A digital textbook optimization system and method based on user interaction behavior

Abstract

The application discloses a digital textbook optimization system and method based on user interaction behavior, and relates to the technical field of path optimization. A textbook knowledge graph is constructed through acquired digital textbooks; nodes in the textbook knowledge graph are aggregated to obtain a graph subset, the topological potential of the graph subset is calculated, and sorting is performed according to the topological potential; nodes in the graph subset with high topological potential are selected to form an initial set, an initial path is generated through an LHK algorithm, path optimization is performed using a self-adaptive large neighborhood search algorithm according to the initial path, and an optimal path is obtained; and the digital textbook is optimized through the optimal path. The method uses the self-adaptive large neighborhood search algorithm in the knowledge graph, introduces specific user interaction behavior data to adjust the learning path in real time, calculates the topological potential of each knowledge point in the knowledge graph, accurately evaluates the priority and learning difficulty of the knowledge point in the learning path, and improves the accuracy and efficiency of textbook optimization.

Description

Technical Field

[0001] This invention relates to the technical field of path optimization, and in particular to a digital teaching material optimization system and method based on user interaction behavior. Background Technology

[0002] In recent years, with the development of educational informatization and the increasing popularity of digital textbooks, existing methods often overlook the logical relationships between knowledge points, failing to form coherent learning paths. Furthermore, path generation relies on fixed rules or simple traversal, lacking dynamic optimization capabilities based on user behavior feedback, making it difficult to balance multiple objectives such as knowledge coverage, learning fluency, and preventing students from falling behind. Simultaneously, the static structure of textbooks cannot be intelligently reorganized based on actual learning data, limiting the personalization and adaptability of the learning experience. Therefore, there is an urgent need for a system and method that can deeply integrate user behavior and knowledge structure, achieving dynamic path optimization and intelligent textbook reorganization.

[0003] Currently, Chinese invention patent application number CN202411784212.3 discloses a data analysis-based textbook content visualization system and method. By dynamically updating the knowledge graph and optimizing the learning path, combined with reinforcement learning and the knowledge graph, the optimal learning path best suited to students' current learning situation can be identified, helping students strengthen their weak knowledge points. Furthermore, visually displaying the optimal learning path enhances students' learning interest and intuitive understanding. Teachers can also use visualized data to accurately grasp students' learning progress and weaknesses, further optimizing teaching content and strategies, thereby improving overall teaching quality. However, this technology lacks in-depth dynamic analysis and real-time feedback of user behavior, and cannot deeply capture students' specific interactive behaviors during the learning process. Summary of the Invention

[0004] The technical problem addressed by this invention is that existing technologies primarily rely on reward functions in reinforcement learning to optimize learning paths, adjusting them based on learning progress, topology changes, and error rates. This over-reliance on static data lacks dynamic analysis and real-time feedback on fine-grained user behavior. While existing technologies can adjust learning paths to some extent through reinforcement learning, they fail to deeply capture students' specific interactive behaviors during the learning process, nor do they have the ability to adjust learning content in real time based on these behaviors. This results in path optimization being primarily a global adjustment, without considering the specific needs and progress of individual learners.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for optimizing digital teaching materials based on user interaction behavior includes the following steps: Step S1: Construct a textbook knowledge graph using digital textbooks and user logs; Step S2: Aggregate the nodes in the knowledge graph of the teaching materials to obtain a subset of the graph, calculate the topological potential of the subset of the graph, and sort it according to the topological potential. Step S3: Select nodes from the graph subset with high topological potential to form an initial set, generate an initial path using the LHK algorithm, and optimize the path using the adaptive large neighborhood search algorithm based on the initial path to obtain the optimal path. Step S4: Reorder the knowledge points and reorganize the jump links according to the optimal path, associate learning resources with high-cost nodes and high-cost edges, and generate optimized digital teaching materials; The learning resources include supplementary learning resources and transitional learning resources.

[0006] Preferably, step S1 specifically includes: Step S11: Obtain digital teaching materials and user logs. The digital teaching materials include a set of knowledge points, a set of knowledge point content, and the prior relationships between knowledge points. The user log includes the user's dwell time, completion rate, number of times the user repeatedly viewed each knowledge point, question accuracy rate, jump frequency, and dwell time after jumping. Step S12: Extract knowledge points from the knowledge point set as nodes, and take the preorder relationship between knowledge points as edges. By vectorizing the digital textbook and user logs respectively, obtain the node feature vector and edge feature vector of each knowledge point. The node feature vector includes behavioral index vector and textbook index vector. The textbook indicator vector includes knowledge point titles, keywords, and concept descriptions; The behavioral indicator vector includes the dwell time, completion rate, number of times repeatedly viewed, and question accuracy rate for each knowledge point; The edge feature vector includes the dwell time after the jump and the jump frequency; Step S13: Bind each knowledge point to the corresponding node feature vector and store it as a unique node in the textbook knowledge graph. Bind each preorder relation to the corresponding edge feature vector and store it as a directed edge connecting two nodes. Determine the direction of the edge through the preorder relation. Based on all the nodes and edges mentioned above, a graph structure is constructed, stored in a graph database, and indexes are created for node features and edge features to obtain the textbook knowledge graph.

[0007] Preferably, step S2 specifically includes: Step S21: Based on the node feature vector, calculate the cosine similarity of the semantics between any two nodes to obtain the node similarity matrix, and use the hierarchical clustering algorithm to aggregate the nodes in the textbook knowledge graph into k graph subsets, where k is an integer greater than 2; Step S22: Calculate the topological potential energy of each graph subset and sort all graph subsets in descending order according to their corresponding topological potential energy values. The quality of each node is obtained by analyzing the question accuracy rate in the node's feature vector. The centroid of the graph subset is obtained by calculating the mean of the eigenvectors of all nodes within the subset. The graph structure distance from each node in a subset of the knowledge graph to the centroid of the subset is calculated based on the edge relationships in the knowledge graph of the textbook. The graph structure distance is the number of edges of the shortest path from the node to the centroid of the subset. The expression for calculating topological potential energy is: ; in, Let be the topological potential of the i-th graph subset. For the i-th subset of the graph, For a subset of the graph The nodes in For nodes quality For nodes To the atlas subset Graph structural distance of the centroid, Impact factor; Step S23: Calculate the behavioral cost of each node based on the node feature vector.

