Personalized recommendation method based on tensor decomposition and neural network fusion

Abstract

The present invention relates to the field of educational resource recommendation technology, specifically a personalized recommendation method based on tensor decomposition and neural network fusion, including analyzing a dynamic learner feature tensor obtained by combining a preset situation tensor, a preset educational resource tensor, and the actual situation of the learner to obtain a fusion tensor; using a high-order singular value decomposition method to reduce the dimension of the fusion tensor, determine the matching relationship between educational resources and learner features under different preset situations, so as to form a matching resource library between learners and educational resources; and filter out recommendation strategies and corresponding educational resources that meet the learner's personal needs from the matching resource library according to preset resource screening rules. Through the above method, the present invention solves the problem that the traditional recommendation method often adopts static recommendation means, and does not combine the learner's learning habits, cognitive level, interests and hobbies and other personal actual situations, resulting in insufficient accuracy and reliability of educational resource recommendation results.

Description

Technical Field

[0001] The present invention relates to the technical field of educational resource recommendation, and in particular to a personalized recommendation method based on tensor decomposition and neural network fusion. Background Art

[0002] At present, with the continuous development of intelligent technology, there are more and more ways to obtain educational resources through the Internet. With the development of the Internet and the popularization of various electronic products, the sharing and recommendation of educational resources have effectively improved the efficiency and convenience of learners' learning.

[0003] However, in the prior art, the recommendation method of educational resources on online learning platforms is usually to recommend a large number of relevant educational resources to learners after simple screening. Learners who receive educational resources often choose to give up because the recommended educational resources are redundant or the content of the educational resources is not suitable for them, which greatly reduces the number of audiences of online learning platforms and causes a waste of educational resources. At the same time, static recommendation methods are often adopted in the recommendation process, and the learners' learning habits, cognitive levels, interests and hobbies and other personal actual conditions are not combined. After the recommended educational resources are determined, the corresponding knowledge points and other contents of the educational resources are not reasonably adjusted and optimized, resulting in the accuracy and reliability of educational resource recommendations being not high enough, thereby reducing the learners' learning efficiency and enthusiasm. In addition, with the passage of time, the learners' learning habits, cognitive levels, interests and hobbies and other personal actual conditions may change accordingly, and the failure to update the learners' information in a timely manner will also reduce the accuracy of the recommendation results.

[0004] Therefore, the present invention provides a personalized recommendation method based on tensor decomposition and neural network fusion to solve the above problems. Summary of the invention

[0005] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a personalized recommendation method based on tensor decomposition and neural network fusion to solve the problem that static recommendation methods are often adopted in the recommendation process, and the learners' learning habits, cognitive levels, interests and hobbies and other personal actual conditions are not taken into consideration, resulting in the recommendation results The accuracy and reliability are not high enough.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A personalized recommendation method based on tensor decomposition and neural network fusion, comprising:

[0008] Obtain the learner's historical learning data, personal data and current input information, and use the learner feature analysis model to process the historical learning data and personal data to obtain an initial learner feature tensor; and adjust the initial learner feature tensor according to the current input information to obtain a dynamic learner feature tensor;

[0009] Adopting the association analysis method to conduct association fusion analysis on the preset situation tensor, the preset educational resource tensor, and the dynamic learner feature tensor to obtain the fusion tensor;

[0010] The fusion tensor is processed by using a high-order singular value decomposition method and a neural network algorithm to determine the matching relationship between educational resources and learner characteristics under different preset scenarios, so as to establish a matching resource library corresponding to the preset scenarios; and a matching reason is added to each matching relationship in the matching resource library;

[0011] Filter out educational resources that meet learners' needs from the matching resource library according to preset resource screening rules;

[0012] Obtain learners' comment data on recommended educational resources, screen and analyze the comment data according to a preset keyword library to obtain question data; use a preset strategy generation model to analyze and process the question data to determine the division attributes of question features in the question data, and determine multiple sub-adjustment strategies according to preset judgment rules, and form a final adjustment strategy with each sub-adjustment strategy; and adjust the dynamic learner feature tensor and / or preset resource screening rules according to the adjustment strategy to adjust the educational resources recommended to the learners.

[0013] Preferably, the use of a learner feature analysis model to process historical learning data and personal data to obtain an initial learner feature tensor includes: using an AIGC cross-modal understanding method to extract features from historical learning data and personal data to obtain learner feature data, and determining high-dimensional representation data of the learner feature data based on a preset data type, a preset dimensional standard, and the learner's dimensional information to obtain an initial learner feature tensor.

[0014] Preferably, the adjusting the initial learner feature tensor according to the current input information to obtain the dynamic learner feature tensor includes: extracting features from the current input information to obtain input features; judging whether the input features have unadded features according to the learner feature data, and if so, analyzing the unadded features based on a preset data type, a preset dimension standard and the learner's dimensional information to add the unadded features to the corresponding positions of the initial learner feature tensor; and weighting the initial learner feature tensor corresponding to the added features in the input features to obtain the dynamic learner feature tensor.

[0015] Preferably, the process of obtaining the preset context tensor includes: obtaining various types of learning context labels marked by experts and users on educational resources; determining the correlation between each learning context based on the type of learning context labels and the number of different contexts of the same educational resource to obtain related data; and performing high-dimensional representation of each learning context and related data to obtain a preset context tensor.

