modeling of content semantics and behavioral structure," aiming to simultaneously leverage course content information and group learning behavior patterns to improve recommendation effectiveness.

[0004] In educational settings, existing solutions still face significant bottlenecks. On one hand, the construction of the semantic structure of course content often relies on KNN nearest neighbors or fixed-threshold similarity rules. When the semantic space density distribution is uneven or the quality of course descriptions varies, noisy connections are easily introduced, leading to unstable course structure representations. Cold-start and long-tail courses are particularly affected. On the other hand, learning behavior structure modeling often remains at the level of neighbor transitions or first-order Markov models, making it difficult to express the higher-order dependencies and stage-based evolutionary patterns embodied in multi-step learning path segments, resulting in insufficient characterization of real learning paths.

[0005] Furthermore, the semantics of course content, the logic of learning paths, and the structure of knowledge concepts are inherently heterogeneous information sources. Many fusion methods simply splice or mix them, lacking collaborative constraints that address perspective differences, which easily leads to problems such as fragmented fusion representations, unstable training, and insufficient generalization ability. Educational scenarios generally exhibit characteristics such as sparse interactions, strong pre- and post-requirement logic, and a significant long-tail distribution, making it urgent to develop a novel sequence recommendation method that can adaptively denoise, characterize high-order path patterns, introduce interpretable conceptual structures, and possess a stable collaborative mechanism. Summary of the Invention

[0006] The main objective of this invention is to provide a method, apparatus, device, and storage medium for recommending educational resource sequences based on three-view structural distillation, aiming to solve the technical problems of unstable course representation, difficulty in characterizing high-order patterns of learning paths, and easy fragmentation of heterogeneous structure fusion in existing technologies under sparse interaction and cold start scenarios.

[0007] To achieve the above objectives, this invention provides a method for recommending educational resource sequences based on three-view structural distillation, the method comprising the following steps:

[0008] A student learning sequence is constructed based on the student-course interaction logs of student users. The student learning sequence contains course indices arranged in chronological order. The course indices are obtained by mapping the course identifiers of the interactive courses through a mapping function.

[0009] Based on the multidimensional features of the student learning sequence, a semantic perspective hypergraph, a path perspective hypergraph, and a concept structure perspective hypergraph are constructed respectively, wherein the multidimensional features include semantic features, path features, and concept structure features;

[0010] After structural encoding of the semantic view hypergraph, the path view hypergraph, and the concept structure view hypergraph, they are fused to generate a three-view course representation matrix.

[0011] The three-view course representation matrix is ​​input into a pre-trained educational resource recommendation model, which outputs a recommended list of educational resource sequences for the student user.

[0012] The educational resource recommendation model is configured to encode the three-view course representation matrix into a student state vector using a Transformer encoder, and to rate courses from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch based on the student state vector. It then performs course recommendation prediction based on the full course rating vector and outputs a recommended list of educational resource sequences for the student user based on the prediction results. The full course rating vector is obtained by fusing the course ratings from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch.

[0013] Optionally, the step of constructing a semantic perspective hypergraph, a path perspective hypergraph, and a concept structure perspective hypergraph based on the multidimensional features of the student learning sequence includes:

[0014] Generate semantic text of the interactive courses corresponding to each course index in the student learning sequence;

[0015] The semantic text is encoded using a text encoder to obtain the course semantic vector;

[0016] The course semantic vectors of each interactive course are stacked to obtain a course semantic vector matrix, which is used to characterize the semantic features of the student learning sequence.

[0017] Construct a semantic perspective hypergraph based on the aforementioned course semantic vector matrix;

[0018] The path structure of interactive courses in the student learning sequence is analyzed to obtain path features, and a path view hypergraph is constructed based on the path features.

[0019] A conceptual structure analysis is performed on the interactive courses in the student learning sequence to obtain conceptual structure features, and a conceptual structure perspective hypergraph is constructed based on the conceptual structure features.

[0020] Optionally, constructing a semantic perspective hypergraph based on the course semantic vector matrix includes:

[0021] Density clustering analysis is performed based on the course semantic vector matrix to obtain the course cluster labels for each interactive course;

[0022] Construct a valid set of cluster tags based on the course cluster tags, and determine the number of semantic prototypes;

[0023] Renumber the valid cluster labels in the set of valid cluster labels as continuous semantic prototype indices;

[0024] The semantic prototype vector is calculated based on the number of semantic prototypes and the continuous semantic prototype index, and the semantic prototype vector is the cluster center.

[0025] The semantic prototype vectors are stacked to obtain a semantic prototype matrix, and normalized vectors and similarity are calculated based on the semantic prototype matrix.

[0026] Based on the normalized vectors and similarity, a course-semantic prototype bipartite association matrix is ​​constructed, and a semantic perspective hypergraph is constructed based on the course-semantic prototype bipartite association matrix.

[0027] Optionally, the step of performing path structure analysis on the interactive courses in the student learning sequence to obtain path features, and constructing a path perspective hypergraph based on the path features, includes:

[0028] Extract continuous path segments of a preset path length from the student learning sequence. The continuous path segments are sequences composed of course indices at corresponding positions in the student learning sequence.

[0029] The continuous path segments are defined as path patterns, and the path patterns are in the form of a sequence of corresponding course indices;

[0030] The frequency of occurrence of each path pattern is counted, where the frequency of occurrence is the sum of the number of consecutive path segments in all student learning sequences that are completely consistent with that path pattern;

[0031] Path patterns that occur more frequently than a preset frequency threshold are classified as frequent path patterns.

[0032] The attenuation weights corresponding to the course indices at each position in the frequent path pattern are calculated based on the preset position attenuation coefficient. The attenuation weights of the same course index at each position in the frequent path pattern are summed to construct the course-path pattern association matrix. The course-path pattern association matrix is ​​used to characterize the path features.

[0033] Based on the course-path pattern association matrix, a path perspective hypergraph is generated.

[0034] Optionally, the step of performing conceptual structure analysis on the interactive courses in the student learning sequence to obtain conceptual structure features, and constructing a conceptual structure viewpoint hypergraph based on the conceptual structure features, includes:

[0035] Construct a course-concept association matrix. In the course-concept association matrix, when a concept is extracted or matched with a corresponding course, the value of the corresponding position in the matrix is ​​1, otherwise it is 0.

[0036] Based on the student learning sequence and the course-concept association matrix, the sequence co-occurrence count of concept prerequisite directions is calculated. The sequence co-occurrence count is the number of co-occurrences of concepts matched in the preceding and following courses in the student learning sequence.

[0037] Based on the course-concept association matrix, the co-occurrence count of concepts in the same course is calculated, where the co-occurrence count is the number of concepts that are simultaneously matched in the same course.

[0038] The sequence co-occurrence count and the co-occurrence count of the same lesson are respectively probabilistically processed to obtain the concept prior relation strength and the concept relevance relation strength. Then, the concept prior relation strength and the concept relevance relation strength are fused by a preset fusion weight coefficient to obtain the concept relation matrix.

[0039] The concept-perspective feature mapping function maps the course semantic vector matrix of the student learning sequence to the concept-perspective course initial feature matrix.

[0040] Based on the initial feature matrix and the concept relation matrix of the concept-view course, concept structure propagation is performed to generate a concept-view course representation matrix. Then, a concept structure-view hypergraph is constructed based on this representation matrix, as shown in the following formula:

[0041]

[0042]

[0043]

[0044] in, Representation matrix of course from the perspective of concepts This is the initial feature matrix for the conceptual perspective course. This represents the initial feature matrix of a trainable concept. and These represent the intermediate feature matrices of concept propagation. Representing the course-concept association matrix, Indicates transpose. Represents the concept relation matrix. Represents the total number of concepts. This indicates the total number of courses.