[0008] Preferably, step S23 specifically includes: Step S231: Obtain the behavior index vector and standardize the behavior index vector using Z-score standardization. Step S232: Based on the standardized behavior index vector, calculate the behavior cost of the node using the Euclidean distance formula. The calculation expression is: ; in, As a consequence of the behavior, The number of behavioral indicator vectors. Let k be the standardized behavioral index vector. Let be the weight coefficient of the k-th behavioral indicator vector, and ; Preferably, step S3 specifically includes: Step S31: From the sorted graph subsets, select the graph subsets with the highest topological potential in sequence until the nodes included in the selected graph subsets satisfy the preset semantic constraints, and obtain the initial node set. Step S32: The initial node set is taken as the required nodes. Based on the knowledge graph of the textbook, the LKH algorithm is used to generate the initial path that traverses all the required nodes, and the sub-path cost of the initial path is calculated according to the sub-path cost calculation rules. The specific rules for calculating subpath costs include: The dwell time and jump frequency after the jump are normalized, and the edge cost distance is calculated by subtracting the normalized dwell time and jump frequency. The expression for calculating the cost of a subpath for any given path is: ; in, Let $C$ be the cost of a subpath of any path. For nodes and The edge cost distance between them For nodes The estimated difficulty value, For nodes The estimated difficulty value, , These are preset weighting coefficients; Step S33: Starting from the initial path, use the adaptive large neighborhood search algorithm for iterative optimization to obtain the optimal path.

[0009] Preferably, step S33 specifically includes: Step S331: The behavior cost of the initial path is obtained by summing the behavior costs of all nodes in the initial path, and the sum of the behavior cost of the initial path and the cost of the sub-paths is calculated as the total cost. The proportion of behavioral cost in the total cost is used as the initial weight of the node replacement operator; The proportion of sub-path cost in the total cost is used as the initial weight of the local rearrangement operator; Step S332: Based on the current operator weights of the node replacement operator and the local rearrangement operator, select the node replacement operator and the local rearrangement operator using the roulette wheel selection method, execute the perturbation operation corresponding to the selected operator, and obtain the perturbed path; Step S333: Using the perturbed path, based on the edge cost distance, perform local optimization using the 2-opt algorithm to obtain the path whose total cost is less than the total cost of the perturbed path as the candidate path; Step S334: If the total cost of the candidate path is less than the total cost of the current path but greater than the total cost of the optimal path, then update the current path using the candidate path. If the total cost of the candidate path is less than the total cost of the current path and less than the total cost of the optimal path, then the optimal path is updated using the candidate path. Step S335: Based on the comparison results between the candidate path and the current path and the optimal path, score the selected operator according to the preset scoring criteria, and update the operator weight of the selected operator based on the score. Repeat steps S332 to S335 until the iteration termination condition is met; The iteration terminates when the maximum number of iterations is reached or the optimal path is not improved in Q consecutive iterations. The optimal path and the corresponding set of knowledge points are then output, where Q is an integer greater than 2.

[0010] Preferably, the perturbation operations corresponding to the selected operator include node replacement operator perturbation operations and local rearrangement operator perturbation operations, specifically including: The perturbation operations of the node replacement operator include: Replace nodes in the current path whose behavior cost is higher than a preset semantic threshold with other nodes in the same graph subset whose behavior cost is lower than that node to obtain the perturbed path. The perturbation operations of the local rearrangement operator include: Remove the subsequence consisting of consecutive nodes from the current path to obtain a partial path. For each node in the subsequence, calculate the insertion cost of inserting the node into both ends of the partial path and between every two adjacent nodes. The calculation expression is as follows: ; in, For part of the path, The total cost of a partial path. For the new partial path, The total cost of the new partial path, Let v be the insertion cost at position pos; For each node, the position with the minimum insertion cost is determined as the optimal insertion position, and the position with the second minimum insertion cost is determined as the suboptimal insertion position. The difference between the insertion cost of the suboptimal insertion position and the insertion cost of the optimal insertion position is the regret value of the node. Nodes in the subsequence are inserted into the optimal insertion positions of the nodes in the partial path in descending order of regret value. The partial path is updated after each node is inserted until all nodes have been inserted, resulting in the perturbed path.

[0011] Preferably, the preset scoring criteria specifically include: The scoring criteria include a first criterion, a second criterion, a third criterion, and a fourth criterion; The first criterion is that the total cost of the candidate path is less than the total cost of the historical best path, and the selected operator is assigned a first preset score. The second criterion is that the total cost of the candidate path is less than the total cost of the current path, but greater than the total cost of the historical best path, and a second preset score is assigned to the selected operator. The third criterion is that if the total cost of a candidate path is greater than the total cost of the historical best path, it is accepted as the new current path according to the preset path acceptance strategy, and the selected operator is assigned a third preset score. The fourth criterion is that if the total cost of the candidate path is greater than the total cost of the current path and it is not accepted by the preset path acceptance strategy, the selected operator is assigned a fourth preset score. Wherein, the first preset score > the second preset score > the third preset score > the fourth preset score; The expression for updating the operator weights based on the scores of the selected operators is as follows: ; in, For the updated operator weights, The operator weights before the update. The fraction of the selected operator, For the number of times the operator is used, Update the coefficients for the weights.

[0012] Preferably, step S4 specifically includes: Step S41: Based on the node order in the optimal path, sort the learning order of the corresponding knowledge points in the digital teaching materials, and reorganize the jump links between knowledge points according to the edge relationships in the knowledge graph. Step S42: For nodes in the optimal path whose behavioral cost is higher than the first knowledge threshold, associate auxiliary learning resources with the knowledge points corresponding to the nodes in the digital teaching materials. For edges in the optimal path whose sub-path cost exceeds the second knowledge threshold, transitional learning resources that connect the corresponding knowledge points are provided in the digital teaching materials. Step S43: Update the above optimization results to the digital textbook platform to form optimized digital textbooks.

[0013] A digital textbook optimization system based on user interaction behavior includes a construction module, a sorting module, an optimization module, and an application module. The building module is used to construct a textbook knowledge graph using digital textbooks and user logs. The sorting module is used to aggregate nodes in the knowledge graph of the textbook to obtain a subset of the graph, calculate the topological potential of the subset of the graph, and sort them according to the topological potential. The optimization module is used to select nodes from the graph subset with high topological potential to form an initial set, generate an initial path through the LHK algorithm, and optimize the path using the adaptive large neighborhood search algorithm based on the initial path to obtain the optimal path. The application module is used to rearrange the order of knowledge points and reorganize jump links according to the optimal path, associate learning resources with high-cost nodes and high-cost edges, and generate optimized digital teaching materials. The learning resources include supplementary learning resources and transitional learning resources.

