Deep learning knowledge point recommendation method and device based on learning demand perception

Abstract

The present invention discloses a deep learning knowledge point recommendation method and device based on learning demand perception, including: converting a plurality of different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network; initializing the features of the user and the features of the knowledge points to obtain the initial embedding expression vectors of the corresponding user and the initial embedding expression vectors of the knowledge points; initializing the meta-path features of the heterogeneous association relationships to obtain the initial feature matrices corresponding to different heterogeneous association relationships; obtaining the static learning demand embedding expression vector and the knowledge point embedding expression vector of the target user; obtaining the dynamic learning demand embedding expression vector of the target user, and obtaining the learning demand embedding expression vector of the target user; obtaining the knowledge point learning times according to the learning demand embedding expression vector of the target user and the knowledge point embedding expression vector, and performing matrix decomposition on the knowledge point learning times to obtain the recommendation result. The present invention can improve the accuracy of the recommendation result.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence online education, and particularly relates to a deep learning knowledge point recommendation method and device based on learning demand perception. Background Art

[0002] The rapid development of information technology has led to a new round of educational reform. The traditional offline education model is gradually transforming into online intelligent education, and online learning has become one of the important ways for users to learn knowledge. Among them, online learning platforms (MOOCs) have successively launched intelligent recommendation functions to provide users with personalized learning resource recommendations to improve users' learning efficiency and learning effects.

[0003] In the prior art, although the provided methods can already calculate the recommendation intensity of knowledge points based on the relationships between knowledge points to improve the performance of recommendations, they lack consideration of various interaction relationships between knowledge points that play a key role in users' learning effects in the educational scenario, such as prerequisite and subsequent, complementary, etc. relationships. The associated relationships between knowledge points are the key to determining users' static learning needs. In addition, although other existing recommendation methods attempt to integrate more heterogeneous interaction relationships, they cannot capture the dynamic changes in learners' learning needs in the educational recommendation scenario, that is, they lack consideration of the dynamics of users' learning needs during the learning process.

[0004] Therefore, it is necessary to continue to improve the above-mentioned defects existing in the prior art. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a deep learning knowledge point recommendation method and device based on learning demand perception. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] In a first aspect, the present invention provides a deep learning knowledge point recommendation method based on learning demand perception, including:

[0007] Converting a number of different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network; wherein, the heterogeneous knowledge network includes multiple entity nodes, namely users, knowledge points, courses, videos or teachers, and heterogeneous association relationships are formed by connecting different entity nodes;

[0008] Initializing the features of the user and the features of the knowledge points in the heterogeneous knowledge network to obtain the initial embedding expression vectors of the corresponding user and the initial embedding expression vectors of the knowledge points;

[0009] Initializing the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network to obtain the initial feature matrices corresponding to different heterogeneous association relationships;

[0010] Obtain the static learning requirement embedding expression vector of the target user and the knowledge point embedding expression vector respectively according to the initial embedding expression vector of the user, the initial embedding expression vector of the knowledge point, and the initial feature matrices corresponding to different heterogeneous association relationships;

[0011] Obtain the dynamic learning requirement embedding expression vector of the target user according to the knowledge point embedding expression vector, and splice the static learning requirement embedding expression vector of the target user and the dynamic learning requirement embedding expression vector of the target user to obtain the learning requirement embedding expression vector of the target user;

[0012] Obtain the knowledge point learning times according to the learning requirement embedding expression vector of the target user and the knowledge point embedding expression vector, perform matrix decomposition on the knowledge point learning times to obtain a knowledge point recommendation list, and obtain a recommendation result from the knowledge point recommendation list.

