An Education Resource Diversified Recommendation Method and System Based on Long Short-Term Memory Neural Network

Abstract

The present invention discloses a method and system for diversified recommendation of educational resources based on long short-term memory neural network, which relates to the technical field of educational resource recommendation. The method includes: S1. Obtaining the historical learning resource information of learners and dividing the learners; S2. Based on the division result, enhancing learner representation and modeling knowledge point mastery; S3. Generating a list of candidate educational resources based on the degree of knowledge point mastery of learners; S4. Calculating the relevance and diversity scores of educational resources based on the list of candidate educational resources; S5. Fusing the relevance and diversity scores of educational resources and recommending a specified number of educational resources to learners in descending order of the final scores. The present invention proposes a sequence enhancement module, which enhances the learning representation of inactive learners by combining the historical learning behaviors of inactive learners and the rich historical data of active learners.

Description

Technical Field

[0001] The present invention relates to the technical field of educational resource recommendation, and particularly relates to a method and system for diversified recommendation of educational resources based on a long short-term memory neural network. Background Art

[0002] With the development of information technology and the popularization of online education, personalized learning has gradually become the research focus in the field of educational technology. Personalized learning dynamically adjusts teaching content to meet learners' abilities, interests, and learning goals, thereby significantly improving learning efficiency. In this context, educational resource recommendation systems have emerged, aiming to use artificial intelligence technology to recommend educational resources that match learners' learning needs to optimize the learning path and enhance learning effects.

[0003] Current educational resource recommendation systems are mainly divided into three categories: knowledge-tracing-based, multi-stage-based, and reinforcement learning-based methods. Knowledge-tracing-based methods recommend by analyzing learners' mastery of knowledge points, but often ignore the diversity of recommended content and are difficult to fully meet learners' needs. Multi-stage-based methods attempt to combine knowledge tracing and neural networks to more efficiently predict learners' knowledge states and recommend corresponding educational resources; some methods use knowledge graphs to capture the complex relationships among educational resources, knowledge points, and learners. However, due to the implicit and variable nature of learners' learning states, these methods rely on a large amount of historical data, and this limitation further magnifies the difficulty of recommending for inactive learners because the learning sequences of such learners are usually short and the content is single, making it difficult for the model to fully capture their learning needs. Reinforcement learning-based methods dynamically generate recommendations through multi-objective optimization, but tend to converge to local optima, resulting in homogeneous recommendations. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for diversified recommendation of educational resources based on a long short-term memory neural network to solve the problems that existing educational resource recommendation technologies tend to overfit active learners, ignore most inactive learners, resulting in a decline in the overall performance of learners' ability assessment, and do not consider learners' actual needs for diversity.

[0005] To achieve the above purpose, the present invention provides a method for diversified recommendation of educational resources based on a long short-term memory neural network, and the steps include:

[0006] S1. Obtain the historical learning resource information of learners and divide the learners;

[0007] S2. Based on the division result, perform learner representation enhancement and knowledge point mastery modeling;

[0008] S3. Generate a list of candidate educational resources based on the learner's knowledge point mastery level;

[0009] S4. Calculate the relevance and diversity scores of educational resources based on the list of candidate educational resources;

[0010] S5. Integrate the relevance of educational resources and the diversity score, and recommend a specified number of educational resources to learners in descending order of the final score.

[0011] Preferably, the S1 includes:

[0012] Obtain the historical learning resource information of learners through an online learning platform, form a sequence for each learner's learning resource information in chronological order, and preprocess the sequence to form an initial data set;

[0013] Among them, learners with the top 5% of educational resource interaction quantities are classified as active learners, and the latter 95% of learners are classified as inactive learners.

[0014] Preferably, the historical learning resource information includes: learner id, serial number of the educational resource learned by the learner, learning time, learning answer result, and knowledge points included in the educational resource; the method of forming the initial data set includes: forming a learning resource sequence, a knowledge point sequence, a learning resource result sequence, and a learning resource time interval sequence with a length of 200 educational resource interactions for each question answered by the user, the knowledge points examined by the question, the answer result, and the answer time in chronological order, and grouping according to the learner id; finally, divide the data, and randomly divide the data into a training set and a test set according to a ratio of 8:2.

