CN116738371A

CN116738371A - User learning portrait construction method and system based on artificial intelligence

Info

Publication number: CN116738371A
Application number: CN202311015144.XA
Authority: CN
Inventors: 黎国权; 朱晖
Original assignee: Guangdong Xinjufeng Technology Co ltd
Current assignee: Guangdong Xinjufeng Technology Co ltd
Priority date: 2023-08-14
Filing date: 2023-08-14
Publication date: 2023-09-12
Anticipated expiration: 2043-08-14
Also published as: CN116738371B

Abstract

The embodiment of the application provides a user learning portrait construction method and a system based on artificial intelligence, which can search heuristic features of learning process description features by acquiring learning process description features corresponding to learning behavior log data and learning evaluation description features of learning behavior interaction data corresponding to the learning behavior log data so as to acquire behavior heuristic information in the learning behavior log data, thereby improving feature expression performance of heuristic search features. The learning evaluation description features and the heuristic search features are converged to generate target construction learning features, the target construction learning features are subjected to relation matrix construction to obtain a learning behavior-test question relation matrix, and user learning portrait generation is further carried out on the learning behavior-test question relation matrix to obtain user learning portrait features corresponding to the learning behavior log data, so that the accuracy of user learning portrait prediction is improved.

Description

User learning portrait construction method and system based on artificial intelligence

Technical Field

The embodiment of the application relates to the technical field of artificial intelligence, in particular to a user learning portrait construction method and system based on artificial intelligence.

Background

With the application of computer software technology, the birth of learning software marks that students learn and really enter the tool age, and plays an irreplaceable important role in improving learning efficiency and learning quality of students in the future. The learning software is to digitally process the traditional education process by using a computer software technology, so that the purpose of assisting learning by taking students as the center is achieved, and the learning software is helpful for students with limited time and space. Meanwhile, the content of the learning software is richer, and various discipline knowledge can be integrated, so that a personalized and differentiated learning path is provided, and student users are helped to quickly master knowledge points. For example, the learning content data can be provided for student users in the learning software by analyzing the learning behavior log data of the student users to construct user learning portraits of the student users. Therefore, how to ensure the accuracy of the user learning portrait prediction is a technical problem to be continuously solved in the technical field.

Disclosure of Invention

In order to at least overcome the defects in the prior art, the embodiment of the application aims to provide a user learning portrait construction method and system based on artificial intelligence.

According to one aspect of the embodiment of the application, there is provided a user learning portrait construction method based on artificial intelligence, including:

acquiring learning behavior log data of a target learning user in target learning software and learning behavior interaction data corresponding to the learning behavior log data, wherein the learning behavior interaction data is used for evaluating state result data of offline learning of the learning behavior log data;

acquiring learning process description features corresponding to the learning behavior log data, and acquiring learning evaluation description features corresponding to the learning behavior interaction data;

performing heuristic feature search on the learning process description features, generating heuristic search features corresponding to the learning behavior log data, and converging the heuristic search features and the learning evaluation description features to generate target construction learning features;

and constructing a relation matrix for the target constructed learning features to obtain a learning behavior-test question relation matrix, generating a user learning portrait for the learning behavior-test question relation matrix, generating user learning portrait features corresponding to the learning behavior log data, constructing a software pushing frame of the target learning user in the target learning software based on the user learning portrait features corresponding to the learning behavior log data, and the learning behavior-test question relation matrix is used for representing an array formed by the learning relation feature data between each learning behavior and each corresponding test question.

In a possible implementation manner of the first aspect, the acquiring the learning process description feature corresponding to the learning behavior log data includes:

performing learning behavior activity separation on the learning behavior log data to obtain a learning behavior activity set, performing knowledge point answer data block division on each learning behavior activity data in the learning behavior activity set to generate a plurality of knowledge point answer data with learning mode categories, and generating a knowledge point answer data sequence corresponding to each learning behavior activity data in the learning behavior activity set;

based on a knowledge point answer data sequence corresponding to learning behavior activity data v contained in the learning behavior activity set, acquiring knowledge point test question answer features corresponding to the learning behavior activity data v, and loading the knowledge point test question answer features into a context characterization unit in a target recurrent neural network; v is a positive integer not greater than the learning behavior activity data amount corresponding to the learning behavior activity set;

based on the context characterization unit, carrying out feature construction on the answer features of the knowledge point test questions, and generating learning description features corresponding to the learning behavior activity data v;

And converging learning description features corresponding to the learning behavior activity data in the learning behavior activity set to generate learning process description features corresponding to the learning behavior log data.

In a possible implementation manner of the first aspect, the feature construction is performed on the answer feature of the knowledge point test question based on the context characterization unit, and the generating the learning description feature corresponding to the learning behavioral activity data v includes:

generating a local context Wen Biaozheng vector corresponding to the knowledge point test question answering feature based on a local context window in the context characterization unit, and converging the knowledge point test question answering feature and the local context Wen Biaozheng vector to generate a target converging feature; and acquiring a covariance feature array and a correlation coefficient array corresponding to a principal component feature analysis subunit in the context characterization unit, determining a linear conversion feature corresponding to the learning behavior activity data v according to the correlation coefficient array and screening features between the covariance feature array and the target convergence feature, and converging the target convergence feature and the linear conversion feature to generate a learning description feature corresponding to the learning behavior activity data v.

In a possible implementation manner of the first aspect, the generating, based on a local context window in the context characterization unit, a local context Wen Biaozheng vector corresponding to the knowledge point test question answer feature includes:

acquiring a global feature decomposition structure corresponding to a local context window in the context characterization unit, and decomposing the knowledge point test question answering feature into a first answering relation vector, a first test question knowledge point vector and a first test question weight vector according to the global feature decomposition structure of the local context window;

performing relationship vector diagnosis on the first answer relationship vector and the knowledge point difficulty point vector of the first test question knowledge point vector to generate an answer diagnosis vector, and acquiring a solution capability factor vector corresponding to the first answer relationship vector;

and carrying out regularized conversion on the co-occurrence feature vector of the answer diagnosis vector and the answer capacity factor vector to generate a first test question answer attention vector, and determining local up-down Wen Biaozheng vectors corresponding to the knowledge point test question answer features based on a fusion vector between the first test question answer attention vector and the first test question weight vector.

In a possible implementation manner of the first aspect, the acquiring learning evaluation description features corresponding to the learning behavior interaction data includes:

performing data partitioning on the learning behavior interaction data to generate K interaction records, and acquiring morpheme representation vectors respectively corresponding to the K interaction records; k is a positive integer;

acquiring paraphrase identification features respectively corresponding to the K interaction records based on natural language processing information of the K interaction records in the learning behavior interaction data;

based on the syntactic structure characteristics of the K interaction records in the learning behavior interaction data, obtaining syntactic structure node vectors respectively corresponding to the K interaction records;

fusing the morpheme representation vector, the paraphrasing recognition feature and the syntactic structure node vector to generate interaction mining features corresponding to the learning behavior interaction data;

and loading the interaction mining features to a document vectorization unit in a target recurrent neural network, and carrying out feature construction on the interaction mining features according to the document vectorization unit to generate learning evaluation description features corresponding to the learning behavior interaction data.