[0013] In a second aspect, the present invention also provides a deep learning knowledge point recommendation device based on learning requirement perception, including:

[0014] A heterogeneous knowledge network construction module, configured to convert several different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network; wherein, the heterogeneous knowledge network includes multiple entity nodes, namely users, knowledge points, courses, videos or teachers respectively, and heterogeneous association relationships are formed by connecting different entity nodes;

[0015] A content feature initialization module, configured to initialize the features of the user and the knowledge points in the heterogeneous knowledge network to obtain the corresponding initial embedding expression vectors of the user and the knowledge points;

[0016] A meta-path feature initialization module, configured to initialize the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network to obtain initial feature matrices corresponding to different heterogeneous association relationships;

[0017] A static learning requirement perception module, configured to obtain the static learning requirement embedding expression vector of the target user and the knowledge point embedding expression vector respectively according to the initial embedding expression vector of the user, the initial embedding expression vector of the knowledge point, and the initial feature matrices corresponding to different heterogeneous association relationships;

[0018] A dynamic learning requirement perception module, configured to obtain the dynamic learning requirement embedding expression vector of the target user according to the knowledge point embedding expression vector, and splice the static learning requirement embedding expression vector of the target user and the dynamic learning requirement embedding expression vector of the target user to obtain the learning requirement embedding expression vector of the target user;

[0019] The knowledge point recommendation module is used to obtain the knowledge point learning times by embedding the expression vector and the knowledge point embedding expression vector according to the learning needs of the target user, perform matrix factorization on the knowledge point learning times, obtain a knowledge point recommendation list, and obtain the recommendation result from the knowledge point recommendation list.

[0020] Advantages of the present invention:

[0021] A deep learning knowledge point recommendation method and device based on learning demand perception provided by the present invention use a meta-path-guided graph convolutional network combined with an attention heterogeneous relationship network to obtain the static learning demand embedding expression vector and the knowledge point embedding expression vector of the target user, aggregate the knowledge point embedding expressions at different times to obtain the knowledge level embedding expressions at different times, and then use a gated recurrent unit network to model the potential dynamic learning demand embedding expressions in the knowledge level embedding expression sequence. It not only considers the complex correlation relationships between knowledge points but also can perceive the dynamic changes of user learning needs, and provides a more fine-grained target user learning demand embedding expression vector in real time and dynamically, effectively ensuring the rationality and accuracy of the recommendation result; adopting a matrix factorization model based on the user's learning times can more accurately mine the user's repeated learning behavior, thereby improving the accuracy of the recommendation result.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0023] Figure 1 is a flowchart of a deep learning knowledge point recommendation method based on learning demand perception provided by an embodiment of the present invention;

[0024] Figure 2 is another schematic diagram of a deep learning knowledge point recommendation method based on learning demand perception provided by an embodiment of the present invention;

[0025] Figure 3 is a schematic diagram of a deep learning knowledge point recommendation device based on learning demand perception provided by an embodiment of the present invention. Detailed Embodiments

