Online course quality evaluation system and method based on knowledge graph

Abstract

The invention discloses an online course quality evaluation system and method based on a knowledge graph. The method specifically comprises the following steps: S1, constructing a semantic relation graph; s2, generating a hyperbolic node vector and a hyperbolic distance; s3, extracting structure connection attributes; s4, generating a path consistency scoring factor; s5, calculating a basic attention weight; s6, generating a behavior disturbance offset; s7, constructing a structure disturbance factor and generating a disturbance fusion attention weight; s8, entropy regularization updating is executed; and S9, generating a structural robustness index and outputting an evaluation result. Unified improvement of course structure relation accurate expression and behavior association evaluation is achieved, and the method can be applied to various online course quality analysis scenes.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph modeling and deep learning evaluation technology, and in particular to an online course quality evaluation system and method based on knowledge graphs. Background Technology

[0002] As online education platforms continue to expand, the number of course resources is growing rapidly, the organization of course content is becoming increasingly complex, and the learning path structure is gradually exhibiting network characteristics. Traditional course quality evaluation methods mainly rely on linear chapter order, manually labeled quality indicators, or simple scoring models based on statistical characteristics. These methods are difficult to reflect the semantic relationships between course content, and cannot handle the structural differences between chapter levels, knowledge dependencies, and learning objectives. The evaluation results often lack structural sensitivity and are difficult to adapt to the rapid changes in online course content.

[0003] Existing research attempts to incorporate course knowledge associations through graph structures, but most graph structures employ Euclidean embedding. Euclidean embedding suffers from limitations in handling multi-level relationships, tree-like structural expansion, and high-curvature semantic spaces. It struggles to effectively distinguish hierarchical differences and semantic directions between nodes, resulting in embedding results that fail to reflect the true structure of the course content. Furthermore, existing scoring models typically rely on simple sequential statistics in path analysis, failing to comprehensively assess chapter changes, difficulty trends, and hierarchical transitions, and thus unable to output path consistency indicators with structural interpretability.

[0004] Online learning behavior records provide a new data source for course quality evaluation, but existing methods generally employ behavioral feature extraction based on global statistics, which fails to map behavioral changes back to semantic structure, resulting in an unclear correspondence between learning behavior and course content structure. Behavioral shifts, rhythm changes, and jump behaviors have little direct impact in most models, leading to a lack of behavioral structural correlation in the model output scores and insufficient reliability.

[0005] Current attention models have been applied in educational evaluation, but traditional attention mechanisms, based on Euclidean space, lack sensitivity to structural factors such as path order, hierarchical span, and difficulty gradient. Furthermore, they cannot handle distance and direction information between nodes in hyperbolic space, making it difficult to obtain stable and interpretable attention weights on hierarchical semantic graphs. Existing models generally lack perturbation modeling mechanisms; learning behavior biases cannot drive dynamic adjustments to graph structure weights, resulting in a lack of responsiveness to behavioral perturbations in the evaluation process.

[0006] Therefore, the field of online course quality assessment urgently needs a comprehensive evaluation method that can express multi-level semantic structures, reflect path change patterns, integrate learning behavior biases, and has a stable training mechanism, in order to improve the structural accuracy, behavioral relevance, and model robustness of course quality assessment. Summary of the Invention

[0007] One objective of this invention is to propose an online course quality evaluation system and method based on knowledge graphs. This invention introduces a hyperbolic space embedding structure and a structural perturbation fusion attention model, which uniformly maps the hierarchical associations, path change patterns, and learning behavior offsets of the course semantic graph to a hyperbolic attention framework. Through path consistency scoring factors, behavioral perturbation offsets, and entropy regularization update mechanisms, a stable course quality evaluation process is constructed, which has the technical advantages of fine structural expression, clear behavioral associations, and robust evaluation results.

[0008] An online course quality evaluation system and method based on knowledge graphs according to an embodiment of the present invention includes the following steps: S1. Establish a set of semantic nodes for the course and construct structural connections between nodes to form a semantic relationship graph; S2. Map the semantic relationship graph to hyperbolic space to generate hyperbolic node vectors, and calculate the hyperbolic distance between nodes based on the hyperbolic node vectors to form a hyperbolic distance set; S3. Extract structural connection attribute values ​​from the semantic relationship graph to form a structural connection attribute set; S4. Construct a path consistency scoring factor, which is generated by node order, difficulty gradient, and hierarchical sequence; S5. Input the hyperbolic distance set, structural connection attribute set, and path consistency scoring factor into the hyperbolic attention structure to generate basic attention weights; S6. Introduce learning behavior sequences and record the temporal sequence in the learning behavior sequence. S7. The behavior perturbation offset is used to construct a structural perturbation factor, and the structural perturbation factor and the basic attention weight are input into the hyperbolic attention structure to generate perturbation fusion attention weights. S8. An entropy-regularized stochastic gradient update model is established. The entropy regularization term comes from the hyperbolic space neighborhood embedding distribution. The stochastic gradient process includes noise injection, neighborhood extraction and hyperbolic space shrinkage mapping. The hyperbolic node vector and perturbation fusion attention parameters are adjusted in the update stage. S9. In the update stage, the attention offset and embedding change are recorded. The attention offset and embedding change are combined to form a structural robustness index and output the online course quality evaluation results.