[0016] Preferably, the process of obtaining the preset educational resource tensor includes: obtaining the first resource features of the educational resources collaboratively annotated by experts and users; using the AIGC cross-modal understanding method to analyze the first resource features to determine the relationship between the first resource features, and adjusting the first resource features according to the relationship to obtain the second resource features; performing feature extraction on the second resource features to obtain key attributes; and performing high-dimensional characterization of the second resource features based on the key attributes to form a preset educational resource tensor.

[0017] Preferably, the association analysis method is used to perform association fusion analysis on the preset situation tensor, the preset education resource tensor, and the dynamic learner feature tensor to obtain a fusion tensor, including: screening based on implicit key features in the preset situation tensor, the preset education resource tensor, and the dynamic learner tensor or two to obtain a first tensor data containing implicit key features; mapping the implicit key features in the first tensor data to each other to obtain a mapping relationship; and performing high-dimensional representation of the mapping relationship and the first tensor data to form a fusion tensor.

[0018] Preferably, the method of using a high-order singular value decomposition method and a neural network algorithm to process the fusion tensor to determine the matching relationship between educational resources and learner characteristics under different preset scenarios to establish a matching resource library corresponding to the preset scenarios, including: using the preset scenario characteristics in the fusion tensor as the dimension, and using a high-order singular value decomposition method to reduce the dimension of the fusion tensor to obtain multiple initial factor matrices under the preset scenario characteristics; using a singular value decomposition method to perform singular value decomposition on each initial factor matrix to obtain a scenario tensor and a factor matrix under the corresponding preset scenario characteristics; using a neural network algorithm to perform correlation analysis on the scenario tensor and the factor matrix under the corresponding preset scenario characteristics to determine the matching relationship between educational resource characteristics and learner characteristics under different preset scenarios, so as to establish a corresponding matching resource library based on the matching relationship.

[0019] Preferably, the method of using a preset strategy generation model to analyze and process the problem data to determine the division attributes of the problem features in the problem data, and determine multiple sub-adjustment strategies according to preset judgment rules, and form a final adjustment strategy with each sub-adjustment strategy, including: extracting features from the problem data to determine the problem features; analyzing the problem features based on preset attribute division rules to determine the classification attributes of the problem features; when the classification attributes of the problem features are learner attributes, comparing the dynamic learner features with the problem features to determine whether the dynamic learner features contain the problem features, if so, obtaining the number of problem features, and determining the adjustment parameters corresponding to the number of problem features based on a preset numerical range to obtain a first adjustment strategy; if not, determining the problem features as features to be added to the dynamic learner features to obtain a second adjustment strategy; when the classification attributes of the problem features are screening attributes, screening the corresponding problem data with the largest number of screening items according to the problem features, and using the screening items as the third adjustment strategy; integrating the first adjustment strategy, the second adjustment strategy and the third screening strategy to obtain an adjustment strategy.

[0020] Preferably, the method of selecting educational resources that meet the needs of learners from a matching resource library according to preset resource screening rules includes: generating a learning strategy that meets the learner's status according to the learner's current learning time and learning preferences, selecting educational resources suitable for learning in each learning time period from the matching resource library according to the learning strategy, and marking the educational resources corresponding to the characteristics of the current input information.

[0021] Preferably, the personalized recommendation method based on tensor decomposition and neural network fusion also includes: after recommending the first educational resource to the learner, obtaining the user's input information, analyzing the input information to obtain an adjusted second educational resource; and when the learner's learning record reaches a preset standard, sending an educational resource update confirmation message to the learner, and after receiving the learner's confirmation message, obtaining the learner's resource demand information, analyzing the learner's resource demand information and the learner's historical record information to update the first recommended resource.

[0022] The beneficial effects of the present invention are:

[0023] 1. The present invention analyzes the dynamic learner feature tensor, preset situation tensor and preset educational resource tensor obtained in combination with the actual situation of the learner to obtain a fusion tensor; then the high-order singular value decomposition method is used to reduce the dimension of the fusion tensor to determine the matching relationship between educational resources and learner features under different preset situations to form a matching resource library between learners and educational resources; and the matching resource library is tailored for learners and can change dynamically with the changes in learner features; finally, according to the preset resource screening rules, the recommendation strategy and corresponding educational resources that meet the learner's personal needs are screened out from the matching resource library. Through the above method, the present invention solves the problem that static recommendation means are often adopted in the recommendation process, and the actual personal situation such as learners' learning habits, cognitive levels, interests and hobbies are not combined, resulting in insufficient accuracy and reliability of the recommendation results. In addition, dimensionality reduction processing helps to explore the potential correlation between the features in the fusion tensor, comprehensively characterize the relationship between learner features, preset situations and preset educational resources, and help improve the robustness of the recommendation results.

[0024] 2. Before recommending educational resources to learners, the present invention will combine the user's historical learning data, registered personal data, and current input information of the user's learning tendency, and use the learner feature analysis model to process the above learner's various data to form a dynamic learner feature tensor unique to the learner. The dynamic learner feature tensor in the present invention combines the learner's learning habits, cognitive level, interests and hobbies, current learning tendencies, etc. After analyzing the learner feature tensor, compared with the traditional recommendation method that does not combine the learner's personal situation, the present invention can make the recommended educational resources meet the actual needs of the learner, thereby improving the accuracy and reliability of the recommendation results.

[0025] 3. The present invention adopts a method of collecting and analyzing learners' comment data in order to discover problems from the comment data; then after formulating targeted adjustment strategies, the learner feature tensor or preset screening rules are adjusted to recommend more accurate educational resources to learners to meet the needs of learners.

