Education resource recommendation method and system

Abstract

The invention discloses an educational resource recommendation method and system. The method comprises the following steps: S1, constructing a multi-layer educational database; S2, calculating a user interest feature vector; S3, carrying out resource matching based on improved cosine similarity; and accurate recommendation and scene adaptability improvement: the system can accurately identify user demands by fusing user portrait analysis, resource semantic feature matching and a dynamic interest model. And cold start optimization and dynamic resource updating: aiming at new users or new resources, the system adopts a group feature fusion strategy, for example, an initial recommendation list is generated through KL divergence and JS divergence analysis of users of the same age group, and recommendation weights are quickly adjusted in combination with real-time behavior data. In addition, the resource library supports dynamic index updating and block hash storage, it is ensured that newly-added resources can be rapidly brought into a recommendation queue, and the response timeliness is improved by 60%.

Description

Technical Field

[0001] The present invention relates to an educational resource recommendation method and system, specifically an educational resource recommendation method and system. Background Art

[0002] With the acceleration of the educational informatization process, the popularization of online education platforms, digital resource libraries and Internet of Things technologies, the number of educational resources has shown an exponential growth. However, the following key problems exist in the personalized recommendation of massive resources in the prior art:

[0003] Insufficient static recommendation and dynamic adaptability:

[0004] Traditional collaborative filtering and content-based recommendation algorithms rely on historical data to construct static user portraits, and cannot capture the dynamic changes of user interests in real time (such as learning stage migration, improvement of knowledge point mastery, etc.). For example, existing systems are difficult to dynamically adjust the recommendation strategy according to user behavior (such as answer accuracy rate, resource browsing duration), resulting in the recommendation results lagging behind the actual needs.

[0005] Cold start problem and data sparsity:

[0006] New users or new resources are prone to fall into the cold start dilemma due to the lack of historical interaction data. Although the prior art alleviates this problem through group feature fusion (such as K-means clustering or user portraits of the same age group), the degree of fusion of cross-disciplinary knowledge associations (such as knowledge point graph construction) and dynamic answering data is insufficient, resulting in low initial recommendation accuracy.

[0007] Resource heterogeneity and label system defects:

[0008] Educational resources are diverse in type (courseware, videos, exercises, etc.), and the proportion of unstructured data (such as teaching videos) is high. Existing methods rely on manual labeling of tags and are difficult to extract fine-grained features through automated semantic analysis (such as BERT model), resulting in the neglect of the logical associations between knowledge points (such as interdisciplinary nature) during resource matching.

[0009] Recommendation homogenization and information cocoon effect:

[0010] A single algorithm (such as collaborative filtering) is prone to the problem of repeatedly recommending similar content. Although a hybrid recommendation model (collaborative filtering + deep learning) has been proposed, it lacks in-depth optimization of user behavior reinforcement (such as click-through rate feedback) and timeliness weight (such as LFU-ARAM cache strategy), resulting in insufficient resource diversity.

[0011] Cross-platform integration and privacy protection challenges:

[0012] Educational resources are scattered across multiple platforms (such as online courses, school databases), and existing systems have technical bottlenecks in cross-platform data integration and permission control. At the same time, the privacy protection mechanism for user behavior data (such as learning trajectories, answer records) is imperfect, making it prone to data leakage risks. Summary of the Invention

[0013] The technical problem to be solved by the present invention is to overcome the above-mentioned technical defects and provide an educational resource recommendation method and system.

[0014] To solve the above technical problems, the technical solution provided by the present invention is an educational resource recommendation method and system:

[0015] Including the following steps:

[0016] S1. Construct a multi-layer educational database, the structure of which satisfies:

[0017] Among them, is the set of resources uploaded manually, is the set of resources crawled from the network, among which, is the set of resources uploaded manually, is the set of resources crawled from the network, among which, is the set of resources uploaded manually, is the set of resources crawled from the network

[0018] Among them is the set of resources uploaded manually, is the set of resources crawled from the network

[0019] S2. Calculate the user interest feature vector:

[0020]

[0021] Among them, w i represents the browsing duration weight of resource r i and α is the time decay factor;

[0022] S3. Resource matching based on improved cosine similarity:

[0023]

[0024] S4. Dynamically update the recommendation queue to satisfy:

[0025]

[0026] As an improvement, the educational database adopts a hybrid storage structure:

[0027] Structured data, unstructured data, structured data, unstructured data, structured data, unstructured data

[0028]

[0029] Among them, the structured data table contains fields:

[0030] Φ = <ResourceID,Type,Keywords,RelationalWeight>

[0031] Unstructured data is stored in chunks:

[0032] where is the hash function where is the hash function where is the hash function where is the SHA-256 hash function.