[0009] Optionally, S2 specifically includes: Create an initial vector for each node in the semantic relationship graph. The initial vector values ​​satisfy the norm being less than one. The radial elements of the vector reflect the node level, and the angular elements reflect the semantic direction. Subsequently, positive samples are formed by extracting connected node pairs from the semantic relationship graph, and negative samples are formed by extracting node pairs from unconnected nodes. Each sample pair enters the embedding learning process, which constrains the positional relationship of nodes in hyperbolic space according to the structural relationship of the sample pair. Construct an embedded objective function, which consists of three parts: distance constraints between connected nodes; separation constraints between unconnected nodes; and hierarchical sequence constraints. Using the objective function as the optimization objective, perform iterative updates. In each iteration, calculate the gradient of the node vector in Euclidean space, project the gradient onto the corresponding tangent space in hyperbolic space, complete the update in the tangent space, and then shrink the update result back into the hyperbolic space through exponential mapping. Once the objective function converges, the node vector becomes a hyperbolic node vector. The distance between any pair of nodes is then calculated using the hyperbolic distance function, forming a hyperbolic distance set.

[0010] Optionally, the hyperbolic distance function in step S2 for calculating the distance between any pair of nodes specifically includes: Generate a vector for each node and keep the length of the vector within a specified range; extract the node vectors for any two nodes, calculate the difference between the two vectors and obtain the square of the difference; calculate the length of the two vectors and take the length difference as the denominator; combine the square of the difference with the denominator to generate the distance ratio, then input the distance ratio into the inverse hyperbolic cosine function and output the corresponding distance, and combine the distances of all node pairs into a hyperbolic distance set.

[0011] Optionally, S3 specifically includes: For each structural connection in the semantic relationship graph, read the node type, level number, and difficulty number of the starting node and the ending node, and generate a connection type label by combining the node types; Calculate the difference between the level number and the difficulty number between the starting node and the ending node. Record the difference between the level numbers as the level span and the difference between the difficulty numbers as the difficulty span. Assign a sequence index tag to each structural connection, which reflects the position of the connection in the course chapter sequence; combine the connection type tag, level span, difficulty span and sequence index tag into a structural connection attribute vector, and collect all structural connection attribute vectors to form a structural connection attribute set.

[0012] Optionally, the path consistency scoring factors in step S4 specifically include: Select a target path containing several nodes in the semantic relationship graph, and generate a node sequence according to the order in which the nodes appear in the path; For each node in the node sequence, read the chapter number, level number, and difficulty number, and arrange the three numbers in order to form the level sequence and difficulty sequence; Assign index number i to any adjacent node pair in the hierarchy sequence and difficulty sequence, and calculate the difference between the hierarchy number and the difficulty number of the node pair. Record the absolute value of the difference between the hierarchy number and the absolute value of the difference between the difficulty number as the hierarchy change value. Each level change value and difficulty change value are classified according to a threshold T1. When the absolute value of the change value is less than the threshold T1, it is recorded as a stable node pair. When the absolute value of the change value is greater than or equal to the threshold T1, it is recorded as a mutation node pair. Using the index number as the scanning order, comparisons are performed from index i to index i+1. When the node pairs corresponding to two indices are consistent in classification, they are divided into the same node segment; when the classifications are inconsistent, they are divided into different node segments. The number of node pairs, the starting index position and the ending index position in the node sequence of each node segment are recorded, and the statistical information of all node segments is used to form path structure change characteristics. Based on the characteristics of path structure changes, a path consistency scoring factor is constructed, which includes three parts: sequence score, difficulty score, and hierarchy score. The sequence score determines whether there are chapter jump events in the path based on the direction of change of chapter numbers in the node sequence. The difficulty score determines the smoothness of the path difficulty change based on the mean, range, and number of abrupt node segments of the difficulty change value. The hierarchy score determines the coordination of the path in the hierarchy direction based on the mean, range, and distribution of the hierarchy change value. The three scores are weighted and aggregated, and the aggregated values ​​are mapped to a preset range to generate a path consistency score factor.

[0013] Optionally, S5 specifically includes: Receive the hyperbolic distance set, the structural connectivity attribute set, and the path consistency scoring factor, arrange the three types of inputs in node pair order, and feed them into the hyperbolic attention structure; In the hyperbolic attention structure, the attention distribution value between node pairs is calculated. The attention distribution value is affected by the hyperbolic distance change, the structural connectivity property change and the path consistency score. During the calculation process, spatial sensitivity terms are generated based on the hyperbolic positional relationship of node pairs, structural sensitivity terms are generated based on the structural connection attributes of node pairs, and path sensitivity terms are generated based on the path consistency score. The three types of sensitivity terms are then weighted and fused within the attention structure to obtain the basic attention weights.

[0014] Optionally, S6 specifically includes: Receive time-series records from the learning behavior sequence, which include dwell time, access frequency, browsing order and jump position; After arranging the time-series records in chronological order, they are aligned with the corresponding nodes in the semantic relationship graph. The alignment process is completed by matching three pieces of information: page access path, node number, and timestamp. After matching is completed, the node access rhythm change, node dwell difference, and node jump offset are calculated. These three types of quantities together constitute the behavior change vector. By comparing the behavior change vector with the node vector in hyperbolic space, and by comparing the degree of deviation between the behavior change direction and the node vector direction, a behavior perturbation offset is generated.

