Recommendation methods, systems, computer equipment, and media based on learning behavior data

Abstract

This invention relates to a recommendation method, system, computer device, and medium based on learning behavior data, belonging to the field of data recommendation technology. The method includes the following steps: collecting learners' online learning behavior sequences and performing time-series analysis on the online learning behavior sequences to obtain a learning duration distribution map; constructing a knowledge graph of learners based on the learning duration distribution map to obtain a learning knowledge distribution map; mining interests based on the learning knowledge distribution map to obtain learner interest feature vectors; performing collaborative filtering matching based on the learner interest feature vectors to obtain a candidate resource set; and performing personalized sorting on the candidate resource set to obtain a recommendation result list. This invention solves the technical problem that traditional methods often separate learning behavior data from knowledge structure, failing to fully integrate the dynamic correlation between the time dimension and knowledge mastery status in the behavior sequence.

Description

Technical Field

[0001] This invention relates to the field of data recommendation technology, and in particular to recommendation methods, systems, computer devices, and media based on learning behavior data. Background Technology

[0002] In existing technologies, although some studies have attempted to introduce knowledge graphs or collaborative filtering mechanisms to improve recommendation performance, they often treat learning behavior data separately from knowledge structures, failing to fully integrate the dynamic correlation between the time dimension of the behavior sequence and the state of knowledge mastery. For example, behavioral details such as the duration of learners' investment in knowledge points, the frequency of repeated access, and the jumping patterns at different times can reflect their depth of understanding and interest transfer trends. However, most current systems do not make sufficient use of this fine-grained temporal information, resulting in coarse knowledge modeling and distorted interest representation. Summary of the Invention

[0003] The purpose of this invention is to at least partially solve one of the technical problems existing in the prior art.

[0004] To achieve the above objectives, the present invention provides a recommendation method based on learning behavior data, comprising the following steps:

[0005] Collect learners' online learning behavior sequences and perform time-series analysis on the online learning behavior sequences to obtain a learning duration distribution map;

[0006] Based on the learning duration distribution map, a knowledge graph is constructed for the learners to obtain a learning knowledge distribution map.

[0007] Interest mining is performed based on the aforementioned knowledge distribution map to obtain learner interest feature vectors;

[0008] Collaborative filtering matching is performed based on the learner's interest feature vector to obtain a candidate resource set, and the candidate resource set is then sorted in a personalized manner to obtain a recommendation result list.

[0009] Furthermore, a time-series analysis is performed on the online learning behavior sequence to obtain a learning duration distribution map, including:

[0010] The online learning behavior sequence is divided into learning periods and the behavior type is identified to obtain a time-series behavior feature set. The time-series behavior feature set is then subjected to time window sliding statistics to obtain a learning engagement time-series matrix.

[0011] Based on the learning engagement time series matrix, the learning behavior sequence is aggregated by time segments to obtain a learning intensity periodic map, and the learning intensity periodic map is decomposed by time series features to obtain a learning duration distribution map.

[0012] Furthermore, based on the learning duration distribution map, a knowledge graph is constructed for the learners to obtain a learning knowledge distribution map, including:

[0013] The knowledge point correlation degree is calculated on the learning time distribution map to obtain the knowledge point time investment matrix, and the knowledge point time investment matrix is ​​decomposed hierarchically to obtain the knowledge unit mastery sequence.

[0014] Based on the knowledge unit mastery sequence, the learner's knowledge point dependency relationship is analyzed to obtain a knowledge point association network, and the topology of the knowledge point association network is reconstructed to obtain a knowledge network diagram.

[0015] The knowledge context map is mapped to a learning trajectory to obtain a knowledge mastery progress map, and dynamic features are extracted from the knowledge mastery progress map to obtain a knowledge mastery feature matrix.

[0016] Based on the knowledge mastery feature matrix, the learner's knowledge structure is analyzed to obtain a knowledge ability vector. The knowledge ability vector is then aggregated in multiple dimensions to obtain a learning knowledge distribution map.

[0017] Furthermore, based on the learned knowledge distribution map, interest mining is performed to obtain learner interest feature vectors, including:

[0018] The node access order of the knowledge distribution map is extracted to obtain the knowledge point temporal access matrix, and the knowledge point temporal access matrix is ​​fused by time decay to obtain the cognitive preference intensity map.

[0019] Node transition sequence analysis is performed on the cognitive preference intensity map to obtain a set of knowledge exploration paths, and hierarchical clustering is performed on the set of knowledge exploration paths to obtain a knowledge interest distribution network.

[0020] The knowledge interest distribution network is subjected to feature extraction and dimensionality reduction to obtain a multidimensional interest feature group, and the multidimensional interest feature group is then mapped into a vector space to obtain the learner interest feature vector.

[0021] Furthermore, collaborative filtering matching is performed based on the learner interest feature vector to obtain a candidate resource set, including:

[0022] The cosine similarity between the learner's interest feature vector and the learning resource feature library is calculated to obtain the interest-resource matching degree matrix. The interest-resource matching degree matrix is ​​then subjected to threshold filtering to obtain the initial matching resource set.

[0023] Based on the initial matched resource set, a resource-resource association graph is constructed. The resource-resource association graph is then clustered and divided using a community detection algorithm to obtain resource interest communities. Density peaks are extracted from the resource interest communities to obtain a core resource subset.

[0024] The Jaccard distance is calculated between the core resource subset and the learner's historical access resource set to obtain a resource novelty score vector. Based on the resource novelty score vector, the core resource subset is further filtered to obtain a candidate resource set.

[0025] Furthermore, the candidate resource set is personalized and sorted to obtain a recommendation result list, including:

[0026] The resource difficulty stratification calculation is performed on the candidate resource set to obtain the resource gradient score matrix, and the difference between the resource gradient score matrix and the learning knowledge distribution map is calculated to obtain the resource suitability sequence.

[0027] Based on the resource suitability sequence, a learning dependency chain analysis is performed on the candidate resource set to obtain a resource pre-relationship network. The resource pre-relationship network is then topologically sorted to obtain a learning progression path graph.

[0028] The learning progression path map is subjected to resource quality assessment and timeliness analysis to obtain a resource weight distribution table. The resource weight distribution table is then sorted by multiple factors to obtain a recommendation result list.