[0014] The beneficial effects of this invention are as follows: This method optimizes learning paths by establishing a knowledge graph. Within the knowledge graph, an adaptive large neighborhood search algorithm is used, and user-specific interaction data is incorporated to adjust the learning path in real time. This interaction data provides insights into the student's current cognitive state and learning habits. Through multiple iterative searches, the order of nodes in the path is flexibly adjusted to minimize the cost of the learning path, ensuring that the order of learning content and knowledge points always aligns with the student's actual needs, thus avoiding ineffective and repetitive learning. Furthermore, the topological potential of each knowledge point is calculated within the knowledge graph to accurately assess its priority and learning difficulty within the learning path. Combined with the adaptive large neighborhood search algorithm for path optimization, a regret value ranking mechanism is used to further optimize path adjustments, ensuring that each path change results in minimal learning loss, thereby improving the accuracy and efficiency of path optimization. Attached Figure Description

[0015] Figure 1 The following is a basic flowchart of a digital teaching material optimization method based on user interaction behavior, provided as an embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example 1, referring to Figure 1 As an embodiment of the present invention, a digital teaching material optimization method based on user interaction behavior is provided, comprising the following steps: Step S1: Construct a textbook knowledge graph using digital textbooks and user logs; Step S2: Aggregate the nodes in the knowledge graph of the textbook to obtain a subset of the graph, calculate the topological potential of the subset of the graph and sort it according to the topological potential. Step S3: Select nodes from the graph subset with high topological potential to form an initial set, generate an initial path using the LHK algorithm, and optimize the path using the adaptive large neighborhood search algorithm based on the initial path to obtain the optimal path. Step S4: Reorder the knowledge points and reorganize the jump links according to the optimal path, associate learning resources with high-cost nodes and high-cost edges, and generate optimized digital teaching materials; Learning resources include supplementary learning resources and transitional learning resources.

[0018] In a specific embodiment, step S1 specifically includes: Step S11: Obtain digital teaching materials and user logs. The digital teaching materials include a set of knowledge points, a set of knowledge point content, and the prior relationships between knowledge points. User logs include the user's dwell time, completion rate, number of times the user repeatedly viewed each knowledge point, question accuracy rate, jump frequency, and dwell time after jumping. Step S12: Extract knowledge points from the knowledge point set as nodes, and take the preorder relationship between knowledge points as edges. By vectorizing the digital textbook and user logs respectively, obtain the node feature vector and edge feature vector of each knowledge point. The node feature vector includes the behavior index vector and the textbook index vector. The textbook indicator vector includes knowledge point titles, keywords, and concept descriptions; The behavioral indicator vector includes the dwell time, completion rate, number of times the question was viewed repeatedly, and the accuracy rate of each knowledge point. The edge feature vector includes the dwell time after the jump and the jump frequency; Specifically, the node feature vector is formed by concatenating the behavioral indicator vector and the textbook indicator vector.

[0019] The behavioral metric vector processes numerical behavioral metrics (duration of stay, completion rate, number of repeated views, and accuracy of answers) from user logs. It standardizes these metrics to a similar numerical range, eliminating the influence of different units. The four standardized metrics are then arranged in a fixed order.

[0020] The textbook indicator vector is a vectorization of the knowledge point content set (title, keywords, concept description) in digital textbooks. All text content for each knowledge point is merged and transformed into a fixed-dimensional dense vector using the BERT model. For example, after word segmentation, the average of the word vectors is taken as the text representation vector for that knowledge point, resulting in the textbook indicator vector.

[0021] The generated behavioral indicator vectors and textbook indicator vectors are concatenated in terms of dimensions to form node feature vectors.

[0022] Edge feature vectors are generated based on user navigation behavior data during the learning process. For each directed edge in the knowledge graph (i.e., the relationship between one knowledge point and another), the frequency of navigation and the duration of stay after navigation are extracted from the user's usage log for all corresponding navigation paths.

[0023] Calculate the average jump frequency and average dwell time after jump for all user behaviors on the edge. Standardize these two metrics and arrange them in a fixed order to form a 2-dimensional edge feature vector.

[0024] Standardization processes include handling missing values, smoothing outliers, and max-min normalization.

[0025] Step S13: Bind each knowledge point to the corresponding node feature vector and store it as a unique node in the textbook knowledge graph. Bind each preorder relation to the corresponding edge feature vector and store it as a directed edge connecting two nodes. Determine the direction of the edge through the preorder relation. Based on all the nodes and edges mentioned above, a graph structure is constructed, stored in a graph database, and indexes are created for node features and edge features to obtain the textbook knowledge graph.

[0026] Specifically, the prerequisite relationship means that in the knowledge system of digital textbooks, there is a logical sequential dependency between knowledge points, that is, some knowledge points are prerequisites for learning other knowledge points.

[0027] In the constructed knowledge graph, each knowledge point is abstracted as a node, and the preorder relationship between knowledge points is represented by a directed edge connecting two nodes. The direction of the directed edge points from the preorder knowledge point to the subsequent knowledge point, thus clarifying the order in the learning path.

[0028] Binding each knowledge point to the corresponding node feature vector specifically includes: Create a data record for each node, which includes a knowledge point ID and a field to store the node's feature vector, ensuring that all feature information of the node can be retrieved directly by the knowledge point ID.

[0029] Binding each preorder relation to its corresponding edge feature vector specifically includes: Each preorder relation corresponds to a directed edge in the knowledge graph. The edge feature vector consists of metrics related to user navigation behavior. The binding operation associates and stores the edge feature vector of each directed edge. In the edge table, each record contains the starting node ID, the target node ID, and a field storing the edge feature vector, thus tightly linking the topological relationships with behavioral data.

[0030] Creating indexes for node and edge features specifically includes: Node Index: A primary key index is created on the knowledge point ID in the node table to enable fast node location.

[0031] Edge Index: A composite index is created on the edge table for the starting node ID and the target node ID, which supports fast lookup of all preorder relationships that start or end with a certain node.

[0032] In a specific embodiment, step S2 specifically includes: Step S21: Based on the node feature vector, calculate the semantic cosine similarity between any two nodes to obtain the node similarity matrix, and use the hierarchical clustering algorithm to aggregate the nodes in the textbook knowledge graph into k graph subsets, where k is an integer greater than 2. Specifically, using hierarchical clustering algorithms to aggregate nodes in the textbook knowledge graph into k subsets includes: Each node is treated as an independent cluster, with the initial number of clusters equal to the total number of nodes. Based on the node similarity matrix, the average linking method is used to calculate the similarity between any two clusters. The average linking method is defined as the average of the similarities of all node pairs in two clusters. Iteration is performed: the pair of clusters with the highest similarity among all current cluster pairs is found, and these two clusters are merged into a new cluster. The similarity between the new cluster and other clusters is then updated.

[0033] The process continues until the number of clusters is reduced to a preset value of k. The resulting k clusters are the final subset of the knowledge point graph.