[0015] Optionally, S7 specifically includes: Arrange the behavioral perturbation offset sequence corresponding to each semantic node in chronological order, and extract intensity features, direction features, and fluctuation features from the offset sequence; Intensity features are generated by statistically analyzing the magnitude of the offset within a given time window, with the magnitude reflecting the order of magnitude of the node's behavioral offset; directional features are generated by comparing the angle between the offset direction and the semantic connection direction; and fluctuation features are generated by calculating the difference in the change of the offset between adjacent time indices. For each node, a perturbation evaluation vector is constructed, mapping the intensity, direction, and fluctuation characteristics to a unified value range, and the perturbation evaluation vector is recorded in a node-dimensional data structure. In the graph structure, for each structural connection, intensity and direction components are extracted from the perturbation evaluation vectors of the nodes at both ends of the connection. Perturbation features of the structural connection dimension are generated by weighting, and the perturbation features are merged with the fluctuation features of the nodes at both ends in sequence to form the structural perturbation factor. Record the values ​​of the structural perturbation factor in the graph structure; when the behavioral perturbation offset of a node within a given time window exceeds a preset offset threshold, the angle between the offset direction and the semantic connection direction exceeds a preset direction threshold, or the difference in the offset between adjacent time indices exceeds a preset fluctuation threshold, the value of the corresponding structural perturbation factor is pushed up to the upper end of the preset interval. When the behavioral disturbance offset exceeds the preset offset threshold, the angle between the offset direction and the semantic connection direction, and the difference in the offset between adjacent time indices are all lower than the corresponding threshold, the value of the structural disturbance factor will be pushed to the lower end of the preset interval. The structural perturbation factor is recorded in the graph structure along with the connection dimension and is input into the hyperbolic attention structure along with the basic attention weights. It participates in the weighting process within the hyperbolic attention structure to generate perturbation fusion attention weights.

[0016] The beneficial effects of the present invention are: (1) By using hyperbolic space mapping and node hierarchical encoding, the hierarchical relationship, difficulty sequence and semantic direction in the course semantic graph can be finely distinguished in the embedding space, thereby improving the expression accuracy of structural information.

[0017] (2) By using the path consistency scoring factor to quantify the changes in chapter order, difficulty gradient, and hierarchy, the path analysis is expanded from linear judgment to structurally sensitive analysis, improving the evaluation process's ability to identify the internal logic of the course. (3) By using the temporal mapping of learning behavior sequences, behavior change vectors, and behavior perturbation offsets, a correspondence mechanism between learning behavior and semantic structure is formed, enabling the evaluation results to reflect the structural impact of learning behavior offsets. (4) By using the structural perturbation factor to drive the secondary weighting of the hyperbolic attention structure, the attention distribution can be dynamically adjusted with behavior offsets, improving the flexibility of node relationship modeling.

[0018] (5) By introducing neighborhood embedding distribution restriction and tangent space update mechanism through entropy regularization stochastic gradient update model, the hyperbolic node vector and attention parameters are kept stable and converged during the training phase, reducing the risk of gradient oscillation.

[0019] (6) By recording the attention offset and embedding change, a structural robustness index is constructed so that the output can reflect the structural change trend and maintain reliability in the structural disturbance scenario. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0021] Figure 1 This is a flowchart illustrating the online course quality evaluation system and method based on knowledge graphs proposed in this invention. Figure 2 This is a schematic diagram of hyperbolic space embedding and path consistency in an online course quality evaluation system and method based on knowledge graphs proposed in this invention. Figure 3 This diagram illustrates the behavioral perturbation factor and perturbation fusion attention of an online course quality evaluation system and method based on knowledge graphs proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figure 1-3 A knowledge graph-based online course quality evaluation system and method, comprising the following steps: S1. Establish a set of semantic nodes for the course and construct structural connections between nodes to form a semantic relationship graph; S2. Map the semantic relationship graph to hyperbolic space to generate hyperbolic node vectors, and calculate the hyperbolic distance between nodes based on the hyperbolic node vectors to form a hyperbolic distance set; S3. Extract structural connection attribute values ​​from the semantic relationship graph to form a structural connection attribute set; S4. Construct a path consistency scoring factor, which is generated by node order, difficulty gradient, and hierarchical sequence; S5. Input the hyperbolic distance set, structural connection attribute set, and path consistency scoring factor into the hyperbolic attention structure to generate basic attention weights; S6. Introduce learning behavior sequences and record the temporal sequence in the learning behavior sequence. S7. The behavior perturbation offset is used to construct a structural perturbation factor, and the structural perturbation factor and the basic attention weight are input into the hyperbolic attention structure to generate perturbation fusion attention weights. S8. An entropy-regularized stochastic gradient update model is established. The entropy regularization term comes from the hyperbolic space neighborhood embedding distribution. The stochastic gradient process includes noise injection, neighborhood extraction and hyperbolic space shrinkage mapping. The hyperbolic node vector and perturbation fusion attention parameters are adjusted in the update stage. S9. In the update stage, the attention offset and embedding change are recorded. The attention offset and embedding change are combined to form a structural robustness index and output the online course quality evaluation results.