[0029] Furthermore, based on the resource suitability sequence, dependency chain analysis is performed on the candidate resource set to obtain a resource pre-relationship network, including:

[0030] The resource adaptability sequence is decomposed into knowledge point elements to obtain a resource knowledge composition table, and the knowledge unit association strength of the resource knowledge composition table is calculated to obtain a knowledge point dependency matrix.

[0031] Based on the knowledge point dependency matrix, the candidate resource set is parsed to obtain a knowledge point co-occurrence sequence, and a learning order constraint analysis is performed on the knowledge point co-occurrence sequence to obtain a resource learning sequence graph.

[0032] The knowledge point transfer path is traced on the resource learning sequence graph to obtain a knowledge transfer chain network, and the critical path is extracted from the knowledge transfer chain network to obtain a knowledge dependency strength graph.

[0033] Based on the knowledge dependency strength graph, the candidate resource set is optimized by resource combination to obtain a resource organization structure tree. The topological dependency relationship of the resource organization structure tree is then reconstructed to obtain a resource pre-relationship network.

[0034] This invention also provides a recommendation system based on learning behavior data, comprising:

[0035] The data acquisition module is used to collect learners' online learning behavior sequences and perform time-series analysis on the online learning behavior sequences to obtain a learning duration distribution map.

[0036] The construction module is used to construct a knowledge graph of learners based on the learning duration distribution map to obtain a learning knowledge distribution map;

[0037] The mining module is used to mine interests based on the learning knowledge distribution map to obtain learner interest feature vectors.

[0038] The matching module is used to perform collaborative filtering matching based on the learner's interest feature vector to obtain a candidate resource set, and to perform personalized sorting of the candidate resource set to obtain a recommendation result list.

[0039] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0040] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0041] This invention provides a recommendation method based on learning behavior data, comprising the following steps: collecting learners' online learning behavior sequences and performing time-series analysis on the online learning behavior sequences to obtain a learning duration distribution map; constructing a knowledge graph of learners based on the learning duration distribution map to obtain a learning knowledge distribution map; mining interests based on the learning knowledge distribution map to obtain learner interest feature vectors; performing collaborative filtering matching based on the learner interest feature vectors to obtain a candidate resource set, and performing personalized sorting of the candidate resource set to obtain a recommendation result list. This method solves the technical problem that traditional methods often separate learning behavior data from knowledge structure, failing to fully integrate the dynamic correlation between the time dimension and knowledge mastery status in the behavior sequence. It enables personalized sorting of the candidate resource set by combining learners' knowledge weaknesses, interest intensity, and learning progress, ensuring that the final recommendation result not only meets user preferences but also serves their long-term learning goals, thus enhancing the educational value and user experience of the recommendation. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of the steps of a recommendation method based on learning behavior data in one embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the steps of a recommendation system based on learning behavior data in one embodiment of the present invention;

[0045] Figure 3 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0046] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0047] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0048] The following describes in detail, with reference to the accompanying drawings, a recommendation method based on learning behavior data according to an embodiment of the present invention. First, the recommendation method based on learning behavior data according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

[0049] Figure 1 This invention provides a recommendation method based on learning behavior data, comprising the following steps:

[0050] Step S1: Collect the learner's online learning behavior sequence and perform time series analysis on the online learning behavior sequence to obtain a learning duration distribution map.

[0051] Specifically, in collecting learners' online learning behavior sequences, the first step is to establish a data collection mechanism to ensure comprehensive recording of all interactions between learners and the learning platform, including but not limited to video viewing, document reading, and completing exercises. Next, when performing time-series analysis on these online learning behavior sequences, the timestamp of each behavior should be used as a key parameter. Algorithms should be used to identify the sequence and intervals between different behaviors, thereby revealing learning patterns and habits. For example, a sliding window technique can be used to capture learning activities within a continuous time period, and then calculate a learning duration distribution map. This map shows the time learners spend on various knowledge points or course modules and their access frequency within a specific time period. Specifically, in an online programming course, if a learner frequently studies Python basic syntax intensively between 8 PM and 10 PM from Monday to Friday, with each session lasting approximately one hour, the above steps can accurately obtain this information, constructing a learning duration distribution map reflecting their learning patterns. Subsequently, this distribution map can be used to further explore the learner's knowledge absorption at different stages, providing a basis for subsequent knowledge graph construction. In this process, it is crucial to ensure the accuracy and completeness of the data so that subsequent steps can be based on reliable and in-depth analysis. At the same time, it is also essential to protect learners' privacy and ensure that data use complies with relevant laws and regulations. In this way, through meticulous analysis of learning behavior sequences, we can not only gain insights into individual learners' preferences and pace but also provide strong support for personalized education.

[0052] Step S2: Based on the learning duration distribution map, construct a knowledge graph for the learners to obtain a learning knowledge distribution map.

[0053] Specifically, after obtaining the learning time distribution map, it needs to be aligned with the pre-defined course or knowledge point structure to construct a knowledge graph. Specifically, firstly, the learning content corresponding to each time segment in the learning time distribution map is mapped to specific nodes in the knowledge system. These nodes represent knowledge points or skill units in the course syllabus. Then, based on indicators such as the learner's cumulative learning time, number of repeated visits, and depth of stay at each node, a weight value is assigned to each knowledge node. This weight reflects the learner's exposure intensity and potential mastery of the knowledge point. Next, according to the knowledge dependencies defined in the knowledge graph (such as prerequisite-successor, contain-included, etc.), the weighted knowledge nodes are connected into a personalized graph structure, i.e., the learning knowledge distribution map. For example, in the aforementioned online programming course scenario, if the learning time distribution map shows that a learner has accumulated 45 minutes, 30 minutes, and 60 minutes of learning on knowledge points such as "variable definition," "data type," and "conditional statement," respectively, and has repeatedly reviewed content related to "conditional statement," then when constructing the knowledge graph, the "conditional statement" node will be given a higher weight and a directed edge will be established with its predecessor, "variable definition," thus forming a learning knowledge distribution map that reflects the learner's current knowledge coverage and focus. This process relies on structured course metadata and standardized knowledge ontology to ensure that the knowledge graphs of different learners can be compared and reasoned within a unified semantic space, thereby providing a computable graph structure input for subsequent interest mining.

[0054] Step S3: Based on the learned knowledge distribution map, interest mining is performed to obtain learner interest feature vectors.