[0015] Preferably, the S2 includes:

[0016] Use the sequence enhancement method to enhance the learner representation, and utilize the historical learning sequence information of active learners. The steps include: intercepting several subsequences from the sequence of the head user, encoding the subsequences through a sequence encoder, and training an enhancer to continuously make the representation of the subsequences approach the representation of the complete user;

[0017] Concatenate the enhanced learner representation with the knowledge point embeddings extracted from the historical learning resources, input them into the matrix long short-term memory network mLstm to model the learner's knowledge point mastery status, and obtain the learner's mastery level of all knowledge points.

[0018] Preferably, a matrix long short-term memory network mLstm is used as the sequence encoder; during the training of the enhancer, a curriculum learning strategy is adopted to control the weight of each user in the loss according to the difficulty of learning knowledge from active learners; after the enhancer training is completed, the sequence representations of inactive learners are passed through the enhancer to obtain the enhanced representation information of inactive learners through the sequence.

[0019] Preferably, the S3 includes: calculating the difficulty corresponding to each educational resource based on the knowledge point mastery degree of the learners, and selecting several educational resources from the educational resource library as the candidate educational resource list for each learner.

[0020] Preferably, the S4 includes:

[0021] Obtain the embeddings of each learner's user characteristics, educational resource characteristics, and knowledge point characteristics from the candidate educational resource list of each learner, input them into a bidirectional long short-term memory network to learn context information, so as to obtain the relevance score of each educational resource for the overall learner;

[0022] Classify the candidate educational resource list of each learner according to knowledge points, obtain the embeddings of knowledge points, educational resources, and learners, input them into the learning diversity estimator to obtain the learning demand distribution of each learner, and calculate the marginal diversity of educational resources. Multiply the element-by-element difference between the learning demand distribution of the learner and the diversity of educational resources to obtain the diversity score of educational resources.

[0023] Preferably, the learning diversity estimator encodes the learner practice sequences of each knowledge point using a long short-term memory network respectively to obtain the learner learning mode representation of each knowledge point; then uses an attention mechanism and a multi-layer perceptron to learn the learning demand distribution of the learner for different knowledge points; finally uses a probability coverage function to calculate the diversity gain of each practice.

[0024] Preferably, the S5 includes: fusing the obtained educational resource relevance and the diversity score to obtain the re-ranking score of each candidate set, and recommending educational resources to learners in descending order according to the final score of educational resources.

[0025] The present invention also provides an educational resource diversified recommendation system based on a long short-term memory neural network. The system is used to implement the above method and includes: an acquisition module, a construction module, a generation module, a calculation module, and a fusion module;

[0026] The acquisition module is used to obtain the historical learning resource information of learners and divide the learners;

[0027] The construction module is used to perform learner representation enhancement and knowledge point mastery modeling based on the division result;

[0028] The generation module is used to generate a list of candidate educational resources based on the learner's knowledge point mastery level;

[0029] The calculation module is used to calculate the educational resource relevance and diversity scores based on the list of candidate educational resources;

[0030] The fusion module is used to fuse the educational resource relevance and the diversity scores, and recommend a specified number of educational resources to the learner in descending order of the final scores.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] (1) Enhanced representation for inactive learners: A major problem with existing educational resource recommendation systems is that they often focus on the data of active learners and ignore the needs of inactive learners. Existing methods rely on a large amount of historical data to train the model. Therefore, the recommendations for inactive learners often have poor effects because their learning sequences are short and single. In contrast, this application proposes a sequence enhancement module to enhance the learning representation of inactive learners by combining the historical learning behaviors of inactive learners and the rich historical data of active learners. This method not only effectively fills in the learning behavior characteristics of inactive learners but also avoids the problem that they cannot obtain accurate recommendations due to data sparsity, thus improving the adaptability of the recommendation system to inactive learners.

[0033] (2) Multi-stage screening and optimization of candidate educational resources: Existing educational resource recommendation systems usually directly select highly relevant educational resources from the list of candidate educational resources without fully considering the difficulty of the educational resources and the distribution of learners' learning needs. In contrast, this application conducts multi-stage screening. First, it screens the candidate educational resources based on the learner's knowledge point mastery level, and then further optimizes them according to the learner's learning needs and the diversity of the educational resources. This method improves the quality of the recommendation results through precise selection of candidate educational resources and diversity optimization, ensuring that the educational resources are not only of moderate difficulty but also have sufficient diversity and challenges.