[0034] Step S22: Calculate the topological potential energy of each graph subset and sort all graph subsets in descending order according to their corresponding topological potential energy values. The quality of each node is obtained by analyzing the question accuracy rate in the node's feature vector. The centroid of the graph subset is obtained by calculating the mean of the eigenvectors of all nodes within the subset. Based on the edge relationships in the knowledge graph in the textbook, calculate the graph structure distance from each node in the graph subset to the centroid of the graph subset. The graph structure distance is the number of shortest path edges from the node to the centroid of the graph subset. The expression for calculating topological potential energy is: ; in, Let be the topological potential of the i-th graph subset. For the i-th subset of the graph, For a subset of the graph The nodes in For nodes quality For nodes To the atlas subset Graph structural distance of the centroid, Impact factor; Specifically, obtaining the quality of each node through its feature vectors includes: The quality of a node is the accuracy rate of the user's answers to the questions for that node. If the node has no answer records, it is assigned a default quality value of 0.5.

[0035] Calculating the graph structure distance from each node within a subset of a graph to the centroid of that subset based on edge relationships in the textbook knowledge graph specifically includes: If a subset of the graph contains only one node, then the graph structure distance from that node to the centroid is defined as 0. Otherwise, for each node in the graph subset, calculate the shortest path length from that node to all other nodes in the graph subset, and take the average of these lengths as the graph structure distance from that node to the centroid of the graph subset. The topological potential of each graph subset is calculated, and the subsets are sorted in descending order based on their topological potential values. The calculation of topological potential integrates the quality of nodes within a graph subset with the structural distance from each node to the centroid of the subset. A higher topological potential indicates that the nodes within a graph subset are not only more active in learning but also highly correlated and cohesive in their knowledge structure. If node behavior costs or sub-path costs are used directly for sorting, these costs primarily reflect the learning difficulty and fluency of users on the current path or local structure, and are local and posterior indicators. However, topological potential is calculated based on the global structure of the textbook knowledge graph, unrestricted by the current path, and can a priori identify important clusters within the knowledge system. This global structural perspective helps anchor the core framework of the knowledge system from the initial stage of path generation, avoiding loose path structures or deviations from the main framework due to abnormal behavior of individual users or local learning difficulties.

[0036] Step S23: Calculate the behavioral cost of each node based on the node feature vector.

[0037] In a specific embodiment, step S23 specifically includes: Step S231: Obtain the behavior index vector and standardize the behavior index vector using Z-score standardization. Step S232: Based on the standardized behavior index vector, calculate the behavior cost of the node using the Euclidean distance formula. The calculation expression is: ; in, As a consequence of the behavior, The number of behavioral indicator vectors. Let k be the standardized behavioral index vector. Let be the weight coefficient of the k-th behavioral indicator vector, and ; Specifically, in step S231, the standardization of the behavioral indicator vector using Z-score standardization also includes taking the inverse of the correctness and completion rates of the questions and making them negative.

[0038] In a specific embodiment, step S3 specifically includes: Step S31: From the sorted graph subsets, select the graph subsets with the highest topological potential in sequence until the nodes included in the selected graph subsets satisfy the preset semantic constraints, and obtain the initial node set. Step S32: Take the initial node set as the required nodes, use the LKH algorithm to generate the initial path that traverses all the required nodes based on the knowledge graph in the textbook, and calculate the sub-path cost of the initial path according to the sub-path cost calculation rules. The specific rules for calculating subpath costs include: The dwell time and jump frequency after the jump are normalized, and the edge cost distance is calculated by subtracting the normalized dwell time and jump frequency. The expression for calculating the cost of a subpath for any given path is: ; in, Let $C$ be the cost of a subpath of any path. For nodes and The edge cost distance between them For nodes The estimated difficulty value, For nodes The estimated difficulty value, , These are preset weighting coefficients; Specifically, the weighting coefficients are set to It is 0.4. The weight is 0.6. Overcoming the inherent difficulty of understanding a single knowledge point typically has a greater impact on learning effectiveness and experience than the transition cost incurred when jumping between two knowledge points. Therefore, assigning a higher weight to the estimated difficulty value of nodes can guide the path optimization algorithm to prioritize avoiding densely packed arrangements of multiple highly difficult knowledge points, thereby effectively reducing continuous cognitive load and preventing learners from interrupting their learning due to frustration. At the same time, assigning a weight with appropriate edge cost distance can take into account the smoothness of jumping between knowledge points, avoiding paths that become structurally loose or jump too frequently due to excessive avoidance of difficult points.

[0039] Step S33: Starting from the initial path, use the adaptive large neighborhood search algorithm for iterative optimization to obtain the optimal path.

[0040] Specifically, the preset semantic constraints are used to ensure that the selected knowledge nodes are semantically sufficiently representative in the early stages of path planning. A semantic rate threshold is set, which indicates that the key concepts covered by the selected node set should reach a certain proportion of the key concepts of all knowledge points. In step S31, nodes are selected sequentially from the graph subset sorted in descending order of topological potential until the keyword set corresponding to the selected node achieves or exceeds the semantic rate threshold in terms of coverage of the keywords of the entire knowledge system.

[0041] The longer the dwell time after a jump and the lower the jump frequency, the greater the edge cost distance.

[0042] The textbook indicator vector of each node serves as a semantic feature vector, and the set of semantic feature vectors of all nodes is considered as the entire semantic space. A representative vector of the semantic feature vectors of the selected node set is calculated, and then the average cosine similarity between the representative vector and all vectors in the entire semantic space is calculated, serving as a measure of coverage. When this average similarity reaches a semantic coverage threshold, the semantic coverage constraint is considered satisfied.

[0043] Generating the initial path for traversing all required nodes using the LKH algorithm specifically includes: The LKH algorithm is used to generate a relatively optimal initial path based on edge cost distance that traverses all required nodes. Its specific steps are as follows: The set of required nodes is mapped to a complete graph. The edge weight between any two nodes in the graph is set as the sum of the edge cost distance of their shortest path in the original knowledge graph (if there is no direct edge between two points, the cumulative edge cost on their shortest path is calculated through graph search), thus forming a weighted complete graph.

[0044] Using the weighted complete graph as input, the goal is to find a path that traverses all required nodes without requiring a return to the starting point, minimizing the total edge cost distance. The LKH algorithm directly searches for and outputs this optimal or near-optimal linear path through its internal mechanism; the output linear path is the initial path. This path uses edge cost distance as the optimization objective, ensuring that the initial solution has a low starting point in terms of transition costs between knowledge points, and providing a well-structured optimization foundation for subsequent adaptive large neighborhood search.

[0045] In a specific embodiment, step S33 specifically includes: Step S331: The behavior cost of the initial path is obtained by summing the behavior costs of all nodes in the initial path, and the sum of the behavior cost of the initial path and the cost of the sub-paths is calculated as the total cost. The proportion of behavioral cost in the total cost is used as the initial weight of the node replacement operator; The proportion of sub-path cost in the total cost is used as the initial weight of the local rearrangement operator; Specifically, the current operator weight refers to the updatable weight value used to guide the selection of operators in this round during the algorithm iteration process (steps S332 to S335).