[0024] In this embodiment, S2 specifically includes: Create an initial vector for each node in the semantic relationship graph. The initial vector values ​​satisfy the norm being less than one. The radial elements of the vector reflect the node level, and the angular elements reflect the semantic direction. Subsequently, positive samples are formed by extracting connected node pairs from the semantic relationship graph, and negative samples are formed by extracting node pairs from unconnected nodes. Each sample pair enters the embedding learning process, which constrains the positional relationship of nodes in hyperbolic space according to the structural relationship of the sample pair. Construct an embedded objective function, which consists of three parts: distance constraints between connected nodes; separation constraints between unconnected nodes; and hierarchical sequence constraints. Using the objective function as the optimization objective, perform iterative updates. In each iteration, calculate the gradient of the node vector in Euclidean space, project the gradient onto the corresponding tangent space in hyperbolic space, complete the update in the tangent space, and then shrink the update result back into the hyperbolic space through exponential mapping. Once the objective function converges, the node vector becomes a hyperbolic node vector. The distance between any pair of nodes is then calculated using the hyperbolic distance function, forming a hyperbolic distance set.

[0025] In this embodiment, the structural relationship of the sample pairs in step S2 specifically refers to the connection state of the two nodes in the semantic relationship graph and their corresponding structural connection attributes. When the sample pair is a positive sample pair, the structural relationship includes whether there is a direct structural connection between the nodes and the node type combination, hierarchical span, difficulty span, and order index label corresponding to this structural connection; when the sample pair is a negative sample pair, the structural relationship is that there is no direct structural connection between the nodes. The embedding learning process constrains the relative positional relationship of node vectors in hyperbolic space through the above structural relationship.

[0026] In this embodiment, the tangent space corresponding to the hyperbolic space in step S2 specifically includes: An update direction is established in the tangent space according to the gradient direction, and a scaling operation is performed on the update direction according to the set step size parameter to generate a linear update increment. The linear update increment and the coordinates of the node vector in the tangent space are synthesized by vector addition to obtain the update result in the tangent space. After obtaining the updated tangent space vector, the linear update result is projected back into the hyperbolic space through exponential mapping.

[0027] In this embodiment, the calculation of the distance between any pair of nodes using the hyperbolic distance function in step S2 specifically includes: Generate a vector for each node and keep the length of the vector within a specified range; extract the node vectors for any two nodes, calculate the difference between the two vectors and obtain the square of the difference; calculate the length of the two vectors and take the length difference as the denominator; combine the square of the difference with the denominator to generate the distance ratio, then input the distance ratio into the inverse hyperbolic cosine function and output the corresponding distance, and combine the distances of all node pairs into a hyperbolic distance set.

[0028] In this embodiment, S3 specifically includes: For each structural connection in the semantic relationship graph, read the node type, level number, and difficulty number of the starting node and the ending node, and generate a connection type label by combining the node types; Calculate the difference between the level number and the difficulty number between the starting node and the ending node. Record the difference between the level numbers as the level span and the difference between the difficulty numbers as the difficulty span. Assign a sequence index tag to each structural connection, which reflects the position of the connection in the course chapter sequence; combine the connection type tag, level span, difficulty span and sequence index tag into a structural connection attribute vector, and collect all structural connection attribute vectors to form a structural connection attribute set.

[0029] In this embodiment, the path consistency scoring factor in step S4 specifically includes: Select a target path containing several nodes in the semantic relationship graph, and generate a node sequence according to the order in which the nodes appear in the path; For each node in the node sequence, read the chapter number, level number, and difficulty number, and arrange the three numbers in order to form the level sequence and difficulty sequence; Assign index number i to any adjacent node pair in the hierarchy sequence and difficulty sequence, and calculate the difference between the hierarchy number and the difficulty number of the node pair. Record the absolute value of the difference between the hierarchy number and the absolute value of the difference between the difficulty number as the hierarchy change value. Each level change value and difficulty change value are classified according to a threshold T1. When the absolute value of the change value is less than the threshold T1, it is recorded as a stable node pair. When the absolute value of the change value is greater than or equal to the threshold T1, it is recorded as a mutation node pair. Using the index number as the scanning order, comparisons are performed from index i to index i+1. When the node pairs corresponding to two indices are consistent in classification, they are divided into the same node segment; when the classifications are inconsistent, they are divided into different node segments. The number of node pairs, the starting index position and the ending index position in the node sequence of each node segment are recorded, and the statistical information of all node segments is used to form path structure change characteristics. Based on the characteristics of path structure changes, a path consistency scoring factor is constructed, which includes three parts: sequence score, difficulty score, and hierarchy score. The sequence score determines whether there are chapter jump events in the path based on the direction of change of chapter numbers in the node sequence. The difficulty score determines the smoothness of the path difficulty change based on the mean, range, and number of abrupt node segments of the difficulty change value. The hierarchy score determines the coordination of the path in the hierarchy direction based on the mean, range, and distribution of the hierarchy change value. The three scores are weighted and aggregated, and the aggregated values ​​are mapped to a preset range to generate a path consistency score factor.