[0033] As an improvement, the file system adopts a dynamic index structure:

[0034] where the update volume where the update volume where the update volume

[0035]

[0036] and constructs a B+ tree index to satisfy:

[0037] where is the maximum number of child nodes of the node and is the total number of records where is the maximum number of child nodes of the node and is the total number of records where is the maximum number of child nodes of the node and is the total number of records where m is the maximum number of child nodes of the node and N is the total number of records.

[0038] As an improvement, the resource association degree calculation adopts:

[0039]

[0040] where represents the resource semantic feature set and t is the keyword vector.

[0041] As an improvement, it includes:

[0042] Heterogeneous data collection module, satisfying:

[0043] M collect = <WebCrawler,API Gateway,Stream Processor>

[0044] Distributed feature calculation engine:

[0045]

[0046] Hybrid recommender implementation:

[0047] RecSys = αCF + βCB + (1 - α - β)DL (α + β ≤ 1).

[0049] As an improvement, the database index adopts an adaptive Bε-tree structure, and the node splitting condition is:

[0050] and and and

[0051]

[0052] Among which, the entropy value calculation:

[0053]

[0054] As an improvement, the user profile update strategy adopts a dual decay model:

[0055] Time decay behavior reinforcement Time decay behavior reinforcement Time decay behavior reinforcement

[0056]

[0057] Among which, CTR_d represents the click-through rate of resource d.

[0058] As an improvement, in the cold start phase, adopt:

[0059]

[0060] Among which is the set of characteristic vectors of the same age group, and KL and JS are the Kullback-Leibler and Jensen-Shannon divergences respectively.

[0061] As an improvement, the cache mechanism adopts the LFU-ARAM strategy, and the cache weight calculation:

[0062] Frequency term Time term Frequency term Time term Frequency term Time term

[0063]

[0064] The elimination condition is:

[0065]

[0066] As an improvement, the anomaly detection module adopts:

[0067]

[0068] When AnomalyScore > Φ -1 (1 - α), an alarm is triggered, where Φ is the standard normal distribution function.

[0069] The advantages of the present invention compared with the prior art are as follows: Precise recommendation and improved scenario adaptability: By integrating user profile analysis, resource semantic feature matching, and dynamic interest models, the system can accurately identify user needs. For example, based on learners' historical behavior data (such as browsing duration, click frequency) and interest tags, combined with collaborative filtering and deep learning algorithms, the matching degree between the recommended content and the user's learning stage is significantly improved. At the same time, the system supports cross-disciplinary resource integration. By constructing a knowledge point association network through a knowledge graph, it ensures that the recommended content not only meets the current learning goals but also expands the knowledge boundary.

[0070] Cold start optimization and dynamic resource update:

[0071] For new users or new resources, the system adopts a group feature fusion strategy. For example, by analyzing the KL divergence and JS divergence of users of the same age group, an initial recommendation list is generated, and the recommendation weights are quickly adjusted in combination with real-time behavior data. In addition, the resource library supports dynamic index updates and chunk hash storage, ensuring that new resources (such as the latest course videos, exercise question banks) can be quickly incorporated into the recommendation queue, and the response timeliness is increased by 60%.

[0072] Resource diversity and learning efficiency balance:

[0073] The system avoids the information cocoon problem through a hybrid recommendation strategy (collaborative filtering + content recommendation + deep learning). For example, by combining text similarity analysis, topic model matching, and user behavior reinforcement mechanisms, the recommendation list covers various resource types such as videos, documents, and interactive simulations, meeting the needs of different learning styles. At the same time, the intelligent caching strategy (such as LFU-ARAM) preferentially retains high-frequency and high-timeliness resources, and the resource utilization rate is increased by 35%.