[0055] Specifically, after obtaining the knowledge distribution map, interest mining is needed based on the weights and topological structure of each knowledge node in the map to generate learner interest feature vectors. Specifically, firstly, the weight value corresponding to each knowledge node in the knowledge distribution map is extracted. This weight comes from the learning input intensity mapped from the learning duration distribution map in the previous step. Then, combined with predefined semantic labels or topic classifications in the knowledge graph (such as "basic grammar," "control structures," "functional programming," etc.), knowledge nodes with the same topic or similar semantics are aggregated, and a weighted average or cumulative score is calculated for each topic dimension. Next, the scores of all topic dimensions are arranged in a fixed order to form a high-dimensional numerical vector, i.e., the learner interest feature vector. For example, in the aforementioned online programming course application scenario, if the knowledge distribution map shows that learners have higher weights on knowledge points belonging to the "control structure" category, such as "conditional statements" and "loop structures," while their weights are close to zero on nodes related to "object-oriented programming," then during the interest mining process, the "control structure" dimension will receive a higher score, while the "object-oriented" dimension will receive a lower score. The resulting interest feature vector might be [0.85, 0.12, 0.67, ...], corresponding to different preset knowledge topics. This vector not only retains the structural information of the original knowledge distribution but also transforms it into a numerical form that can be directly processed by subsequent collaborative filtering algorithms, thereby ensuring that the recommendation process can accurately reflect the learner's current interest focus and knowledge preferences. The entire process relies on a unified knowledge topic system and standardized vectorization rules to ensure that the interest feature vectors of different learners are comparable and consistent.

[0056] Step S4: Perform collaborative filtering matching based on the learner interest feature vector to obtain a candidate resource set, and then perform personalized sorting on the candidate resource set to obtain a recommendation result list.

[0057] Specifically, after obtaining the learner's interest feature vector, it needs to be used for collaborative filtering to generate a candidate resource set. First, the similarity between the learner's interest feature vector and the interest feature vectors of other learners in the system is calculated. Common methods include cosine similarity or Pearson correlation coefficient, thereby identifying several neighboring users with similar interests. Then, learning resources that these neighboring users have interacted with but that the current learner has not yet accessed are collected, forming a preliminary candidate resource set. Simultaneously, an item-based collaborative filtering strategy can be combined to map the candidate resources themselves to the same knowledge topic dimension space as the interest feature vector, calculating the matching score between the resource content vector and the learner's interest feature vector, further expanding or filtering the candidate resource set. Next, the resulting candidate resource set is personalized and ranked. The ranking process considers not only the collaborative filtering matching score but also multi-dimensional features such as the resource's difficulty level, knowledge coverage, and the strength of its association with the learner's current knowledge weaknesses. A weighted linear combination or ranking model (such as Learning-to-Rank) is used to calculate the comprehensive ranking score for each resource, and these resources are ranked from highest to lowest, finally outputting a recommendation list. For example, in the aforementioned online programming course scenario, if a learner's interest feature vector shows a high level of interest in "control structures" type knowledge points, collaborative filtering might match resources such as "nested loop practice exercises" or "comprehensive case studies of conditional statements" that other users with similar preferences have studied. During the ranking phase, video courses on "application of loops combined with conditions" that are closely related to the learner's existing content and of appropriate difficulty are prioritized, thus forming a recommendation list that conforms to both group behavior patterns and individual cognitive states. The entire process strictly relies on the learner interest feature vector generated in the preceding steps and ensures that all operations are completed within a unified knowledge topic and vector space to maintain the consistency and executability of the recommendation logic.

[0058] In a specific embodiment, time-series analysis is performed on the online learning behavior sequence to obtain a learning duration distribution map, including:

[0059] The online learning behavior sequence is divided into learning periods and the behavior type is identified to obtain a time-series behavior feature set. The time-series behavior feature set is then subjected to time window sliding statistics to obtain a learning engagement time-series matrix.

[0060] Based on the learning engagement time series matrix, the learning behavior sequence is aggregated by time segments to obtain a learning intensity periodic map, and the learning intensity periodic map is decomposed by time series features to obtain a learning duration distribution map.

[0061] Specifically, in the process of performing time-series analysis on the online learning behavior sequence to obtain a learning duration distribution map, it is first necessary to perform learning period segmentation and behavior type identification on the original online learning behavior sequence. Specifically, each record in the behavior sequence is assigned to a preset time granularity unit (such as by hour, by day, or by course session) based on its timestamp, forming several continuous or discrete learning periods. Simultaneously, each behavior is classified and labeled according to the operation type field in the behavior log (such as "video playback," "exercise submission," "document reading," "pause," "fast forward," etc.), thereby assigning a clear behavior type label to the behavior within each period. Through the above processing, a structured dataset containing information such as time location, behavior type, and duration can be constructed, which is the time-series behavior feature set. Next, a time window sliding statistical method is applied to the set of temporal behavioral features: a fixed-length time window (e.g., 30 minutes) is set, and the window slides along the time axis in steps of a certain size (e.g., 5 minutes). The total duration of various effective learning behaviors within the window (excluding non-engagement behaviors such as "fast forward" and "long periods of inactivity") is statistically analyzed and normalized to a learning engagement intensity value per unit time. This generates a two-dimensional matrix, where rows correspond to different sliding window positions and columns correspond to different behavior types or knowledge point categories. This matrix is ​​the learning engagement temporal matrix. Subsequently, the learning behavior sequence is aggregated into time segments based on the learning engagement temporal matrix. In this step, based on the changes in engagement in the matrix, clustering or threshold segmentation methods are used to merge windows with consecutive high engagement into a single learning segment, while intervals with low or zero engagement are considered as intervals or interruptions. This divides the entire learning process into several semantically meaningful time segments. Each time segment is associated with its start and end times, the range of knowledge points covered, and the cumulative engagement time, thus integrating them into a visual structure reflecting the learner's learning activity level in different periods, i.e., a learning intensity periodic map. For example, in the aforementioned online programming course application scenario, if a learner repeatedly watches videos related to "loop structures" and completes the accompanying exercises from 7:00 PM to 9:00 PM Monday through Wednesday, but does not engage in any activity on Thursday, this process will be aggregated into two high-intensity learning segments and a low-intensity window period, forming a typical periodic learning pattern graph. Finally, the learning intensity periodic graph is decomposed into temporal features to extract more representative learning duration distribution patterns. This decomposition process can employ time-series signal processing techniques such as Fourier transform, wavelet analysis, or principal component analysis to decompose the learning intensity signal that fluctuates over time in the graph into trend terms, periodic terms, and random terms. Among these, the trend term reflects the changing trend of long-term learning investment, the periodic term reveals learning habits under fixed rhythms such as daily and weekly, and the weight distribution of each knowledge point in these components is summarized into the final numerical expression.By weighting and integrating these components and reorganizing them according to knowledge point dimensions, a learning time distribution chart can be generated. This chart, with knowledge points on the horizontal axis and standardized cumulative effective learning time or frequency on the vertical axis, clearly shows the learner's time allocation across different knowledge units. For example, in this programming course, the learning time distribution chart might show that "conditional statements" have a cumulative effective learning time of 75 minutes, "loop structures" 92 minutes, and "function definitions" only 18 minutes, intuitively reflecting the learner's focus of interest and depth of knowledge exposure, providing precise input for subsequent knowledge graph construction.