[0034] (3) Balance between Relevance and Diversity: Traditional educational resource recommendation methods usually only focus on the relevance of educational resources, which easily leads to the homogenization of recommendation results and lacks attention to the diversity of learners' needs. Especially in methods based on reinforcement learning, the system is prone to converge to a local optimum, ignoring learners' needs for diverse and extensive educational resources. In contrast, this application considers both the relevance and diversity of educational resources during the recommendation process through a relevance estimator and a diversity estimator, followed by neural rearrangement. This method dynamically adjusts the recommendation results by calculating the diversity scores of educational resources and combining the learning need distribution of learners, ensuring that the recommendations not only conform to the learners' knowledge point mastery but also meet their need for diversity, thus avoiding the situation of homogeneous recommendations. Brief Description of the Drawings

[0035] To more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a schematic flowchart of the method according to an embodiment of the present invention;

[0037] Figure 2 It is a schematic diagram for generating a list of candidate educational resources according to an embodiment of the present invention;

[0038] Figure 3 It is a schematic diagram for generating a recommended list of educational resources according to an embodiment of the present invention. Detailed Embodiments

[0039] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0040] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] As Figure 1 shown, it is a schematic flowchart of the method in this embodiment, and the steps include:

[0042] S1. Obtain the historical learning resource information of the learner and divide the learners.

[0043] Obtain the historical learning resource information of learners from the online learning platform, form a sequence of the learning resource information of each learner in chronological order, and perform specific preprocessing on the sequence. The preprocessing process includes multiple steps such as data statistics, data cleaning, grouping processing, and serialization processing. Specifically, first analyze the original data, count key indicators such as the number of interactions, the number of learners, the number of educational resources, and the number of knowledge points, and at the same time calculate the average number of interactions of each learner, the average number of knowledge points covered by each educational resource, and the number of missing values. In the data cleaning stage, delete the missing values and outliers in key fields such as learner id, educational resource id, and knowledge point id. Immediately afterwards, group the data by learner id and construct an independent learning record set for each learner. In the serialization processing stage, arrange the learning resources, knowledge points, answer results, etc. of each learner in chronological order and unify them into a sequence with a fixed length of 200 for easy model processing. Subsequently, calculate the maximum number of knowledge points in each learner's interaction, and map learner id, educational resource id, knowledge point id, etc. to integer indexes to simplify data representation and form an initial data set.

[0044] In this embodiment, the historical learning resource information includes: learner id, the serial number of the educational resource learned by the learner, learning time, learning answer result, and the knowledge points included in the educational resource; the method for forming the initial data set includes: arranging the questions answered by each user, the knowledge points examined by the questions, the answer results, and the answer time in chronological order to form a learning resource sequence, a knowledge point sequence, a learning resource result sequence, and a learning resource time interval sequence with 200 educational resource interactions in length, and grouping according to the learner id; finally, divide the data, and randomly divide the data into a training set and a test set according to a ratio of 8:2.

[0045] In this embodiment, learners with educational resource interaction quantities in the top 5% are classified as active learners, and the latter 95% of learners are classified as inactive learners.

[0046] S2. Based on the division results, perform learner representation enhancement and knowledge point mastery modeling.

[0047] Through the sequence enhancement method (implemented by constructing a sequence enhancement module in this embodiment, the structure of which is as Figure 2 shown), utilize the rich historical learning sequence information of active learners to enhance the representation of inactive learners; among them, the sequence enhancement method refers to intercepting a part of the subsequence from the sequence of the head user, encoding the subsequence through a sequence encoder, and training an enhancer to make the representation of the subsequence continuously approach the representation of the complete user.

[0048] Specifically, from the historical learning sequence P of active learners sThe sequence containing the I most recent interactions is intercepted as a subsequence ,Will As the educational resource interaction sequence of inactive learners, it is intended to simulate the situation where inactive learners have insufficient interaction. This embodiment uses mLstm as the sequence encoder to learn to generate active learners P s Sequence representation of h s , and the subsequence Representation , capturing the learner’s recent learning needs. In order to transfer the knowledge learned by the model from the active learner sequence to the inactive learner, the goal of this embodiment is to generate a complete representation close to the active learner by minimizing the following loss:

[0049] ,

[0050] in, Represents the learner representation enhancer, which is responsible for Outputs a complete sequence representation.

[0051] Since knowledge is transferred from active learners to inactive learners, the performance of the model depends on the quality of the active learner representations, which further depends on the number of relevant interactions among them. Specifically, this embodiment adopts a curriculum learning strategy to gradually train the model from simple to difficult:

[0052] ,

[0053] in represents the loss coefficient of an active learner s; Indicates the current training round; Indicates the maximum round; and Respectively represent the training data The maximum and minimum values of .