[0045] Optionally, the step of structurally encoding and fusing the semantic perspective hypergraph, the path perspective hypergraph, and the concept structure perspective hypergraph to generate a three-view curriculum representation matrix includes:

[0046] The semantic perspective hypergraph is subjected to semantic perspective propagation encoding to obtain the semantic perspective course representation matrix, as shown in the following formula:

[0047]

[0048] in, The course represents the representation matrix from a semantic perspective. Representation layer normalization, This represents the element-wise activation function. Represents the random deactivation function. The input feature matrix represents the semantic view hypergraph. Represents the course-semantic prototype bipartite association matrix;

[0049] The path-view hypergraph is subjected to path-view propagation encoding to obtain the path-view course representation matrix, as shown in the following formula:

[0050]

[0051] in, The path-view course representation matrix The input feature matrix represents the path-view hypergraph. Represents the course-path pattern association matrix;

[0052] The semantic perspective course representation matrix, the path perspective course representation matrix, and the concept perspective course representation matrix corresponding to the concept structure perspective hypergraph are fused to obtain a three-perspective course representation matrix, as shown in the following formula:

[0053]

[0054] in, A matrix representing a three-view course. Represents the fusion mapping function, The three-view course represents the dimension of the matrix. Representation matrix of course from the perspective of concepts This indicates a splicing operation based on feature dimensions.

[0055] Optionally, the educational resource recommendation model is further configured to construct a fixed-length input sequence and supervision labels, obtain sequence embeddings by looking up a table according to a unified curriculum representation, and superimpose positional embeddings to generate Transformer input sequence representation features, as shown in the following formula:

[0056]

[0057]

[0058]

[0059] in, For students A fixed-length input sequence, The maximum sequence length, To monitor the index of labeled courses, For sequence embedding matrix, Let represent the t-th row of the sequence embedding matrix for student u, corresponding to the embedding vector of the t-th course in the input sequence. For the position embedding matrix, This indicates the dimension of the position embedding matrix. Input sequences to the Transformer to represent features;

[0060] The educational resource recommendation model is further configured to encode the features of the Transformer input sequence using a Transformer encoder to obtain a student state vector, as shown in the following formula:

[0061]

[0062] in, For Transformer encoders; for The student state vector, The dimension representing the student's state vector;

[0063] The educational resource recommendation model is further configured to score courses based on the student state vector from semantic perspective, path perspective, and conceptual structure perspective to obtain a full course score vector.

[0064] The educational resource recommendation model is further configured to predict course recommendations based on the full set of course rating vectors, and obtain the fusion prediction probability, as shown in the following formula:

[0065]

[0066] in, Let represent the full course scoring vector under the three-view fusion representation. Indicates the fusion prediction probability. It is an exponential function. and Indicates a course index;

[0067] The educational resource recommendation model is further configured to construct a label smoothing target distribution based on supervised labels, and generate a recommendation loss term based on the label smoothing target distribution and the fusion prediction probability, as shown in the following formula:

[0068]

[0069]

[0070] in, This indicates that the label smooths the target distribution. The label smoothing coefficient. Represents the student index. For the set of student indices in the training batch, The batch size, It is the natural logarithm. Indicates the recommended loss term;

[0071] The educational resource recommendation model is further configured to perform temperature softening processing on the full course rating vector and the branch course rating vectors corresponding to the semantic perspective branch, path perspective branch, and conceptual structure perspective branch, respectively, to obtain the fusion distribution and the branch distribution under the three perspective branches. Based on the fusion distribution and the branch distribution under the three perspective branches, a cross-perspective consistency distillation loss term is constructed, as shown in the following formula:

[0072]

[0073] in, This represents the cross-perspective consistent distillation loss term. This is the distillation temperature coefficient. This indicates the fusion distribution after temperature softening. , and These represent the branch distribution under the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch, respectively. Let KL divergence be a metric.

[0074] The educational resource recommendation model is further configured to construct a total loss term based on the recommendation loss term and the cross-perspective consistency distillation loss term, and to update the model parameters through backpropagation based on the total loss term to complete joint training;

[0075] The educational resource recommendation model is further configured to calculate the target course score under the fusion perspective representation based on the updated model parameters, select a preset number of recommended courses according to the target course score, and output the recommended courses as the course recommendation prediction result to output the educational resource sequence recommendation list for student users.

[0076] Furthermore, to achieve the above objectives, the present invention also proposes an educational resource sequence recommendation device based on three-view structure distillation, wherein the educational resource sequence recommendation device based on three-view structure distillation includes:

[0077] The learning sequence construction module is used to construct student learning sequences based on student-course interaction logs of student users. The student learning sequence contains course indices arranged in chronological order, and the course indices are obtained by mapping the course identifiers of the interactive courses through a mapping function.

[0078] The three-view hypergraph construction module is used to construct a semantic view hypergraph, a path view hypergraph, and a concept structure view hypergraph based on the multidimensional features of the student learning sequence, wherein the multidimensional features include semantic features, path features, and concept structure features.

[0079] The hypergraph fusion module is used to fuse the semantic view hypergraph, the path view hypergraph and the concept structure view hypergraph after structural encoding to generate a three-view course representation matrix.

[0080] The educational resource recommendation module is used to input the three-view course representation matrix into the pre-trained educational resource recommendation model and output the educational resource sequence recommendation list for the student user.

[0081] The educational resource recommendation model is configured to encode the three-view course representation matrix into a student state vector using a Transformer encoder, and to rate courses from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch based on the student state vector. It then performs course recommendation prediction based on the full course rating vector and outputs a recommended list of educational resource sequences for the student user based on the prediction results. The full course rating vector is obtained by fusing the course ratings from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch.

[0082] Furthermore, to achieve the above objectives, this application also proposes an educational resource sequence recommendation device based on three-view structure distillation. The device includes: a memory, a processor, and an educational resource sequence recommendation program based on three-view structure distillation stored in the memory. The processor is used to run the educational resource sequence recommendation program based on three-view structure distillation, and the computer program is configured to implement the steps of the educational resource sequence recommendation method based on three-view structure distillation as described above.

[0083] Furthermore, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the educational resource sequence recommendation method based on three-view structure distillation as described above.

[0084] This invention, through the construction and encoding fusion of a three-view hypergraph, fully explores the multi-dimensional features and complex relationships of courses, effectively compensating for the limitations of single-view features, and improving the completeness and discriminativeness of course feature expression. Through the multi-branch scoring and fusion mechanism of the educational resource recommendation model, multi-view features are accurately transformed into personalized recommendation results, strengthening the rationality and adaptability of recommendation predictions. It accurately captures course representations and effectively characterizes higher-order patterns of learning paths in sparse interaction and cold-start scenarios, thereby effectively mining student learning trajectories and multi-dimensional course features, improving the accuracy and personalization of educational resource recommendations, optimizing student learning experience, standardizing the recommendation process, improving recommendation efficiency, reducing resource waste caused by ineffective recommendations, and achieving precise matching of educational resources with student learning needs. This provides an efficient and feasible technical path for personalized education recommendations. Attached Figure Description

[0085] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0086] Figure 1 This is a schematic diagram of the structure of an educational resource sequence recommendation device based on three-view structural distillation, which is part of the hardware operating environment of the embodiment of the present invention.

[0087] Figure 2 This is a flowchart illustrating the first embodiment of the educational resource sequence recommendation method based on three-view structural distillation of the present invention.

[0088] Figure 3 This is a schematic diagram of the educational resource recommendation model in one embodiment of the educational resource sequence recommendation method based on three-view structural distillation of the present invention;

[0089] Figure 4 This is a schematic diagram of the educational resource sequence recommendation process in one embodiment of the educational resource sequence recommendation method based on three-view structural distillation of the present invention;

[0090] Figure 5 This is a flowchart illustrating the second embodiment of the educational resource sequence recommendation method based on three-view structural distillation of the present invention.

[0091] Figure 6 This is a structural block diagram of the first embodiment of the educational resource sequence recommendation device based on three-view structural distillation of the present invention.

[0092] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0093] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0094] Reference Figure 1 , Figure 1 This is a schematic diagram of the educational resource sequence recommendation device based on three-view structural distillation, which is part of the hardware operating environment of the embodiment of the present invention.

[0095] like Figure 1 As shown, the educational resource sequence recommendation device based on three-view structure distillation may 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 enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0096] Those skilled in the art will understand that Figure 1The structure shown does not constitute a limitation on the educational resource sequence recommendation device based on three-view structural distillation, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0097] like Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and an educational resource sequence recommendation program based on a three-view structure distillation.

[0098] exist Figure 1 In the educational resource sequence recommendation device based on three-view structure distillation shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the educational resource sequence recommendation device based on three-view structure distillation of the present invention can be set in the educational resource sequence recommendation device based on three-view structure distillation. The educational resource sequence recommendation device based on three-view structure distillation calls the educational resource sequence recommendation program based on three-view structure distillation stored in the memory 1005 through the processor 1001 and executes the educational resource sequence recommendation method based on three-view structure distillation provided in the embodiment of the present invention.