[0030] In this embodiment, S5 specifically includes: Receive the hyperbolic distance set, the structural connectivity attribute set, and the path consistency scoring factor, arrange the three types of inputs in node pair order, and feed them into the hyperbolic attention structure; In the hyperbolic attention structure, the attention distribution value between node pairs is calculated. The attention distribution value is affected by the hyperbolic distance change, the structural connectivity property change and the path consistency score. During the calculation process, spatial sensitivity terms are generated based on the hyperbolic positional relationship of node pairs, structural sensitivity terms are generated based on the structural connection attributes of node pairs, and path sensitivity terms are generated based on the path consistency score. The three types of sensitivity terms are then weighted and fused within the attention structure to obtain the basic attention weights.

[0031] In this embodiment, the spatial sensitivity item in step S5 specifically includes: Read the hyperbolic distance of the target node pair from the hyperbolic distance set; the hyperbolic distance describes the relative position of the node pair in hyperbolic space; input the distance value into the interval mapping unit, perform linear or piecewise linear transformation on the distance value in the interval mapping unit according to the preset threshold and compression rules, and output the mapped spatial magnitude; record the spatial magnitude as a spatial sensitivity term to characterize the degree of difference between the node pairs in spatial dimensions.

[0032] In this embodiment, the structurally sensitive items in step S5 specifically include: Read the connection type marker, level span, difficulty span, and order index of node pairs from the structural connection attribute set. Convert the connection type marker into a structural type code, use the level span and difficulty span as structural distance quantities, and the order index as positional quantities. Input these three quantities into the structural combination unit, and perform weighted synthesis according to preset weights in the structural combination unit to obtain the structural difference value.

[0033] In this embodiment, the path-sensitive items in step S5 specifically include: The system reads the path consistency score corresponding to the target node pair from the path consistency scoring factor. This score is then input into the path mapping unit, where interval compression or expansion is performed to ensure the path consistency score aligns with the internal scale used by the attention structure. The mapped score serves as a path sensitivity term, reflecting the importance of the node pair in the path pattern.

[0034] In this embodiment, S6 specifically includes: Receive time-series records from the learning behavior sequence, which include dwell time, access frequency, browsing order and jump position; After arranging the time-series records in chronological order, they are aligned with the corresponding nodes in the semantic relationship graph. The alignment process is completed by matching three pieces of information: page access path, node number, and timestamp. After matching is completed, the node access rhythm change, node dwell difference, and node jump offset are calculated. These three types of quantities together constitute the behavior change vector. By comparing the behavior change vector with the node vector in hyperbolic space, and by comparing the degree of deviation between the behavior change direction and the node vector direction, a behavior perturbation offset is generated.

[0035] In this embodiment, S7 specifically includes: Arrange the behavioral perturbation offset sequence corresponding to each semantic node in chronological order, and extract intensity features, direction features, and fluctuation features from the offset sequence; Intensity features are generated by statistically analyzing the magnitude of the offset within a given time window, with the magnitude reflecting the order of magnitude of the node's behavioral offset; directional features are generated by comparing the angle between the offset direction and the semantic connection direction; and fluctuation features are generated by calculating the difference in the change of the offset between adjacent time indices. For each node, a perturbation evaluation vector is constructed, mapping the intensity, direction, and fluctuation characteristics to a unified value range, and the perturbation evaluation vector is recorded in a node-dimensional data structure. In the graph structure, for each structural connection, intensity and direction components are extracted from the perturbation evaluation vectors of the nodes at both ends of the connection. Perturbation features of the structural connection dimension are generated by weighting, and the perturbation features are merged with the fluctuation features of the nodes at both ends in sequence to form the structural perturbation factor. Record the values ​​of the structural perturbation factor in the graph structure; when the behavioral perturbation offset of a node within a given time window exceeds a preset offset threshold, the angle between the offset direction and the semantic connection direction exceeds a preset direction threshold, or the difference in the offset between adjacent time indices exceeds a preset fluctuation threshold, the value of the corresponding structural perturbation factor is pushed up to the upper end of the preset interval. When the behavioral disturbance offset exceeds the preset offset threshold, the angle between the offset direction and the semantic connection direction, and the difference in the offset between adjacent time indices are all lower than the corresponding threshold, the value of the structural disturbance factor will be pushed to the lower end of the preset interval. The structural perturbation factor is recorded in the graph structure along with the connection dimension and is input into the hyperbolic attention structure along with the basic attention weights. It participates in the weighting process within the hyperbolic attention structure to generate perturbation fusion attention weights.

[0036] In this embodiment, the upper end advancement in step S7 specifically includes: In the graph structure, a structural perturbation factor is maintained for each structural connection, and a value range is set for the structural perturbation factor. When detecting behavioral perturbation offsets, a perturbation enhancement process is initiated when the behavioral perturbation offset of a node within a given time window exceeds a preset offset threshold, the angle between the offset direction and the semantic connection direction exceeds a preset direction threshold, or the difference in the offset between adjacent time indices exceeds a preset fluctuation threshold. The perturbation enhancement process includes: reading existing values ​​from the current structural perturbation factor, inputting the values ​​into the increment generation unit, and calculating the perturbation increment in the increment generation unit according to the magnitude exceeding the threshold; performing a linear superposition operation on the perturbation increment and the existing values ​​to obtain the updated perturbation value; comparing the updated perturbation value with the upper end of the preset range, and truncating it to the upper end value when the updated perturbation value exceeds the upper end value, and directly recording it as a new perturbation factor value when it does not exceed the upper end value.