[0054] In the early stages of training (i.e. ), the model learns more from active learners with longer sequence lengths than from active learners with shorter sequence lengths. As training progresses, the model starts to learn more from active learners with shorter sequence lengths than from active learners with longer sequence lengths. is a learner representation enhancer based on the input Generate a complete learner representation. By minimizing the loss function ,train Contains enough knowledge to generate complete representations from short practice sequences of learners, and finally acquires inactive learners through sequence enhancers Enhanced representation:

[0055] ,

[0056] wherein represents inactive learners Enhanced representation; represents the representation of inactive learners obtained based on the entire educational resource sequence of the learner ; is a hyperparameter used to control the contribution of the original representation of inactive learners. In addition to the original representation in addition, representations generated by considering the knowledge obtained from active learners through the sequence characterization enhancer are also considered, so as to supplement and enhance the low-quality representations of inactive learners caused by the lack of educational resource interaction. After the enhancer training is completed, the sequence representation of the inactive learner is passed through the enhancer to obtain the representation information of the inactive learner after sequence enhancement , which is concatenated with the knowledge point embeddings extracted from the learner's historical learning resources

[0057] to obtain the learner's feature embedding , which is input into the matrix long short-term memory network mLstm to model the learner's knowledge point mastery status, so as to obtain the learner's mastery degree of all knowledge points:

[0058]

[0059] ,

[0060] Then the cost function L of the optimized mLstm module K can be expressed as:

[0061] , where N represents the number of learners; T i represents the time period of educational resources before the i-th learner answers; represents the one-hot encoding of the knowledge point answered at time t+1, and r t+1 represents the corresponding answer; y t then represents the output corresponding to the mLstm module at time t. Finally, the loss function for training the knowledge point mastery degree prediction module is:

[0062] ,

[0063] where λ represents the hyperparameter; S H represents the set of active learners.

[0064] S3. Generate a list of candidate educational resources based on the learner's mastery of knowledge points.

[0065] As Figure 2 shown, based on the learner's mastery of each knowledge point, calculate the difficulty corresponding to each educational resource, and select N educational resources from the educational resource library as the candidate educational resource list for each learner.

[0066] The specific steps include: after the knowledge point mastery prediction module is trained, use it to predict the learner's mastery of different knowledge points, and then derive the difficulty of each exercise for different learners according to the following formula:

[0067] ,

[0068] where represents the difficulty of educational resource ; represents the learner's mastery of the i-th knowledge point.

[0069] Finally, select the candidate educational resource list Ω according to the difficulty of the educational resources, and select 150 candidate educational resources for each learner.

[0070] S4. Calculate the relevance and diversity scores of educational resources based on the candidate educational resource list.

[0071] As Figure 3 shown, first, obtain the embeddings of each learner's user characteristics, educational resource characteristics, and knowledge point characteristics from the candidate educational resource list of each learner, and input them into a bidirectional long short-term memory network to learn the context information, so as to obtain the relevance score of each educational resource for the overall learner.

[0072] Specifically, let represent the l-th (l = 1, 2,..., L) educational resource in the candidate educational resource list (candidate set) Ω. The forward output state of the l-th educational resource can be obtained through the bidirectional LSTM network as , and the reverse output state is . Finally, the relevance of educational resource is expressed as the connection of the forward and reverse output states . Then stack the relevance representation vectors of each educational resource to obtain the relevance representation matrix R C , so as to model the relationship between each candidate educational resource and each knowledge point in the candidate subset for the learner.

[0073] After that, classify the candidate educational resource list of each learner according to knowledge points, obtain the embeddings of knowledge points, educational resources and learners, input them into the learning diversity estimator to obtain the learning demand distribution of each learner, and calculate the marginal diversity of educational resources. Multiply the element-by-element difference between the learning demand distribution of the learner and the diversity of educational resources to obtain the diversity score of educational resources.

[0074] The above learning diversity estimator encodes the learner exercise sequence of each knowledge point using a long short-term memory network to obtain the learner learning pattern representation of each knowledge point; then uses an attention mechanism and a multi-layer perceptron to learn the learning demand distribution of the learner for different knowledge points; finally, uses a probability coverage function to calculate the diversity gain of each exercise.