[0062] In a specific embodiment, a knowledge graph is constructed for learners based on the learning duration distribution map to obtain a learning knowledge distribution map, including:

[0063] The knowledge point correlation degree is calculated on the learning time distribution map to obtain the knowledge point time investment matrix, and the knowledge point time investment matrix is ​​decomposed hierarchically to obtain the knowledge unit mastery sequence.

[0064] Based on the knowledge unit mastery sequence, the learner's knowledge point dependency relationship is analyzed to obtain a knowledge point association network, and the topology of the knowledge point association network is reconstructed to obtain a knowledge network diagram.

[0065] The knowledge context map is mapped to a learning trajectory to obtain a knowledge mastery progress map, and dynamic features are extracted from the knowledge mastery progress map to obtain a knowledge mastery feature matrix.

[0066] Based on the knowledge mastery feature matrix, the learner's knowledge structure is analyzed to obtain a knowledge ability vector. The knowledge ability vector is then aggregated in multiple dimensions to obtain a learning knowledge distribution map.

[0067] Specifically, after obtaining the learning time distribution map, the first step is to calculate the correlation between knowledge points to quantify the learner's time investment in each knowledge point. Specifically, the effective learning time corresponding to each knowledge point in the learning time distribution map is used as the raw input. Combined with the predefined knowledge point IDs and hierarchical structure in the course knowledge system, a matrix is ​​constructed, with knowledge points as rows and learning periods or knowledge categories as columns—this is the knowledge point time investment matrix. Each element in this matrix represents the learner's cumulative learning time or standardized investment intensity on a specific knowledge point. Next, this knowledge point time investment matrix is ​​hierarchically decomposed: based on the hierarchical relationship of the knowledge system (e.g., "Programming Basics" includes sub-nodes such as "Variables," "Data Types," and "Operators"), the time investment values ​​of sub-knowledge points are aggregated layer by layer from bottom to top, and the comprehensive mastery index of the parent node is calculated, thus forming a sequence organized by knowledge hierarchy—this is the knowledge unit mastery sequence. Subsequently, based on the knowledge unit mastery sequence, a knowledge point dependency analysis is performed on the learner. In this step, the pre-defined prerequisite-successor and precondition-result logical dependency rules in the course body are used to determine whether learners are learning in a reasonable knowledge order. For example, if the learning time for a "loop structure" is high but the time spent on its prerequisite knowledge point "conditional statement" is extremely low, it may indicate a break in the knowledge chain. By traversing the dependencies between all knowledge point pairs and combining the mastery level values ​​in the knowledge unit mastery sequence, a weighted directed graph can be constructed, where nodes are knowledge points, edges represent dependencies, and edge weights reflect the coordination of the mastery levels of preceding and following knowledge points. This graph is the knowledge point association network. Next, the topology of this knowledge point association network is reconstructed: isolated nodes are removed, strongly coupled subgraphs are merged, and the edge directions are dynamically adjusted according to the mastery level, ultimately generating a clear and logically coherent knowledge network graph to represent the knowledge connection path that the learner is currently constructing. Subsequently, the learning trajectory is mapped onto the knowledge network graph. This process projects the learner's historical behavioral sequence (such as video viewing order and exercise completion order) onto the node paths of the knowledge network map in chronological order, identifying the actual knowledge access paths they have traversed. Combined with the mastery strength of each node, the learning status of each node on the path is labeled (e.g., "initial contact," "repeated review," "mastered"), thus forming a knowledge mastery progress map. Based on this, dynamic features are extracted from this knowledge mastery progress map: indicators such as the rate of change in mastery, access frequency, and backtracking depth for each knowledge point are extracted over time and organized into a structured matrix, namely the knowledge mastery feature matrix, where rows correspond to knowledge points and columns correspond to different dynamic feature dimensions. Finally, a knowledge structure analysis is performed on the learner based on the knowledge mastery feature matrix.By employing methods such as clustering, principal component analysis, or weighted summation, multidimensional features in a matrix are fused into a low-dimensional vector. Each dimension of this vector represents a specific type of knowledge ability (e.g., "grammar comprehension," "logical control," "problem-solving ability"), resulting in a knowledge ability vector. This knowledge ability vector is then aggregated in multiple dimensions—for example, by knowledge topic (e.g., "basic syntax," "control structures," "functional programming")—and the relevant dimensions are summed or averaged, aligned with the original knowledge point structure. This ultimately generates a structured graph that encompasses both the breadth and depth of knowledge coverage, i.e., a learning knowledge distribution map. For instance, in the aforementioned online programming course scenario, if a learner demonstrates high frequency of review and significant time commitment for "conditional statements" and "loop structures," but almost no record of "function definitions," their learning knowledge distribution map will show highlighted nodes and dense connections in the "control structures" area, while appearing sparse or blank in the "functional programming" area. This visually reflects the completeness and bias of their current knowledge structure, providing a precise basis for subsequent interest mining.

[0068] In a specific embodiment, interest mining is performed based on the learned knowledge distribution map to obtain learner interest feature vectors, including:

[0069] The node access order of the knowledge distribution map is extracted to obtain the knowledge point temporal access matrix, and the knowledge point temporal access matrix is ​​fused by time decay to obtain the cognitive preference intensity map.

[0070] Node transition sequence analysis is performed on the cognitive preference intensity map to obtain a set of knowledge exploration paths, and hierarchical clustering is performed on the set of knowledge exploration paths to obtain a knowledge interest distribution network.