[0099] This invention provides a method for recommending educational resource sequences based on three-view structural distillation, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the educational resource sequence recommendation method based on three-view structural distillation of the present invention.

[0100] In this embodiment, the educational resource sequence recommendation method based on three-view structural distillation includes the following steps:

[0101] Step S10: Construct a student learning sequence based on the student-course interaction logs of student users. The student learning sequence contains course indices arranged in chronological order. The course indices are obtained by mapping the course identifiers of the interactive courses through a mapping function.

[0102] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a terminal electronic device capable of performing the above functions. The following description uses an educational resource sequence recommendation device based on three-view structural distillation (hereinafter referred to as the recommendation device) as an example to illustrate this embodiment and the following embodiments.

[0103] It should be noted that the student-course interaction log is a structured data set automatically recorded by the system when students interact with various courses on the education and learning platform. It covers core information such as student identifier, course identifier, interaction time, and interaction type (such as clicking to learn, completing assignments, participating in quizzes, etc.).

[0104] A student learning sequence can be an ordered sequence of courses, organized by time as the core dimension, which arranges all the course interactions of a single student in chronological order. Its core function is to present the student's learning trajectory and behavioral logic.

[0105] It should be noted that the course index refers to the conversion of the original course identifier into a standardized numerical (or vector) form that the model can recognize and calculate. It is an intermediate carrier connecting the original course information and the data processed by the model.

[0106] Course identifiers are unique identifiers used to distinguish different courses. They can be in string form (such as course name abbreviation + number) or numeric form (such as unique serial number), and serve as the "identity identifier" of a course in the system.

[0107] In the specific implementation, the raw data acquisition and index mapping construction include:

[0108] Recommended device to establish a mapping function from the original course identifier to the course index:

[0109]

[0110] in: This is a mapping function from course identifiers to course indices; This is the original identifier for the course; For course indexing; This represents the total number of courses.

[0111] Student-course interaction log collection and learning sequence construction, including:

[0112] Map the original course identifiers in the interaction logs to course indexes:

[0113]

[0114] Constructing a student learning sequence in chronological order:

[0115]

[0116] in: Index for students and ; Total number of students; For sequence position index; For students In position The corresponding original course identifier; This is the mapped course index; For students The course learning sequence; For students The sequence length.

[0117] In some embodiments, the recommended device can extract all students' student-course interaction logs from the database of the education and learning platform, filter valid data (removing invalid interactions such as accidental clicks, duplicate records, and abnormal timestamps), and group the data by student identifier to ensure that each group of data corresponds to the complete interaction record of a single student.

[0118] Step S20: Construct a semantic perspective hypergraph, a path perspective hypergraph, and a concept structure perspective hypergraph based on the multidimensional features of the student learning sequence, wherein the multidimensional features include semantic features, path features, and concept structure features.

[0119] It should be noted that multidimensional features can be a collection of multidimensional information extracted from students' learning sequences that reflects the attributes of the course itself and the students' learning patterns. The core features include semantic features, path features, and conceptual structure features, which are used to comprehensively depict course associations and learning trajectories.

[0120] Semantic features refer to the features used to describe the content attributes of the course itself, reflecting the core semantic information of the course such as knowledge topics, content categories, and teaching objectives, and embodying "what" the course is.

[0121] Path features refer to the transitional association features between courses extracted based on the temporal sequence of students' learning. They reflect the jump logic and sequential dependencies of courses during the student's learning process, embodying "how students learn".

[0122] Conceptual structural features refer to the knowledge hierarchy and logical connections between courses extracted based on the knowledge system corresponding to the course. They reflect the position and interrelationship of the course in the overall knowledge framework and embody "what the knowledge connections are".

[0123] It should be noted that a semantic hypergraph can be a hypergraph structure constructed with courses as nodes and semantic relationships (such as semantic similarity and content relevance) between courses as hyperedges, which is used to characterize the semantic relationships between courses, with a core focus on the relationships at the level of course content.

[0124] A path-view hypergraph can be a hypergraph structure constructed with courses as nodes and the sequential jump paths and association strengths of courses in the student's learning sequence as hyperedges, used to depict the transition relationships of courses in the student's learning trajectory, with a core focus on the associations at the level of student learning behavior.

[0125] A hypergraph from a conceptual structure perspective can be a hypergraph structure that uses courses as nodes and knowledge levels and conceptual connections (such as belonging to the same knowledge module or having a progressive relationship) between courses as hyperedges to depict the connections between course knowledge structures, with a core focus on connections at the knowledge system level.

[0126] In some embodiments, the recommended device extracts three types of features for the constructed student learning sequence: semantic features can be extracted using text embedding techniques (such as TF-IDF and Word2Vec) through course titles, description text, knowledge point tags, etc.; path features can be extracted by statistical methods such as analyzing the course jump frequency and the co-occurrence frequency of adjacent courses in the student learning sequence; and conceptual structure features can be obtained by combining the knowledge graph corresponding to the course to extract information such as the course's hierarchy in the knowledge system and related courses.

[0127] In some embodiments, the recommending device may construct corresponding hypergraphs based on three types of features:

[0128] Semantic Hypergraph: Treat courses as nodes, calculate the semantic similarity between any two courses, and connect courses with similarity reaching a set threshold through hyperedges, with the weight of the hyperedge corresponding to the semantic similarity.

[0129] Path-view hypergraph: Treating courses as nodes, and based on the jump relationship between adjacent courses in the student's learning sequence, multiple courses that frequently co-occur and jump consecutively are connected through hyperedges, with the weight of the hyperedge corresponding to the jump frequency;

[0130] Hypergraph from a conceptual structure perspective: Courses are treated as nodes. Based on the conceptual associations of courses in the knowledge graph (such as modules or progressive relationships), multiple related courses are connected by hyperedges, and the weight of the hyperedge corresponds to the strength of the conceptual association.

[0131] In some embodiments, the recommended device can simplify the three types of hypergraphs constructed, removing hyperedges with too low weight and isolated nodes to ensure the sparsity and effectiveness of the hypergraph, which facilitates subsequent structural encoding.

[0132] Understandably, this embodiment breaks through the limitations of traditional sequence recommendation that only focuses on time order, and comprehensively explores the complex relationships between courses and student learning patterns from three dimensions: semantics, behavioral paths, and knowledge structure. By replacing the traditional graph structure with a hypergraph structure, it can better depict the multi-dimensional relationships between multiple courses and enrich the expressive dimensions of course features. The three types of perspective hypergraphs focus on information at different levels, providing diversified and targeted structural inputs for subsequent multi-perspective fusion and feature encoding, effectively avoiding the limitations of single-perspective features.

[0133] Step S30: After structural encoding of the semantic perspective hypergraph, the path perspective hypergraph, and the concept structure perspective hypergraph, they are fused to generate a three-view course representation matrix.

[0134] It is understood that the structural encoding in this embodiment can employ a specific encoding algorithm to transform the structural information of the hypergraph (node ​​associations, hyperedge weights, etc.) into a vector form that the model can compute and process, thereby extracting the structural features of the hypergraph and quantifying them.

[0135] It should be noted that the three-view course representation matrix can be a unified course feature matrix that integrates the feature matrices obtained by structurally encoding the three types of hypergraphs through a fusion strategy. This matrix contains information from the three types of perspectives: semantics, path, and conceptual structure, and is used to quantitatively express the course features.

[0136] In some embodiments, the recommended device may employ a hypergraph structure encoding algorithm to encode the semantic view hypergraph, the path view hypergraph, and the concept structure view hypergraph respectively, transforming the structural information and associated features of each course in each type of hypergraph into corresponding feature vectors, thereby obtaining course feature matrices of three single perspectives.

[0137] Furthermore, in order to effectively integrate the course features from the three perspectives of semantics, path, and conceptual structure, eliminate the one-sidedness and limitations of single-perspective features, and achieve complementary and synergistic effects of multi-dimensional features, in one embodiment, the above step S30 may include:

[0138] Step S301: Perform semantic perspective propagation encoding on the semantic perspective hypergraph to obtain the semantic perspective curriculum representation matrix;

[0139] Step S302: Perform path-view propagation encoding on the path-view hypergraph to obtain the path-view course representation matrix;

[0140] Step S303: Merge the semantic perspective course representation matrix, the path perspective course representation matrix, and the concept perspective course representation matrix corresponding to the concept structure perspective hypergraph to obtain a three-perspective course representation matrix.