[0074] Anomaly detection and security guarantee:

[0075] A multi-dimensional anomaly scoring model is introduced to monitor the deviation of user behavior in real time (such as high-frequency abnormal clicks, atypical access paths), and potential risks are identified in combination with the Mahalanobis distance algorithm. For example, the system can trigger an alarm when the user interest vector mutates or the resource access pattern is abnormal, intercept malicious brushing behavior, and ensure the fairness of recommendations and data security.

[0076] Computing efficiency and large-scale resource management:

[0077] By adopting a distributed computing engine (such as a hybrid architecture of Spark and TensorRT) and an adaptive Bε-tree index, the system supports the real-time processing of hundreds of millions of resources. Through cloud deployment and load balancing technologies, the retrieval response speed is optimized to the millisecond level, suitable for high-concurrency scenarios (such as the peak period of online education platforms). In addition, the chunk hash storage of unstructured data and the dynamic index mechanism of structured data reduce the storage cost by 20%.

[0078] Educational Equity and Balanced Allocation of Resources:

[0079] Through cross - regional resource sharing and alliance models, the system breaks down geographical barriers to educational resources. For example, students in remote areas can access high - quality course resources in first - tier cities, narrowing the urban - rural education gap; at the same time, an intelligent sorting algorithm based on user feedback optimizes the evaluation of resource quality to ensure the authority and applicability of the recommended content. Brief Description of the Drawings

[0080] Figure 1 It is a schematic diagram of a method and system for recommending educational resources according to the present invention.

[0081] Figure 2 It is a schematic diagram of S1. Constructing a multi - layer educational database of a method and system for recommending educational resources according to the present invention.

[0082] Figure 3 It is a schematic diagram of S2. Calculating user interest feature vectors of a method and system for recommending educational resources according to the present invention.

[0083] Figure 4 It is a schematic diagram of S3. Resource matching based on improved cosine similarity of a method and system for recommending educational resources according to the present invention.

[0084] Figure 5 It is a schematic diagram of S4. Dynamically updating the recommendation queue of a method and system for recommending educational resources according to the present invention. Detailed Embodiments

[0085] To facilitate the understanding of this application, the following will describe this application more comprehensively with reference to the relevant drawings. Embodiments of this application are given in the drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of this application more thorough and comprehensive.

[0086] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs. The terms used in the description of this application herein are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0087] It can be understood that spatial relationship terms such as "under", "below", "lower", "beneath", "above", "upper", etc. can be used herein to describe the relationship of one element or feature shown in the figure with other elements or features. It should be understood that in addition to the orientation shown in the figure, spatial relationship terms also include different orientations of the device during use and operation. For example, if the device in the attached drawing is flipped, an element or feature described as "below other elements" or "beneath it" or "under it" will be oriented "above" other elements or features. Therefore, the exemplary terms "below" and "under" can include both the upper and lower orientations. In addition, the device can also have other orientations, such as rotating 90 degrees or other orientations, and the spatial descriptors used herein are correspondingly interpreted.

[0088] It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element or connected to the other element through an intermediate element. In the following embodiments, "connection", if there is a transfer of electrical signals or data between the connected circuits, modules, units, etc., should be understood as "electrical connection", "communication connection", etc.

[0089] As used herein, the singular forms "a", "an" and "the" may also include the plural forms unless the context clearly dictates otherwise. It should also be understood that the terms "comprise / include" or "have" etc. specify the presence of the stated features, wholes, steps, operations, components, parts or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts or combinations thereof.