[0071] The knowledge interest distribution network is subjected to feature extraction and dimensionality reduction to obtain a multidimensional interest feature group, and the multidimensional interest feature group is then mapped into a vector space to obtain the learner interest feature vector.

[0072] Specifically, after obtaining the knowledge distribution map, the first step is to extract the node access order to reconstruct the learner's actual browsing or learning trajectory in the knowledge space. Specifically, based on the timestamp information and connections recorded by each knowledge point node in the knowledge distribution map, the accessed knowledge points are arranged in chronological order to form an ordered sequence. This sequence is then converted into a matrix structure, where rows represent time steps (e.g., 1st visit, 2nd visit, etc.), columns represent all possible knowledge points, and matrix elements indicate whether a knowledge point was accessed at the corresponding time step and its duration of access, thus obtaining the knowledge point temporal access matrix. Next, the knowledge point temporal access matrix undergoes time decay fusion: an exponential decay function (e.g., decay factor λ=0.9) is introduced, assigning lower weights to access records at earlier time steps and retaining higher weights for more recent visits. The cumulative intensity value of each knowledge point after decay is calculated through weighted summation, and these values ​​are reorganized into a structured graph reflecting current cognitive preferences, i.e., a cognitive preference intensity map. Subsequently, node transfer sequence analysis is performed based on the cognitive preference intensity map. This process identifies all consecutive pairs of knowledge point transitions by traversing the jump records between knowledge points (e.g., from "conditional statements" to "loop structures"), and statistically analyzes their transition frequency and directionality, thereby constructing several semantically coherent knowledge exploration paths. For example, in a programming course, a path might appear: "variable definition → data type → conditional statement → loop structure → nested loop application". These paths are integrated into a set, namely the knowledge exploration path set. Next, this knowledge exploration path set is subjected to hierarchical clustering: first, coarse-grained grouping is performed based on the knowledge topics contained in the paths (e.g., "control structures" and "basic syntax"), and then fine-grained clustering is performed within each group based on path length and transition pattern similarity (e.g., using edit distance or DTW algorithm), ultimately forming a network structure composed of multiple interest clusters, where each cluster represents a typical knowledge exploration pattern, nodes are knowledge points, and edges are high-frequency transition relationships. This network is the knowledge interest distribution network. Subsequently, feature extraction and dimensionality reduction processing are performed on the knowledge interest distribution network. In the feature extraction stage, the central knowledge points, intra-cluster node density, cross-cluster connection strength, path diversity, and other indicators of each interest cluster are extracted from the network to form an initial high-dimensional feature set, i.e., a multi-dimensional interest feature group. Since this feature group has a high dimension and may contain redundancy, dimensionality reduction processing is required. Principal component analysis (PCA), t-SNE, or autoencoders can be used to compress the feature dimension while retaining the main interest structure information.Finally, the multidimensional interest feature groups after dimensionality reduction are mapped into a vector space: each dimension is mapped to a predefined interest semantic axis (such as "grammar mastery tendency", "logical reasoning preference", "practical application interest", etc.), and they are unified to the [0,1] interval through standardization (such as Min-Max normalization), ultimately generating a fixed-length numerical vector, i.e., the learner interest feature vector. For example, in the aforementioned online programming course application scenario, if a learner frequently switches from "conditional statements" to "loop structures" and repeatedly reviews "nested loops", while hardly involving "functions" or "object-oriented" content, then their knowledge exploration path set will mainly focus on the control structure domain, forming a high-density interest cluster after hierarchical clustering; after feature extraction and dimensionality reduction, this cluster scores significantly higher than other dimensions in the "logical control interest" dimension, and the final generated learner interest feature vector may be [0.12, 0.89, 0.76, 0.05, …], corresponding to different preset interest axes.

[0073] In a specific embodiment, collaborative filtering matching is performed based on the learner interest feature vector to obtain a candidate resource set, including:

[0074] The cosine similarity between the learner's interest feature vector and the learning resource feature library is calculated to obtain the interest-resource matching degree matrix. The interest-resource matching degree matrix is ​​then subjected to threshold filtering to obtain the initial matching resource set.

[0075] Based on the initial matched resource set, a resource-resource association graph is constructed. The resource-resource association graph is then clustered and divided using a community detection algorithm to obtain resource interest communities. Density peaks are extracted from the resource interest communities to obtain a core resource subset.

[0076] The Jaccard distance is calculated between the core resource subset and the learner's historical access resource set to obtain a resource novelty score vector. Based on the resource novelty score vector, the core resource subset is further filtered to obtain a candidate resource set.

[0077] Specifically, after obtaining the learner's interest feature vector, it is first necessary to calculate its cosine similarity with the learning resource feature library to quantify the degree of matching between the learner's interests and various learning resources. Specifically, each resource in the learning resource feature library (such as video courses, exercises, documents, etc.) has been pre-represented as a resource feature vector with the same dimension as the learner's interest feature vector. These vectors are constructed by mapping the knowledge points covered by the resources to a topic-dimensional space consistent with the interest feature vector. Subsequently, the cosine similarity is calculated between the learner's interest feature vector and each resource feature vector in the library, resulting in a matching score between 0 and 1. All scores are then organized into a matrix form according to resource index, i.e., the interest-resource matching degree matrix. Next, a preset threshold (e.g., 0.6) is set for this matrix, retaining only resource entries with a matching degree higher than this threshold, thereby filtering out a list of resources highly relevant to the learner's current interests, forming an initial matching resource set. Subsequently, a resource-resource association graph is constructed based on the initial matching resource set. The nodes of this graph represent each resource in the initial matching resource set. Edges are established based on co-occurrence relationships or semantic similarity between resources: if two resources are frequently and continuously accessed by the same user in historical user behavior data, or if the cosine similarity between their resource feature vectors exceeds a certain threshold, a weighted edge is added between them. The weight can be set as co-occurrence frequency or similarity value. Based on this, a community detection algorithm (such as the Louvain algorithm or label propagation algorithm) is used to cluster the resource-resource association graph, identifying several subgraph structures with tightly connected internal structures and sparse external structures. Each subgraph corresponds to a resource interest community, representing a cluster of resources under a specific knowledge topic or learning style. Then, density peak extraction is performed on each resource interest community: the local density (such as the sum of edge weights in the neighborhood) and relative distance of each node within the community are calculated. Resources with high density and farthest from nodes with even higher density are selected as the core nodes of the community, thus extracting the most representative resources from each community to form a core resource subset. Next, Jaccard distance is calculated between the core resource subset and the learner's historical resource set to evaluate the novelty of the recommended content. Specifically, each resource in the core resource subset and the historical resource set are treated as two separate sets. The ratio of their intersection to their union, i.e., the Jaccard similarity coefficient, is calculated. This coefficient is then subtracted from 1 to obtain the Jaccard distance, which serves as the resource's novelty score. A higher score indicates a lower overlap between the resource and the learner's existing content, and thus stronger novelty. All novelty scores for the core resources are organized into a vector, i.e., the resource novelty score vector. Finally, the core resource subset is further filtered based on this vector: a lower novelty threshold is set (e.g., Jaccard distance ≥ 0.4), eliminating overly repetitive or already accessed resources, and retaining resources that combine interest matching and content novelty, ultimately forming a candidate resource set.For example, in the aforementioned online programming course application scenario, if the learner's interest feature vector shows a high preference for topics related to "control structures," the initial matching resource set may include resources such as "detailed explanation of conditional statements," "practical application of nested loops," and "comprehensive exercises in flow control." By constructing a resource-resource association graph and performing community discovery, a resource interest community centered on "application of loops combined with conditions" can be identified. After density peak extraction, "nested loop debugging techniques" is selected as the core resource of this community. Furthermore, Jaccard distance calculation reveals that the learner has not yet accessed this resource (there is no intersection in the historical set), and its novelty score is 1.0, so it is retained in the candidate resource set.