[0141] In practical implementation, the three-view structure encoding and fusion yield a unified course representation, including:

[0142] Semantic perspective input feature construction:

[0143]

[0144] Path-view input feature construction:

[0145]

[0146] The semantic perspective two-step propagation yields the semantic perspective course representation:

[0147]

[0148] The two-step propagation from the path perspective yields the path perspective course representation:

[0149]

[0150] The fusion of three perspectives yields a unified course representation:

[0151]

[0152] in: It is a semantic feature mapping function; Embed functions for the course index; Indicates taking the first matrix. OK; Represents the course-semantic prototype bipartite association matrix; For layer normalization; Element-wise activation function; For random inactivation; A semantic perspective course representation matrix; The path-view course representation matrix; Represents the course-path pattern association matrix; This is a splicing operation based on feature dimensions; For fusion mapping functions; The unified curriculum representation matrix after integration; The three-view course represents the dimension of the matrix.

[0153] Understandably, this embodiment effectively integrates course features from three perspectives—semantics, path, and conceptual structure—through a matrix fusion strategy, eliminating the one-sidedness and limitations of single-perspective features and achieving complementary synergy among multi-dimensional features. The generated three-perspective course representation matrix centrally retains the three core features of the course: semantic content, learning path, and knowledge structure. It comprehensively and accurately portrays course attributes and student learning patterns, enhances the distinguishability and expressive power of course features, and provides comprehensive and high-quality feature input for the accurate reasoning and personalized recommendation of the subsequent recommendation model, laying the core foundation for improving recommendation accuracy.

[0154] Step S40: Input the three-view course representation matrix into the pre-trained educational resource recommendation model and output the educational resource sequence recommendation list for the student user.

[0155] It should be noted that the educational resource recommendation model is configured to encode the three-view course representation matrix into a student state vector using a Transformer encoder, and to score courses from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch based on the student state vector, respectively. Course recommendation prediction is performed based on the full course score vector, and a recommended list of educational resource sequences for student users is output based on the course recommendation prediction results. The full course score vector is obtained by fusing the course scores from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch.

[0156] In its implementation, the educational resource recommendation model performs a series of reasoning operations, such as encoding, scoring, and prediction, on the input three-view course representation matrix to automatically filter out courses that meet students' learning needs. Based on the reasoning results, the recommendation priorities are arranged from high to low to generate an ordered list of recommended educational resources. Courses that students have already studied are removed (the option to retain or remove courses can be set according to requirements), and the final output is given to the student user.

[0157] Furthermore, in order to improve the accuracy of educational resource recommendations, in one embodiment, the educational resource recommendation model is further configured to construct a fixed-length input sequence and supervision labels, obtain the sequence embedding by looking up a table according to a unified course representation, and superimpose the position embedding to generate Transformer input sequence representation features;

[0158] The educational resource recommendation model is further configured to encode the features of the Transformer input sequence using a Transformer encoder to obtain a student state vector.

[0159] The educational resource recommendation model is further configured to score courses based on the student state vector from semantic perspective, path perspective, and conceptual structure perspective to obtain a full course score vector.

[0160] The educational resource recommendation model is further configured to perform course recommendation prediction based on the full set of course rating vectors to obtain the fusion prediction probability;

[0161] The educational resource recommendation model is further configured to construct a label smoothing target distribution based on supervised labels, and generate a recommendation loss term based on the label smoothing target distribution and the fusion prediction probability;

[0162] The educational resource recommendation model is further configured to perform temperature softening processing on the full course rating vector and the branch course rating vectors corresponding to the semantic perspective branch, path perspective branch and conceptual structure perspective branch respectively, to obtain the fusion distribution and the branch distribution under the three perspective branches, and to construct a cross-perspective consistency distillation loss term based on the fusion distribution and the branch distribution under the three perspective branches.

[0163] The educational resource recommendation model is further configured to construct a total loss term based on the recommendation loss term and the cross-perspective consistency distillation loss term, and to update the model parameters through backpropagation based on the total loss term to complete joint training;

[0164] The educational resource recommendation model is further configured to calculate the target course score under the fusion perspective representation based on the updated model parameters, select a preset number of recommended courses according to the target course score, and output the recommended courses as the course recommendation prediction result to output the educational resource sequence recommendation list for student users.

[0165] In the specific implementation, refer to Figure 3 , Figure 3 This is a schematic diagram of the framework of an educational resource recommendation model. The model is configured to perform Transformer sequence modeling, full-scale scoring, and cross-perspective consistency distillation as joint training objectives, specifically including:

[0166] Constructing a fixed-length input sequence and supervision labels:

[0167]

[0168] The sequence embedding is obtained by looking up a table using the unified curriculum representation:

[0169]

[0170] Overlay position embedding:

[0171]

[0172] Transformer encoding yields the student state vector:

[0173]

[0174] Full course rating (integrated representation):

[0175]

[0176] Full course rating (each perspective branch):

[0177]

[0178] The fused prediction probability is obtained using softmax:

[0179]

[0180] Construct a label smoothing target distribution:

[0181]

[0182] Recommended loss:

[0183]

[0184] Fusion distribution after temperature softening (teacher distribution):

[0185]

[0186] Branch distribution of the three-view branches after temperature softening (student distribution):

[0187]

[0188]

[0189]

[0190] Define KL divergence:

[0191]

[0192] Cross-perspective consistency distillation loss (teachers are fused distributions, students are branch distributions):

[0193]

[0194] Total loss:

[0195]

[0196] in, For students A fixed-length input sequence, The maximum sequence length, To monitor the index of labeled courses, For sequence embedding matrix, Let represent the t-th row of the sequence embedding matrix for student u, corresponding to the embedding vector of the t-th course in the input sequence. For the position embedding matrix, This indicates the dimension of the position embedding matrix. Input sequences to the Transformer to represent features; For Transformer encoders, for The student state vector, The dimension representing the student's state vector; The full course scoring vector under the fusion representation; , , These are the scoring vectors for each perspective branch; To fuse predicted probabilities, and Indicates a course index; It is an exponential function; Smooth the target distribution of labels; This is the label smoothing coefficient; For the set of student indices in the training batch; The batch size; It is the natural logarithm; This is the distillation temperature coefficient; This represents the fusion distribution probability after temperature softening. These represent the branch distribution probabilities under semantic perspective branches, path perspective branches, and conceptual structure perspective branches, respectively. Use a minimal constant for numerical stability; Let KL divergence be a metric. Recommended loss; For cross-perspective consistent distillation loss; This is the weighting coefficient for distillation loss; This represents the total loss.

[0197] Reference Figure 4 , Figure 4 This is a schematic diagram of an educational resource sequence recommendation process based on joint optimization of semantic prototype soft hypergraphs, path-level transition structures, and concept structure perspective consistency distillation in one embodiment. The recommendation device constructs standardized student learning sequences based on student-course interaction logs through preprocessing, time sorting, and index mapping. Multidimensional features (semantics, path, and concept structure) of the sequences are extracted, and corresponding three types of perspective hypergraphs are constructed and optimized. The three types of hypergraphs are propagated and encoded, and the resulting three single-perspective feature matrices are fused to generate a three-perspective course representation matrix. This matrix is ​​input into a pre-trained recommendation model, which generates and outputs a personalized educational resource sequence recommendation list for students through inference. After obtaining course metadata and student-course interaction log data, this method first constructs course semantic text using course names and course descriptions, and obtains course semantic vectors through a text encoder. Based on this, density clustering is used to adaptively obtain several semantic prototypes (cluster centers), and many-to-many connections are established between courses and semantic prototypes according to similarity, forming a semantic perspective course-semantic prototype bipartite soft hypergraph structure. This structure is used to express the semantic attribution of multiple course topics and reduce KNN-style graph noise.