[0075] Specifically, classify the candidate educational resource list of each learner obtained according to knowledge points, obtain the embeddings of knowledge points, educational resources and learners, and combine LSTM to explicitly encode the time dependence recorded by the learner in each knowledge point to model the interaction within the knowledge point. The final output state of Lstm can be regarded as the encoding of all information contained in the sequence, which has the greatest impact on future predictions. After encoding the learning patterns of the learner for each knowledge point separately, the signals specific to the knowledge points are aggregated, and the learning demand distribution of the learner on the knowledge points is obtained through a self-attention mechanism. Then, all the knowledge points representation vectors are superimposed to obtain matrix W. Then use a multi-layer perceptron to generate the personalized learning demand distribution ω for knowledge points:

[0076] ,

[0077] where ω represents the possibility that the learner is interested in each knowledge point; represents the self-attention network.

[0078] Given a candidate subset, the diversity of educational resources depends on the differences or novelty between the current educational resource and other educational resources in the candidate subset. Therefore, use the probability coverage function as the diversity function, and the specific formula is as follows:

[0079] ,

[0080] where, represents the knowledge point coverage of educational resources; L represents the number of educational resources; represents the probability that at least one educational resource in the candidate list Ω covers knowledge point k.

[0081] In addition, this embodiment defines the marginal diversity of educational resources as , that is, whether there is educational resources in the candidate list in terms of the diversity difference. Thus, for each educational resource in the candidate list Ω , the diversity representation vector for l = 1, 2,..., L can be obtained .

[0082] ,

[0083] Among them, represents element-wise multiplication. Therefore, the k-th element of represents the diversity gain of educational resources in the knowledge points.

[0084] S5. Integrate the relevance and diversity scores of educational resources, and recommend a specified number of educational resources to learners in descending order of the final scores.

[0085] The obtained relevance of the educational resources and the diversity scores are subjected to feature fusion to obtain the re-ranking scores of each candidate set, and educational resources are recommended to learners in descending order of the final scores of the educational resources.

[0086] By connecting the correlation matrix and the diversity matrix, and then using MLP to fuse the correlation and diversity to predict the scores:

[0087] ,

[0088] Among them, represents the re-ranking score matrix of educational resources.

[0089] Based on the relevance scores and diversity scores of each educational resource, the re-ranking scores of each educational resource for each learner are obtained, and the educational resources are sorted according to the re-ranking scores; finally, according to the recommended list length K of educational resources, the top K sorted educational resources are recommended to learners.

[0090] Embodiment 2

[0091] This embodiment also provides an educational resource diversified recommendation system based on a long short-term memory neural network, including: an acquisition module, a construction module, a generation module, a calculation module, and a fusion module; the acquisition module is used to obtain the historical learning resource information of learners and divide the learners; the construction module is used to perform learner representation enhancement and knowledge point mastery modeling based on the division results; the generation module is used to generate a candidate educational resource list based on the learner's knowledge point mastery level; the calculation module is used to calculate the relevance and diversity scores of educational resources based on the candidate educational resource list; the fusion module is used to fuse the relevance and diversity scores of educational resources, and recommend a specified number of educational resources to learners in descending order of the final scores.

[0092] Next, in combination with this embodiment, how the present invention solves technical problems in real life will be described in detail.

[0093] First, use the acquisition module to obtain the historical learning resource information of the learner and divide the learners.

[0094] Obtain the historical learning resource information of the learner from the online learning platform, form a sequence of the learning resource information of each learner in chronological order, and perform specific preprocessing on the sequence. The preprocessing process includes multiple steps such as data statistics, data cleaning, grouping processing, and serialization processing. Specifically, first analyze the original data, count key indicators such as the number of interactions, the number of learners, the number of educational resources, and the number of knowledge points, and at the same time calculate the average number of interactions of each learner, the average number of knowledge points covered by each educational resource, and the number of missing values. In the data cleaning stage, delete the missing values and outliers in the key fields such as learner id, educational resource id, and knowledge point id. Immediately afterwards, group the data by learner id and construct an independent learning record set for each learner. In the serialization processing stage, arrange the learning resources, knowledge points, answer results, etc. of each learner in chronological order and unify them into a sequence with a fixed length of 200 for easy model processing. Subsequently, calculate the maximum number of knowledge points in each learner's interaction, and map the learner id, educational resource id, knowledge point id, etc. to integer indexes to simplify data representation and form an initial data set.

[0095] In this embodiment, the historical learning resource information includes: learner id, serial number of the educational resource learned by the learner, learning time, learning answer result, and knowledge points included in the educational resource; the method for forming the initial data set includes: arranging the questions answered by each user, the knowledge points examined by the questions, the answer results, and the answer time in chronological order to form a learning resource sequence, a knowledge point sequence, a learning resource result sequence, and a learning resource time interval sequence with 200 educational resource interactions in length, and grouping according to the learner id; finally, divide the data, and randomly divide the data into a training set and a test set according to a ratio of 8:2.