In a possible implementation manner of the first aspect, the performing a heuristic feature search on the learning process description feature, generating a heuristic search feature corresponding to the learning behavior log data, includes:

carrying out knowledge point labeling on the learning process description features, generating knowledge point labeling information corresponding to the learning process description features, and converging the learning process description features and the knowledge point labeling information to generate knowledge point response system structure information;

the method comprises the steps of obtaining initial heuristic features with the same element knowledge points as the learning evaluation description features, and loading the initial heuristic features and the knowledge point answering system structure information into a heuristic feature searching unit in a target recurrent neural network, wherein the initial heuristic features are used for representing answering feature data of element answering knowledge points, corresponding precursor associated answering knowledge points and subsequent associated answering knowledge points and heuristic relations among the element answering knowledge points, corresponding precursor associated answering knowledge points and subsequent associated answering knowledge points;

and circularly updating the initial heuristic features according to the heuristic feature searching unit and the knowledge point answering system structure information to generate heuristic searching features corresponding to the learning behavior log data.

In a possible implementation manner of the first aspect, the learning process description features include M learning description features, where M is a positive integer;

the step of labeling the knowledge points of the learning process description features, and generating knowledge point labeling information corresponding to the learning process description features, includes:

acquiring triggering knowledge points of the M learning description features in the learning behavior log data, and performing knowledge point answering data blocking on the triggering knowledge points of the M learning description features to generate forward triggering knowledge points and backward triggering knowledge points;

performing front knowledge point labeling on the forward trigger knowledge points in the learning process description characteristics, and generating front knowledge point labeling information corresponding to the forward trigger knowledge points;

performing post-knowledge point labeling on the backward triggering knowledge points in the learning process description characteristics, and generating post-knowledge point labeling information corresponding to the backward triggering knowledge points;

and outputting the pre-knowledge point labeling information and the post-knowledge point labeling information as knowledge point labeling information corresponding to the learning process description characteristics.

In a possible implementation manner of the first aspect, the heuristic feature searching unit in the target recurrent neural network includes X first observation subunits and X second observation subunits, where the X first observation subunits and the X second observation subunits are alternately cascaded, X is a positive integer, the first observation subunits are used for performing feature observation of cross-sequence data, and the second observation subunits are used for performing feature observation of local end sequence data;

The step of circularly updating the initial heuristic features according to the heuristic feature searching unit and the knowledge point answering system structure information to generate heuristic searching features corresponding to the learning behavior log data, comprising the following steps:

acquiring data to be observed of an r first observation subunit in the heuristic feature searching unit; when r is 1, the data to be observed of the first observation subunit of the r th comprises the knowledge point response system structure information and the initial heuristic features; when r is not 1, the data to be observed of the first observation subunit with the r number comprises the knowledge point response system structure information and the observation result data of the second observation subunit with the r-1 number; r is a positive integer not greater than X;

acquiring a first observation index array, a second observation index array and a third observation index array corresponding to the r first observation subunit, and selecting feature vectors of the observation result data of the first observation index array and the r-1 second observation subunit to generate a second response relation vector;

performing feature vector selection on the second observation index array and the knowledge point serving as answer system structure information to generate a second test question knowledge point vector, and performing feature vector selection on the third observation index array and the knowledge point serving as answer system structure information to generate a second test question weight vector;

Determining observation result data of the r first observation subunit based on the second answer relation vector, the second test question knowledge point vector and the second test question weight vector;

loading the observation result data of the r first observation subunit into an r second observation subunit in the heuristic feature search unit, and performing self-attention feature extraction on the observation result data of the r first observation subunit according to the r second observation subunit to generate the observation result data of the r second observation subunit;

and outputting the observed result data of the X second observed subunit in the heuristic feature searching unit as heuristic search features corresponding to the learning behavior log data.

In a possible implementation manner of the first aspect, the building a relationship matrix of the target building learning feature to obtain a learning behavior-test question relationship matrix, and generating a user learning portrait for the learning behavior-test question relationship matrix, to generate a user learning portrait feature corresponding to the learning behavior log data, includes:

the relation matrix construction unit is used for loading the target construction learning characteristics into a target recurrent neural network, and carrying out combined characteristic mapping on the target construction learning characteristics according to the relation matrix construction unit to generate a learning behavior-test question relation matrix;

Loading the learning behavior-test question relation matrix into a weak knowledge point estimation unit in the target recurrent neural network, and carrying out weak knowledge point estimation on the learning behavior-test question relation matrix according to the weak knowledge point estimation unit to generate weak knowledge point estimation index distribution;

and performing full-connection output on the weak knowledge point estimation index distribution to obtain a user learning weak point thermodynamic diagram, and determining user learning portrait features corresponding to the learning behavior log data based on the user learning weak point thermodynamic diagram.

For example, in a possible implementation manner of the first aspect, the method further includes:

acquiring template learning behavior data, template interaction data corresponding to the template learning behavior data and template user learning portrait features;

generating template learning process description features corresponding to the template learning behavior data through a context characterization unit in a basic recurrent neural network, and generating template learning evaluation description features corresponding to the template interaction data according to a document vectorization unit in the basic recurrent neural network; the context characterization unit and the document vectorization unit are generated by updating weight parameters according to a plurality of template training sample combinations, wherein one template training sample combination comprises template learning behavior activity data and learning behavior interaction data;

Acquiring a heuristic walk window in a basic heuristic feature search unit contained in the basic recurrent neural network, performing heuristic feature search on the template learning process description feature based on the heuristic walk window and a first observation subunit and a second observation subunit in the basic heuristic feature search unit to generate a template heuristic search feature, and converging the template heuristic search feature and the template learning evaluation description feature to generate a template construction learning feature;

performing feature construction on the template construction learning features according to a basic relation matrix construction unit in the basic recurrent neural network to generate a template learning behavior-test question relation matrix, and generating a user learning portrait prediction sequence corresponding to the template learning behavior-test question relation matrix according to a basic weak knowledge point estimation unit in the basic recurrent neural network; the user learning portrayal prediction sequence reflects the confidence level corresponding to each user learning weak knowledge point in the template user learning portrayal characteristics;

updating weight information of the basic heuristic feature searching unit, the basic relation matrix constructing unit and the basic weak knowledge point estimating unit based on the heuristic walk window, the user learning portrait prediction sequence and the template user learning portrait features, and outputting a basic recurrent neural network containing the updated weight information as a target recurrent neural network; the target recurrent neural network is used for generating user learning portraits for the loaded learning behavior log data and the learning behavior interaction data.