[0198] Meanwhile, focusing on the patterns of learning behavior, this invention no longer relies solely on adjacent transition statistics. Instead, it mines high-frequency path segments of fixed length from the learning sequences of a group of students, structuring the joint occurrence and sequential information of multiple courses within these path segments into path-level transition hyperedges. This forms a path-perspective structure used to characterize the higher-order dependencies and stage-based evolution of learning paths. Furthermore, this invention introduces a conceptual structure perspective: based on automatically extracting or matching concept (knowledge point) nodes from course texts, it constructs course-concept inclusion relationships. Combining sequence sequence statistics and co-occurrence statistics of the same course, it automatically generates prerequisite and relevance relationships between concepts, forming a heterogeneous conceptual structure network of "course node-concept node-course node," providing interpretable prior knowledge and supplementary information for cold starts.

[0199] During the model learning phase, this invention performs structural representation learning from semantic, path, and conceptual perspectives to obtain three sets of course representations, which are then combined into a unified course representation through a nonlinear fusion module. This unified course representation serves as the input embedding for the Transformer sequence encoder, modeling the dynamic evolution of student learning states and outputting a score prediction for the entire course. To replace the role of contrastive learning in "cross-perspective alignment, improving representation consistency, and training stability," this invention employs a cross-perspective consistency distillation mechanism: using the prediction distribution output by the fusion model as a "soft teacher," it constrains the prediction distributions output by each perspective branch to maintain consistency with the fusion distribution, thereby achieving cross-perspective information collaboration and consistency constraints, and avoiding training instability caused by negative sample construction and temperature sensitivity.

[0200] Compared with existing technologies, the present invention can achieve the following technical effects: (1) By using density clustering semantic prototype and soft connection bipartite structure, the semantics of course texts are injected into the sequence recommendation model in a noise-resistant structured manner, thereby improving the representation and transferability of cold start and long-tail courses; (2) By using path-level transfer structure to model the high-order segments of the learning sequence, the "multi-step path pattern" is explicitly depicted rather than just the adjacent pattern, thereby improving the ability to depict the real learning path and the stability of sequence prediction; (3) By introducing interpretable knowledge point associations and prerequisite / related relationships from the perspective of conceptual structure, the sparsity of pure behavioral statistics is supplemented, so that the recommendation results have knowledge structure interpretability and structural consistency in addition to the semantic consistency of the content; (4) By replacing contrastive learning with cross-perspective consistency distillation, cross-perspective collaborative alignment and fusion prediction stability are improved, and the training noise and fusion representation fragmentation risk caused by negative sample sampling are reduced; (5) By directly embedding the fused course representation as the input of Transformer, the sequence modeling utilizes semantic structure information, path structure information and conceptual structure information at the same time, thereby improving the recommendation accuracy and generalization ability. This invention is applicable to personalized recommendation of courses and educational resources in scenarios such as online education platforms and smart campus systems.

[0201] refer to Figure 5 , Figure 5 This is a flowchart illustrating the second embodiment of the educational resource sequence recommendation method based on three-view structural distillation of the present invention.

[0202] Based on the first embodiment described above, in this embodiment, step S20 further includes:

[0203] Step S201: Generate the semantic text of the interactive courses corresponding to each course index in the student learning sequence.

[0204] It should be noted that semantic text can be a collection of textual information used to describe the core content, knowledge topics, teaching objectives, and content summary of an interactive course, which can intuitively reflect the content attributes of the course.

[0205] In practice, the recommended device can construct semantic text for each course:

[0206]

[0207] in, For the course Semantic text; For the course Course name text; For the course Course description text; This is a text concatenation operation.

[0208] In some embodiments, the recommending device can obtain all course indices in the student's learning sequence, and use a previously defined mapping function to reverse map and obtain the original course identifier corresponding to each course index; retrieve the basic course information corresponding to each original course identifier from the course database of the education and learning platform, and filter out the core text used to represent the course semantics (such as course title, course introduction, knowledge point tags, syllabus summary, etc.); perform simple cleaning on the retrieved course semantic-related text, remove meaningless characters, duplicate text and redundant information, and integrate it into semantic text of a unified format to ensure that each course index corresponds to a unique and complete semantic text.

[0209] Step S202: Encode the semantic text using a text encoder to obtain the course semantic vector.

[0210] In practical implementation, the recommended device can use a text encoder to obtain the course semantic vector:

[0211]

[0212] in, For text encoders; For the course semantic vectors, For vector dimensions; It is the field of real numbers.

[0213] In some embodiments, the recommended device may select a suitable text encoder (such as a pre-trained BERT encoder) and adjust the encoder's input parameters (such as text truncation length, embedding dimension, etc.) according to the length and complexity of the course semantic text. The semantic text of each course is input into the adjusted text encoder one by one, and the encoder automatically mines the semantic information and extracts semantic features from the text. The output layer of the encoder obtains the quantization vector corresponding to each semantic text, performs standardization processing on the vector to eliminate the difference in units, and finally obtains the course semantic vector corresponding to each course.

[0214] Step S203: Stack the course semantic vectors of each interactive course to obtain a course semantic vector matrix, which is used to represent the semantic features of the student learning sequence.

[0215] It should be noted that the course semantic vector matrix can be a two-dimensional matrix formed by stacking the course semantic vectors corresponding to all interactive courses in the student learning sequence one by one in the chronological order of the courses in the sequence. The rows of the matrix correspond to the semantic vector of a single course, and the columns correspond to the semantic feature dimensions. It is a centralized quantitative representation of the semantic features of the student learning sequence.

[0216] In the specific implementation, the recommendation device stacks the course semantic vectors of each interactive course to obtain a course semantic vector matrix:

[0217]

[0218] in, This is the semantic vector matrix of the course; For vector dimensions; For the real number field; This indicates transpose, and N represents the number of courses.

[0219] Step S204: Construct a semantic perspective hypergraph based on the course semantic vector matrix.

[0220] In some embodiments, the recommended device uses the course corresponding to each row vector in the course semantic vector matrix as a node of the semantic view hypergraph to determine the node set of the hypergraph; calculates the similarity between any two course semantic vectors in the course semantic vector matrix (using methods such as cosine similarity and Euclidean distance), and uses the similarity as a measure of the strength of semantic association between courses; sets a semantic similarity threshold, connects multiple courses with similarity higher than the threshold through hyperedges, and sets the weight of the hyperedges to the average semantic similarity between the corresponding courses; simplifies the constructed initial semantic view hypergraph by removing hyperedges with too low weights and isolated nodes to ensure that the hypergraph structure is concise and effective, and finally completes the construction of the semantic view hypergraph.

[0221] Furthermore, in order to accurately model the semantic view hypergraph, step S204 above may include:

[0222] Step S2041: Perform density clustering analysis based on the course semantic vector matrix to obtain the course cluster labels for each interactive course;

[0223] Step S2042: Construct a set of valid cluster tags based on the course cluster tags, and determine the number of semantic prototypes;

[0224] Step S2043: Renumber the valid cluster labels in the set of valid cluster labels as continuous semantic prototype indices;

[0225] Step S2044: Calculate the semantic prototype vector based on the number of semantic prototypes and the continuous semantic prototype index, where the semantic prototype vector is the cluster center;

[0226] Step S2045: Stack the semantic prototype vectors to obtain a semantic prototype matrix, and calculate the normalized vector and similarity based on the semantic prototype matrix;

[0227] Step S2046: Construct a course-semantic prototype bipartite association matrix based on the normalized vector and similarity, and construct a semantic perspective hypergraph based on the course-semantic prototype bipartite association matrix.

[0228] In the specific implementation, semantic prototype construction and semantic prototype soft hypergraph generation include:

[0229] Density clustering yields course cluster labels:

[0230]

[0231] Construct a valid set of cluster labels and define the number of semantic prototypes:

[0232]

[0233] Renumber the valid cluster labels as continuous semantic prototype indices:

[0234]

[0235] Calculate the semantic prototype vector (cluster center):

[0236]

[0237] Stacking yields a semantic prototype matrix:

[0238]

[0239] Calculate normalized vectors and similarity:

[0240]

[0241] Constructing Courses - Semantic Prototype Bipartite Association Matrix (Top- for each course) (One prototype):

[0242]

[0243] in: It is a density clustering algorithm; For the course Cluster labels; For noise labeling; A valid set of cluster labels; Represents the cardinality of a set; The number of semantic prototypes; Re-number and map cluster labels; It is the inverse mapping; For semantic prototype index; It is a semantic prototype vector; This is a semantic prototype matrix; It is a norm 2; , A normalized vector; Similarity; This is an operator that takes the first few maximum values. The number of semantic prototypes connected to each course; For indicator functions; The course-semantic prototype bipartite association matrix.