[0037] In this embodiment, S8 specifically includes: Based on the semantic relationship graph, a set of corresponding neighboring nodes is determined for each semantic node, and the hyperbolic node vector distribution of the neighboring nodes is read. A neighborhood embedding distribution is formed according to the relative distance relationship between the neighboring nodes, and the dispersion of the distribution is statistically analyzed. The statistical results are used as an entropy regularization term and introduced into the update process.

[0038] In each update round, after the basic gradient calculation is completed, a random perturbation vector is generated and superimposed with the gradient direction to form a gradient input with random perturbation. Subsequently, a subset of nodes and their structural connections are selected according to a preset neighborhood extraction strategy to form a local subgraph, and the update calculation for this round is performed only within this subgraph.

[0039] Since hyperbolic space does not allow for direct linear updates, the gradient with random perturbation is projected onto the tangent space corresponding to the node vector's location. Within the tangent space, the step size update of the node vector is calculated linearly. After completing the tangent space update, the update result is reverted back to the hyperbolic space using an exponential mapping.

[0040] In this embodiment, S9 specifically includes: During the update phase, the changes in perturbation fusion attention weights and hyperbolic node vectors in two adjacent update rounds are recorded, forming attention offsets and embedding changes. After scaling these two types of quantities, the attention offsets of the corresponding connections and the embedding changes of the nodes at both ends of the connection are combined in the structural connectivity dimension to obtain the structural change value at the connectivity level. The structural change values ​​of all structural connections are summarized to form an overall structural robustness index.

[0041] Example 1: To verify the feasibility and effectiveness of this invention in practical applications, it was applied to an online course quality analysis scenario. In this scenario, the course content consists of multiple chapters, knowledge points, and corresponding learning objectives, and the course structure exhibits clear hierarchical and path-dependent relationships. Traditional evaluation methods based on human experience or simple statistical rules are insufficient to accurately reflect the rationality of the course's internal structure, nor can they effectively characterize the relationship between learning behavior and the course structure, resulting in insufficient stability and weak interpretability of course quality evaluation results. In this embodiment, the system first performs semantic modeling on the course content, abstracting chapters, knowledge points, and learning objectives into semantic nodes, and establishing structural connections between nodes to form a semantic relationship graph. This graph structure clearly describes the organization of the course content, hierarchical relationships, and dependencies between knowledge points. Subsequently, the system maps the semantic relationship graph to hyperbolic space, generates a hyperbolic node vector for each semantic node, and calculates the hyperbolic distance between nodes to characterize the relative relationships between nodes in the hierarchical and semantic directions.

[0042] In actual operation, the system further extracts structural connection attributes between nodes, including information such as chapter order relationships, hierarchical span, and difficulty span. Simultaneously, the system collects learning behavior sequences from the course learning process, mapping browsing order, dwell time changes, and jump behaviors back to a semantic relationship graph to generate behavioral perturbation offsets. In this way, learning behaviors no longer exist merely as independent statistical data, but form a clear correspondence with the course structure. During the course quality evaluation phase, the system performs consistency analysis on the course path, generating a path consistency scoring factor by comprehensively considering changes in node order, difficulty trends, and hierarchical characteristics. This scoring factor reflects the coherence of the course content in its structural arrangement. The system further inputs the path consistency score, structural connection attributes, and hyperbolic distance information into a hyperbolic attention structure to generate basic attention weights. After introducing behavioral perturbation offsets, the system constructs a structural perturbation factor to dynamically adjust the attention weights, ensuring that the evaluation results reflect the actual feedback of learning behaviors to the course structure.

[0043] Table 1: Comparison of Course Quality Result Evaluation

[0044] As can be seen from the data in the table, different evaluation methods produce significant differences in course quality analysis results. Traditional statistical methods mainly rely on superficial information such as the number of chapters and access frequency for aggregation. The evaluation results fluctuate in terms of structural consistency and path coherence, and the numerical distribution is relatively scattered, making it difficult to reflect the true organizational state of the course's internal structure. When the course content has a span of levels or changes in difficulty, this type of method is prone to evaluation instability.

[0045] The rule-based scoring method introduces some structural rule constraints on top of traditional statistics, which improves the coherence of the path and the smoothness of changes in difficulty compared to traditional methods. However, the rules themselves rely on manual setting, limiting their adaptability. When the course structure is adjusted or learning behaviors change, the evaluation results still fluctuate significantly, and the improvement in structural consistency and behavioral matching is limited.

[0046] In contrast, the method of this invention exhibits more stable evaluation results across all indicators. The structural consistency score remains at a high level, indicating that the approach based on semantic relationship graphs and hyperbolic space modeling can more accurately depict the hierarchical relationships and semantic connections between course content, avoiding excessive influence of local structural changes on the overall evaluation. In the path coherence indicator, this invention can effectively identify the relationship between changes in chapter order, hierarchical jumps, and difficulty gradients, making the evaluation results closer to the actual course arrangement logic.