For example, in a possible implementation manner of the first aspect, the performing, based on the heuristic walking window and the first observation subunit and the second observation subunit in the basic heuristic feature searching unit, a heuristic feature search on the template learning process description feature, and generating a template heuristic search feature includes:

acquiring initial template heuristic features with the same meta-knowledge points as the template learning evaluation description features, performing relationship vector diagnosis on the initial template heuristic features and a first feature vector mapping array corresponding to a first observation subunit in the basic heuristic feature search unit, and generating a third response relationship vector;

performing relationship vector diagnosis on the template learning process description feature and a second feature vector mapping array corresponding to a first observation subunit in the basic heuristic feature searching unit to generate a third test question knowledge point vector, and performing relationship vector diagnosis on the template learning process description feature and a third feature vector mapping array corresponding to the first observation subunit in the basic heuristic feature searching unit to generate a third test question weight vector;

Determining a learning behavior focusing vector associated with the template learning process description characteristic based on the third answer relation vector, the third test question knowledge point vector and the third test question weight vector;

activating the heuristic walk window to obtain a learning behavior activation vector, and outputting a fusion vector of the learning behavior activation vector and the learning behavior focusing vector as observation result data of a first observation subunit in the basic heuristic feature search unit;

and loading the observation result data of the first observation subunit in the basic heuristic feature searching unit to a second observation subunit in the basic heuristic feature searching unit, and carrying out self-attention feature extraction on the observation result data of the first observation subunit in the basic heuristic feature searching unit according to the second observation subunit in the basic heuristic feature searching unit to generate template heuristic search features corresponding to the template learning process description features.

For example, in one possible implementation manner of the first aspect, the updating the weight information of the basic heuristic feature searching unit, the basic relation matrix constructing unit and the basic weak knowledge point estimating unit based on the heuristic walking window, the user learning portrait prediction sequence and the template user learning portrait feature, and outputting a basic recurrent neural network including the updated weight information as a target recurrent neural network includes:

Outputting the sum of heuristic parameter values of all the wandering nodes in the heuristic wandering window as a heuristic wandering activation value, and outputting a fusion vector between the heuristic wandering activation value and a set constraint parameter value as a target influence coefficient corresponding to the basic heuristic feature searching unit;

determining initial training error values associated with the basic weak knowledge point estimation unit based on the number of the template learning behavior data and the confidence level of each user learning weak knowledge point in the template user learning portrait characteristic in the user learning portrait prediction sequence;

outputting the sum of the target influence coefficient and the initial training error value as a target training error value corresponding to the basic recurrent neural network;

and optimizing the weight information of the basic heuristic feature searching unit, the basic relation matrix constructing unit and the basic weak knowledge point estimating unit according to the target training error value until the target training error value meets the convergence requirement, and generating the context characterizing unit, the document vectorizing unit, the basic heuristic feature searching unit, the basic relation matrix constructing unit and the basic weak knowledge point estimating unit as a target recurrent neural network when the target training error value is ended.

For example, in a possible implementation manner of the first aspect, the determining the initial training error value associated with the basic weak knowledge point estimation unit based on the number of the template learning behavior data and the confidence level of each user learning weak knowledge point in the template user learning portrayal feature in the user learning portrayal prediction sequence includes:

acquiring confidence degrees corresponding to the weak knowledge points of each user in the template user learning portrait features from the user learning portrait prediction sequence, carrying out logarithmic operation on the confidence degrees corresponding to the weak knowledge points of each user in the template user learning portrait features, and generating logarithmic probability values corresponding to the weak knowledge points of each user in the template user learning portrait features;

and accumulating the logarithmic probability values corresponding to the weak knowledge points learned by each user in the template user learning portrait features to generate a logarithmic probability total value corresponding to the template learning behavior data, and determining the initial training error value associated with the basic weak knowledge point estimation unit based on the ratio between the logarithmic probability total value and the number of the template learning behavior data.

According to one aspect of an embodiment of the present application, there is provided an artificial intelligence based user learning portrayal construction system comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement an artificial intelligence based user learning portrayal construction method in any of the foregoing possible implementations.

According to an aspect of embodiments of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The computer instructions are read from a computer-readable storage medium by a processor of a computer device, and executed by the processor, cause the computer device to perform the methods provided in the various alternative implementations of the above aspects.

In the technical scheme provided by some embodiments of the present application, by acquiring the learning process description feature corresponding to the learning behavior log data and acquiring the learning evaluation description feature of the learning behavior interaction data corresponding to the learning behavior log data, further, heuristic feature search can be performed on the learning process description feature to acquire behavior heuristic information in the learning behavior log data, so that feature expression performance of the learning behavior data feature (such as heuristic search feature) can be improved. The learning evaluation description features and the heuristic search features are converged to generate target construction learning features, the target construction learning features are subjected to relation matrix construction to obtain a learning behavior-test question relation matrix, and user learning portrait generation is further carried out on the learning behavior-test question relation matrix to obtain user learning portrait features corresponding to the learning behavior log data, so that the accuracy of user learning portrait prediction is improved.

Drawings

For a clearer description of the technical solutions of the embodiments of the present application, reference will be made to the accompanying drawings, which are needed to be activated in the embodiments, and it should be understood that the following drawings only illustrate some embodiments of the present application, and therefore should not be considered as limiting the scope, and other related drawings can be extracted by those skilled in the art without the inventive effort.

FIG. 1 is a schematic flow chart of a user learning portrait construction method based on artificial intelligence according to an embodiment of the present application;

fig. 2 is a schematic block diagram of a user learning image construction system based on artificial intelligence for implementing the user learning image construction method based on artificial intelligence according to an embodiment of the present application.

Detailed Description

The following description is presented to enable one of ordinary skill in the art to make and use the application and is provided in the context of a particular application and its requirements. It will be apparent to those having ordinary skill in the art that various changes can be made to the disclosed embodiments and that the general principles defined herein may be applied to other embodiments and applications without departing from the principles and scope of the application. Therefore, the present application is not limited to the described embodiments, but is to be accorded the widest scope consistent with the claims.

FIG. 1 is a flow chart of an artificial intelligence based user learning portrayal construction method according to an embodiment of the present application, and the detailed description of the artificial intelligence based user learning portrayal construction method is provided below.

Step101, learning behavior log data and learning behavior interaction data corresponding to the learning behavior log data are obtained.

For example, the embodiment may obtain learning behavior log data and learning behavior interaction data corresponding to the learning behavior log data from different data sources. The data source may represent a manner of acquiring learning behavior log data and learning behavior interaction data.

Step102, learning process description features corresponding to the learning behavior log data are obtained, and learning evaluation description features corresponding to the learning behavior interaction data are obtained.

In an alternative embodiment, after learning behavior log data and corresponding learning behavior interaction data thereof are acquired, a target recurrent neural network meeting network convergence conditions may be acquired, where the target recurrent neural network may perform feature expression and prediction on the learning behavior log data and the learning behavior interaction data to generate corresponding user learning portrayal features, where the user learning portrayal features may be weak knowledge points existing in the learning behavior log data of the relevant target learning user.