[0244] Step S205: Perform path structure analysis on the interactive courses in the student learning sequence to obtain path features, and construct a path view hypergraph based on the path features.

[0245] It should be noted that path structure analysis can be based on the temporal sequence of courses in a student's learning sequence, analyzing the relationships between courses such as jump logic, continuous interaction patterns, and co-occurrence frequency, thereby uncovering path patterns in the student's learning behavior trajectory.

[0246] It should be noted that path features refer to quantitative information extracted through path structure analysis that reflects the course jump associations and continuous interaction characteristics in the student's learning trajectory, including course jump frequency, number of co-occurrences of adjacent courses, path length, path repetition rate, etc.

[0247] In some embodiments, the recommended device can retrieve the student's learning sequence and, based on the chronological order of courses in the sequence, analyze the jump relationships between each course and its preceding and succeeding courses to form a set of course jump paths. Statistical analysis is performed on the set of course jump paths to extract path features: the jump frequency of any two courses and the co-occurrence frequency of adjacent courses are statistically analyzed, and the length and repetition rate of each path are calculated. These statistical information are quantified into a path feature vector. Using interactive courses in the student's learning sequence as nodes, and using the jump frequency and co-occurrence frequency in the path features as hyperedge weights, multiple courses with frequent jumps and high-frequency co-occurrence are connected through hyperedges. The hypergraph structure is optimized by removing hyperedges with too low jump frequency or no actual connection, adjusting the quantification standard of hyperedge weights, and finally completing the construction of the path-view hypergraph.

[0248] Furthermore, in order to accurately model the path-view hypergraph, step S205 above may include:

[0249] Step S2051: Extract a continuous path segment with a preset path length from the student learning sequence. The continuous path segment is a sequence composed of course indices at corresponding positions in the student learning sequence.

[0250] Step S2052: Define the continuous path segment as a path pattern, wherein the path pattern is a sequence of the corresponding course index;

[0251] Step S2053: Count the frequency of occurrence of each path pattern, where the frequency of occurrence is the sum of the number of consecutive path segments in all student learning sequences that are completely consistent with the path pattern;

[0252] Step S2054: Path patterns that occur more frequently than a preset frequency threshold are identified as frequent path patterns;

[0253] Step S2055: Calculate the attenuation weight corresponding to the course index at each position in the frequent path pattern based on the preset position attenuation coefficient, sum the attenuation weights of the same course index at each position in the frequent path pattern, and construct the course-path pattern association matrix. The course-path pattern association matrix is ​​used to characterize the path features.

[0254] Step S2056: Generate a path perspective hypergraph based on the course-path pattern association matrix.

[0255] In the specific implementation, path-level transition structure mining and path-view hypergraph generation include:

[0256] Extracting a length of [length] from the student learning sequence Continuous path segments:

[0257]

[0258] Select the frequent path patterns and number them. Each pattern is written as:

[0259]

[0260] Define the frequency of occurrence for each path pattern:

[0261]

[0262] Constructing a Path-Based Course - Path Pattern Incidence Matrix (Introducing Position Attenuation Coefficient) (Reflecting priority)

[0263]

[0264] in: The length of the path segment and ; For students In position Extracted path segments; The number of frequent path patterns selected; For path pattern index and ; For the first One path pattern; For the path pattern Course index for each location; Frequency of occurrence of path patterns; The position attenuation coefficient and ; This is a course-path pattern association matrix. This indicates the course index.

[0265] Step S206: Perform conceptual structure analysis on the interactive courses in the student learning sequence to obtain conceptual structure features, and construct a conceptual structure perspective hypergraph based on the conceptual structure features.

[0266] It should be noted that conceptual structure analysis refers to the process of analyzing the hierarchical position of interactive courses in the knowledge framework and the types of knowledge associations (such as similar knowledge, progressive knowledge, and related knowledge) in the student's learning sequence by combining the knowledge system and knowledge graph corresponding to the course, thereby uncovering the knowledge structure associations behind the course.

[0267] It should be noted that conceptual structure features can be quantitative features extracted through conceptual structure analysis that reflect the course's hierarchy in the knowledge system, the strength of knowledge connections, and the affiliation of knowledge modules, and are the core manifestation of the course's knowledge attributes.

[0268] In some embodiments, the recommended device can access the knowledge graph corresponding to the course to clarify the hierarchical position of each interactive course in the knowledge graph, its knowledge module, and related courses (such as courses in the same module, courses with progressive relationships, and courses with complementary relationships). It performs conceptual structure analysis on the interactive courses in the student's learning sequence, extracting conceptual structure features: quantifying the knowledge level of the course, calculating the knowledge association strength between the course and related courses, and labeling the knowledge module affiliation of the course. This information is then transformed into a quantified conceptual structure feature vector. Using interactive courses as nodes and the knowledge association strength in the conceptual structure features as the hyperedge weight, multiple courses in the same module, with progressive or complementary relationships, are connected through hyperedges, clarifying the knowledge association type corresponding to the hyperedge. The hypergraph is optimized by removing hyperedges with too low knowledge association strength, standardizing the quantification criteria of the hyperedge weights, and ensuring that the hypergraph can accurately depict the knowledge structure relationships of the courses, ultimately completing the construction of the hypergraph from a conceptual structure perspective.

[0269] Furthermore, in order to accurately model the conceptual structure view hypergraph, step S206 above may include:

[0270] Step S2061: Construct a course-concept association matrix. In the course-concept association matrix, when a concept is extracted or matched with a corresponding course, the value of the corresponding position in the matrix is ​​1, otherwise it is 0.

[0271] Step S2062: Based on the student learning sequence and the course-concept association matrix, count the sequence co-occurrence count of the concept prerequisite direction, where the sequence co-occurrence count is the number of co-occurrences of concepts matched in the preceding and following courses in the student learning sequence;

[0272] Step S2063: Based on the course-concept association matrix, count the co-occurrence of concepts in the same course, where the co-occurrence count is the number of concepts that are simultaneously matched in the same course;

[0273] Step S2064: Perform probabilistic processing on the sequence co-occurrence count and the co-occurrence count of the same lesson respectively to obtain the concept prior relation strength and the concept relevance relation strength. Then, fuse the concept prior relation strength and the concept relevance relation strength through a preset fusion weight coefficient to obtain the concept relation matrix.

[0274] Step S2065: Map the course semantic vector matrix of the student learning sequence to the initial feature matrix of the course from the concept perspective using the concept perspective feature mapping function;

[0275] Step S2066: Based on the initial feature matrix of the concept perspective course and the concept relation matrix, perform concept structure propagation to generate a concept perspective course representation matrix, and construct a concept structure perspective hypergraph based on the concept perspective course representation matrix.

[0276] In its specific implementation, the construction of the concept structure perspective and the learning of relation-aware representations include:

[0277] Constructing a Course - Concept Association Matrix:

[0278]

[0279] Co-occurrence count of statistical concepts based on prerequisites (concept sequence induced by course sequence):

[0280]

[0281] Counting the co-occurrence of concepts related to each other in the same course (correlation induced by the co-occurrence of concepts in the same course):

[0282]

[0283] The strength of concept relationships is obtained by probabilizing the counts:

[0284]

[0285] By integrating prerequisites and related relationships, we obtain the concept relationship matrix:

[0286]

[0287] Constructing a conceptual perspective: Initial characteristics of a course and initial characteristics of concepts:

[0288]

[0289] The conceptual structure is propagated and fed back to the curriculum side to obtain a conceptual perspective curriculum representation:

[0290]

[0291]

[0292]

[0293] in, The total number of concepts; For concept index and ; For concept indexing; Course-concept association matrix; , Count the conceptual relations; , The strength of the concept relationship after probabilization; Use a minimal constant for numerical stability; , For fusion weighting coefficients; A concept relation matrix; For conceptual perspective feature mapping function; This is the initial feature matrix for the conceptual perspective course. Representation matrix of course from the perspective of concepts This represents the initial feature matrix of a trainable concept. and These represent the intermediate feature matrices of concept propagation. Representing the course-concept association matrix, Indicates transpose. Represents the concept relation matrix. Represents the total number of concepts. This indicates the total number of courses.