[0026] 4. The present invention also adopts a method of analyzing the personality traits and learning characteristics of learners and matching suitable instructors from a preset guidance matching library. This method can match learners with corresponding professional instructors who meet the learners' requirements for one-on-one guidance, thereby meeting the learners' personalized learning needs and increasing the learners' sense of trust and dependence. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic flow chart of a personalized recommendation method based on tensor decomposition and neural network fusion according to the present invention;

[0028] Figure 2 A schematic flow chart of obtaining a dynamic learner feature tensor according to the present invention. DETAILED DESCRIPTION

[0029] The following will refer to the attached Figure 1 To Attachment Figure 2 The embodiments of the present invention are described in detail. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0030] A personalized recommendation method based on tensor decomposition and neural network fusion, as shown in the attached Figure 1 As shown, the following steps are included:

[0031] Step S11: Obtain the learner's historical learning data, personal data and current input information, and use the learner feature analysis model to process the historical learning data and personal data to obtain an initial learner feature tensor; and adjust the initial learner feature tensor according to the current input information to obtain a dynamic learner feature tensor.

[0032] Preferably, historical learning data includes the learner's historical learning courses, historical learning progress, historical grades, historical search records, etc.; personal data includes the learner's cognitive level, historical learning preferences, etc.; current input information includes the learner's learning tendencies, current learning preferences, learning habits, etc.

[0033] Step S12: using an association analysis method to perform association fusion analysis on the preset situation tensor, the preset educational resource tensor, and the dynamic learner feature tensor to obtain a fusion tensor.

[0034] Specifically, when performing the association fusion analysis, a suitable association analysis method is used to perform association analysis on the features in the preset context tensor, the preset educational resource tensor, and the dynamic learner feature tensor, and to mine the implicit relationship between the features to obtain a fusion tensor with an association relationship, for example, the Apriori algorithm and the FP-growth algorithm.

[0035] Step S13: Use high-order singular value decomposition method and neural network algorithm to process the fusion tensor, determine the matching relationship between educational resources and learner characteristics under different preset scenarios, so as to establish a matching resource library corresponding to the preset scenarios; and add matching reasons to each matching relationship in the matching resource library.

[0036] In step S13, a high-order singular value decomposition method is used to obtain important information in the fusion tensor; then a suitable neural network algorithm is used to analyze the important information to obtain the matching relationship between educational resources and learner characteristics under different preset scenarios, so as to establish a matching resource library corresponding to the preset scenario based on the matching relationship.

[0037] Step S14: Filter out educational resources that meet learners' needs from the matching resource library according to preset resource screening rules.

[0038] Step S15: Obtain learners' comment data on the recommended educational resources, and screen and analyze the comment data according to a preset keyword library to obtain question data; use a preset strategy generation model to analyze and process the question data to determine the division attributes of the question features in the question data, and determine multiple sub-adjustment strategies according to preset judgment rules, and form a final adjustment strategy with each sub-adjustment strategy; and adjust the dynamic learner feature tensor and / or the preset resource screening rules according to the adjustment strategy to adjust the educational resources recommended to the learners.

[0039] In step S15, the learners' comment data on the recommended educational resources are obtained, and the comment data are screened and analyzed according to the preset keyword library to obtain problem data; the adjustment strategy determined by the reviewer based on the problem data is obtained, or the adjustment strategy is determined by generating a model based on the preset strategy; then the relevant parameters of the dynamic learner feature tensor are adjusted according to the adjustment strategy, or screening conditions are added or adjusted to the preset resource screening rules to adjust the educational resources recommended to the learners, so that the recommended educational resources are more in line with the actual needs of the learners.

[0040] In addition, the present invention can also screen and collect excellent feedback data in the comment data, the problem data is used for sub-strategy adjustment and pruning of the neural network model, and the excellent feedback data is used to strengthen the recommendation strategy and the neural network model in the strengthening method, so that the personalized recommendation method of the present invention is adjusted by combining the problem data and the excellent feedback data, and the robustness and accuracy of the recommendation results are gradually improved.

[0041] An alternative to review data is to use a reinforcement learning mechanism based on human feedback to analyze changes in learners’ needs in real time, so as to dynamically adjust the educational resources recommended to learners to meet their educational resource needs.

[0042] Preferably, the dynamic learner feature is a feature in a dynamic learner feature tensor. Each sub-adjustment strategy includes a first adjustment strategy, a second adjustment strategy and a third screening strategy.

[0043] Among them, the preset strategy generation model includes: extracting features from problem data to determine problem features; analyzing the problem features based on preset attribute division rules to determine the classification attributes of the problem features; when the classification attributes of the problem features are learner attributes, comparing the dynamic learner features with the problem features to determine whether the dynamic learner features contain the problem features, if so, obtaining the number of problem features, and determining the adjustment parameters corresponding to the number of problem features based on a preset numerical range to obtain a first adjustment strategy; if not, determining the problem features as features to be added to the dynamic learner features to obtain a second adjustment strategy; when the classification attributes of the problem features are screening attributes, screening the largest number of screening items in the same or similar problem data according to the problem features, and using the screening items as the third adjustment strategy; integrating the first adjustment strategy, the second adjustment strategy and the third screening strategy to obtain an adjustment strategy.

[0044] The process of the auditor determining the adjustment strategy based on the problem data includes: subjectively analyzing the problem data through expert review, screening out typical problem data from the problem data, and determining the classification attributes of the typical problem data, and then determining the adjustment content for the typical problem data. The adjustment content includes relevant parameters of dynamic learner characteristics and screening conditions of preset screening rules, and then obtaining the adjustment strategy for the typical problem data.