In an alternative embodiment, the context characterization unit may perform feature extraction on each learning behavior activity data in the learning behavior log data, so as to generate a learning process description feature corresponding to the learning behavior log data, where the learning process description feature may reflect a behavior feature in a learning process in the learning behavior log data, for example, a description feature in a test question answering process, a description feature in a course learning process, a description feature in a test question explanation process, and so on. In an alternative embodiment, in the process of extracting features of the learning behavior log data by using the context characterization unit, learning behavior activities of the learning behavior log data may be separated to obtain a learning behavior activity set, so that knowledge point answer data of each learning behavior activity data in the learning behavior activity set may be partitioned to generate a plurality of knowledge point answer data with learning mode types, and a knowledge point answer data sequence corresponding to each learning behavior activity data in the learning behavior activity set may be generated, where the learning mode types may include a pre-learning type, a review type, a test type, and the like. A context characterization unit for acquiring knowledge point test question answering characteristics corresponding to the learning behavior activity data v based on a knowledge point answer data sequence corresponding to the learning behavior activity data v contained in the learning behavior activity set, and loading the knowledge point test question answering characteristics into a target recurrent neural network; the knowledge point test question answering feature can be formed by fusing feature vectors of all knowledge point answering data in a knowledge point answering data sequence corresponding to learning behavior activity data v, wherein v is a positive integer not greater than the number of learning behavior activity data corresponding to a learning behavior activity set. Based on the context characterization unit, carrying out feature construction on the knowledge point test question answer features corresponding to the learning behavior activity data v, and generating learning description features corresponding to the learning behavior activity data v; and further, the learning description features corresponding to the learning behavior activity data in the learning behavior activity set can be aggregated to generate learning process description features corresponding to the learning behavior log data.

In an alternative embodiment, feature extraction may be performed on the learning behavior interaction data by the document vectorization unit, so as to generate learning evaluation description features corresponding to the learning behavior interaction data, where the learning evaluation description features may reflect interaction features in the learning behavior interaction data, for example, evaluation features for guiding a teacher to learn learning behavior results of the user with respect to related targets. In an alternative embodiment, the annotation vectors corresponding to each interaction record contained in the learning behavior interaction data can be fused, so that the interaction mining characteristics corresponding to the learning behavior interaction data can be obtained; and loading the interaction mining features into a document vectorization unit, and carrying out feature construction on the interaction mining features through the document vectorization unit, thereby generating learning evaluation description features corresponding to learning behavior interaction data.

For example, after learning behavior log data and learning behavior interaction data Ab corresponding thereto are acquired, learning behavior log data may be subjected to learning behavior activity separation to obtain learning behavior activity sets, and feature extraction processes in the context characterization unit of each learning behavior activity data in the learning behavior activity sets are identical, and any one of the learning behavior activity data (for example, learning behavior activity data Ba) in the learning behavior activity sets is described below as an example. The learning behavior activity data Ba can be decomposed into knowledge point answer data of a plurality of learning mode categories, a knowledge point answer data sequence Dc corresponding to the learning behavior activity data Ba is generated, and then feature extraction can be performed on each knowledge point answer data in the knowledge point answer data sequence Dc, for example, each knowledge point answer data in the knowledge point answer data sequence Dc can be expressed into a feature sequence, feature sequences corresponding to all knowledge point answer data in the knowledge point answer data sequence Dc are fused into a global feature sequence, and the global feature sequence can be used as a feature of a knowledge point test answer corresponding to the learning behavior activity data Ba. And loading the answer characteristics of the knowledge point test questions corresponding to the learning behavior activity data Ba to a context characterization unit, and performing characteristic construction on the answer characteristics of the knowledge point test questions corresponding to the learning behavior activity data Ba by the context characterization unit to generate learning description characteristics Ed corresponding to the learning behavior activity data Ba. Similarly, each learning behavior activity data in the learning behavior log data can be extracted to the learning description characteristic corresponding to each learning behavior activity data after passing through the context characterization unit, and then the learning description characteristics corresponding to each learning behavior activity data are converged to generate the learning process description characteristic corresponding to the learning behavior log data.

Step103, heuristic feature searching is carried out on the learning process description features, heuristic search features corresponding to the learning behavior log data are generated, and the heuristic search features and the learning evaluation description features are converged to generate target construction learning features.

In an alternative embodiment, since the learning behavior log data is formed by a plurality of continuous learning behavior activity data, the heuristic feature search unit in the target recurrent neural network may perform heuristic feature search on the learning process description feature, so as to generate heuristic search features corresponding to the learning behavior log data.

The heuristic feature search unit may be formed by alternately cascading a plurality of first observation subunits (mutual attention components) and a plurality of second observation subunits (self attention components), wherein the number of first observation subunits is the same as the number of second observation subunits. The first observation subunit in the heuristic feature search unit may introduce an initial heuristic feature, where both the initial heuristic feature and the learning process description feature may be loaded into the heuristic feature search unit, and the initial heuristic feature may be cyclically updated in the heuristic feature search unit according to the learning process description feature, and finally the heuristic search feature is output. The learning process description features corresponding to the learning behavior log data can be loaded into all first observation subunits in the heuristic feature searching unit, and the loaded data of the first observation subunits in the heuristic feature searching unit can comprise observation result data of a previous second observation subunit besides the learning process description features; of course, the loading data of the first observation subunit in the heuristic feature search unit includes a learning process description feature and an initial heuristic feature, and the dimension of the initial heuristic feature is the dimension of the heuristic search feature finally generated by the heuristic search feature. The loading data of each second observation subunit in the heuristic feature search unit is the observation data of the previous first observation subunit.

The heuristic feature searching unit in the target recurrent neural network can fuse the learning process description feature generated by the context characterization unit and the learning evaluation description feature generated by the document vectorization unit, namely the dimension of the heuristic search feature generated by the heuristic feature searching unit is the same as the dimension of the learning evaluation description feature, so that the heuristic search feature and the learning evaluation description feature can be fused and output as a target to construct the learning feature.

Step104, constructing a relation matrix of the target construction learning characteristics to obtain a learning behavior-test question relation matrix, generating a user learning portrait of the learning behavior-test question relation matrix, generating user learning portrait characteristics corresponding to learning behavior log data, and constructing a software pushing frame of the target learning user in the target learning software based on the user learning portrait characteristics corresponding to the learning behavior log data.

In an alternative embodiment, the learning behavior-test question relation matrix is used for representing an array of learning relation feature data between each learning behavior and each corresponding test question, and the learning relation feature data may be used for representing relation features between the learning behavior and knowledge points of each corresponding test question, and may include, for example, root cause features corresponding to correct answer relations, incorrect answer relations, no answer relations, abnormal answer relation relations, and the like. In an alternative embodiment, the target construction learning feature may be loaded to a relationship matrix construction unit in the target recurrent neural network, and the relationship matrix construction unit performs combined feature mapping on the target construction learning feature to generate a learning behavior-test question relationship matrix; loading the learning behavior-test question relation matrix into a weak knowledge point estimation unit in the target recurrent neural network, and carrying out weak knowledge point estimation on the learning behavior-test question relation matrix through the weak knowledge point estimation unit to generate weak knowledge point estimation index distribution; and fully connecting and outputting the weak knowledge point estimation index distribution to obtain a user learning weak point thermodynamic diagram, and determining user learning portrait features corresponding to the learning behavior log data based on the user learning weak point thermodynamic diagram. For example, a target user learning weak point with a thermal value greater than a set thermal value may be loaded into a pushing policy of the target learning user in a software pushing framework in the target learning software, so that content data such as reinforcement learning data and topic data corresponding to the loaded target user learning weak point may be subsequently pushed to the target learning user.