[0294] This embodiment uses conceptual structure analysis to uncover the knowledge structure connections behind the course, transforming the surface interactions of the course into deep knowledge connections, and realizing the extraction of knowledge-level features. The hypergraph constructed based on the conceptual structure features can clearly depict the position and relationship of the course in the knowledge system, highlight the knowledge attributes of the course, provide knowledge-level feature support for subsequent hypergraph encoding, further improve the multi-perspective feature system, and ensure that the three types of perspective hypergraphs can fully cover the different attributes and relationships of the course.

[0295] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing an educational resource sequence recommendation program based on three-view structure distillation. When the educational resource sequence recommendation program based on three-view structure distillation is executed by a processor, it implements the steps of the educational resource sequence recommendation method based on three-view structure distillation as described above.

[0296] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0297] The aforementioned computer-readable storage medium may be included in an educational resource sequence recommendation device based on three-view structure distillation; or it may exist independently and not assembled into an educational resource sequence recommendation device based on three-view structure distillation.

[0298] Furthermore, this invention also proposes a computer program product, including an educational resource sequence recommendation program based on three-view structure distillation, wherein when the educational resource sequence recommendation program based on three-view structure distillation is executed by a processor, it implements the steps of the educational resource sequence recommendation method based on three-view structure distillation as described above.

[0299] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the educational resource sequence recommendation method based on three-view structure distillation described above, and will not be repeated here.

[0300] Reference Figure 6 , Figure 6 This is a structural block diagram of the first embodiment of the educational resource sequence recommendation device based on three-view structural distillation of the present invention.

[0301] like Figure 6 As shown, the educational resource sequence recommendation device based on three-view structural distillation proposed in this embodiment of the invention includes:

[0302] The learning sequence construction module 10 is used to construct a student learning sequence based on the student-course interaction log of the student user. The student learning sequence contains course indexes arranged in chronological order. The course indexes are obtained by mapping the course identifiers of the interactive courses through a mapping function.

[0303] The three-view hypergraph construction module 20 is used to construct a semantic view hypergraph, a path view hypergraph, and a concept structure view hypergraph based on the multidimensional features of the student learning sequence, wherein the multidimensional features include semantic features, path features, and concept structure features.

[0304] The hypergraph fusion module 30 is used to fuse the semantic view hypergraph, the path view hypergraph and the concept structure view hypergraph after structural encoding to generate a three-view course representation matrix.

[0305] The educational resource recommendation module 40 is used to input the three-view course representation matrix into the pre-trained educational resource recommendation model and output the educational resource sequence recommendation list for the student user.

[0306] The educational resource recommendation model is configured to encode the three-view course representation matrix into a student state vector using a Transformer encoder, and to rate courses from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch based on the student state vector. It then performs course recommendation prediction based on the full course rating vector and outputs a recommended list of educational resource sequences for the student user based on the prediction results. The full course rating vector is obtained by fusing the course ratings from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch.

[0307] This embodiment fully explores the multi-dimensional features and complex relationships of courses through the construction and encoding fusion of a three-view hypergraph, effectively compensating for the limitations of single-view features and improving the completeness and discriminativeness of course feature expression. Through the multi-branch scoring and fusion mechanism of the educational resource recommendation model, multi-view features are accurately transformed into personalized recommendation results, enhancing the rationality and adaptability of recommendation prediction. It accurately captures course representations and effectively portrays higher-order rules of learning paths in sparse interaction and cold-start scenarios, thereby effectively mining student learning trajectories and multi-dimensional features of courses, improving the accuracy and personalization of educational resource recommendations, optimizing student learning experience, standardizing the recommendation process, improving recommendation efficiency, reducing resource waste caused by ineffective recommendations, and achieving precise matching of educational resources with student learning needs. This provides an efficient and feasible technical path for personalized education recommendations.

[0308] The educational resource sequence recommendation device based on three-view structure distillation provided in this application adopts the educational resource sequence recommendation method based on three-view structure distillation in the above embodiments, and can solve the technical problems of educational resource sequence recommendation based on three-view structure distillation. Compared with the prior art, the beneficial effects of the educational resource sequence recommendation device based on three-view structure distillation provided in this application are the same as the beneficial effects of the educational resource sequence recommendation method based on three-view structure distillation provided in the above embodiments, and other technical features in the educational resource sequence recommendation device based on three-view structure distillation are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0309] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0310] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0311] In addition, for technical details not described in detail in this embodiment, please refer to the educational resource sequence recommendation method based on three-view structural distillation provided in any embodiment of the present invention, which will not be repeated here.

[0312] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0313] It should be noted that the user information (including but not limited to user device information, user personal information, user location information, user behavior information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0314] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0315] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0316] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for recommending educational resource sequences based on three-view structural distillation, characterized in that, The educational resource sequence recommendation method based on three-view structure distillation includes: A student learning sequence is constructed based on the student-course interaction logs of student users. The student learning sequence contains course indices arranged in chronological order. The course indices are obtained by mapping the course identifiers of the interactive courses through a mapping function. Based on the multidimensional features of the student learning sequence, a semantic perspective hypergraph, a path perspective hypergraph, and a concept structure perspective hypergraph are constructed respectively. The multidimensional features include semantic features, path features, and concept structure features. The semantic features are used to describe the content attributes of the course itself. The path features reflect the jump logic and sequential dependencies of the course during the student's learning process. The concept structure features reflect the position and interrelationship of the course in the overall knowledge framework. After structural encoding of the semantic view hypergraph, the path view hypergraph, and the concept structure view hypergraph, they are fused to generate a three-view course representation matrix. The three-view course representation matrix is ​​input into a pre-trained educational resource recommendation model, which outputs a recommended list of educational resource sequences for the student user. The educational resource recommendation model is configured to encode the three-view course representation matrix into a student state vector using a Transformer encoder, and to score courses based on the student state vector from the semantic view branch, the path view branch, and the conceptual structure view branch, respectively. It then performs course recommendation prediction based on the full course score vector and outputs a recommended list of educational resource sequences for the student user based on the course recommendation prediction results. The full course score vector is obtained by fusing the course scores from the semantic view branch, the path view branch, and the conceptual structure view branch. The educational resource recommendation model is further configured to construct a fixed-length input sequence and supervision labels, obtain sequence embeddings by looking up a table according to a unified curriculum representation, and superimpose positional embeddings to generate Transformer input sequence representation features, as shown in the following formula: in, For students A fixed-length input sequence, The maximum sequence length, To monitor the index of labeled courses, For sequence embedding matrix, Let represent the t-th row of the sequence embedding matrix for student u, corresponding to the embedding vector of the t-th course in the input sequence. For the position embedding matrix, This indicates the dimension of the position embedding matrix. Input sequences to the Transformer to represent features; The educational resource recommendation model is further configured to encode the features of the Transformer input sequence using a Transformer encoder to obtain a student state vector, as shown in the following formula: in, For Transformer encoders, for The student state vector, The dimension representing the student's state vector; The educational resource recommendation model is further configured to score courses based on the student state vector from semantic perspective, path perspective, and conceptual structure perspective to obtain a full course score vector. The educational resource recommendation model is further configured to predict course recommendations based on the full set of course rating vectors, and obtain the fusion prediction probability, as shown in the following formula: in, Let represent the full course scoring vector under the three-view fusion representation. Indicates the fusion prediction probability. It is an exponential function. and Indicates a course index; The educational resource recommendation model is further configured to construct a label smoothing target distribution based on supervised labels, and generate a recommendation loss term based on the label smoothing target distribution and the fusion prediction probability, as shown in the following formula: in, This indicates that the label smooths the target distribution. The label smoothing coefficient. Represents the student index. For the set of student indices in the training batch, The batch size, It is the natural logarithm. Indicates the recommended loss term; The educational resource recommendation model is further configured to perform temperature softening processing on the full course rating vector and the branch course rating vectors corresponding to the semantic perspective branch, path perspective branch, and conceptual structure perspective branch, respectively, to obtain the fusion distribution and the branch distribution under the three perspective branches. Based on the fusion distribution and the branch distribution under the three perspective branches, a cross-perspective consistency distillation loss term is constructed, as shown in the following formula: in, This represents the cross-perspective consistent distillation loss term. This is the distillation temperature coefficient. This indicates the fusion distribution after temperature softening. , and These represent the branch distribution under the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch, respectively. Let KL divergence be a metric. The educational resource recommendation model is further configured to construct a total loss term based on the recommendation loss term and the cross-perspective consistency distillation loss term, and to update the model parameters through backpropagation based on the total loss term to complete joint training; The educational resource recommendation model is further configured to calculate the target course score under the fusion perspective representation based on the updated model parameters, select a preset number of recommended courses according to the target course score, and use the recommended courses as the course recommendation prediction result to output a recommended list of educational resource sequences for student users. The construction of semantic view hypergraph, path view hypergraph, and concept structure view hypergraph based on the multidimensional features of the student learning sequence includes: Generate semantic text of the interactive courses corresponding to each course index in the student learning sequence; The semantic text is encoded using a text encoder to obtain the course semantic vector; The course semantic vectors of each interactive course are stacked to obtain a course semantic vector matrix, which is used to characterize the semantic features of the student learning sequence. Construct a semantic perspective hypergraph based on the aforementioned course semantic vector matrix; The path structure of interactive courses in the student learning sequence is analyzed to obtain path features, and a path view hypergraph is constructed based on the path features. A conceptual structure analysis is performed on the interactive courses in the student learning sequence to obtain conceptual structure features, and a conceptual structure perspective hypergraph is constructed based on the conceptual structure features.