[0026] The present invention will be further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0027] In the related art, the patent "Knowledge Point Recommendation System Based on Deep Learning Method" (Application No.: 202110168685.0) discloses a real-time knowledge point recommendation system that can evaluate the ability values of the user's recent knowledge points from multiple dimensions and update the database in real time to provide knowledge point recommendations; the patents "Knowledge Point Recommendation Method, Device, Storage Medium and Electronic Device Based on Knowledge Graph" (Application No.: 202110084853.8) and "Knowledge Point Recommendation Method, Device, Storage Medium and Electronic Device" (Application No.: 202110084854.2) both disclose knowledge point recommendation methods based on knowledge graphs of knowledge points, which can model the user's interest preferences based on the knowledge points learned by the user in the past and determine the knowledge connection degree between the learned knowledge points and the unlearned knowledge points based on the knowledge graph of knowledge points, and finally provide knowledge point recommendations through linear weight fusion; the patent "A Key Knowledge Point Recommendation Method and Its System" (Application No.: 201310456539.3) discloses a key knowledge point recommendation method that can calculate the recommendation intensity of knowledge points based on the knowledge point relationship between knowledge points to improve the performance of recommendations. The patent "A Knowledge Point Recommendation Method Based on Implicit Attribute and Implicit Relationship Mining" (Application No.: 201711107927.5) discloses a knowledge point recommendation method based on implicit attribute and implicit relationship mining. This method integrates more implicit attributes of users and implicit association relationships between users, and can more accurately model the characteristics of users and knowledge points; the patent "A Knowledge Point Recommendation Method and System for Subject Questions Based on Clustering Algorithm" (Application No.: 202110039634.8) discloses a knowledge point recommendation method for subject questions based on clustering algorithm. This method can enrich the feature expression of knowledge points by supplementing the knowledge point characteristics of previous years and this year, as well as the knowledge point trend characteristics in real questions; the patent "A Knowledge Point Recommendation Method, Device, Electronic Equipment and Storage Medium" (Application No.: 202110208825.2) discloses a knowledge point recommendation method that can extract additional knowledge point characteristics from teaching and research videos and use the attention mechanism to score the importance of knowledge points, improving the performance of recommendations; the patent "Knowledge Point Recommendation Method, Device, Terminal and Computer Readable Storage Medium" (Application No.: 202110872968.3) discloses a method for recommending knowledge points using wrong questions in test questions. This method supplements the user's wrong question information to enrich the user's feature expression; the patent "Knowledge Point Recommendation Method, Device and Storage Medium" (Application No.: 202210081389.1) discloses a knowledge point recommendation method that supplements practice questions and the user's cognitive state characteristics in practice questions in addition to the interaction data between users and knowledge points to achieve recommendations. The recommendation methods disclosed in the above patents supplement more additional attribute and relationship characteristics on the basis of the interaction data between users and knowledge points to improve the performance of recommendations.

[0028] In summary, on the one hand, existing methods lack consideration of the multiple entities and heterogeneous relationships in the user learning interaction process when perceiving learning needs, especially the precedence and succession relationships between knowledge points. Although some patents have supplemented the association relationship between knowledge points, they only consider the historical interaction data between users and knowledge points, and lack consideration of the heterogeneous interaction relationships between users, courses, videos, knowledge point content, and teachers in the learning process, especially the precedence and succession relationships between knowledge points, resulting in the recommendation performance of previous recommendation methods still need to be improved. On the other hand, although other existing recommendation methods try to integrate more heterogeneous interaction relationships, they lack consideration of the dynamic nature of users' learning needs in the learning process; these recommendation methods regard the user's historical learning process as a static process, lack consideration of the sequential nature of user learning behaviors, and cannot model the dynamic nature of user learning needs.

[0029] In view of this, the present invention provides a deep learning knowledge point recommendation method based on learning need perception, which initializes the content features of entities and the meta-path features of heterogeneous relationships based on the heterogeneous knowledge network of the user's historical learning behavior; adopts the meta-path-based attention graph convolutional network to perceive the static learning needs in the user's historical learning behavior, and then adopts the gated recurrent unit network to perceive the user's dynamic learning needs in different time stages; based on the user's historical learning interaction, the user's knowledge point learning frequency matrix is ​​matrix decomposed to obtain subsequent knowledge point set recommendations; thereby improving the accuracy of the user's learning need representation, and being able to more accurately mine the user's repeated learning behavior, thereby improving the accuracy of the recommendation results.

[0030] See also Figure 1 and Figure 2 As shown, Figure 1 is a flow chart of a deep learning knowledge point recommendation method based on learning needs perception provided by an embodiment of the present invention, Figure 2 is another schematic diagram of a deep learning knowledge point recommendation method based on learning needs perception provided by an embodiment of the present invention. A deep learning knowledge point recommendation method based on learning needs perception provided by the present invention includes:

[0031] S101. Convert several different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network; wherein the heterogeneous knowledge network includes multiple entity nodes, namely users, knowledge points, courses, videos or teachers, and different entity nodes are connected to form heterogeneous association relationships.