[0045] Among them, social tagging is a user-driven data organization method that allows users to use tags to tag resources on the Internet, such as web pages, videos, and pictures. These tags are added by users on their own initiative and reflect the user's understanding, opinions, and interest preferences of the resources. The social tagging system allows any user to participate, conveniently organize their own resources in the form of tags, and find like-minded users. The advantages of the social tagging system include reducing retrieval costs and quickly adapting to changes in network vocabulary. Social tagging is a data structure that is completely driven by users. Users can use tags to tag resources on the Internet. Tags are added by users on their own initiative, expressing the user's understanding and opinions of the resources, and also reflecting the user's interest preferences to a certain extent. This interaction between users and the system forms a three-dimensional data structure unique to the social tagging system. Users can add tags to educational resources on these platforms. This practice is usually called "folk classification", and relevant educational resources can be found through tags.

[0046] Specifically, the present invention analyzes the dynamic learner feature tensor, preset situation tensor and preset educational resource tensor obtained in combination with the actual situation of the learner to obtain a fusion tensor; then the high-order singular value decomposition method is used to reduce the dimension of the fusion tensor to determine the matching relationship between educational resources and learner features in different preset situations to form a matching resource library between learners and educational resources; and the matching resource library is tailored for learners and can change dynamically with the changes in learner features; finally, according to the preset resource screening rules, the recommendation strategy and corresponding educational resources that meet the learner's personal needs are screened out from the matching resource library. Through the above method, the present invention solves the problem that static recommendation means are often adopted in the recommendation process, and the personal actual situation such as the learner's learning habits, cognitive level, interests and hobbies are not combined, resulting in the problem that the accuracy and reliability of the recommendation results are not high enough. In addition, the dimensionality reduction process helps to explore the potential correlation between the features in the fusion tensor, comprehensively characterize the relationship between learner features, preset situations and preset educational resources, and help improve the robustness of the recommendation results.

[0047] Furthermore, before recommending educational resources to learners, the present invention will combine the learner's historical learning data, registered personal data, and current input information of the learner's learning tendencies, and use a learner feature analysis model to process the above-mentioned learner's various data to form a dynamic learner feature tensor unique to the learner. Taking into account the possibility that the learner's interests and current learning tendencies will change during the learning process, the present invention sets a function that the learner can input current demand information, that is, to obtain the learner's current input information in order to obtain the learner's latest dynamic learner feature tensor. The dynamic learner feature tensor in the present invention combines the learner's learning habits, cognitive level, interests and hobbies, current learning tendencies, etc. After analyzing the learner feature tensor, compared with traditional recommendation methods that do not combine the learner's personal situation, the present invention can make the recommended educational resources meet the actual needs of the learner, thereby improving the accuracy and reliability of the recommendation results.

[0048] In addition, most recommendation methods in the prior art have the problems of "information cocoon" and "recommendation homogeneity". In order to improve the diversity of digital educational resource recommendations, the present invention can capture learners' long-term interests and respond to new changes in learners' data in a timely manner, add new changes to the analysis, and enhance the accuracy and reliability of educational resource recommendation results.

[0049] In one embodiment of the present invention, the use of a learner feature analysis model to process historical learning data and personal data to obtain an initial learner feature tensor includes: using an AIGC cross-modal understanding method to extract features from historical learning data and personal data to obtain learner feature data, and determining high-dimensional representation data of the learner feature data based on a preset data type, a preset dimensional standard, and the learner's dimensional information to obtain an initial learner feature tensor.

[0050] In this embodiment, first, according to the preset dimensional standard, the historical learning data and personal data are dimensionally processed through a cross-modal method to unify the formats of the historical learning data and personal data into the same format; then, the AIGC cross-modal understanding method is used to extract features of the historical learning data and personal data in the same format to obtain learner feature data that can reflect the learner's learning characteristics; then, the learner feature data is classified according to the preset data type and the learner's dimensional information to form a multi-dimensional initial learner feature tensor.

[0051] Furthermore, when the learner just starts using the online learning platform applied by the present invention, the learner's personal information has not yet been obtained. The present invention will use AIGC simulation to generate the learner's initial learner feature tensor to compensate for the user's cold start problem, and then continuously adjust and obtain the dynamic learner feature tensor in the process of obtaining user input information.

[0052] Furthermore, during the learner's learning process, with the learner's consent, the learner's demeanor, eye movements, mouse, keyboard and other monitoring data are obtained through monitoring, and then AIGC is used to process and analyze the monitoring data to adjust the learner's feature tensor and obtain the latest dynamic learner feature tensor.

[0053] Through the setting method of this embodiment, the present invention can process the learner's historical learning data and personal data to obtain an initial learner feature tensor that characterizes the learner's personal situation; so as to subsequently analyze the learner feature tensor and match the learner with educational resources that meet the learner's personalized needs.

[0054] Furthermore, in one embodiment of the present invention, the learner's dimensional information includes user portrait, situational awareness and social network, and a visualization tool is used to visualize the dimensional information for visual display.

[0055] Among them, the specific analysis of user portraits, situational awareness and social networks in the dimensional information includes comprehensively capturing learners' learning styles, cognitive levels, emotional states, learning situations, behavioral dynamics, etc. The current learning situation of learners is visualized in each dimension of the dimensional information, so that learners can intuitively understand their own learning situation and know their own learning level more clearly.

[0056] In one embodiment of the present invention, as shown in the attached Figure 2 As shown, the method of adjusting the initial learner feature tensor according to the current input information to obtain the dynamic learner feature tensor includes the following steps:

[0057] Step S21: extract features from the current input information to obtain input features.