[0047] Regarding the smoothness of difficulty changes, this invention analyzes the trend of difficulty changes along the path, enabling the evaluation results to better distinguish sudden changes in difficulty. This reflects the impact of difficulty changes without distorting the overall score due to changes in individual nodes. The behavioral matching index reflects the correlation between learning behavior and course structure. By mapping learning behavior to a semantic structure, this invention can identify the impact of behavioral deviations on course quality evaluation, providing a behavioral-level interpretive basis for the evaluation results.

[0048] The comprehensive evaluation stability index further verifies the advantages of the method of this invention. Under the simultaneous presence of structural changes and behavioral perturbations, the evaluation results output by the method of this invention exhibit a small fluctuation range and a clear convergence trend, indicating that the constructed attention fusion mechanism and structural robustness analysis can maintain the reliability of the evaluation results under complex conditions. Data comparison in the table confirms that this invention has good practical application effects in terms of structural expressiveness, behavioral correlation analysis, and overall evaluation stability.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based online course quality evaluation system and method, characterized in that, Includes the following steps: S1. Establish a set of semantic nodes for the course and construct structural connections between nodes to form a semantic relationship graph; S2. Map the semantic relationship graph to hyperbolic space to generate hyperbolic node vectors, and calculate the hyperbolic distance between nodes based on the hyperbolic node vectors to form a hyperbolic distance set; S3. Extract structural connection attribute values ​​from the semantic relationship graph to form a structural connection attribute set; S4. Construct a path consistency scoring factor, which is generated from node order, difficulty gradient, and hierarchical sequence; S5. Input the hyperbolic distance set, structural connectivity attribute set, and path consistency scoring factor into the hyperbolic attention structure to generate basic attention weights; S6. Introduce learning behavior sequences, map the temporal records in the learning behavior sequences to the semantic relationship graph and generate behavior perturbation offsets; S7. The behavioral perturbation offset is used to construct the structural perturbation factor, and the structural perturbation factor and the basic attention weight are input into the hyperbolic attention structure to generate the perturbation fusion attention weight. S8. Establish an entropy-regularized stochastic gradient update model. The entropy regularization term comes from the hyperbolic space neighborhood embedding distribution. The stochastic gradient process includes noise injection, neighborhood extraction, and hyperbolic space shrinkage mapping. In the update stage, the hyperbolic node vector and the perturbation fusion attention parameter are adjusted. S9. During the update phase, record the attention shift and embedding change, combine the attention shift and embedding change to form a structural robustness index, and output the online course quality evaluation results.

2. The knowledge graph-based online course quality evaluation system and method according to claim 1, characterized in that, S2 specifically includes: Create an initial vector for each node in the semantic relationship graph. The initial vector values ​​satisfy the norm being less than one. The radial elements of the vector reflect the node level, and the angular elements reflect the semantic direction. Positive samples are formed by extracting connected node pairs from the semantic relationship graph, and negative samples are formed by extracting node pairs from unconnected nodes. Each sample pair enters the embedding learning process, which constrains the positional relationship of nodes in hyperbolic space based on the structural relationship of the sample pair. Construct an embedded objective function, which consists of three parts: distance constraints between connected nodes; separation constraints between unconnected nodes; and hierarchical sequence constraints. Using the objective function as the optimization objective, perform iterative updates. In each iteration, calculate the gradient of the node vector in Euclidean space, project the gradient onto the corresponding tangent space in hyperbolic space, complete the update in the tangent space, and then shrink the update result back into the hyperbolic space through exponential mapping. Once the objective function converges, the node vector becomes a hyperbolic node vector. The distance between any pair of nodes is then calculated using the hyperbolic distance function, forming a hyperbolic distance set.

3. The knowledge graph-based online course quality evaluation system and method according to claim 1, characterized in that, The hyperbolic distance function in step S2 calculates the distance between any pair of nodes, specifically including: Generate a vector for each node and keep the length of the vector within a specified range; extract the node vectors for any two nodes, calculate the difference between the two vectors and obtain the square of the difference; calculate the length of the two vectors and take the length difference as the denominator; combine the square of the difference with the denominator to generate the distance ratio, then input the distance ratio into the inverse hyperbolic cosine function and output the corresponding distance, and combine the distances of all node pairs into a hyperbolic distance set.

4. The knowledge graph-based online course quality evaluation system and method according to claim 1, characterized in that, S3 specifically includes: For each structural connection in the semantic relationship graph, read the node type, level number, and difficulty number of the starting node and the ending node, and generate a connection type label by combining the node types; Calculate the difference between the level number and the difficulty number between the starting node and the ending node. Record the difference between the level numbers as the level span and the difference between the difficulty numbers as the difficulty span. Assign a sequence index tag to each structural connection, which reflects the position of the connection in the course chapter sequence; combine the connection type tag, level span, difficulty span and sequence index tag into a structural connection attribute vector, and collect all structural connection attribute vectors to form a structural connection attribute set.