[0032] Specifically, in this embodiment, a heterogeneous knowledge network is constructed by using the heterogeneous association relationships and semantic feature information contained in several different types of entity interaction data in the user's historical learning behavior. Among them, the heterogeneous knowledge network includes multiple nodes, and the nodes represent different entities, namely users, knowledge points, courses, videos, teachers, etc. The connections between the nodes in the heterogeneous knowledge network represent the heterogeneous association relationships between various entities. It should be noted that users, knowledge point content, courses, videos, and teachers all belong to entities, and the content to be learned can be obtained through users, knowledge point content, courses, videos, and teachers.

[0033] S102. Initialize the features of the user and the knowledge points in the heterogeneous knowledge network to obtain the initial embedded expression vectors of the corresponding user and the knowledge points.

[0034] Specifically, in this embodiment, according to the features of the user in the heterogeneous knowledge network, the first meta-path-guided graph convolutional network is used for processing to obtain the initial embedded expression vector of the user. According to the initial embedded expression vector of the knowledge points, the second meta-path-guided graph convolutional network is used for processing to obtain the initial embedded expression vector of the knowledge points. For example, multi-hop encoding (Multi-hot encoding) is performed on the "user-course" interaction subgraph to initialize the features of the user and obtain the initial embedded expression vector of the user. At the same time, the name of the knowledge point is encoded using Word2Vec to initialize the features of the knowledge point and obtain the initial embedded expression vector of the knowledge point.

[0035] In this embodiment, the expression of the initial embedded expression vector of the user is:

[0036] f user = D -1 · A uc ;

[0037] D = diag(A uc 1);

[0038] The expression of the initial embedded expression vector of the knowledge points is:

[0039] f concept = Word2vec(concept);

[0040] Among them, f user is the content feature embedded expression vector of the user, A uc is the adjacency matrix for interacting with the course. If the user has learned the course, the corresponding element value is 1. If the user has not learned the course, the corresponding element is 0. D is the degree matrix of the adjacency matrix, and all the diagonal elements in this matrix represent the number of edges associated with the user node corresponding to this element. (·) -1 is to take the inverse, fconcept Let the content feature embedding expression vector of the knowledge point be \(\mathbf{v}\), Word2vec(·) be the method of converting the name of the knowledge point into a vector using the Word2vector word vectorization method, concept be the name of the knowledge point, diag(·) be the matrix diagonalization operation, and \(\mathbf{1}\) be the all-ones vector.

[0041] S103. Initialize the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network to obtain the initial feature matrices corresponding to different heterogeneous association relationships.

[0042] Specifically, in this embodiment, the initialization of the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network includes the initialization of the meta-path of the heterogeneous association relationship and the initialization of the meta-path of the prerequisite and successor knowledge points. Among them,

[0043] The initialization of the meta-path of the heterogeneous association relationship includes initializing the representation matrix of the association relationship between different types of entities to obtain the initial feature matrix \(P\) of different types of entity pairs uku , and its expression is:

[0044]

[0045] where \(A\) uku is the relationship matrix corresponding to the specific meta-path \(u_kv_k\), \(I\) is the identity matrix, and \(D_1\) is the degree matrix of \(A\) uku + \(I\);

[0046] The initialization of the meta-path of the prerequisite and successor knowledge points includes initializing the prerequisite and successor relationship between knowledge points to obtain the feature matrix \(P\) of knowledge point pairs kk , and its expression is:

[0047]

[0048] where \(A\) kk is the first-order prerequisite and successor adjacency matrix between knowledge points, \(I\) is the identity matrix, and \(D_2\) is the degree matrix of \(A\) kk + \(I\).

[0049] It should be noted that in this embodiment, the initialization of the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network includes performing broadcast encoding on the "user-course", "user-video", "user-knowledge point", "user-course-teacher", and "knowledge point-knowledge point" adjacency interaction relationships to obtain the initial feature matrices of "user-course-user", "user-video-user", "user-knowledge point-user", "user-course-teacher-course-user", "knowledge point-user-knowledge point", and "knowledge point-knowledge point".