[0058] Among them, the feature extraction method can be an autoencoder or a feature extraction model built based on a convolutional neural network, and there is no limitation on the feature extraction method.

[0059] Step S22: Determine whether the input features have unadded features according to the learner feature data, if yes, execute step S23, if not, execute step S24.

[0060] Step S23: Analyze the unadded features based on the preset data type, the preset dimension standard and the learner's dimension information to add the unadded features to the corresponding positions of the initial learner feature tensor to obtain the dynamic learner feature tensor.

[0061] Step S24: weight adjustment is performed on the initial learner feature tensor corresponding to the added features in the input features to obtain a dynamic learner feature tensor.

[0062] Specifically, when the acquired input features already exist in the learner feature data, the number of corresponding features in the initial learner feature tensor is adjusted, etc., so as to adjust the weights of the corresponding features; when there are unadded features in the acquired input features, the type of the unadded features is determined based on the preset data type and represented in the form of a label; and the format of the unadded features is adjusted according to the preset dimensional standard for unified processing, and then according to the relationship between the unadded features and the learner's dimensional information, the unadded features after dimensional conversion are added to the appropriate positions of the initial learner feature tensor to obtain the dynamic learner feature tensor.

[0063] Alternatively, multiple algorithms in AIGC may be used to automatically adjust the learner feature tensor while the AI ​​robot is interacting with the learner, so as to obtain a dynamic learner feature tensor.

[0064] Through the setting method of this embodiment, before recommending educational resources to learners, the present invention will combine the user's historical learning data, registered personal data, and current input information of the user's learning tendency, and use the learner feature analysis model to process the various data of the above learners to form a dynamic learner feature tensor unique to the learner. The dynamic learner feature tensor in the present invention combines the learner's learning habits, cognitive level, interests and hobbies, current learning tendencies, etc. After analyzing the learner feature tensor, compared with the traditional recommendation method that does not combine the learner's personal situation, the present invention can make the recommended educational resources meet the actual needs of the learner, thereby improving the accuracy and reliability of the recommendation results.

[0065] In one embodiment of the present invention, the process of obtaining the preset context tensor includes: obtaining various types of learning context labels marked by experts and users on educational resources; determining the association between each learning context based on the type of learning context label and the number of different contexts of the same educational resource to obtain associated data; and performing high-dimensional representation of each learning context and associated data to obtain a preset context tensor.

[0066] Preferably, the learning situation includes suspense, lifelikeness, authenticity, complexity, emotion, typicality, subjectivity and variability.

[0067] Specifically, the correlation between the learning situations contained in each educational resource of the same type and the learning situations that appear in the same educational resource is determined. The correlation is expressed in the form of the probability of other learning situations appearing when one learning situation is known. When labeling educational resources, experts will label the learning situations of educational resources, and learners who are suitable for learning the educational resources will be labeled to improve the accuracy of labeling educational resources in terms of learning situations, so as to improve the accuracy of the matching relationship between learner characteristics and educational resource characteristics in the future when the preset situation is used as a dimension, so as to achieve refined matching.

[0068] Through the setting method of this embodiment, the present invention can form preset situation data according to the learning situation and the correlation between each learning situation, and perform high-dimensional representation of the preset situation data to obtain a preset situation tensor, so as to subsequently combine the preset educational resource tensor and the learner feature tensor for analysis to obtain the learner's educational resource recommendation results.

[0069] In one embodiment of the present invention, the process of obtaining a preset educational resource tensor includes: obtaining first resource features of educational resources collaboratively annotated by experts and users; analyzing the first resource features using the AIGC cross-modal understanding method to determine the relationships between the first resource features, and adjusting the first resource features according to the relationships to obtain second resource features; extracting features from the second resource features to obtain key attributes; and performing high-dimensional characterization of the second resource features based on the key attributes to form a preset educational resource tensor.

[0070] In this embodiment, after the first resource feature is analyzed by the AIGC cross-modal understanding method, the mutual relationship between each first resource feature is obtained, and the mutual relationship is added to the first resource feature to obtain a second resource feature containing the mutual relationship. Then, an autoencoder or a feature extraction model based on a neural network is used to extract the second resource feature to obtain the key attributes of the second resource feature; then, the second resource feature is characterized in high dimensions according to each key attribute to form a preset educational resource tensor.

[0071] Through the setting method of this embodiment, the present invention can classify the second educational resource features into different dimensions according to key attributes, forming a subdivided preset educational resource tensor, so that the learner can be matched with accurate educational resources according to the learner feature tensor, the preset situation tensor and the preset educational resource tensor, helping to improve the learner's learning efficiency and learning progress. In addition, the first resource feature of the educational resource is determined by the collaborative annotation of cutting-edge experts and users, and during the determination of the resource features, the resource features annotated by the user will be pre-processed to prevent malicious annotation by the user, which will affect the subsequent analysis and help avoid reducing the accuracy of the recommendation results.

[0072] In one embodiment of the present invention, an association analysis method is used to perform association fusion analysis on a preset situation tensor, a preset education resource tensor, and a dynamic learner feature tensor to obtain a fusion tensor, including: screening based on implicit key features in the preset situation tensor, the preset education resource tensor, and the dynamic learner tensor or two to obtain first tensor data containing implicit key features; mapping the implicit key features in the first tensor data to each other to obtain a mapping relationship; and performing high-dimensional representation of the mapping relationship and the first tensor data to form a fusion tensor.