[0078] In a specific embodiment, the candidate resource set is personalized and sorted to obtain a recommendation result list, including:

[0079] The resource difficulty stratification calculation is performed on the candidate resource set to obtain the resource gradient score matrix, and the difference between the resource gradient score matrix and the learning knowledge distribution map is calculated to obtain the resource suitability sequence.

[0080] Based on the resource suitability sequence, a learning dependency chain analysis is performed on the candidate resource set to obtain a resource pre-relationship network. The resource pre-relationship network is then topologically sorted to obtain a learning progression path graph.

[0081] The learning progression path map is subjected to resource quality assessment and timeliness analysis to obtain a resource weight distribution table. The resource weight distribution table is then sorted by multiple factors to obtain a recommendation result list.

[0082] Specifically, after obtaining the candidate resource set, the first step is to perform a stratified calculation of resource difficulty to quantify the cognitive load level of each resource. Specifically, based on predefined difficulty labels in the course knowledge system (such as "beginner," "intermediate," and "advanced") or by extracting indicators such as lexical complexity, syntactic depth, and knowledge point density from the resource text content (such as video subtitles and document text) using natural language processing technology, each resource is mapped to a standardized difficulty value range (e.g., 0.0–1.0), thus constructing a structured representation where rows correspond to resources and columns correspond to difficulty dimensions—the resource gradient scoring matrix. Next, the difference between this resource gradient scoring matrix and the learning knowledge distribution map generated in the previous steps is calculated: each knowledge point in the learning knowledge distribution map carries a mastery level weight, which is mapped to the set of knowledge points covered by the resource. The absolute difference or KL divergence between the overall difficulty of the resource and the learner's current knowledge mastery level is calculated as the degree of mismatch between the resource and the learner's cognitive state. The complement or negative value is then used to form an adaptation score, which is finally organized into a resource adaptation sequence according to resource order. Subsequently, a learning dependency chain analysis is performed on the candidate resource set based on the resource suitability sequence. This process utilizes the pre-defined knowledge point prerequisite relationships in the course ontology (e.g., "Mastering conditional statements is a prerequisite for learning loop structures") to compare the prerequisite knowledge points that each candidate resource depends on with their mastery status in the learning knowledge distribution graph. If the weight of a key prerequisite knowledge point of a resource in the learning knowledge distribution graph is lower than a threshold, it is determined that the resource has a missing dependency, and the required prerequisite resources are recorded. By traversing all candidate resources and their dependencies, a directed graph structure is constructed, where nodes represent resources and edges represent logical dependencies such as "A must be learned before B can be effectively learned." This graph is the resource prerequisite relationship network. Then, the resource prerequisite relationship network is topologically sorted: using the Kahn algorithm or a depth-first search strategy, resource nodes are linearly arranged according to dependency order, eliminating loops and ensuring that all prerequisite resources are placed before subsequent resources, thereby generating a learning progression path graph that conforms to the laws of cognitive development. Next, the learning progression path graph is subjected to resource quality assessment and timeliness analysis. Resource quality assessment is based on multi-source indicators, including average user ratings, completion rate, error rate, and teacher rating levels. These indicators are normalized and then weighted to form a quality score. Timeliness analysis examines the resource's release date, content update history, and technology stack version (e.g., Python 3.8 vs. Python 2.7), assigning lower timeliness scores to outdated content. The quality score and timeliness score are then combined according to a preset ratio to assign a comprehensive weight to each resource in the learning progression path, thus forming a resource weight distribution table.Finally, the resource weight distribution table is sorted using a multi-factor fusion method: in addition to weights, factors such as the adaptation score in the resource fit sequence, the position number in the learning progression path graph (the earlier the better), and the novelty score (from previous steps) are introduced. A weighted linear combination or a Learning-to-Rank model (such as RankNet) is used to calculate the final ranking score for each resource, and these resources are then arranged from highest to lowest to output a recommendation list. For example, in the aforementioned online programming course application scenario, the candidate resource set includes resources such as "Nested Loop Practice," "Function Parameter Passing," and "Recursion Basics." A comparison of difficulty stratification and learning knowledge distribution revealed that learners had mastered "loop structures" but had not yet encountered "functions," resulting in a lower suitability for "function parameter passing." Dependency chain analysis further confirmed that "recursion basics" require "functions" as a prerequisite, thus placing it at the end of the path. Meanwhile, "nested loop practice" received the highest ranking score after considering multiple factors due to mastery of prerequisite knowledge points, a high quality score (completion rate of 92%), and strong timeliness (updated in 2024), making it the top recommendation.

[0083] In a specific embodiment, dependency chain analysis is performed on the candidate resource set based on the resource suitability sequence to obtain a resource pre-relationship network, including:

[0084] The resource adaptability sequence is decomposed into knowledge point elements to obtain a resource knowledge composition table, and the knowledge unit association strength of the resource knowledge composition table is calculated to obtain a knowledge point dependency matrix.