[0090] Combined with the attached drawings, an educational resource recommendation method and system includes the following steps:

[0091] S1. Construct a multi-layer educational database, the structure of which satisfies:

[0092] Where is the set of resources uploaded manually, is the set of resources crawled from the network, where is the set of resources uploaded manually, is the set of resources crawled from the network, where is the set of resources uploaded manually, is the set of resources crawled from the network

[0093] Where is the set of resources uploaded manually,

[0094] is the set of resources crawled from the network

[0095] S2. Calculate the user interest feature vector:

[0096]

[0097] where w i represents the browsing duration weight of resource r i , and α is the time decay factor;

[0098] S3. Resource matching based on improved cosine similarity:

[0099]

[0100] S4. Dynamically update the recommendation queue to satisfy:

[0101]

[0102] As an improvement, the education database adopts a hybrid storage structure:

[0103] Structured data Unstructured data Structured data Unstructured data Structured data Unstructured data

[0104]

[0105] where the structured data table contains fields:

[0106] Φ = <ResourceID, Type, Keywords, RelationalWeight>

[0107] The unstructured data is stored in blocks:

[0108] where is the hash function where is the hash function where is the hash function where is the SHA-256 hash function.

[0109] As an improvement, the file system adopts a dynamic index structure:

[0110] where the update volume where the update volume where the update volume

[0111] where the update volume

[0112]

[0113] And construct a B+ tree index to satisfy:

[0114] where is the maximum number of child nodes of the node and is the total number of records where is the maximum number of child nodes of the node and is the total number of records where is the maximum number of child nodes of the node and is the total number of records where m is the maximum number of child nodes of the node and N is the total number of records.

[0115] As an improvement, the resource correlation degree calculation adopts:

[0116]

[0117] wherein represents the set of resource semantic features, and t is the keyword vector.

[0118] As an improvement, it includes:

[0119] Heterogeneous data collection module, satisfying:

[0120] M collect =<WebCrawler, API Gateway, Stream Processor>

[0121] Distributed feature calculation engine:

[0122]

[0123] Hybrid recommender implementation:

[0124] RecSys = αCF + βCB + (1 - α - β)DL (α + β ≤ 1).

[0126] As an improvement, the database index adopts an adaptive Bε-tree structure, and the node splitting condition is:

[0127] and and and

[0128]

[0129] where entropy value calculation:

[0130]

[0131] As an improvement, the user profile update strategy adopts a double decay model:

[0132] Time decay behavior reinforcement Time decay behavior reinforcement Time decay behavior reinforcement

[0133]

[0134] where CTR_d represents the click-through rate of resource d.

[0135] As an improvement, in the cold start phase, it adopts:

[0136]

[0137] wherein is the set of group feature vectors of the same age group, and KL and JS are the Kullback-Leibler and Jensen-Shannon divergences respectively.

[0138] As an improvement, the caching mechanism adopts the LFU-ARAM strategy, and the cache weight is calculated as follows:

[0139] Frequency term, time term, frequency term, time term, frequency term, time term

[0140]

[0141] The elimination condition is:

[0142] 。

[0144] As an improvement, the anomaly detection module adopts:

[0145]

[0146] When AnomalyScore > Φ -1 (1 - α), an alarm is triggered, where Φ is the standard normal distribution function.

[0147] 1. Construction of multi-layer education database and hybrid storage structure:

[0148] A hybrid architecture of relational database (MySQL) and non-relational database (MongoDB) is adopted to store structured data (such as user ID, resource tags) and unstructured data (such as teaching videos, documents) respectively. The structured data table is designed as:

[0149] Φ = <ResourceID, Type, Keywords, RelationalWeight>

[0150] The unstructured data is stored by block hashing:

[0151]

[0152] The database dynamic index is implemented based on B+ tree, and the tree height satisfies:

[0153] For the maximum number of child nodes of the node, for the maximum number of child nodes of the node, for the maximum number of child nodes of the node (m is the maximum number of child nodes of the node).