[0046] In the first iteration, the current operator weights are the initial weights.

[0047] Starting from the second iteration, the current operator weight is calculated using a weight update formula based on the operator's score obtained according to the scoring criteria and its historical usage frequency in the previous iteration. Its core function is to adaptively adjust the probability of each operator being selected based on its historical performance. Operators with excellent performance will have their weights increased, increasing their chances of being selected in subsequent iterations; conversely, operators with poor performance will have their weights decreased.

[0048] Step S332: Based on the current operator weights of the node replacement operator and the local rearrangement operator, select the node replacement operator and the local rearrangement operator using the roulette wheel selection method, execute the perturbation operation corresponding to the selected operator, and obtain the perturbed path; Specifically, the node replacement operator and local rearrangement operator selected using the roulette wheel selection method include: Based on the current weights of the node replacement operator and the local rearrangement operator, calculate the probability that the node replacement operator and the local rearrangement operator will be selected next time. The calculation expression is as follows: ; in, Let be the probability that operator i is selected. Let i be the current weight of operator i. N The total number of operators, The sum of all weights is used for normalization, so that the sum of the selection probabilities of all operators is 1.

[0049] Step S333: Using the perturbed path, based on the edge cost distance, perform local optimization using the 2-opt algorithm to obtain the path whose total cost is less than the total cost of the perturbed path as the candidate path; Step S334: If the total cost of the candidate path is less than the total cost of the current path but greater than the total cost of the optimal path, then update the current path using the candidate path. If the total cost of the candidate path is less than the total cost of the current path and less than the total cost of the optimal path, then the optimal path is updated using the candidate path. Step S335: Based on the comparison results between the candidate path, the current path, and the optimal path, score the selected operator according to the preset scoring criteria, and update the operator weight of the selected operator based on the score. Repeat steps S332 to S335 until the iteration termination condition is met; The iteration terminates when the maximum number of iterations is reached or the optimal path is not improved in Q consecutive iterations. The optimal path and the corresponding set of knowledge points are then output, where Q is an integer greater than 2.

[0050] Specifically, based on edge cost distance, the 2-opt algorithm for local optimization includes: Each time a new path is evaluated, the 2-opt algorithm iteratively attempts to swap two non-adjacent edges in the path. After each swap, the total cost of the new path is calculated. If the total cost decreases, the swap is accepted and the iteration continues until the total cost cannot be reduced through swaps. The resulting locally optimal path is then used as the candidate path for this iteration. Since edge cost distance directly reflects the user's learning fluency at each knowledge point transition, local optimization targeting this can effectively and directly improve the smoothness of the path's micro-connections.

[0051] The iteration termination condition includes: Reaching the maximum number of iterations: Set the maximum number of iterations to 100 to prevent the algorithm from getting stuck in an excessively long search in extreme cases.

[0052] The historical best path has not been improved in Q consecutive iterations: Setting the parameter Q to 5 prevents premature stopping due to a few accidental iterations failing to improve the path, and can also identify relatively quickly when the algorithm has entered a plateau period, thus terminating unnecessary searches in time.

[0053] In a specific embodiment, the perturbation operation corresponding to the selected operator includes a node replacement operator perturbation operation and a local rearrangement operator perturbation operation, specifically including: The perturbation operations of the node replacement operator include: Replace nodes in the current path whose behavior cost is higher than a preset semantic threshold with other nodes in the same graph subset whose behavior cost is lower than that node to obtain the perturbed path. Specifically, the preset semantic threshold is set as the median of the behavioral costs of all nodes in the current path. The median is not sensitive to extreme values, represents the typical level of the behavioral costs of the current path, and can be adjusted according to the current path.

[0054] The node appearing in the perturbation operation of the node replacement operator only refers to the node whose behavioral cost is higher than the preset semantic threshold.

[0055] The perturbation operations of the local rearrangement operator include: Remove the subsequence consisting of consecutive nodes from the current path to obtain a partial path. For each node in the subsequence, calculate the insertion cost of inserting the node into both ends of the partial path and between every two adjacent nodes. The calculation expression is as follows: ; in, For part of the path, The total cost of a partial path. For the new partial path, The total cost of the new partial path, Let v be the insertion cost at position pos; For each node, the position with the minimum insertion cost is determined as the optimal insertion position, and the position with the second minimum insertion cost is determined as the suboptimal insertion position. The difference between the insertion cost of the suboptimal insertion position and the insertion cost of the optimal insertion position is the regret value of the node. Nodes in the subsequence are inserted into the optimal insertion positions of the nodes in the partial path in descending order of regret value. The partial path is updated after each node is inserted until all nodes have been inserted, resulting in the perturbed path.

[0056] Specifically, in the perturbation operation of the local rearrangement operator, the regret value is used to determine the order of node re-insertion, prioritizing nodes most sensitive to their insertion position to reduce the overall reconstruction cost. If the cost difference between the optimal and second-best insertion positions of a node is large (i.e., the regret value is high), it indicates that the node's position choice has a significant impact on the total path cost, and its optimal position should be determined as early as possible. Conversely, the insertion order of nodes with lower regret values ​​has less impact on the overall result and can be processed later. By inserting nodes sequentially in descending order of regret value, this strategy can more effectively approximate the globally optimal solution during the local reconstruction process.

[0057] This application constructs a textbook knowledge graph by binding the semantic content of knowledge points with user interaction behavior data to nodes and edges of the knowledge graph. It uses topological potential to filter a subset of the core knowledge graph and introduces semantic constraints to ensure the representativeness of the initial node set. Then, using the edge cost distance revealed by user behavior as the optimization target, it uses the LKH algorithm to generate an initial path and integrates an adaptive large neighborhood search algorithm with custom operators to continuously optimize the path. The node replacement operator uses dynamic thresholds to locate and replace high behavioral cost nodes within the same semantic graph subset to accurately reduce the local learning difficulty. The local reordering operator introduces a regret value sorting strategy to prioritize the nodes most sensitive to the insertion position to significantly improve the quality of local reconstruction. At the same time, it dynamically adjusts the usage weight of the operator based on its historical performance and sets reasonable termination conditions to generate and continuously optimize a learning path that combines knowledge logic, learning fluency, and individual adaptability.