Therefore, learning behavior log data can obtain learning process description features after passing through a context characterization unit in a target recurrent neural network, learning behavior interaction data corresponding to the learning behavior log data can obtain learning evaluation description features after passing through a document vectorization unit, and heuristic feature search is carried out on the learning process description features through a heuristic feature search unit introduced into the target recurrent neural network, so that heuristic search features are generated, and the learning evaluation description features and the heuristic search features are converged to generate target construction learning features; and after the target construction learning features pass through the relation matrix construction unit, determining the user learning portrait features by using the weak knowledge point estimation unit. The heuristic feature searching unit is used for acquiring the behavior heuristic information among the learning behavior activity data in the learning behavior log data, so that the feature expression performance of the learning behavior data features can be improved, and the accuracy of user learning portrait prediction is improved.

Further embodiments of artificial intelligence based user learning portrayal construction methods are described below, which may include steps 201 through 209 below:

step201, learning behavior log data and learning behavior interaction data corresponding to the learning behavior log data are obtained.

Step202, obtaining knowledge point test question answering features corresponding to learning behavior activity data v in learning behavior log data, generating local context Wen Biaozheng vectors corresponding to the knowledge point test question answering features based on a local context window in a context characterization unit, and converging the knowledge point test question answering features and the local context Wen Biaozheng vectors to generate target converging features.

In an alternative embodiment, the local context token vector may refer to a feature having a context in the knowledge point test question answering feature, for example, a context content feature in the knowledge point test question answering process, such as a question reference relationship content feature.

In an alternative embodiment, learning behavior log data may be subjected to learning behavior activity separation to obtain a learning behavior activity set, and knowledge point answer data of each learning behavior activity data in the learning behavior activity set is subjected to knowledge point answer data partitioning to generate a plurality of knowledge point answer data with learning mode types, so as to generate a knowledge point answer data sequence corresponding to each learning behavior activity data in the learning behavior activity set; furthermore, based on the knowledge point answer data sequence corresponding to the learning behavior activity data v in the learning behavior activity set, the knowledge point test question answer characteristics corresponding to the learning behavior activity data v can be obtained, wherein v is a positive integer not greater than the number of the learning behavior activity data contained in the learning behavior activity set.

The following description will take the example that the local context window contains only one second observation subunit. After the knowledge point test question answering feature corresponding to the learning behavior activity data v is input into the context characterization unit, a global feature decomposition structure corresponding to a local context window in the context characterization unit can be obtained, and the knowledge point test question answering feature is decomposed into a first answering relation vector, a first test question knowledge point vector and a first test question weight vector according to the global feature decomposition structure of the local context window.

In an alternative embodiment, a relationship vector diagnosis is performed on the first answer relationship vector and the knowledge point difficulty point vector of the first test question knowledge point vector, and an answer diagnosis vector is generated. The solution capacity factor vector corresponding to the first solution relation vector Q1 can be obtained (the first solution relation vector Q1 and the first test question knowledge point vector K1 have the same solution capacity factor vector); and further, the answer diagnosis vector and the answer capacity factor vector can be subjected to regularized conversion to generate a first test question answer attention vector, and local up-down Wen Biaozheng vectors corresponding to the knowledge point test question answer characteristics are determined based on the fusion vector between the first test question answer attention vector and the first test question weight vector.

In an alternative embodiment, the screening feature between the first test question answering attention vector and the first test question weight vector V1 is output as the observation data of the local context window, where the observation data of the local context window may be used as the local context Wen Biaozheng vector (the local context characterization vector may be referred to as the learning behavior activity data local context Wen Biaozheng vector) corresponding to the knowledge point test question answering feature.

In an alternative embodiment, if the local context window in the context characterization unit may include a plurality of second observation subunits, each of the second observation subunits in the local context window may correspond to one observation data, such as observation data O1, observation data O2, and observation data O3, … …; and then the observation result data corresponding to the plurality of second observation subunits can be fused into local up-down Wen Biaozheng vectors of the learning behavior activity data corresponding to the knowledge point test question answering feature. In an alternative embodiment, the learning behavior activity data generated by the knowledge point test question answer feature and the local context window may be locally aggregated by Wen Biaozheng vectors to generate the target aggregate feature.

Step203, acquiring a covariance feature array and a correlation coefficient array corresponding to a principal component feature analysis subunit in a context characterization unit, determining linear conversion features corresponding to learning behavior activity data v according to the correlation coefficient array and screening features between the covariance feature array and target convergence features, converging the target convergence features and the linear conversion features to generate learning description features corresponding to the learning behavior activity data v, converging learning description features corresponding to each learning behavior activity data in a learning behavior activity set, and generating learning process description features corresponding to learning behavior log data.

In an alternative embodiment, the target convergence feature may be subjected to regularization conversion to generate a regularized target convergence feature, and further, based on the covariance feature array and the correlation coefficient array corresponding to the principal component feature analysis subunit in the context characterization unit, linear conversion is performed on the regularized target convergence feature to obtain a linear conversion feature; or the target convergence feature can be directly subjected to linear conversion based on the covariance feature array and the correlation coefficient array corresponding to the principal component feature analysis subunit in the context characterization unit, so as to obtain a linear conversion feature. In an alternative embodiment, the target convergence feature and the linear transformation feature may be converged to generate a learning description feature corresponding to the learning behavior activity data v.

Therefore, learning description features corresponding to the learning behavior activity data in the learning behavior log data can be obtained, and the learning description features corresponding to the learning behavior activity data in the learning behavior activity set are aggregated to generate learning process description features corresponding to the learning behavior log data.

Step204, obtaining a morpheme representation vector, a paraphrase identification feature and a syntactic structure node vector corresponding to the learning behavior interaction data, and fusing the morpheme representation vector, the paraphrase identification feature and the syntactic structure node vector to generate interaction mining features corresponding to the learning behavior interaction data.

The processing of the learning behavior interaction data by the document vectorization unit may be, for example: carrying out data partitioning on the learning behavior interaction data to generate K interaction records, and obtaining morpheme representation vectors corresponding to the K interaction records respectively; based on natural language processing information of the K interaction records in the learning behavior interaction data, acquiring paraphrasing recognition features respectively corresponding to the K interaction records; based on the syntactic structure characteristics of the K interaction records in the learning behavior interaction data, obtaining syntactic structure node vectors respectively corresponding to the K interaction records; fusing the morpheme representation vector, the paraphrasing recognition feature and the syntactic structure node vector to generate interaction mining features corresponding to the learning behavior interaction data; and loading the interaction mining features into a document vectorization unit in the target recurrent neural network, and performing feature construction on the interaction mining features through the document vectorization unit to generate learning evaluation description features corresponding to learning behavior interaction data.

Step205, loading the interactive mining features into a document vectorization unit in the target recurrent neural network, and performing feature construction on the interactive mining features through the document vectorization unit to generate learning evaluation description features corresponding to learning behavior interactive data.

Step206, carrying out knowledge point labeling on the learning process description features, generating knowledge point labeling information corresponding to the learning process description features, and converging the learning process description features and the knowledge point labeling information to generate knowledge point response system structure information.