[0096] In this embodiment, learners with educational resource interaction quantities in the top 5% are classified as active learners, and the latter 95% of learners are classified as inactive learners.

[0097] After that, use the construction module to perform learner representation enhancement and knowledge point mastery modeling based on the division results.

[0098] Through the sequence enhancement method (this embodiment is implemented by constructing a sequence enhancement module, and its structure is as Figure 2As shown in [figure], the representation of inactive learners is enhanced by leveraging the rich historical learning sequence information of active learners. Among them, the sequence enhancement method refers to intercepting a part of the subsequence from the sequence of the head user, encoding the subsequence through a sequence encoder, and training an enhancer to continuously approximate the representation of the complete user.

[0099] Specifically, from the historical learning sequence P of active learners s Intercept the sequence containing the I most recent interactions as a subsequence , and use As the educational resource interaction sequence of inactive learners, aiming to simulate the situation of insufficient interaction of inactive learners. In this embodiment, mLstm is used as the sequence encoder to learn and generate the sequence representation h of active learner P s , as well as the representation of the subsequence s , to capture the most recent learning needs of learners. In order to transfer the knowledge learned by the model from the active learner sequence to the inactive learner, the goal of this embodiment is to generate a complete representation close to the active learner by minimizing the following loss: representation , capturing the learner's most recent learning needs. In order to transfer the knowledge learned by the model from the active learner sequence to the inactive learner, the goal of this embodiment is to generate a complete representation close to the active learner by minimizing the following loss:

[0100] ,

[0101] Among them, Represents the learner representation enhancer, which is responsible for outputting the complete sequence representation according to the input .

[0102] Since knowledge is transferred from active learners to inactive learners, the performance of the model depends on the quality of the active learner representation, which further depends on the number of relevant interactions therein. Specifically, this embodiment adopts a curriculum learning strategy to train the model step by step from simple to difficult:

[0103] ,

[0104] Among them Represents the loss coefficient of an active learner s; Represents the current training round; Represents the maximum number of rounds; And Respectively represent the maximum and minimum values of In the training data.

[0105] In the early stage of training (i.e., ) The model learns more from active learners with longer sequence lengths than from those with shorter sequence lengths. As training progresses, the model begins to learn more from active learners with shorter sequence lengths than from those with longer sequence lengths. is a learner representation enhancer based on the input to generate a complete learner representation. By minimizing the loss function , train to contain sufficient knowledge to generate a complete representation based on the short practice sequences of the learners, and finally obtain the inactive learners through the sequence enhancer Enhanced representation:

[0106] ,

[0107] where represents the inactive learner Enhanced representation; represents the representation of the inactive learner obtained based on the entire educational resource sequence of the learner ; is a hyperparameter used to control the contribution of the original representation of the inactive learner . In addition to the original representation , representations generated by considering the knowledge obtained from active learners through the sequence representation enhancer are also considered, so that the low-quality representations of inactive learners caused by the lack of educational resource interaction are supplemented and enhanced.

[0108] After the enhancer training is completed, the sequence representation of the inactive learner is passed through the enhancer to obtain the representation information of the inactive learner after sequence enhancement , which is concatenated with the knowledge point embeddings extracted from the learner's historical learning resources to obtain the feature embedding of the learner , which is input into the matrix long short-term memory network mLstm to model the learner's knowledge point mastery status, so as to obtain the learner's mastery degree of all knowledge points:

[0109] ,

[0110] Then the cost function L of the optimized mLstm module K can be expressed as:

[0111] ,

[0112] where N represents the number of learners; T i represents the time period of the educational resources before the i-th learner answers; One-hot encoding of the knowledge points answered at time t + 1, and use r t+1 to represent the corresponding answer; y t represents the output corresponding to the mLstm module at time t. Finally, the loss function for training the knowledge point mastery prediction module is:

[0113] ,

[0114] where λ represents the hyperparameter; S H represents the set of active learners.

[0115] The generation module generates a list of candidate educational resources based on the learner's knowledge point mastery.

[0116] As Figure 2 shown, based on the learner's mastery of each knowledge point, calculate the difficulty corresponding to each educational resource, and select N educational resources from the educational resource library as the candidate educational resource list for each learner.

[0117] The specific steps include: after the knowledge point mastery prediction module is trained, use it to predict the learner's mastery of different knowledge points, and then derive the difficulty of each exercise for different learners according to the following formula:

[0118] ,

[0119] where represents the difficulty of the educational resource ; represents the learner's mastery of the i-th knowledge point.