[0073] Preferably, the implicit key features may be identical features or similar features, wherein the similarity of the similar features exceeds 90% or exceeds 80%, and the similarity may be adjusted according to the parameter data amount or parameter type of the fusion tensor.

[0074] In this embodiment, the present invention determines the implicit key features through the common features or similar features of the preset situation tensor, the preset education resource tensor and the dynamic learner tensor, or both; then extracts and fuses the preset situation tensor, the preset education resource tensor and the dynamic learner tensor linked by the implicit key features to obtain the first tensor data; and constructs the mapping relationship of the first tensor data with the implicit key features, so that the features in the first tensor data can be associated. The first tensor data is characterized in high dimensions using the invisible key features of the mapping relationship as the dimension parameter to form a fused tensor. Through the setting method of this embodiment, the present invention can deeply mine the preset situation tensor, the preset education resource tensor and the dynamic learner tensor, and establish a fused tensor with an associated relationship between the three, so that the data obtained by the subsequent decomposition of the fused tensor is more in line with the actual situation of the learner; and the matching relationship between the learner characteristics and the educational resources is established more accurately.

[0075] In one embodiment of the present invention, a high-order singular value decomposition method and a neural network algorithm are used to process the fusion tensor to determine the matching relationship between educational resources and learner characteristics under different preset scenarios, so as to establish a matching resource library corresponding to the preset scenarios, including: using the preset scenario characteristics in the fusion tensor as the dimension, and using a high-order singular value decomposition method to reduce the dimension of the fusion tensor to obtain multiple initial factor matrices under the preset scenario characteristics; using a singular value decomposition method to perform singular value decomposition on each initial factor matrix to obtain a scenario tensor and a factor matrix under the corresponding preset scenario characteristics; using a neural network algorithm to perform correlation analysis on the scenario tensor and the factor matrix under the corresponding preset scenario characteristics, to determine the matching relationship between educational resource characteristics and learner characteristics under different preset scenarios, so as to establish a corresponding matching resource library according to the matching relationship.

[0076] Preferably, the neural network algorithm can be a generative adversarial network algorithm, a convolutional neural network algorithm, a feedforward neural network algorithm, a recurrent neural network algorithm, a generative adversarial network algorithm, etc.

[0077] In this embodiment, one way to perform high-order singular value decomposition on the fused tensor is: taking a third-order tensor as an example, , , The three dimensions are situational data , learner characteristic data and educational resource characteristics data , the formula is:

[0078] ,

[0079] in, It is the high-dimensional data after the third-order tensor decomposition; is a second-order tensor, represented in matrix form; is a vector, , , , , Indices representing arrays of tensor decomposition, subarrays formed.

[0080] For the high-dimensional data existing in learner feature data and educational resource data, For high-dimensional data existing in situational data and educational resource data, After decomposing the high-dimensional data in the context data and learner feature data, the neural network algorithm is used to process the low-dimensional data. That is, the neural network algorithm is used to perform correlation analysis on the context tensor and the factor matrix under the corresponding preset context features, and the matching relationship between the educational resource features and the learner features under different preset contexts is determined, so as to establish the corresponding matching resource library according to the matching relationship.

[0081] Through the setting method of this embodiment, the present invention can use the high-order singular value decomposition method to reduce the dimension of the fusion vector, and then use the neural network algorithm to perform correlation analysis on the reduced-dimensional data to determine the matching relationship between the characteristics of educational resources and the characteristics of learners in different preset situations, so as to establish a matching resource library that meets the characteristics of learners. Therefore, it is possible to filter out educational resources that meet the learning habits and personal preferences of learners from the matching resource library according to the preset screening rules, thereby improving the accuracy of recommending educational resources to learners, as well as the diversity of recommending corresponding educational resources to learners in different preset situations. The present invention solves the problem that static recommendation methods are often adopted in the recommendation process, and the actual personal conditions such as learners' learning habits, cognitive levels, interests and hobbies are not combined, resulting in insufficient accuracy and reliability of recommendation results.

[0082] In one embodiment of the present invention, a matching reason is added to each matching relationship in the matching resource library, including: using the AIGC cross-modal understanding method to perform correlation analysis on the learner characteristics, preset contexts and educational resource characteristics of each item in the matching resource library to generate the matching reason of the matching relationship.

[0083] Through the setting method of this embodiment, the present invention adopts the AIGC cross-modal understanding method to perform causal analysis on learner characteristics and matching educational resources under each preset scenario, and determines the reason why the educational resource is matched in the matching resource library under the preset scenario, thereby increasing the interpretability of the educational resource recommendation results, and thereby improving the reliability of the recommendation results of the present invention.

[0084] In one embodiment of the present invention, a preset strategy generation model is used to analyze and process the problem data to determine the division attributes of the problem features in the problem data, and multiple sub-adjustment strategies are determined according to preset judgment rules, and the final adjustment strategy is formed with each sub-adjustment strategy, including: feature extraction of the problem data to determine the problem features; analyzing the problem features based on preset attribute division rules to determine the classification attributes of the problem features; when the classification attributes of the problem features are learner attributes, comparing the dynamic learner features with the problem features to determine whether the dynamic learner features contain the problem features, if so, obtaining the number of problem features, and determining the adjustment parameters corresponding to the number of problem features based on a preset numerical range to obtain a first adjustment strategy; if not, determining the problem features as features to be added to the dynamic learner features to obtain a second adjustment strategy; when the classification attributes of the problem features are screening attributes, screening the largest number of screening items in the corresponding problem data according to the problem features, and using the screening items as the third adjustment strategy; integrating the first adjustment strategy, the second adjustment strategy and the third screening strategy to obtain an adjustment strategy.