In an alternative embodiment, after the context characterization unit in the target recurrent neural network obtains the learning process description feature corresponding to the learning behavior log data, the heuristic feature search unit in the target recurrent neural network may be used to obtain behavior heuristic information between the learning behavior activity data included in the learning behavior log data. In order to keep the behavior heuristic information of the learning process description feature, the learning process description feature can be labeled with knowledge points, knowledge point labeling information corresponding to the learning process description feature is generated, and the learning process description feature and the knowledge point labeling information are added to obtain knowledge point response system structure information.

Wherein, the learning process description feature is assumed to comprise M learning description features, M represents the number of the learning description features contained in the learning process description feature; the learning process describes the feature and the positive post knowledge point labeling process can comprise the following steps: acquiring triggering knowledge points of M learning description features in learning behavior log data, and partitioning knowledge point answering data by the triggering knowledge points of the M learning description features to generate forward triggering knowledge points and backward triggering knowledge points; performing front knowledge point labeling on the forward trigger knowledge points in the learning process description characteristics, and generating front knowledge point labeling information corresponding to the forward trigger knowledge points; the backward triggering knowledge points in the description characteristics of the learning process are marked, and backward knowledge point marking information corresponding to the backward triggering knowledge points is generated; and outputting the front knowledge point labeling information and the rear knowledge point labeling information as knowledge point labeling information corresponding to the learning process description characteristics.

Step207, obtaining initial heuristic features with the same element knowledge points as the learning evaluation description features, and loading the initial heuristic features and knowledge point answer system structure information into a heuristic feature searching unit in the target recurrent neural network.

In an alternative embodiment, an initial heuristic feature with the same meta-knowledge point as the learning evaluation description feature may be obtained in the heuristic feature searching unit, and the initial heuristic feature that is finally circularly updated in the heuristic feature searching unit may be used as a heuristic search feature corresponding to the learning behavior log data. The initial heuristic features and knowledge point answering architecture information may be loaded into a heuristic feature search unit in a target recurrent neural network.

Step208, circularly updating the initial heuristic features through the heuristic feature searching unit and the knowledge point answering system structure information to generate heuristic searching features corresponding to the learning behavior log data.

In an alternative embodiment, the heuristic feature searching unit in the target recurrent neural network may include X first observation subunits and X second observation subunits, where the X first observation subunits and the X second observation subunits in the heuristic feature searching unit are alternately cascaded, and the initial heuristic feature may be cyclically updated according to knowledge point answer system structure information including learning process description features and knowledge point labeling information, and the X first observation subunits and the X second observation subunits in the heuristic feature searching unit, so as to finally generate heuristic search features corresponding to learning behavior log data.

The processing steps of the knowledge point answering system structure information and the initial heuristic features in the heuristic feature searching unit can be as follows: acquiring data to be observed of an r first observation subunit in the heuristic feature searching unit; when r is 1, the data to be observed of the r first observation subunit (namely, the first observation subunit in the heuristic feature searching unit) comprises knowledge point answering architecture information and initial heuristic features; when r is not 1, the data to be observed of the r first observation subunit (any one of the second first observation subunit to the X first observation subunit in the heuristic feature search unit) comprises knowledge point response architecture information and observation result data of the r-1 second observation subunit; r is a positive integer not greater than X. Then, a first observation index array, a second observation index array and a third observation index array corresponding to the r first observation subunit are obtained, feature vector selection is carried out on observation result data of the first observation index array and the r-1 second observation subunit, and a second response relation vector is generated; and selecting the feature vector of the second observation index array and the knowledge point answering system structure information to generate a second test question knowledge point vector, and selecting the feature vector of the third observation index array and the knowledge point answering system structure information to generate a second test question weight vector. Based on the second answer relation vector, the second test question knowledge point vector and the second test question weight vector, observation data of the r first observation subunit can be determined. The observation result data of the r first observation subunit is loaded to the r second observation subunit in the heuristic feature searching unit, and the self-attention feature extraction is carried out on the observation result data of the r first observation subunit through the r second observation subunit, so that the observation result data of the r second observation subunit is generated; and outputting the observed result data of the X second observed subunit in the heuristic feature searching unit as heuristic search features corresponding to the learning behavior log data.

Assuming that the heuristic feature search unit of the target recurrent neural network may be alternately cascaded by the X first observation subunits and the X second observation subunits, for example, when the value of X is 6, the 6 first observation subunits and the 6 second observation subunits may be alternately cascaded. The X first observation subunits may be denoted as first observation subunit 1, first observation subunit 2, … …, first observation subunit X; the X second observation subunits may be denoted as second observation subunit 1, second observation subunit 2, … …, second observation subunit X.

Step209, constructing a relation matrix of the target construction learning characteristic to obtain a learning behavior-test question relation matrix, and generating a user learning portrait of the learning behavior-test question relation matrix to generate a user learning portrait characteristic corresponding to the learning behavior log data.

The learning behavior log data can obtain the learning process description characteristic after passing through a context characterization unit in the target recurrent neural network, and the learning behavior interaction data corresponding to the learning behavior log data can obtain the learning evaluation description characteristic after passing through a document vectorization unit; furthermore, heuristic feature searching is carried out on the description features of the learning process through a heuristic feature searching unit introduced into the target recurrent neural network so as to obtain heuristic search features, and the learning evaluation description features and the heuristic search features are converged to generate target construction learning features; and after the target construction learning features pass through the relation matrix construction unit, determining the user learning portrait features by using the weak knowledge point estimation unit. The heuristic feature searching unit is used for acquiring the behavior heuristic information among the learning behavior activity data in the learning behavior log data, so that the feature expression performance of the learning behavior data features can be improved, and the accuracy of user learning portrait prediction is improved; according to the user learning image characteristics of the learning behavior log data, the transmission performance of the learning behavior log data can be improved.

Further embodiments of artificial intelligence based user learning portrayal construction methods are described below, including steps 301 through 305 below:

step301, obtaining template learning behavior data, template interaction data corresponding to the template learning behavior data and template user learning portrait features.

In an alternative embodiment, during the training process of the basic recurrent neural network, a sample sequence for training the basic recurrent neural network may be obtained, and each sample in the sample sequence may be composed of a template learning behavior data, a template interaction data and a template user learning portrait feature; the template learning behavior data and the template interaction data in the sample can be used as loading data of the basic recurrent neural network, and the template user learning image characteristics can be used as training label information of the sample to update weight information of the basic recurrent neural network.

Step302, generating template learning process description features corresponding to template learning behavior data through a context characterization unit in the basic recurrent neural network, and generating template learning evaluation description features corresponding to template interaction data through a document vectorization unit in the basic recurrent neural network; the context characterization unit and the document vectorization unit are generated by updating weight parameters according to a plurality of template training sample combinations, wherein one template training sample combination comprises template learning behavior activity data and learning behavior interaction data.