[0120] Finally, screen out the candidate educational resource list Ω according to the difficulty of the educational resources, and select 150 candidate educational resources for each learner.

[0121] The calculation module calculates the educational resource relevance and diversity scores based on the candidate educational resource list.

[0122] As Figure 3 shown, first, obtain the embeddings of each learner's user characteristics, educational resource characteristics, and knowledge point characteristics from the candidate educational resource list of each learner, input them into the bidirectional long short-term memory network to learn the context information, and thus obtain the relevance score of each educational resource for the overall learner.

[0123] Specifically, let represent the l-th (l = 1, 2,..., L) educational resource in the candidate educational resource list (candidate set) Ω. The forward output state of the l-th educational resource can be obtained through the bidirectional LSTM network as , the reverse output state is . Finally, the relevance of educational resources is expressed as the connection between the forward and reverse output states . Then stack the relevance representation vectors of each educational resource to obtain the relevance representation matrix R C , thereby modeling the relationship between each candidate educational resource and each knowledge point in the candidate subset for the learner.

[0124] After that, classify the candidate educational resource list of each learner according to knowledge points, obtain the embeddings of knowledge points, educational resources and learners, input them into the learning diversity estimator to obtain the learning demand distribution of each learner, and calculate the marginal diversity of educational resources. Multiply the element-by-element difference between the learning demand distribution of the learner and the diversity of educational resources to obtain the diversity score of educational resources.

[0125] The above learning diversity estimator encodes the learner practice sequences of each knowledge point using a long short-term memory network to obtain the learner learning pattern representation of each knowledge point; then uses an attention mechanism and a multi-layer perceptron to learn the learning demand distribution of the learner for different knowledge points; finally uses a probability coverage function to calculate the diversity gain of each practice.

[0126] Specifically, classify the candidate educational resource list of each learner obtained according to knowledge points, obtain the embeddings of knowledge points, educational resources and learners, and combine LSTM to explicitly encode the time-dependent relationship recorded by the learner in each knowledge point to model the interaction within the knowledge point. The final output state of Lstm can be regarded as the encoding of all information contained in the sequence, which has the greatest impact on future predictions. After encoding the learning patterns of the learner for each knowledge point separately, aggregate the signals specific to the knowledge point, and obtain the learning demand distribution of the learner on the knowledge point through a self-attention mechanism. Then, stack all the representation vectors to obtain the matrix W. Then use a multi-layer perceptron to generate the personalized learning demand distribution ω for knowledge points:

[0127] ,

[0128] where ω represents the possibility that the learner is interested in each knowledge point; represents the self-attention network.

[0129] Given a candidate subset, the diversity of educational resources depends on the differences or novelty between the current educational resource and other educational resources in the candidate subset. Therefore, use the probability coverage function As a diversity function, the specific formula is as follows:

[0130] ,

[0131] where, represents the knowledge point coverage of educational resources; L represents the number of educational resources; represents the probability that at least one educational resource in the candidate list Ω covers the knowledge point k.

[0132] In addition, the marginal diversity of educational resources is defined in this embodiment as , that is, the diversity difference when the educational resource appears or not in the candidate list. Thus, the diversity representation vector of each educational resource in the candidate list Ω, l = 1, 2,..., L can be obtained.

[0133] ,

[0134] where, represents the element-wise product. Therefore, the k-th element of represents the diversity gain of educational resources in the knowledge points.

[0135] Finally, the fusion module fuses the educational resource relevance and diversity scores, and recommends a specified number of educational resources to the learner in descending order of the final scores.

[0136] The obtained educational resource relevance and the diversity scores are subjected to feature fusion to obtain the re-ranking scores of each candidate set, and educational resources are recommended to the learner in descending order of the final scores of the educational resources.

[0137] By connecting the correlation matrix and the diversity matrix, and then using MLP to fuse the relevance and diversity to predict the scores:

[0138] ,

[0139] where, represents the re-ranking score matrix of educational resources.

[0140] Based on the relevance scores and diversity scores of each educational resource, the re-ranking scores of each educational resource for each learner are obtained, and the educational resources are sorted according to the re-ranking scores; finally, according to the recommended list length K of educational resources, the top K educational resources in the ranking are recommended to the learner.