5. The knowledge graph-based online course quality evaluation system and method according to claim 1, characterized in that, The path consistency scoring factors in step S4 specifically include: Select a target path containing several nodes in the semantic relationship graph, and generate a node sequence according to the order in which the nodes appear in the path; For each node in the node sequence, read the chapter number, level number, and difficulty number, and arrange the three numbers in order to form the level sequence and difficulty sequence; Assign index number i to any adjacent node pair in the hierarchy sequence and difficulty sequence, and calculate the difference between the hierarchy number and the difficulty number of the node pair. Record the absolute value of the difference between the hierarchy number and the absolute value of the difference between the difficulty number as the hierarchy change value. Each level change value and difficulty change value are classified according to a threshold T1. When the absolute value of the change value is less than the threshold T1, it is recorded as a stable node pair. When the absolute value of the change value is greater than or equal to the threshold T1, it is recorded as a mutation node pair. Using the index number as the scanning order, comparisons are performed from index i to index i+1. When the node pairs corresponding to two indices are consistent in classification, they are divided into the same node segment; when the classifications are inconsistent, they are divided into different node segments. The number of node pairs, the starting index position and the ending index position in the node sequence of each node segment are recorded, and the statistical information of all node segments is used to form path structure change characteristics. Based on the characteristics of path structure changes, a path consistency scoring factor is constructed, which includes three parts: sequence score, difficulty score, and hierarchy score. The sequence score determines whether there are chapter jump events in the path based on the direction of change of chapter numbers in the node sequence. The difficulty score determines the smoothness of the path difficulty change based on the mean, range, and number of abrupt node segments of the difficulty change value. The hierarchy score determines the coordination of the path in the hierarchy direction based on the mean, range, and distribution of the hierarchy change value. The three scores are weighted and aggregated, and the aggregated values ​​are mapped to a preset range to generate a path consistency score factor.

6. The knowledge graph-based online course quality evaluation system and method according to claim 1, characterized in that, S5 specifically includes: Receive the hyperbolic distance set, the structural connectivity attribute set, and the path consistency scoring factor, arrange the three types of inputs in node pair order, and feed them into the hyperbolic attention structure; In the hyperbolic attention structure, the attention distribution value between node pairs is calculated. The attention distribution value is affected by the hyperbolic distance change, the structural connectivity property change and the path consistency score. During the calculation process, spatial sensitivity terms are generated based on the hyperbolic positional relationship of node pairs, structural sensitivity terms are generated based on the structural connection attributes of node pairs, and path sensitivity terms are generated based on the path consistency score. The three types of sensitivity terms are then weighted and fused within the attention structure to obtain the basic attention weights.

7. The knowledge graph-based online course quality evaluation system and method according to claim 1, characterized in that, S6 specifically includes: Receive time-series records from the learning behavior sequence, which include dwell time, access frequency, browsing order and jump position; After arranging the time-series records in chronological order, they are aligned with the corresponding nodes in the semantic relationship graph. The alignment process is completed by matching three pieces of information: page access path, node number, and timestamp. After matching is completed, the node access rhythm change, node dwell difference, and node jump offset are calculated. These three types of quantities together constitute the behavior change vector. By comparing the behavior change vector with the node vector in hyperbolic space, and by comparing the degree of deviation between the behavior change direction and the node vector direction, a behavior perturbation offset is generated.

8. The knowledge graph-based online course quality evaluation system and method according to claim 1, characterized in that, Specifically, S7 includes: Arrange the behavioral perturbation offset sequence corresponding to each semantic node in chronological order, and extract intensity features, direction features, and fluctuation features from the offset sequence; Intensity features are generated by statistically analyzing the magnitude of the offset within a given time window; directional features are generated by comparing the angle between the offset direction and the semantic connection direction; and fluctuation features are generated by calculating the difference in the change of the offset between adjacent time indices. For each node, a perturbation evaluation vector is constructed, mapping the intensity, direction, and fluctuation characteristics to a unified value range, and the perturbation evaluation vector is recorded in a node-dimensional data structure. In the graph structure, for each structural connection, intensity and direction components are extracted from the perturbation evaluation vectors of the nodes at both ends of the connection. Perturbation features of the structural connection dimension are generated by weighting, and the perturbation features are merged with the fluctuation features of the nodes at both ends in sequence to form the structural perturbation factor. Record the values ​​of the structural perturbation factor in the graph structure; when the behavioral perturbation offset of a node within a given time window exceeds a preset offset threshold, the angle between the offset direction and the semantic connection direction exceeds a preset direction threshold, or the difference in the offset between adjacent time indices exceeds a preset fluctuation threshold, the value of the corresponding structural perturbation factor is pushed up to the upper end of the preset interval. When the behavioral disturbance offset exceeds the preset offset threshold, the angle between the offset direction and the semantic connection direction, and the difference in the offset between adjacent time indices are all lower than the corresponding threshold, the value of the structural disturbance factor will be pushed to the lower end of the preset interval. The structural perturbation factor is recorded in the graph structure along with the connection dimension and is input into the hyperbolic attention structure along with the basic attention weights. It participates in the weighting process within the hyperbolic attention structure to generate perturbation fusion attention weights.

Images