[0058] In a specific embodiment, the preset scoring criteria include: The scoring criteria include a first criterion, a second criterion, a third criterion, and a fourth criterion; The first criterion is that the total cost of the candidate path is less than the total cost of the historical best path, and the selected operator is assigned a first preset score. The second criterion is that the total cost of the candidate path is less than the total cost of the current path, but greater than the total cost of the historical best path, and a second preset score is assigned to the selected operator. The third criterion is that if the total cost of a candidate path is greater than the total cost of the historical best path, it is accepted as the new current path according to the preset path acceptance strategy, and the selected operator is assigned a third preset score. The fourth criterion is that if the total cost of a candidate path is greater than the total cost of the current path and it is not accepted by the preset path acceptance strategy, the selected operator is assigned a fourth preset score. Among them, the first preset score > the second preset score > the third preset score > the fourth preset score; Specifically, the preset path acceptance strategy is the Metropolis acceptance criterion, which accepts the candidate path of the i-th iteration as the current path with a certain probability. The calculation expression is: ; in, Let be the probability of the i-th iteration. The total cost of the current path, The total cost of the candidate paths, T This is the current annealing temperature, and its value decreases at a preset annealing rate during the iteration process.

[0059] Annealing temperature T initial value Based on the overall cost setting of the initial path, and after each iteration... Update The annealing coefficient is less than 1.

[0060] This method sets a scoring standard and assigns a corresponding operator score to each scoring result. This score is used to quantify the quality of the current operator perturbation operation and participates in the subsequent calculation of operator weight updates.

[0061] The rules for assigning fraction values ​​are as follows: If the total cost of a candidate path is less than the total cost of the historical best path, then the path is the best solution found so far. In this case, the operator used in this round is assigned a first preset score, which is the highest score among the four levels. This reward is given to the operator that can discover a new and better solution, and encourages it to be used more frequently in subsequent searches.

[0062] If the total cost of a candidate path is less than the total cost of the current path, but greater than or equal to the total cost of the historical best path, then the path, although not surpassing the historical best, is better than the current solution. In this case, the operator used in this round is assigned a second preset score, which is moderately high. This score is used to reward operators that can continuously improve the current solution and promote the algorithm to converge towards a better direction.

[0063] If the total cost of a candidate path is greater than or equal to the total cost of the historical best path, but is accepted as the new current path according to the preset simulated annealing acceptance criterion, the path is not better, but is adopted because the algorithm needs to explore. In this case, the operator used in this round is assigned a third preset score. This score is low and moderately rewards operators that do not directly improve the search during the exploration process but help maintain the diversity of the search.

[0064] If the total cost of a candidate path is greater than the total cost of the current path and it is not accepted as a new path (i.e., the path is of poor quality and not adopted), then the operator used in this round is assigned a fourth preset score, which is the lowest score among the four levels. It is usually set to zero to penalize operators that fail to produce effective improvements and reduce their probability of use in subsequent iterations.

[0065] In this embodiment, the first preset score is set to 10, the second preset score is set to 5, the third preset score is set to 2, and the fourth preset score is set to 0.

[0066] The expression for updating the operator weights based on the scores of the selected operators is as follows: ; in, For the updated operator weights, The operator weights before the update. The fraction of the selected operator, For the number of times the operator is used, Update the coefficients for the weights.

[0067] Specifically, the operator weights are updated using a smooth update strategy. The weight update coefficient is set to 0.2 to control the smoothness of the weight update. The cumulative number of times the operator is used is used to balance the cumulative effect of the score of the frequently used operator and to avoid premature saturation of the weights.

[0068] In a specific embodiment, step S4 specifically includes: Step S41: Based on the node order in the optimal path, sort the learning order of the corresponding knowledge points in the digital teaching materials, and reorganize the jump links between knowledge points according to the edge relationships in the knowledge graph. Step S42: For nodes in the optimal path whose behavioral cost is higher than the first knowledge threshold, associate auxiliary learning resources with the knowledge points corresponding to the nodes in the digital teaching materials. For edges in the optimal path where the cost of a sub-path exceeds the second knowledge threshold, transitional learning resources that connect the corresponding knowledge points are provided in the digital teaching materials. Step S43: Update the above optimization results to the digital textbook platform to form optimized digital textbooks.

[0069] Specifically, the first knowledge threshold is the median of the behavioral costs of all nodes in the optimal path.

[0070] The second knowledge threshold is the median cost of all sub-paths in the optimal path.

[0071] Based on the edge relationships in the knowledge graph, the reorganization of the jump links between knowledge points specifically includes: Based on the node sequence determined by the optimal path, navigation links between adjacent nodes in the path are strengthened, while other links outside the path may be weakened, hidden, or rearranged. For example, on the learning page for knowledge point A, the link pointing to the next knowledge point B in the optimal path is set as the primary recommendation or the only default "Next" button. The direct technical effect of this optimization is to guide and constrain the user's browsing flow, keeping it consistent with the recommended learning path.

[0072] Supplementary learning resources refer to the binding relationship established between nodes identified in the optimal path as having a behavioral cost exceeding the knowledge threshold (i.e., knowledge points that users generally find difficult to learn) and supplementary learning resources within the digital textbook platform. Supplementary learning resources include explanatory prompts, entry points for reviewing prior knowledge, and adaptive practice questions.

[0073] Transitional learning resources refer to the binding relationship established between edges identified as having high sub-path costs in the optimal path (i.e., connections where users experience a poor transition between two consecutive knowledge points) and specific guiding content within the digital textbook platform. Transitional learning resources include transitional instructions, learning strategy suggestions, and bridging examples.

[0074] Supplementary and transitional learning resources are stored in the digital textbook platform and associated with corresponding knowledge points and their IDs (i.e., the start and end points of edges). This method optimizes and updates these associations to achieve dynamic resource recommendation.

[0075] Example 2, another embodiment of the present invention, is a digital textbook optimization system based on user interaction behavior, including a construction module, a sorting module, an optimization module, and an application module: The building module is used to construct a textbook knowledge graph using digital textbooks and user logs; The sorting module is used to aggregate nodes in the knowledge graph of the textbook, obtain a subset of the graph, calculate the topological potential of the subset of the graph, and sort them according to the topological potential. The optimization module is used to select nodes from the graph subset with high topological potential to form an initial set, generate an initial path using the LHK algorithm, and optimize the path using the adaptive large neighborhood search algorithm based on the initial path to obtain the optimal path. The application module is used to rearrange the order of knowledge points and reorganize jump links according to the optimal path, associate learning resources with high-cost nodes and high-cost edges, and generate optimized digital teaching materials. Learning resources include supplementary learning resources and transitional learning resources.