[0085] Based on the knowledge point dependency matrix, the candidate resource set is parsed to obtain a knowledge point co-occurrence sequence, and a learning order constraint analysis is performed on the knowledge point co-occurrence sequence to obtain a resource learning sequence graph.

[0086] The knowledge point transfer path is traced on the resource learning sequence graph to obtain a knowledge transfer chain network, and the critical path is extracted from the knowledge transfer chain network to obtain a knowledge dependency strength graph.

[0087] Based on the knowledge dependency strength graph, the candidate resource set is optimized by resource combination to obtain a resource organization structure tree. The topological dependency relationship of the resource organization structure tree is then reconstructed to obtain a resource pre-relationship network.

[0088] Specifically, after obtaining the resource fit sequence, the first step is to decompose the sequence into knowledge point elements to identify the specific knowledge content covered by each candidate resource. Specifically, based on the predefined knowledge point tagging system in the course knowledge ontology, the metadata associated with each resource in the resource fit sequence (such as the syllabus, subtitle text, and exercise knowledge point annotations) is parsed into a set of discrete knowledge units. The frequency or coverage depth of each unit in the resource is recorded, thus constructing a structured table, namely the resource knowledge composition table. Rows correspond to candidate resources, columns correspond to all possible knowledge points, and elements in the table represent the degree to which the resource contains a certain knowledge point. Next, the knowledge unit association strength is calculated for this resource knowledge composition table: using co-occurrence statistics or conditional probability methods, the frequency of any two knowledge points appearing simultaneously in the same resource is calculated. Combined with course logic rules (such as "if A is a prerequisite for B, then the association strength between A and B is enhanced"), a symmetric or directed numerical matrix, namely the knowledge point dependency matrix, is generated to quantify the inherent dependencies between knowledge points. Subsequently, based on the knowledge point dependency matrix, the candidate resource set is analyzed to further refine the co-occurrence patterns among knowledge points. In this step, the knowledge point set of each resource in the resource knowledge composition table is traversed and arranged according to its order of appearance in the original learning behavior or instructional design (e.g., video explanation order), forming an ordered knowledge point sequence. Such sequences for all resources are merged to form a knowledge point co-occurrence sequence set. Then, a learning order constraint analysis is performed on this knowledge point co-occurrence sequence: the prerequisite-successor rules in the course knowledge graph are introduced as hard constraints to check whether each co-occurrence sequence violates cognitive logic (e.g., "function call" appears before "function definition"), and violating sequences are corrected or marked. Finally, sequences conforming to educational logic are integrated into a graph structure with directional edges, where nodes are knowledge points and edges represent legitimate learning sequence relationships. This graph is the resource learning sequence graph. Next, the knowledge point transmission path is traced on the resource learning sequence graph. This process employs a graph traversal algorithm (such as depth-first search), starting from the basic knowledge points contained in each resource and proceeding forward along the directed edges of the resource learning sequence graph, recording all reachable knowledge point paths, thereby constructing multiple transfer chains reflecting the gradual expansion and deepening of knowledge. These transfer chains are merged and deduplicated to form a network structure covering the overall knowledge evolution path of the candidate resources, i.e., a knowledge transfer chain network. Based on this, critical paths are extracted from the knowledge transfer chain network: the sum of the weights of each edge on each path (provided by the knowledge point dependency matrix) is calculated, and combined with the path length and the number of covered resources, core paths with high weights and high educational value are selected. Then, the frequency and importance of each knowledge point in the critical paths are summarized to generate a graph with knowledge points as nodes and dependency strength as values, i.e., a knowledge dependency strength graph. Finally, resource combination optimization is performed on the candidate resource set based on the knowledge dependency strength graph.This optimization process uses clusters of knowledge points with high dependency intensity in the knowledge dependency strength graph as aggregation units. Candidate resources covering the same or adjacent clusters are grouped into the same logical group, and parent-child relationships are constructed based on the dependency direction, forming a hierarchical tree structure, i.e., a resource organization structure tree. The root node represents basic knowledge point resources, and the leaf nodes represent higher-order comprehensive application resources. Then, the topological dependency relationships of this resource organization structure tree are reconstructed: traversing each level of the tree, the parent-child connections between resources are adjusted according to the directed dependency edges in the knowledge dependency strength graph, ensuring that all prerequisite knowledge points of the child resource are covered in its ancestor nodes, and eliminating circular dependencies or redundant branches. Finally, a directed acyclic graph that accurately reflects "which resources must be learned first, which can be parallelized, and which need to be applied later" is output, i.e., a resource prerequisite relationship network. For example, in the aforementioned online programming course application scenario, candidate resources include "variable assignment examples," "conditional judgment practice," "detailed explanation of for loops," and "comprehensive problems with nested loops." After decomposing the knowledge points, the "Comprehensive Problem with Nested Loops" includes knowledge units such as "conditional statements," "for loops," and "nested logic." The knowledge point dependency matrix shows that "for loops" have a strong dependency on "variable assignment." Resource content analysis reveals that its knowledge point co-occurrence sequence is [variable assignment → for loop → nested logic]. Learning order constraint analysis confirms that this sequence conforms to the teaching logic. Path tracing reveals that the migration chain from "variable assignment" to "nested logic" runs through multiple resources. After extracting the critical path, "for loops" is identified as a node with high dependency strength. Resource combination optimization sets "variable assignment example" as the parent node, "detailed explanation of for loops" as its child node, and places "Comprehensive Problem with Nested Loops" at a deeper level. The resource prerequisite relationship network formed after topology reconstruction clearly requires that the former two must be completed before the latter can be recommended.

[0089] The above describes the recommendation method based on learning behavior data in the embodiments of the present invention. The following describes the recommendation system based on learning behavior data in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the recommendation system based on learning behavior data in this invention includes:

[0090] The acquisition module 21 is used to acquire the learner's online learning behavior sequence and perform time series analysis on the online learning behavior sequence to obtain a learning duration distribution map;

[0091] Module 22 is used to construct a knowledge graph of learners based on the learning duration distribution map to obtain a learning knowledge distribution map;

[0092] Mining module 23 is used to perform interest mining based on the learning knowledge distribution map to obtain learner interest feature vectors;

[0093] The matching module 24 is used to perform collaborative filtering matching based on the learner's interest feature vector to obtain a candidate resource set, and to perform personalized sorting on the candidate resource set to obtain a recommendation result list.