[0050] S104. Obtain the static learning demand embedding expression vector of the target user and the knowledge point embedding expression vector respectively according to the initial embedding expression vector of the user, the initial embedding expression vector of the knowledge point, and the initial feature matrices corresponding to different heterogeneous association relationships.

[0051] Specifically, in this embodiment, according to the initial embedding expression vector of the user, combined with the initial feature matrices corresponding to the different heterogeneous association relationships, the first attention heterogeneous relationship network is used for weighted fusion to obtain the static learning demand embedding expression vector of the target user; according to the initial embedding expression vector of the knowledge point, combined with the initial feature matrices corresponding to the different heterogeneous association relationships, the second attention heterogeneous relationship network is used for weighted fusion to obtain the knowledge point embedding expression vector; thus, the first attention heterogeneous relationship network and the second attention heterogeneous relationship network are used to mine the prerequisite and successor relationships of knowledge points, the learning relationships between users and courses, the mastery relationships between users and knowledge points, the inclusion relationships between courses and knowledge points, the teaching relationships between teachers and courses, etc., and extract the static learning demand embedding expression vector of the target user and the knowledge point embedding expression vector.

[0052] It should be noted that the first attention heterogeneous relationship network performs automatic attention weighted fusion on multiple user embedding expression vectors obtained by the "user - course - user", "user - video - user", "user - knowledge point - user", "user - course - teacher - course - user" meta - path guided graph convolutional network to obtain the static learning demand embedding expression vector of the target user; the second attention heterogeneous relationship network performs automatic attention weighted fusion on multiple knowledge point embedding expression vectors obtained by the "knowledge point - user - knowledge point", "knowledge point - knowledge point" meta - path guided graph convolutional network to obtain the embedding expression vector of the knowledge point.

[0053] In this embodiment, a graph convolutional deep learning algorithm based on meta - path is used to explicitly model the complex association relationships between heterogeneous entities, perceive the learning demands of users at the fine - grained knowledge point level from all the historical learning behaviors of users, and provide static learning demand guidance for subsequent recommendation result prediction.

[0054] S105. Obtain the dynamic learning demand embedding expression vector of the target user according to the knowledge point embedding expression vector, and splice the static learning demand embedding expression vector of the target user and the dynamic learning demand embedding expression vector of the target user to obtain the learning demand embedding expression vector of the target user.

[0055] Specifically, in this embodiment, taking the course learning as a unit, the knowledge point embedding expression vectors included in the courses learned at different times are cumulatively aggregated to obtain the knowledge level embedding expression vectors of the user at different times, and they are sorted by time into a knowledge level embedding expression vector sequence; among them, the expression of the knowledge level embedding expression vector of the user at different times is:

[0056]

[0057] Among them, is the improved knowledge level of user u after learning course , is the number of knowledge points included in course , is the i-th knowledge point embedding expression vector in course , represents the t-th course in the sequence of all courses learned by user u, and t represents the order of the course sequence arranged by time;

[0058] According to the knowledge level embedding expression vector sequence, a gated recurrent unit grid is used for sequential modeling, mining and transformation to obtain the dynamic learning demand embedding expression vector v dynamic of the target user, and its expression is:

[0059] v dynamic = h m ;

[0060] Among them, h m is the hidden state output at the last moment of the gated recurrent unit network, and m is the length of the learning sequence of the target user's course set.

[0061] The expression of the dynamic learning demand embedding expression vector v user of the target user is:

[0062] v user = concat(v static , v dynamic );

[0063] Among them, concat(i,j) is the concatenation operation on the embedding expression vectors i and j, v static is the static learning demand embedding expression vector of the target user, and v dynamic is the dynamic learning demand embedding expression vector of the target user.