[0141] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for diversifying the recommendation of educational resources based on a long short-term memory neural network, characterized by the steps Including: S1. Obtain the historical learning resource information of learners and divide the learners; S2. Based on the division results, perform learner representation enhancement and knowledge point mastery modeling; S3. Generate a list of candidate educational resources based on the degree of learners' knowledge point mastery; S4. Based on the list of candidate educational resources, calculate the relevance and diversity scores of educational resources; S5. Integrate the relevance of the educational resources and the diversity scores, and recommend a specified number of educational resources to learners in descending order of the final scores; The S1 includes: obtaining the historical learning resource information of learners through an online learning platform, forming a sequence of each learner's learning resource information in chronological order, and preprocessing the sequence to form an initial data set; among them, learners with educational resource interaction quantities in the top 5% are divided into active learners, and the latter 95% of learners are divided into inactive learners; The S2 includes: using a sequence enhancement method to enhance the learner representation, and using the historical learning sequence information of active learners. The steps include: intercepting several subsequences from the sequence of the head user, encoding the subsequences through a sequence encoder, and training an enhancer to continuously make the representation of the subsequences approach the representation of the complete user; splicing the enhanced learner representation with the knowledge point embeddings extracted from the historical learning resources, and inputting them into the matrix long short-term memory network mLstm to model the knowledge point mastery status of learners, and obtain the degree of learners' mastery of all knowledge points; Using the matrix long short-term memory network mLstm as the sequence encoder; during the process of training the enhancer, adopt a curriculum learning strategy to control the weight of each user in the loss according to the difficulty of learning knowledge from active learners; after the enhancer training is completed, pass the sequence representation of inactive learners through the enhancer to obtain the representation information of inactive learners after sequence enhancement; The S4 includes: obtaining the embeddings of each learner's user characteristics, educational resource characteristics, and knowledge point characteristics from the list of candidate educational resources of each learner, inputting them into a bidirectional long short-term memory network to learn context information, so as to obtain the relevance score of each educational resource for the overall learner; classifying the list of candidate educational resources of each learner according to knowledge points, obtaining the embeddings of knowledge points, educational resources, and learners, inputting them into a learning diversity estimator to obtain the learning demand distribution of each learner, and calculating the marginal diversity of educational resources, and multiplying the element-by-element difference between the learning demand distribution of learners and the diversity of educational resources to obtain the diversity score of educational resources; The learning diversity estimator is to use a long short-term memory network to encode the learner practice sequences of each knowledge point respectively to obtain the learner learning mode representation of each knowledge point; then use an attention mechanism and a multi-layer perceptron to learn the learning demand distribution of learners for different knowledge points; finally, use a probability coverage function to calculate the diversity gain of each practice.

2. The educational resource diversification recommendation method based on a long short-term memory neural network according to claim 1, wherein The historical learning resource information includes: learner ID, serial number of the educational resources learned by the learner, learning time, learning answer results, and knowledge points included in the educational resources; the method for forming the initial data set includes: arranging the questions answered by each user, the knowledge points examined by the questions, the answer results, and the answer time in a learning resource sequence, knowledge point sequence, learning resource result sequence, and learning resource time interval sequence with a length of 200 educational resource interactions in chronological order, and grouping them according to the learner ID; finally, dividing the data, and randomly dividing the data into a training set and a test set according to a ratio of 8:

2.

3. The educational resource diversification recommendation method based on a long short-term memory neural network according to claim 1, wherein S3 includes: calculating the difficulty corresponding to each educational resource based on the learner's knowledge point mastery level, and selecting several educational resources from the educational resource library as the candidate educational resource list for each learner.

4. The educational resource diversification recommendation method based on a long short-term memory neural network according to claim 1, wherein S5 includes: subjecting the obtained educational resource relevance and the diversity score to feature fusion to obtain the re-ranking score of each candidate set, and recommending educational resources to the learner from high to low according to the final score of the educational resources.

5. A diversified educational resource recommendation system based on a long short-term memory neural network, the system being used to implement the method described in any one of claims 1-4, characterized in that, Includes: Collection module, construction module, generation module, calculation module, and fusion module; The collection module is used to obtain the historical learning resource information of the learner and divide the learner. The construction module is used to perform learner representation enhancement and knowledge point mastery modeling based on the division result. The generation module is used to generate a candidate educational resource list based on the learner's knowledge point mastery level. The calculation module is used to calculate the educational resource relevance and diversity score based on the candidate educational resource list. The fusion module is used to fuse the educational resource relevance and the diversity score, and recommend a specified number of educational resources to the learner from high to low according to the final score.

Images