[0076] This invention optimizes learning paths by establishing a knowledge graph. Within the knowledge graph, an adaptive large neighborhood search algorithm is used, incorporating user-specific interaction data to adjust the learning path in real time. This interaction data provides insights into the student's current cognitive state and learning habits. Through multiple iterative searches, the order of nodes within the path is flexibly adjusted to minimize the cost of the learning path, ensuring that the order of learning content and knowledge points always aligns with the student's actual needs and avoiding ineffective and repetitive learning. Furthermore, the topological potential of each knowledge point is calculated within the knowledge graph to accurately assess its priority and learning difficulty within the learning path. Combined with the adaptive large neighborhood search algorithm, a regret value ranking mechanism is used to optimize path adjustments, ensuring that each path change results in minimal learning loss, thereby improving the accuracy and efficiency of path optimization.

[0077] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for optimizing digital teaching materials based on user interaction behavior, characterized in that, Includes the following steps: Step S1 involves constructing a textbook knowledge graph using digital teaching materials and user logs, specifically including: Step S11: Obtain digital teaching materials and user logs. The digital teaching materials include a set of knowledge points, a set of knowledge point content, and the prior relationships between knowledge points. The user log includes the user's dwell time, completion rate, number of times the user repeatedly viewed each knowledge point, question accuracy rate, jump frequency, and dwell time after jumping. Step S12: Extract knowledge points from the knowledge point set as nodes, and take the preorder relationship between knowledge points as edges. By vectorizing the digital textbook and user logs respectively, obtain the node feature vector and edge feature vector of each knowledge point. The node feature vector includes behavioral index vector and textbook index vector. The textbook indicator vector includes knowledge point titles, keywords, and concept descriptions; The behavioral indicator vector includes the dwell time, completion rate, number of times repeatedly viewed, and question accuracy rate for each knowledge point; The edge feature vector includes the dwell time after the jump and the jump frequency; Step S13: Bind each knowledge point to the corresponding node feature vector and store it as a unique node in the textbook knowledge graph. Bind each preorder relation to the corresponding edge feature vector and store it as a directed edge connecting two nodes. Determine the direction of the edge through the preorder relation. Based on all the nodes and edges mentioned above, a graph structure is constructed, stored in a graph database, and indexes are created for node features and edge features to obtain the textbook knowledge graph. Step S2 involves aggregating the nodes in the textbook knowledge graph to obtain a subset of the graph, calculating the topological potential of the subset, and sorting it according to the topological potential. Specifically, this includes: Step S21: Based on the node feature vector, calculate the cosine similarity of the semantics between any two nodes to obtain the node similarity matrix, and use the hierarchical clustering algorithm to aggregate the nodes in the textbook knowledge graph into k graph subsets, where k is an integer greater than 2; Step S22: Calculate the topological potential energy of each graph subset and sort all graph subsets in descending order according to their corresponding topological potential energy values. The quality of each node is obtained by analyzing the question accuracy rate in the node's feature vector. The centroid of the graph subset is obtained by calculating the mean of the eigenvectors of all nodes within the subset. The graph structure distance from each node in a subset of the knowledge graph to the centroid of the subset is calculated based on the edge relationships in the knowledge graph of the textbook. The graph structure distance is the number of edges of the shortest path from the node to the centroid of the subset. The expression for calculating topological potential energy is: ； in, Let be the topological potential of the i-th graph subset. For the i-th subset of the graph, For a subset of the graph The nodes in For nodes quality For nodes To the atlas subset Graph structural distance of the centroid, Impact factor; Step S23, based on the node feature vector, calculate the behavioral cost of each node, specifically including: Step S231: Obtain the behavior index vector and standardize the behavior index vector using Z-score standardization. Step S232: Based on the standardized behavior index vector, calculate the behavior cost of the node using the Euclidean distance formula. The calculation expression is: ； in, As a consequence of the behavior, The number of behavioral indicator vectors. Let k be the standardized behavioral index vector. Let be the weight coefficient of the k-th behavioral indicator vector, and ; Step S3: Select nodes from the graph subset with high topological potential to form an initial set, generate an initial path using the LHK algorithm, and optimize the path using the adaptive large neighborhood search algorithm based on the initial path to obtain the optimal path. Step S4: Reorder the knowledge points and reorganize the jump links according to the optimal path, associate learning resources with high-cost nodes and high-cost edges, and generate optimized digital teaching materials; The learning resources include supplementary learning resources and transitional learning resources.

2. The digital textbook optimization method based on user interaction behavior as described in claim 1, characterized in that, Step S3 specifically includes: Step S31: From the sorted graph subsets, select the graph subsets with the highest topological potential in sequence until the nodes included in the selected graph subsets satisfy the preset semantic constraints, and obtain the initial node set. Step S32: Take the initial node set as the required nodes, use the LKH algorithm to generate the initial path that traverses all the required nodes based on the knowledge graph in the textbook, and calculate the sub-path cost of the initial path according to the sub-path cost calculation rules. The specific rules for calculating subpath costs include: The dwell time and jump frequency after the jump are normalized, and the edge cost distance is calculated by subtracting the normalized dwell time and jump frequency. The expression for calculating the cost of a subpath for any given path is: ； in, Let $C$ be the cost of a subpath of any path. For nodes and The edge cost distance between them For nodes The estimated difficulty value, For nodes The estimated difficulty value, , These are preset weighting coefficients; Step S33: Starting from the initial path, use the adaptive large neighborhood search algorithm for iterative optimization to obtain the optimal path.

3. The digital textbook optimization method based on user interaction behavior as described in claim 2, characterized in that, Step S33 specifically includes: Step S331: The behavior cost of the initial path is obtained by summing the behavior costs of all nodes in the initial path, and the sum of the behavior cost of the initial path and the cost of the sub-paths is calculated as the total cost. The proportion of behavioral cost in the total cost is used as the initial weight of the node replacement operator; The proportion of sub-path cost in the total cost is used as the initial weight of the local rearrangement operator; Step S332: Based on the current operator weights of the node replacement operator and the local rearrangement operator, select the node replacement operator and the local rearrangement operator using the roulette wheel selection method, execute the perturbation operation corresponding to the selected operator, and obtain the perturbed path; Step S333: Using the perturbed path, based on the edge cost distance, perform local optimization using the 2-opt algorithm to obtain the path whose total cost is less than the total cost of the perturbed path as the candidate path; Step S334: If the total cost of the candidate path is less than the total cost of the current path but greater than the total cost of the optimal path, then update the current path using the candidate path. If the total cost of the candidate path is less than the total cost of the current path and less than the total cost of the optimal path, then the optimal path is updated using the candidate path. Step S335: Based on the comparison results between the candidate path and the current path and the optimal path, score the selected operator according to the preset scoring criteria, and update the operator weight of the selected operator based on the score. Repeat steps S332 to S335 until the iteration termination condition is met; The iteration terminates when the maximum number of iterations is reached or the optimal path is not improved in Q consecutive iterations. The optimal path and the corresponding set of knowledge points are then output, where Q is an integer greater than 2.