[0064] In this embodiment, a method for modeling knowledge level based on courses is proposed, and a recurrent neural network is used to model the dynamic changes of the user's knowledge level at different time stages, perceive the sequential dependence relationship between the learning needs of the user at different moments, capture the dynamic learning needs of the user at the next moment, and provide dynamic learning need guidance for subsequent recommendation result prediction.

[0065] S106. Obtain the number of times of learning a knowledge point according to the embedded expression vector of the learning needs of the target user and the embedded expression vector of the knowledge point, perform matrix factorization on the number of times of learning the knowledge point to obtain a knowledge point recommendation list, and obtain a recommendation result from the knowledge point recommendation list.

[0066] Specifically, in this embodiment, obtaining the number of times of learning a knowledge point according to the embedded expression vector of the learning needs of the target user and the embedded expression vector of the knowledge point, performing matrix factorization on the number of times of learning the knowledge point to obtain a knowledge point recommendation list, and obtaining a recommendation result from the knowledge point recommendation list includes:

[0067] Multiply the embedded expression vector of the learning needs of the target user by the embedded expression vector of the knowledge point to obtain the number of times of learning the knowledge point;

[0068] Perform matrix factorization on the number of times of learning the knowledge point to obtain a knowledge point recommendation list, and obtain a recommendation result from the knowledge point recommendation list. Its expression is:

[0069]

[0070] Among them, R u,k is the number of times user u learns knowledge point k in the knowledge point recommendation list, [·] T is to find the transpose, x u is the embedded expression vector of the latent feature of user u, y k is the embedded expression vector of the latent feature of knowledge point k, d is the dimension size of the embedded expression vector x u of the latent feature of user u and the embedded expression vector y k of the latent feature of knowledge point k.

[0071] The embedded expression vector x u of the latent feature of user u and the embedded expression vector y k of the latent feature of knowledge point k are obtained by optimizing and training the objective function of the matrix factorization model based on the user's historical learning behavior using the gradient descent algorithm; among them, the expression of the objective function of the matrix factorization model is:

[0072]

[0073]

[0074] where r u,k is the actual learning times of user u for knowledge point k, is the predicted learning times of user u for knowledge point k, U is the set of users, K is the set of knowledge points, |U| is the total number of users, |K| is the total number of knowledge points, λ is the regularization parameter, λ(||x u ||2 + ||y k ||2 + ||t u ||2 + ||t k ||2) is for regularization processing, v user is the embedding expression vector of the learning needs of the target user, e k ∈E is the embedding expression vector of the knowledge point, T is for transpose, W u and W k are trainable weight parameters, β u and β k are adjustment parameters.

[0075] In summary, a deep learning knowledge point recommendation method based on learning need perception provided by the present invention uses a meta-path-guided graph convolutional network combined with an attention heterogeneous relationship network to obtain the static learning need embedding expression vector of the target user and the embedding expression vector of the knowledge point, aggregates the embedding expressions of the knowledge points at different times to obtain the knowledge level embedding expressions at different times, and then uses a gated recurrent unit network to model the potential dynamic learning need embedding expressions in the knowledge level embedding expression sequence. It not only considers the complex correlation relationships between knowledge points but also can perceive the dynamic changes of user learning needs, and provides a more fine-grained embedding expression vector of the target user's learning needs in real time and dynamically, effectively ensuring the rationality and accuracy of the recommendation results; adopting a matrix factorization model based on the user's learning times can more accurately mine the user's repeated learning behavior, thereby improving the accuracy of the recommendation results.