[0085] Preferably, the preset attribute classification rules include the correspondence between preset words and preset attributes, and the association between preset words and similar words. The classification attributes include learner attributes and screening attributes. The learner attributes are the classification attributes corresponding to the learner characteristics, and the screening attributes are the classification attributes corresponding to the preset resource screening rules.

[0086] In this embodiment, the present invention adopts a method of analyzing and processing the problem data to determine the division attributes of the problem features in the problem data, and determining multiple sub-adjustment strategies according to preset judgment rules, and forming a final adjustment strategy with each sub-adjustment strategy. Through the above method, the present invention can find problems from the comment data based on the collection of learners' comment data; then after formulating targeted adjustment strategies, the learner feature tensor or preset screening rules are adjusted to recommend more accurate educational resources to learners to meet the needs of learners.

[0087] In one embodiment of the present invention, educational resources that meet the needs of learners are screened out from a matching resource library according to preset resource screening rules, including: generating a learning strategy that meets the learner's status according to the learner's current learning time and learning preferences; screening educational resources suitable for learning in each learning time period from the matching resource library according to the learning strategy, and marking the educational resources corresponding to the characteristics of the current input information.

[0088] In this embodiment, the type of educational resources that are in line with the current learning time period are formulated according to the learner's learning time and learning preferences, and then the corresponding type of educational resources are screened out from the matching resource library according to the type of educational resources to meet the learner's personalized needs. For example, using an ant colony algorithm, a simulated annealing algorithm, or a genetic algorithm to screen out the corresponding type of educational resources from the matching resource library according to the type of educational resources and the learner's key features helps to improve the matching speed and matching accuracy. In the recommended educational resources, the educational resources corresponding to the characteristics of the learner's current input information are highlighted and marked, so that the learner can intuitively see the educational resources that he urgently needs, further meeting the learner's personalized needs. Through the setting method of this embodiment, the present invention can recommend appropriate educational resources according to the learner's learning time and learning preferences to meet the learner's personalized needs.

[0089] In one embodiment of the present invention, the personalized recommendation method based on tensor decomposition and neural network fusion also includes: after recommending the first educational resource to the learner, obtaining the user's input information, analyzing the input information to obtain an adjusted second educational resource; and when the learner's learning record reaches a preset standard, sending an educational resource update confirmation message to the learner, and after receiving the learner's confirmation message, obtaining the learner's resource demand information, analyzing the learner's resource demand information and the learner's historical record information to update the first educational resource.

[0090] Preferably, the first educational resource refers to an educational resource recommended to the learner. The preset standard refers to the evaluation standard of each learning course, and different evaluation standards are set according to the learning requirements of different learners. For example, when learning the basic knowledge of a certain industry, the evaluation is based on the standard of preliminary understanding; when learning a certain professional course, the evaluation is based on the standard of proficiency.

[0091] The method of analyzing the learner resource demand information and the learner historical record information can be used as an analysis model constructed by a neural network algorithm.

[0092] In this embodiment, the present invention sets a method for adjusting the first recommended resource according to the learner's requirements so that the recommended educational resources can further meet the learner's educational resource needs. In addition, a method for tracking and evaluating the learner's learning situation in real time is set, that is, when the learner's learning record reaches the preset standard, an educational resource update confirmation message is sent to the learner, and after receiving the learner's confirmation message, the learner's resource demand information is obtained, and the learner's resource demand information and learner's historical record information are analyzed to update the first educational resource, so as to provide learners with recommendation services in real time and meet the diverse needs of learners.

[0093] In one embodiment of the present invention, the personalized recommendation method based on tensor decomposition and neural network fusion also includes: when the learner determines to customize educational resources, obtaining the learner's personality characteristics and learning characteristics screened from the dynamic learner feature tensor; matching the learner with a corresponding instructor from a preset guidance matching library based on the personality characteristics and learning characteristics to establish a consultation window between the learner and the instructor.

[0094] In this embodiment, when learners want to obtain educational resources in a targeted manner, the present invention adopts a method of analyzing the learners' personality traits and learning characteristics and matching suitable instructors from a preset guidance matching library. This method can match learners with corresponding professional instructors who meet the learners' requirements for one-on-one guidance, thereby meeting the learners' personalized learning needs and increasing the learners' sense of trust and dependence.

[0095] In one embodiment of the present invention, the present invention analyzes the matching relationship between learner characteristics and educational resources from the perspective of each preset scenario, increases the diversity of matching results, and makes the educational resources matched for learners flexible and diverse, which helps learners get rid of the traditional boring learning method and allows learners to enjoy the joy of learning while learning, thereby helping to improve learners' learning efficiency and learning interest.