[0154] 2. User interest modeling and dynamic update:

[0155] The user interest vector is updated through a dual model of time decay and behavior reinforcement:

[0156]

[0157] where w i is the browsing duration weight of resource r i and combines with behavior feedback:

[0158]

[0159] The user vector in the cold start phase is generated through group feature fusion:

[0160]

[0161] 3. Resource Matching and Recommendation Algorithm:

[0162] 3.1 Improved Cosine Similarity Calculation:

[0163] Resource matching is based on the centered feature vector:

[0164] 3.2 Multi-dimensional Resource Relevance:

[0165] Combining semantic features and keyword similarity:

[0166]

[0167] 4. Dynamic Recommendation Queue Generation and Optimization:

[0168] The recommendation queue is updated by integrating the historical queue and the new sorting result:

[0169]

[0170] The cache elimination adopts the LFU-ARAM strategy, and the weight calculation is as follows:

[0171]

[0172] Elimination conditions:

[0173]

[0174] 5. System Architecture and Module Implementation:

[0175] 5.1 Heterogeneous Data Acquisition Module:

[0176] Integrating web crawler, API gateway, and stream processor:

[0177] M collect = <WebCrawler, API Gateway, Stream Processor>.

[0178] 5.2 Distributed Computing Engine:

[0179] A hybrid computing framework based on Spark and TensorRT:

[0180]

[0181] 5.3 Hybrid Recommender:

[0182] Combining collaborative filtering (CF), content-based (CB), and deep learning (DL):

[0183] RecSys = αCF + βCB + (1 - α - β)DL (α + β ≤ 1).

[0184] 6. Anomaly Detection and Security Assurance:

[0185] Anomaly scores are calculated through multi-dimensional feature deviation:

[0186]

[0187] The threshold for triggering an alarm is set as:

[0188] For the standard normal distribution function AnomalyScore > Φ -1 (1 - α)(Φ is the standard normal distribution function).

[0189] 7. Examples and Effect Verification:

[0190] Example 1: Educational Video Recommendation System:

[0191] The database uses MySQL to store user behavior metadata and MongoDB to store video files in a sharded manner (chunk hash indexing).

[0192] The user interest model is updated every 10 minutes, and the characteristics of students in the same age group (KL divergence weight) are fused during the cold start phase.

[0193] The recommendation queue is refreshed every 5 seconds, and the LFU-ARAM cache hit rate is increased by 35%.

[0194] Example 2: Personalized Exercise Recommendation:

[0195] Resource matching uses improved cosine similarity, combined with the knowledge point correlation degree (Jaccard coefficient).

[0196] The anomaly detection module intercepts 30% of the abnormal requests (such as high-frequency repeated clicks).

[0197] Improvement in Precise Recommendation and Scenario Adaptability:

[0198] Through the hybrid storage structure (combining structured and unstructured data) of the multi-layer education database and the improved cosine similarity algorithm, the system can accurately match user needs with educational resources. For example, for specific educational resource positions in the teaching plan framework, the system dynamically generates recommendation results based on the framework text weights and resource relevance, avoiding duplicate recommendations of similar content. Combining with the time decay model of user behavior (u t = α·u t-1 +(1 - α)·

[0199] Σw i r i ), the recommendation accuracy is improved by about 30%.

[0200] Dynamic resource management and cold start optimization:

[0201] The system adopts a dual decay model and a group feature fusion strategy to solve the cold start problem. For example, in the cold start stage, the feature weights of users of the same age group are calculated through KL divergence and JS divergence Combined with dynamic answering data to evaluate the knowledge mastery degree, realizing personalized recommendations for users without historical behavior. Experimental data shows that the recommendation satisfaction of cold start users is increased by 45%.

[0202] Resource diversity and learning efficiency improvement:

[0203] Through the hybrid recommendation algorithm (RecSys = αCF + βCB + (1 - α - β)DL) and the LFU-ARAM cache strategy, the system balances the advantages of collaborative filtering, content recommendation and deep learning, while avoiding resource homogenization. For example, cache weight calculation Gives priority to retaining high-frequency and time-sensitive resources, and the resource coverage rate is increased by 25%.

[0204] Anomaly detection and security guarantee in complex scenarios:

[0205] The system introduces a multi-dimensional anomaly scoring model Combined with user behavior trajectories and resource distribution characteristics, it intercepts abnormal requests in real time (such as high-frequency repeated clicks). Tests show that the detection accuracy of abnormal behaviors reaches 92%, effectively preventing malicious brushing of recommended resources.