2. The educational resource sequence recommendation method based on three-view structural distillation as described in claim 1, characterized in that, The construction of the semantic perspective hypergraph based on the course semantic vector matrix includes: Density clustering analysis is performed based on the course semantic vector matrix to obtain the course cluster labels for each interactive course; Construct a valid set of cluster tags based on the course cluster tags, and determine the number of semantic prototypes; Renumber the valid cluster labels in the set of valid cluster labels as continuous semantic prototype indices; The semantic prototype vector is calculated based on the number of semantic prototypes and the continuous semantic prototype index, and the semantic prototype vector is the cluster center. The semantic prototype vectors are stacked to obtain a semantic prototype matrix, and normalized vectors and similarity are calculated based on the semantic prototype matrix. Based on the normalized vectors and similarity, a course-semantic prototype bipartite association matrix is ​​constructed, and a semantic perspective hypergraph is constructed based on the course-semantic prototype bipartite association matrix.

3. The educational resource sequence recommendation method based on three-view structural distillation as described in claim 2, characterized in that, The step of performing path structure analysis on interactive courses in the student learning sequence to obtain path features, and constructing a path perspective hypergraph based on the path features, includes: Extract continuous path segments of a preset path length from the student learning sequence. The continuous path segments are sequences composed of course indices at corresponding positions in the student learning sequence. The continuous path segments are defined as path patterns, and the path patterns are in the form of a sequence of corresponding course indices; The frequency of occurrence of each path pattern is counted, where the frequency of occurrence is the sum of the number of consecutive path segments in all student learning sequences that are completely consistent with that path pattern; Path patterns that occur more frequently than a preset frequency threshold are classified as frequent path patterns. The attenuation weights corresponding to the course indices at each position in the frequent path pattern are calculated based on the preset position attenuation coefficient. The attenuation weights of the same course index at each position in the frequent path pattern are summed to construct the course-path pattern association matrix. The course-path pattern association matrix is ​​used to characterize the path features. Based on the course-path pattern association matrix, a path perspective hypergraph is generated.

4. The educational resource sequence recommendation method based on three-view structural distillation as described in claim 3, characterized in that, The step of performing conceptual structure analysis on the interactive courses in the student learning sequence to obtain conceptual structure features, and constructing a conceptual structure view hypergraph based on the conceptual structure features, includes: Construct a course-concept association matrix. In the course-concept association matrix, when a concept is extracted or matched with a corresponding course, the value of the corresponding position in the matrix is ​​1, otherwise it is 0. Based on the student learning sequence and the course-concept association matrix, the sequence co-occurrence count of concept prerequisite directions is calculated. The sequence co-occurrence count is the number of co-occurrences of concepts matched in the preceding and following courses in the student learning sequence. Based on the course-concept association matrix, the co-occurrence count of concepts in the same course is calculated, where the co-occurrence count is the number of concepts that are simultaneously matched in the same course. The sequence co-occurrence count and the co-occurrence count of the same lesson are respectively probabilistically processed to obtain the concept prior relation strength and the concept relevance relation strength. Then, the concept prior relation strength and the concept relevance relation strength are fused by a preset fusion weight coefficient to obtain the concept relation matrix. The concept-perspective feature mapping function maps the course semantic vector matrix of the student learning sequence to the concept-perspective course initial feature matrix. Based on the initial feature matrix and the concept relation matrix of the concept-view course, concept structure propagation is performed to generate a concept-view course representation matrix. Then, a concept structure-view hypergraph is constructed based on this representation matrix, as shown in the following formula: in, Representation matrix of course from the perspective of concepts This is the initial feature matrix for the conceptual perspective course. This represents the initial feature matrix of a trainable concept. and These represent the intermediate feature matrices of concept propagation. Representing the course-concept association matrix, Indicates transpose. Represents the concept relation matrix. Represents the total number of concepts. This indicates the total number of courses.

5. The educational resource sequence recommendation method based on three-view structural distillation as described in claim 4, characterized in that, The semantic perspective hypergraph, the path perspective hypergraph, and the concept structure perspective hypergraph are structurally encoded and then fused to generate a three-view course representation matrix, including: The semantic perspective hypergraph is subjected to semantic perspective propagation encoding to obtain the semantic perspective course representation matrix, as shown in the following formula: in, The course represents the representation matrix from a semantic perspective. Representation layer normalization, This represents the element-wise activation function. Represents the random deactivation function. The input feature matrix represents the semantic view hypergraph. Represents the course-semantic prototype bipartite association matrix; The path-view hypergraph is subjected to path-view propagation encoding to obtain the path-view course representation matrix, as shown in the following formula: in, The path-view course representation matrix The input feature matrix represents the path-view hypergraph. Represents the course-path pattern association matrix; The semantic perspective course representation matrix, the path perspective course representation matrix, and the concept perspective course representation matrix corresponding to the concept structure perspective hypergraph are fused to obtain a three-perspective course representation matrix, as shown in the following formula: in, A matrix representing a three-view course. Represents the fusion mapping function, The three-view course represents the dimension of the matrix. Representation matrix of course from the perspective of concepts This indicates a splicing operation based on feature dimensions.

6. An educational resource sequence recommendation device based on three-view structural distillation, characterized in that, The apparatus is configured to implement the educational resource sequence recommendation method based on three-view structural distillation as described in any one of claims 1 to 5, the apparatus comprising: The learning sequence construction module is used to construct student learning sequences based on student-course interaction logs of student users. The student learning sequence contains course indices arranged in chronological order, and the course indices are obtained by mapping the course identifiers of the interactive courses through a mapping function. The three-view hypergraph construction module is used to construct a semantic view hypergraph, a path view hypergraph, and a concept structure view hypergraph based on the multidimensional features of the student learning sequence, wherein the multidimensional features include semantic features, path features, and concept structure features. The hypergraph fusion module is used to fuse the semantic view hypergraph, the path view hypergraph and the concept structure view hypergraph after structural encoding to generate a three-view course representation matrix. The educational resource recommendation module is used to input the three-view course representation matrix into the pre-trained educational resource recommendation model and output the educational resource sequence recommendation list for the student user. The educational resource recommendation model is configured to encode the three-view course representation matrix into a student state vector using a Transformer encoder, and to rate courses from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch based on the student state vector. It then performs course recommendation prediction based on the full course rating vector and outputs a recommended list of educational resource sequences for the student user based on the prediction results. The full course rating vector is obtained by fusing the course ratings from the semantic perspective branch, the path perspective branch, and the conceptual structure perspective branch.

7. An educational resource sequence recommendation device based on three-view structural distillation, characterized in that, The educational resource sequence recommendation device based on three-view structure distillation includes: a memory, a processor, and an educational resource sequence recommendation program based on three-view structure distillation stored in the memory. The processor is used to run the educational resource sequence recommendation program based on three-view structure distillation, and the educational resource sequence recommendation program based on three-view structure distillation is configured to implement the educational resource sequence recommendation method based on three-view structure distillation as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an educational resource sequence recommendation program based on three-view structure distillation, which, when executed by a processor, implements the educational resource sequence recommendation method based on three-view structure distillation as described in any one of claims 1 to 5.

Images