4. The digital textbook optimization method based on user interaction behavior as described in claim 3, characterized in that, The perturbation operations corresponding to the selected operators include node replacement operator perturbation operations and local rearrangement operator perturbation operations, specifically including: The perturbation operations of the node replacement operator include: Replace nodes in the current path whose behavior cost is higher than a preset semantic threshold with other nodes in the same graph subset whose behavior cost is lower than that node to obtain the perturbed path. The perturbation operations of the local rearrangement operator include: Remove the subsequence consisting of consecutive nodes from the current path to obtain a partial path. For each node in the subsequence, calculate the insertion cost of inserting the node into both ends of the partial path and between every two adjacent nodes. The calculation expression is as follows: ； in, For part of the path, The total cost of a partial path. For the new partial path, The total cost of the new partial path, Let v be the insertion cost at position pos; For each node, the position with the minimum insertion cost is determined as the optimal insertion position, and the position with the second minimum insertion cost is determined as the suboptimal insertion position. The difference between the insertion cost of the suboptimal insertion position and the insertion cost of the optimal insertion position is the regret value of the node. Nodes in the subsequence are inserted into the optimal insertion positions of the nodes in the partial path in descending order of regret value. The partial path is updated after each node is inserted until all nodes have been inserted, resulting in the perturbed path.

5. The digital textbook optimization method based on user interaction behavior as described in claim 4, characterized in that, The preset scoring criteria specifically include: The scoring criteria include a first standard, a second standard, a third standard, and a fourth standard; The first criterion is that the total cost of the candidate path is less than the total cost of the historical best path, and the selected operator is assigned a first preset score. The second criterion is that the total cost of the candidate path is less than the total cost of the current path, but greater than the total cost of the historical best path, and a second preset score is assigned to the selected operator. The third criterion is that if the total cost of a candidate path is greater than the total cost of the historical best path, it is accepted as the new current path according to the preset path acceptance strategy, and the selected operator is assigned a third preset score. The fourth criterion is that if the total cost of the candidate path is greater than the total cost of the current path and it is not accepted by the preset path acceptance strategy, the selected operator is assigned a fourth preset score. Wherein, the first preset score > the second preset score > the third preset score > the fourth preset score; The expression for updating the operator weights based on the scores of the selected operators is as follows: ； in, For the updated operator weights, The operator weights before the update. The fraction of the selected operator, For the number of times the operator is used, Update the coefficients for the weights.

6. The digital textbook optimization method based on user interaction behavior as described in claim 5, characterized in that, Step S4 specifically includes: Step S41: Based on the node order in the optimal path, sort the learning order of the corresponding knowledge points in the digital teaching materials, and reorganize the jump links between knowledge points according to the edge relationships in the knowledge graph. Step S42: For nodes in the optimal path whose behavioral cost is higher than the first knowledge threshold, associate auxiliary learning resources with the knowledge points corresponding to the nodes in the digital teaching materials. For edges in the optimal path whose sub-path cost exceeds the second knowledge threshold, transitional learning resources that connect the corresponding knowledge points are provided in the digital teaching materials. Step S43: Update the above optimization results to the digital textbook platform to form optimized digital textbooks.

7. A digital textbook optimization system based on user interaction behavior, characterized in that, It includes a build module, a sorting module, an optimization module, and an application module: The construction module is used to build a textbook knowledge graph using digital textbooks and user logs, specifically including: Obtain digital teaching materials and user logs, wherein the digital teaching materials include a set of knowledge points, a set of knowledge point content, and the prior relationships between knowledge points; The user log includes the user's dwell time, completion rate, number of times the user repeatedly viewed each knowledge point, question accuracy rate, jump frequency, and dwell time after jumping. The knowledge points are extracted from the knowledge point set as nodes, and the preorder relationships between knowledge points are used as edges. By vectorizing the digital textbook and user logs respectively, the node feature vector and edge feature vector of each knowledge point are obtained. The node feature vector includes behavioral index vector and textbook index vector. The textbook indicator vector includes knowledge point titles, keywords, and concept descriptions; The behavioral indicator vector includes the dwell time, completion rate, number of times repeatedly viewed, and question accuracy rate for each knowledge point; The edge feature vector includes the dwell time after the jump and the jump frequency; Each knowledge point is bound to the corresponding node feature vector and stored as a unique node in the textbook knowledge graph. Each preorder relation is bound to the corresponding edge feature vector and stored as a directed edge connecting two nodes. The direction of the edge is determined by the preorder relation. Based on all the nodes and edges mentioned above, a graph structure is constructed, stored in a graph database, and indexes are created for node features and edge features to obtain the textbook knowledge graph. The sorting module is used to aggregate nodes in the textbook knowledge graph to obtain a subset of the graph, calculate the topological potential of the subset of the graph, and sort them according to the topological potential. Specifically, it includes: Based on the node feature vectors, the semantic cosine similarity between any two nodes is calculated to obtain the node similarity matrix. Then, the nodes in the textbook knowledge graph are aggregated into k graph subsets using a hierarchical clustering algorithm, where k is an integer greater than 2. Calculate the topological potential energy of each graph subset, and sort all graph subsets in descending order according to their corresponding topological potential energy values; The quality of each node is obtained by analyzing the question accuracy rate in the node's feature vector. The centroid of the graph subset is obtained by calculating the mean of the eigenvectors of all nodes within the subset. The graph structure distance from each node in a subset of the knowledge graph to the centroid of the subset is calculated based on the edge relationships in the knowledge graph of the textbook. The graph structure distance is the number of edges of the shortest path from the node to the centroid of the subset. The expression for calculating topological potential energy is: ； in, Let be the topological potential of the i-th graph subset. For the i-th subset of the graph, For a subset of the graph The nodes in For nodes quality For nodes To the atlas subset Graph structural distance of the centroid, Impact factor; Based on the node feature vectors, the behavioral cost of each node is calculated, specifically including: Obtain the behavioral indicator vector and standardize it using Z-score standardization. Based on the standardized behavior index vector, the behavior cost of a node is calculated using the Euclidean distance formula. The calculation expression is as follows: ； in, As a consequence of the behavior, The number of behavioral indicator vectors. Let k be the standardized behavioral index vector. Let be the weight coefficient of the k-th behavioral indicator vector, and ; The optimization module is used to select nodes from the graph subset with high topological potential to form an initial set, generate an initial path using the LHK algorithm, and optimize the path using the adaptive large neighborhood search algorithm based on the initial path to obtain the optimal path. The application module is used to rearrange the order of knowledge points and reorganize jump links according to the optimal path, associate learning resources with high-cost nodes and high-cost edges, and generate optimized digital teaching materials. The learning resources include supplementary learning resources and transitional learning resources.

Images