[0096] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs, which may be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general programmable processor, which may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device. It should be noted that in the description of the present invention, the terms "first", "second", and "third" are used only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0097] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0098] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. More specific examples of machine-readable storage media may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0099] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0100] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0101] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A personalized recommendation method based on tensor decomposition and neural network fusion, characterized in that: include: Obtain the learner's historical learning data, personal data and current input information, and use the learner feature analysis model to process the historical learning data and personal data to obtain the initial learner feature tensor; And adjust the initial learner feature tensor according to the current input information to obtain the dynamic learner feature tensor; Adopting the association analysis method to perform association fusion analysis on the preset situation tensor, the preset education resource tensor, and the dynamic learner feature tensor to obtain a fusion tensor, including: screening based on the implicit key features of the preset situation tensor, the preset education resource tensor, and the dynamic learner tensor, to obtain the first tensor data containing the implicit key features; mapping each other according to the implicit key features in the first tensor data to obtain a mapping relationship; high-dimensionally representing the mapping relationship and the first tensor data to form a fusion tensor; The fusion tensor is processed by a high-order singular value decomposition method and a neural network algorithm to determine the matching relationship between educational resources and learner characteristics under different preset scenarios, so as to establish a matching resource library corresponding to the preset scenarios, including: using the preset scenario characteristics in the fusion tensor as the dimension, and using a high-order singular value decomposition method to reduce the dimension of the fusion tensor to obtain multiple initial factor matrices under the preset scenario characteristics; using a singular value decomposition method to perform singular value decomposition on each initial factor matrix to obtain a scenario tensor and a factor matrix under the corresponding preset scenario characteristics; using a neural network algorithm to perform correlation analysis on the scenario tensor and the factor matrix under the corresponding preset scenario characteristics to determine the matching relationship between educational resource characteristics and learner characteristics under different preset scenarios, so as to establish a corresponding matching resource library according to the matching relationship; and adding a matching reason to each matching relationship in the matching resource library; Filter out educational resources that meet learners' needs from the matching resource library according to preset resource screening rules; Obtain learners' comment data on recommended educational resources, screen and analyze the comment data to obtain problem data; use a preset strategy generation model to analyze and process the problem data to determine the division attributes of problem features in the problem data, and determine multiple sub-adjustment strategies based on preset judgment rules, and form a final adjustment strategy with each sub-adjustment strategy; adjust the dynamic learner feature tensor and / or preset resource screening rules according to the adjustment strategy to adjust the educational resources recommended to the learners.

2. The personalized recommendation method according to claim 1, characterized in that: The method of using a learner feature analysis model to process historical learning data and personal data to obtain an initial learner feature tensor includes: using an AIGC cross-modal understanding method to extract features from historical learning data and personal data to obtain learner feature data, and determining high-dimensional representation data of the learner feature data based on a preset data type, a preset dimensional standard, and the learner's dimensional information to obtain an initial learner feature tensor.

3. The personalized recommendation method according to claim 2, characterized in that: The adjusting the initial learner feature tensor according to the current input information to obtain the dynamic learner feature tensor includes: extracting features from the current input information to obtain input features; judging whether the input features have unadded features according to the learner feature data, and if so, analyzing the unadded features based on a preset data type, a preset dimension standard and the learner's dimensional information to add the unadded features to the corresponding positions of the initial learner feature tensor; and weighting the initial learner feature tensor corresponding to the added features in the input features to obtain the dynamic learner feature tensor.

4. The personalized recommendation method according to claim 1, characterized in that: The process of obtaining the preset situation tensor includes: obtaining various types of learning situation labels marked by experts and users on educational resources; determining the correlation between various learning situations according to the types of learning situation labels and the number of different situations of the same educational resource to obtain related data; performing high-dimensional representation of each learning situation and related data to obtain a preset situation tensor, wherein the learning situation includes suspense, life, authenticity, complexity, emotion, typicality, subjectivity and variability.

5. The personalized recommendation method according to claim 1, characterized in that: The process of acquiring the preset educational resource tensor includes: acquiring the first resource features of the educational resources collaboratively annotated by experts and users; using the AIGC cross-modal understanding method to analyze the first resource features to determine the relationship between the first resource features, and adjusting the first resource features according to the relationship to obtain the second resource features; extracting features from the second resource features to obtain key attributes; and performing high-dimensional representation of the second resource features based on the key attributes to form a preset educational resource tensor.

6. The personalized recommendation method according to claim 1, characterized in that: The method of using a preset strategy generation model to analyze and process the problem data to determine the division attributes of the problem features in the problem data, and determine multiple sub-adjustment strategies according to preset judgment rules, and form a final adjustment strategy with each sub-adjustment strategy, including: extracting features from the problem data to determine the problem features; analyzing the problem features based on preset attribute division rules to determine the classification attributes of the problem features; when the classification attributes of the problem features are learner attributes, comparing the dynamic learner features with the problem features to determine whether the dynamic learner features contain the problem features, if so, obtaining the number of problem features, and determining the adjustment parameters corresponding to the number of problem features based on a preset numerical range to obtain a first adjustment strategy; if not, determining the problem features as features to be added to the dynamic learner features to obtain a second adjustment strategy; when the classification attributes of the problem features are screening attributes, screening the largest number of screening items in the corresponding problem data according to the problem features, and using the screening items as the third adjustment strategy; integrating the first adjustment strategy, the second adjustment strategy and the third screening strategy to obtain an adjustment strategy.

7. The personalized recommendation method according to claim 1, characterized in that: The method of selecting educational resources that meet the needs of learners from a matching resource library according to preset resource selection rules includes: Based on the learner's current learning time and learning preferences, a learning strategy that suits the learner's status is generated. Based on the learning strategy, educational resources suitable for learning in each learning time period are screened from the matching resource library, and the educational resources corresponding to the characteristics of the current input information are marked.

8. The personalized recommendation method according to claim 1, characterized in that: Also includes: After recommending the first educational resource to the learner, obtaining the user's input information, and analyzing the input information to obtain an adjusted second educational resource; And when the learner's learning record reaches a preset standard, an educational resource update confirmation message is sent to the learner. After receiving the learner's confirmation message, the learner's resource demand information is obtained, and the learner's resource demand information and the learner's historical record information are analyzed to update the first educational resource.

Images