[0094] In this embodiment, the specific implementation of each module in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0095] Reference Figure 3 This invention also provides a computer device whose internal structure can be as follows: Figure 3 As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0096] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0097] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0098] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0099] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0100] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A recommendation method based on learning behavior data, characterized in that, Includes the following steps: Collect learners' online learning behavior sequences and perform time-series analysis on the online learning behavior sequences to obtain a learning duration distribution map; Based on the learning duration distribution map, a knowledge graph is constructed for the learners to obtain a learning knowledge distribution map. Interest mining is performed based on the aforementioned knowledge distribution map to obtain learner interest feature vectors; Collaborative filtering matching is performed based on the learner's interest feature vector to obtain a candidate resource set, and the candidate resource set is then sorted in a personalized manner to obtain a recommendation result list.

2. The recommendation method based on learning behavior data according to claim 1, characterized in that, A time-series analysis was performed on the online learning behavior sequence to obtain a learning duration distribution map, including: The online learning behavior sequence is divided into learning periods and the behavior type is identified to obtain a time-series behavior feature set. The time-series behavior feature set is then subjected to time window sliding statistics to obtain a learning engagement time-series matrix. Based on the learning engagement time series matrix, the learning behavior sequence is aggregated by time segments to obtain a learning intensity periodic map, and the learning intensity periodic map is decomposed by time series features to obtain a learning duration distribution map.

3. The recommendation method based on learning behavior data according to claim 1, characterized in that, Based on the learning duration distribution map, a knowledge graph is constructed for the learners to obtain a learning knowledge distribution map, including: The knowledge point correlation degree is calculated on the learning time distribution map to obtain the knowledge point time investment matrix, and the knowledge point time investment matrix is ​​decomposed hierarchically to obtain the knowledge unit mastery sequence. Based on the knowledge unit mastery sequence, the learner's knowledge point dependency relationship is analyzed to obtain a knowledge point association network, and the topology of the knowledge point association network is reconstructed to obtain a knowledge network diagram. The knowledge context map is mapped to a learning trajectory to obtain a knowledge mastery progress map, and dynamic features are extracted from the knowledge mastery progress map to obtain a knowledge mastery feature matrix. Based on the knowledge mastery feature matrix, the learner's knowledge structure is analyzed to obtain a knowledge ability vector. The knowledge ability vector is then aggregated in multiple dimensions to obtain a learning knowledge distribution map.

4. The recommendation method based on learning behavior data according to claim 1, characterized in that, Based on the aforementioned knowledge distribution map, interest mining is performed to obtain learner interest feature vectors, including: The node access order of the knowledge distribution map is extracted to obtain the knowledge point temporal access matrix, and the knowledge point temporal access matrix is ​​fused by time decay to obtain the cognitive preference intensity map. Node transition sequence analysis is performed on the cognitive preference intensity map to obtain a set of knowledge exploration paths, and hierarchical clustering is performed on the set of knowledge exploration paths to obtain a knowledge interest distribution network. The knowledge interest distribution network is subjected to feature extraction and dimensionality reduction to obtain a multidimensional interest feature group, and the multidimensional interest feature group is then mapped into a vector space to obtain the learner interest feature vector.

5. The recommendation method based on learning behavior data according to claim 1, characterized in that, Based on the learner interest feature vector, collaborative filtering matching is performed to obtain a candidate resource set, including: The cosine similarity between the learner's interest feature vector and the learning resource feature library is calculated to obtain the interest-resource matching degree matrix. The interest-resource matching degree matrix is ​​then subjected to threshold filtering to obtain the initial matching resource set. Based on the initial matched resource set, a resource-resource association graph is constructed. The resource-resource association graph is then clustered and divided using a community detection algorithm to obtain resource interest communities. Density peaks are extracted from the resource interest communities to obtain a core resource subset. The Jaccard distance is calculated between the core resource subset and the learner's historical access resource set to obtain a resource novelty score vector. Based on the resource novelty score vector, the core resource subset is further filtered to obtain a candidate resource set.

6. The recommendation method based on learning behavior data according to claim 1, characterized in that, The candidate resource set is then sorted in a personalized manner to obtain a list of recommended results, including: The resource difficulty stratification calculation is performed on the candidate resource set to obtain the resource gradient score matrix, and the difference between the resource gradient score matrix and the learning knowledge distribution map is calculated to obtain the resource suitability sequence. Based on the resource suitability sequence, a learning dependency chain analysis is performed on the candidate resource set to obtain a resource pre-relationship network. The resource pre-relationship network is then topologically sorted to obtain a learning progression path graph. The learning progression path map is subjected to resource quality assessment and timeliness analysis to obtain a resource weight distribution table. The resource weight distribution table is then sorted by multiple factors to obtain a recommendation result list.

7. The recommendation method based on learning behavior data according to claim 6, characterized in that, Based on the resource suitability sequence, dependency chain analysis is performed on the candidate resource set to obtain a resource pre-relationship network, including: The resource adaptability sequence is decomposed into knowledge point elements to obtain a resource knowledge composition table, and the knowledge unit association strength of the resource knowledge composition table is calculated to obtain a knowledge point dependency matrix. Based on the knowledge point dependency matrix, the candidate resource set is parsed to obtain a knowledge point co-occurrence sequence, and a learning order constraint analysis is performed on the knowledge point co-occurrence sequence to obtain a resource learning sequence graph. The knowledge point transfer path is traced on the resource learning sequence graph to obtain a knowledge transfer chain network, and the critical path is extracted from the knowledge transfer chain network to obtain a knowledge dependency strength graph. Based on the knowledge dependency strength graph, the candidate resource set is optimized by resource combination to obtain a resource organization structure tree. The topological dependency relationship of the resource organization structure tree is then reconstructed to obtain a resource pre-relationship network.

8. A recommendation system based on learning behavior data, characterized in that, The method for performing recommendation based on learning behavior data according to any one of claims 1 to 7 includes: The data acquisition module is used to collect learners' online learning behavior sequences and perform time-series analysis on the online learning behavior sequences to obtain a learning duration distribution map. The construction module is used to construct a knowledge graph of learners based on the learning duration distribution map to obtain a learning knowledge distribution map; The mining module is used to mine interests based on the learning knowledge distribution map to obtain learner interest feature vectors. The matching module is used to perform collaborative filtering matching based on the learner's interest feature vector to obtain a candidate resource set, and to perform personalized sorting of the candidate resource set to obtain a recommendation result list.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images