In an alternative embodiment, the network framework of the basic recurrent neural network may include a context characterization unit, a document vectorization unit, a basic heuristic feature search unit, a basic relationship matrix construction unit, and a basic weak knowledge point estimation unit. The template learning behavior data and the template interaction data which are input into the basic recurrent neural network are respectively extracted by a context characterization unit and a document vectorization unit, the context characterization unit and the document vectorization unit are generated by updating weight parameters in advance according to a plurality of template training sample combinations, and one template training sample combination comprises one template learning behavior activity data and one learning behavior interaction data.

After the template learning behavior data, the corresponding template interaction data and the template user learning portrait characteristic are obtained, the template learning behavior data can be loaded to a pre-trained context characterization unit in the basic recurrent neural network, feature extraction is carried out on each learning behavior activity data in the template learning behavior data through the context characterization unit, learning description features respectively corresponding to each learning behavior activity data in the template learning behavior data are generated, and then the learning description features corresponding to each learning behavior activity data in the template learning behavior data are converged to generate template learning process description features.

Step303, obtaining a heuristic walk window in a basic heuristic feature search unit contained in the basic recurrent neural network, performing heuristic feature search on template learning process description features based on the heuristic walk window and a first observation subunit and a second observation subunit in the basic heuristic feature search unit, generating template heuristic search features, and converging the template heuristic search features and template learning evaluation description features to generate template construction learning features.

In an alternative embodiment, the basic heuristic feature searching unit in the basic recurrent neural network may include X first observation subunits and X second observation subunits, where the X first observation subunits and the X second observation subunits in the heuristic feature searching unit are alternately cascaded. After the template learning process description feature is loaded to the basic heuristic feature searching unit, the specific implementation flow in the basic heuristic feature searching unit may include: the method comprises the steps that a heuristic walk window in a basic heuristic feature searching unit contained in a basic recurrent neural network can be obtained, initial template heuristic features with the same element knowledge points as template learning evaluation description features are obtained, relationship vector diagnosis is conducted on the initial template heuristic features and a first feature vector mapping array corresponding to a first observation subunit in the basic heuristic feature searching unit, and a third response relationship vector is generated; performing relationship vector diagnosis on the template learning process description feature and a second feature vector mapping array corresponding to a first observation subunit in the basic heuristic feature searching unit to generate a third test question knowledge point vector, and performing relationship vector diagnosis on the template learning process description feature and a third feature vector mapping array corresponding to the first observation subunit in the basic heuristic feature searching unit to generate a third test question weight vector; determining a learning behavior focusing vector associated with the description features of the template learning process based on the third answer relation vector, the third test question knowledge point vector and the third test question weight vector; activating the heuristic walk window to obtain a learning behavior activation vector, outputting a fusion vector of the learning behavior activation vector and a learning behavior focusing vector as observation result data of a first observation subunit in the basic heuristic feature search unit; and loading the observation result data of the first observation subunit in the basic heuristic feature searching unit to the second observation subunit in the basic heuristic feature searching unit, and carrying out self-attention feature extraction on the observation result data of the first observation subunit in the basic heuristic feature searching unit through the second observation subunit in the basic heuristic feature searching unit to generate template heuristic search features corresponding to the template learning process description features.

Step304, performing feature construction on the template construction learning features through a basic relation matrix construction unit in the basic recurrent neural network to generate a template learning behavior-test question relation matrix, and generating a user learning portrait prediction sequence corresponding to the template learning behavior-test question relation matrix through a basic weak knowledge point estimation unit in the basic recurrent neural network; the user learning portrayal prediction sequence reflects the confidence level corresponding to each user learning weak knowledge point in the template user learning portrayal features.

In an alternative embodiment, after the template construction learning features corresponding to the template learning behavior data and the template interaction data are obtained, the template construction learning features may be loaded into a generating sub-network in the basic recurrent neural network, and the template construction learning features are subjected to feature construction by a basic relation matrix construction unit in the generating sub-network to generate a template learning behavior-test question relation matrix. And then, loading the template learning behavior-test question relation matrix into a basic weak knowledge point estimation unit in a generation sub-network, and generating a user learning portrait prediction sequence corresponding to the template learning behavior-test question relation matrix through the basic weak knowledge point estimation unit, wherein the user learning portrait prediction sequence reflects the confidence degree corresponding to each user learning weak knowledge point in the template user learning portrait characteristics. In other words, the basic weak knowledge point estimation unit may perform user learning portrait generation on the template learning behavior-test question relation matrix, output a user learning portrait prediction sequence by a terminal parameter layer (such as a softmax layer) of the basic weak knowledge point estimation unit, and aggregate weak mastery knowledge points corresponding to the maximum confidence degrees at each weak mastery knowledge point to generate estimated weak mastery knowledge points corresponding to the template learning behavior data by the confidence degrees corresponding to each weak mastery knowledge point included in the user learning portrait prediction sequence.

Step305, updating weight information of the basic heuristic feature searching unit, the basic relation matrix constructing unit and the basic weak knowledge point estimating unit based on the heuristic walk window, the user learning portrayal prediction sequence and the template user learning portrayal feature, and outputting a basic recurrent neural network containing the updated weight information as a target recurrent neural network.

In an alternative embodiment, the sum of heuristic parameter values of each wandering node in the heuristic wandering window may be output as a heuristic wandering activation value, and the fusion vector between the heuristic wandering activation value and the set constraint parameter value may be output as a target influence coefficient corresponding to the heuristic feature searching unit.

The initial training error value associated with the underlying weak knowledge point estimation unit may be determined based on the number of template learning behavior data and the confidence level of each user learning weak knowledge point in the template user learning portrayal feature in the user learning portrayal prediction sequence. For example, a confidence level corresponding to each user learning weak knowledge point in the template user learning portraits feature may be obtained in the user learning portraits prediction sequence, and a log operation may be performed on the confidence level corresponding to each user learning weak knowledge point in the template user learning portraits feature to generate a log probability value corresponding to each user learning weak knowledge point in the template user learning portraits feature; and accumulating the logarithmic probability values corresponding to the weak knowledge points learned by each user in the template user learning portrait features to generate a logarithmic probability total value corresponding to the template learning behavior data, and determining the initial training error value associated with the basic weak knowledge point estimation unit based on the ratio between the logarithmic probability total value and the number of the template learning behavior data.

Fig. 2 illustrates a hardware structural intent of an artificial intelligence based user learning portrayal construction system 100 for implementing the above-described artificial intelligence based user learning portrayal construction method according to an embodiment of the present application, as shown in fig. 2, the artificial intelligence based user learning portrayal construction system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130 and a communication unit 140.

In an alternative embodiment, the artificial intelligence based user learning portrayal construction system 100 may be a single server or a group of servers. The set of servers may be centralized or distributed (e.g., the artificial intelligence based user learning portrayal construction system 100 may be a distributed system). In an alternative embodiment, the artificial intelligence based user learning portrayal construction system 100 may be local or remote. For example, the artificial intelligence based user learning portrayal construction system 100 can access information and/or data stored in the machine readable storage medium 120 via a network. As another example, the artificial intelligence based user learning portrayal construction system 100 can be directly connected to the machine readable storage medium 120 to access stored information and/or data. In an alternative embodiment, the artificial intelligence based user learning portrayal construction system 100 may be implemented on a cloud platform. For example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

The machine-readable storage medium 120 may store data and/or instructions. In an alternative embodiment, the machine-readable storage medium 120 may store data acquired from an external terminal. In an alternative embodiment, machine-readable storage medium 120 may store data and/or instructions for use by artificial intelligence based user learning portrayal construction system 100 to perform or use to perform the exemplary methods described herein. In alternative embodiments, machine-readable storage medium 120 may include mass storage, removable storage, volatile read-write memory, read-only memory, and the like, or any combination thereof. Exemplary mass storage devices may include magnetic disks, optical disks, solid state disks, and the like. Exemplary removable memory may include flash drives, floppy disks, optical disks, memory cards, compact disks, tape, and the like.