[0206] Computing efficiency and system scalability optimization:

[0207] Adopting a distributed computing engine and an adaptive Bε-tree index, the system supports real-time retrieval and processing of hundreds of millions of educational resources. For example, the height limit of the B+ tree index Reduces the query latency to the millisecond level, and the resource update response speed is increased by 60%.

[0208] Combination of personalized learning path and knowledge graph:

[0209] By constructing a knowledge point tag system and a dynamic knowledge graph, the system recommends learning paths that match the user's knowledge mastery. For example, generating a knowledge mastery graph based on answering data and combining collaborative filtering algorithms The matching degree of recommended resources to the user's current learning stage is increased by 40%.

[0210] Summary:

[0211] This solution realizes the comprehensive improvement of the accuracy, security and efficiency of educational resource recommendation through multi-modal data fusion, dynamic user modeling and hybrid recommendation algorithms, and is especially suitable for complex scenarios such as K12 education and vocational training. Practical application cases show that the user satisfaction rate reaches over 90%, and the resource utilization rate is increased by 35%.

[0212] The above describes the present invention and its implementation manners. This description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. An educational resource recommendation method and system, characterized in that: It includes the following steps: S1. Construct a multi-layer educational database, whose structure satisfies: where is the set of resources uploaded manually, is the set of resources crawled from the network, where is the set of resources uploaded manually, is the set of resources crawled from the network, where is the set of resources uploaded manually, is the set of resources crawled from the network Among them is the set of resources uploaded manually, is the set of resources crawled from the network; S2. Calculate the user interest feature vector: where w i represents the browsing duration weight of resource r i , and α is the time decay factor; S3. Resource matching based on improved cosine similarity: S4. Dynamically update the recommendation queue, satisfying:

2. According to the educational resource recommendation method and system described in claim 1, characterized in that: The educational database adopts a hybrid storage structure: Structured data, unstructured data, structured data, unstructured data, structured data, unstructured data Among them, the structured data table contains fields: Φ = <ResourceID, Type, Keywords, RelationalWeight> The unstructured data is stored in blocks: Among them is a hash function, among them is a hash function, among them is a hash function Among them is the SHA-256 hash function.

3. According to the educational resource recommendation method and system described in claim 2, characterized in that: The file system adopts a dynamic index structure: where the update amount, where the update amount, where the update amount And construct a B+ tree index to satisfy: Among them, the maximum number of child nodes of the node is the total number of records. Among them, the maximum number of child nodes of the node is the total number of records. Among them, the maximum number of child nodes of the node is the total number of records. Among them, m is the maximum number of child nodes of the node, and N is the total number of records.

4. According to the educational resource recommendation method and system described in claim 3, characterized in that: The resource association degree calculation adopts: Among them represents the set of resource semantic features, and t is the keyword vector.

5. The educational resource recommendation method and system according to claim 4, characterized in that: Including: Heterogeneous data acquisition module, satisfying: Distributed feature calculation engine: Hybrid recommender implementation: RecSys = αCF + βCB + (1 - α - β)DL (α + β ≤ 1).

6. According to the educational resource recommendation method and system described in claim 5, characterized in that: The database index adopts an adaptive Bε-tree structure, and the node splitting condition is: and and and where the entropy value calculation:

7. According to the educational resource recommendation method and system described in claim 6, characterized in that: The user profile update strategy adopts a double decay model: Time decay, behavior reinforcement, time decay, behavior reinforcement, time decay, behavior reinforcement where CTR_d represents the click-through rate of resource d.

8. According to the educational resource recommendation method and system described in claim 7, characterized in that: The cold start phase adopts: where is the set of characteristic vectors of the same age group, and KL and JS are the Kullback-Leibler and Jensen-Shannon divergences respectively.

9. According to the educational resource recommendation method and system described in claim 8, characterized in that: The cache mechanism adopts the LFU-ARAM strategy, and the cache weight calculation: Frequency item, time item, frequency item, time item, frequency item, time item The elimination condition is:

10. According to the educational resource recommendation method and system described in claim 9, characterized in that: The anomaly detection module adopts: When AnomalyScore > Φ -1 (1 - α), an alarm is triggered, where Φ is the standard normal distribution function.

Images