In a specific implementation, the plurality of processors 110 execute computer executable instructions stored by the machine-readable storage medium 120, so that the processors 110 may execute the user learning portrait construction method based on artificial intelligence according to the method embodiment, where the processors 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processors 110 may be used to control the transceiving actions of the communication unit 140.

The specific implementation process of the processor 110 may refer to the above embodiments of the method executed by the user learning portrait construction system 100 based on artificial intelligence, and the implementation principle and technical effect are similar, which is not described herein again.

In addition, the embodiment of the application also provides a readable storage medium, wherein computer executable instructions are preset in the readable storage medium, and when a processor executes the computer executable instructions, the user learning portrait construction method based on artificial intelligence is realized.

Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof. Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof.

Claims

1. The method for constructing the user learning image based on the artificial intelligence is characterized by comprising the following steps of:

2. The method for constructing a user learning image based on artificial intelligence according to claim 1, wherein the obtaining the learning process description feature corresponding to the learning behavior log data includes:

3. The method for constructing a user learning image based on artificial intelligence according to claim 2, wherein the feature construction is performed on the answer feature of the knowledge point test question based on the context characterization unit, and the learning description feature corresponding to the learning behavior activity data v is generated, and the method comprises the following steps: generating a local context Wen Biaozheng vector corresponding to the knowledge point test question answering feature based on a local context window in the context characterization unit, and converging the knowledge point test question answering feature and the local context Wen Biaozheng vector to generate a target converging feature; acquiring a covariance feature array and a correlation coefficient array corresponding to a principal component feature analysis subunit in the context characterization unit, determining a linear conversion feature corresponding to the learning behavior activity data v according to the correlation coefficient array and screening features between the covariance feature array and the target convergence feature, and converging the target convergence feature and the linear conversion feature to generate a learning description feature corresponding to the learning behavior activity data v; the generating a local context Wen Biaozheng vector corresponding to the answer feature of the knowledge point test question based on the local context window in the context characterization unit includes: acquiring a global feature decomposition structure corresponding to a local context window in the context characterization unit, and decomposing the knowledge point test question answering feature into a first answering relation vector, a first test question knowledge point vector and a first test question weight vector according to the global feature decomposition structure of the local context window; performing relationship vector diagnosis on the first answer relationship vector and the knowledge point difficulty point vector of the first test question knowledge point vector to generate an answer diagnosis vector, and acquiring a solution capability factor vector corresponding to the first answer relationship vector; and carrying out regularized conversion on the co-occurrence feature vector of the answer diagnosis vector and the answer capacity factor vector to generate a first test question answer attention vector, and determining local up-down Wen Biaozheng vectors corresponding to the knowledge point test question answer features based on a fusion vector between the first test question answer attention vector and the first test question weight vector.

4. The method for constructing a learning image of a user based on artificial intelligence according to claim 1, wherein the obtaining learning evaluation description features corresponding to the learning behavior interaction data includes:

5. The method for constructing a user learning image based on artificial intelligence according to claim 1, wherein the performing heuristic feature search on the learning process description feature to generate heuristic search features corresponding to the learning behavior log data includes:

6. The method for constructing an artificial intelligence based user learning image according to claim 5, wherein the learning process description features include M learning description features, M being a positive integer;

7. The method for constructing a user learning image based on artificial intelligence according to claim 5, wherein the heuristic feature search unit in the target recurrent neural network comprises X first observation subunits and X second observation subunits, the X first observation subunits and the X second observation subunits are alternately cascaded, X is a positive integer, the first observation subunits are used for performing feature observation of cross-sequence data, and the second observation subunits are used for performing feature observation of local-end sequence data;

8. The method for constructing a user learning portrait based on artificial intelligence according to claim 1, wherein the constructing a relation matrix for the target constructed learning feature to obtain a learning behavior-test question relation matrix, generating a user learning portrait for the learning behavior-test question relation matrix, and generating a user learning portrait feature corresponding to the learning behavior log data includes:

9. The artificial intelligence based user learning image construction method of claim 1, further comprising:

performing feature construction on the template construction learning features according to a basic relation matrix construction unit in the basic recurrent neural network to generate a template learning behavior-test question relation matrix, and generating a user learning portrait prediction sequence corresponding to the template learning behavior-test question relation matrix according to a basic weak knowledge point estimation unit in the basic recurrent neural network; the user learning portrayal prediction sequence reflects the confidence level corresponding to each user learning weak knowledge point in the template user learning portrayal characteristics; updating weight information of the basic heuristic feature searching unit, the basic relation matrix constructing unit and the basic weak knowledge point estimating unit based on the heuristic walk window, the user learning portrait prediction sequence and the template user learning portrait features, and outputting a basic recurrent neural network containing the updated weight information as a target recurrent neural network; the target recurrent neural network is used for generating user learning portraits for the loaded learning behavior log data and the learning behavior interaction data; the heuristic feature searching for the template learning process description feature based on the heuristic walk window and the first observation subunit and the second observation subunit in the basic heuristic feature searching unit, generating a template heuristic search feature, includes: acquiring initial template heuristic features with the same meta-knowledge points as the template learning evaluation description features, performing relationship vector diagnosis on the initial template heuristic features and a first feature vector mapping array corresponding to a first observation subunit in the basic heuristic feature search unit, and generating a third response relationship vector; performing relationship vector diagnosis on the template learning process description feature and a second feature vector mapping array corresponding to a first observation subunit in the basic heuristic feature searching unit to generate a third test question knowledge point vector, and performing relationship vector diagnosis on the template learning process description feature and a third feature vector mapping array corresponding to the first observation subunit in the basic heuristic feature searching unit to generate a third test question weight vector; determining a learning behavior focusing vector associated with the template learning process description characteristic based on the third answer relation vector, the third test question knowledge point vector and the third test question weight vector; activating the heuristic walk window to obtain a learning behavior activation vector, and outputting a fusion vector of the learning behavior activation vector and the learning behavior focusing vector as observation result data of a first observation subunit in the basic heuristic feature search unit; and loading the observation result data of the first observation subunit in the basic heuristic feature searching unit to a second observation subunit in the basic heuristic feature searching unit, and carrying out self-attention feature extraction on the observation result data of the first observation subunit in the basic heuristic feature searching unit according to the second observation subunit in the basic heuristic feature searching unit to generate template heuristic search features corresponding to the template learning process description features.

10. An artificial intelligence based user learning portrayal construction system comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the artificial intelligence based user learning portrayal construction method of any one of claims 1-9.