[0076] Based on the same inventive concept, please continue to refer to Figure 3 shown in Figure 3 is a schematic diagram of a deep learning knowledge point recommendation device based on learning need perception provided by an embodiment of the present invention. The present invention also provides a deep learning knowledge point recommendation device based on learning need perception, which is applied to the deep learning knowledge point recommendation method provided by the above embodiment of the present invention. For the specific method, please refer to the above embodiment, and the repeated parts will not be elaborated; the device includes:

[0077] A heterogeneous knowledge network construction module 201, configured to convert several different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network; wherein, the heterogeneous knowledge network includes multiple entity nodes, which are users, knowledge points, courses, videos or teachers respectively, and heterogeneous association relationships are formed by connecting different entity nodes;

[0078] A content feature initialization module 202, configured to initialize the features of users and knowledge points in the heterogeneous knowledge network to obtain the initial embedded expression vectors of the corresponding users and the initial embedded expression vectors of the knowledge points;

[0079] A meta-path feature initialization module 203, configured to initialize the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network to obtain the initial feature matrices corresponding to different heterogeneous association relationships;

[0080] A static learning demand perception module 204, configured to respectively obtain the static learning demand embedded expression vector and the knowledge point embedded expression vector of the target user according to the initial embedded expression vector of the user, the initial embedded expression vector of the knowledge point, and the initial feature matrices corresponding to different heterogeneous association relationships;

[0081] A dynamic learning demand perception module 205, configured to obtain the dynamic learning demand embedded expression vector of the target user according to the knowledge point embedded expression vector, and splice the static learning demand embedded expression vector of the target user and the dynamic learning demand embedded expression vector of the target user to obtain the learning demand embedded expression vector of the target user;

[0082] A knowledge point recommendation module 206, configured to obtain the knowledge point learning times according to the learning demand embedded expression vector of the target user and the knowledge point embedded expression vector, perform matrix factorization on the knowledge point learning times to obtain a knowledge point recommendation list, and obtain a recommendation result from the knowledge point recommendation list.

[0083] Specifically, in this embodiment, the heterogeneous knowledge network construction module 201 converts several different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network. The content feature initialization module 202 and the meta-path feature initialization module 203 initialize the content feature embedded expression vectors of users and knowledge point entities, as well as the heterogeneous relationship meta-path features, based on the heterogeneous knowledge network. The static learning demand perception module 204 and the dynamic learning demand perception module 205 perform learning demand perception modeling based on the heterogeneous knowledge network, entity content feature embedding expressions, and relationship meta-path features, successively perform static learning demand mining and dynamic learning demand mining, extract the final learning demand embedded expression vector of the user and the embedded expression vector of the knowledge point, and the knowledge point recommendation module 206 constructs a learning times matrix of the user and the knowledge point according to the learning times of the user for the knowledge point, and performs matrix factorization on the learning times matrix based on the final learning demand embedded expression vector of the user and the embedded expression vector of the knowledge point to calculate the recommendation result.

[0084] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant are intended to cover non-exclusive inclusion, so that an article or device comprising a series of elements not only includes those elements, but also includes other elements not explicitly listed. Without further limitation, an element defined by the statement "comprising one..." does not exclude the existence of additional identical elements in the article or device comprising the said element. Similar words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "upper", "lower", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation on the present invention.

[0085] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0086] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A deep learning knowledge point recommendation method based on learning demand perception, characterized in that, Including: Converting several different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network; wherein, the heterogeneous knowledge network includes multiple entity nodes, namely users, knowledge points, courses, videos or teachers respectively, and heterogeneous association relationships are formed by connecting different entity nodes; Initializing the features of the user and the features of the knowledge points in the heterogeneous knowledge network to obtain the initial embedding expression vectors of the corresponding user and the initial embedding expression vectors of the knowledge points; Initializing the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network to obtain the initial feature matrices corresponding to different heterogeneous association relationships; Respectively obtaining the static learning demand embedding expression vector and the knowledge point embedding expression vector of the target user according to the initial embedding expression vector of the user, the initial embedding expression vector of the knowledge point, and the initial feature matrices corresponding to different heterogeneous association relationships; wherein, the obtaining process of the static learning demand embedding expression vector of the target user includes: processing according to the features of the user in the heterogeneous knowledge network using a first meta-path guided graph convolutional network to obtain the initial embedding expression vector of the user; according to the initial embedding expression vector of the user, combining with the initial feature matrices corresponding to different heterogeneous association relationships, using a first attention heterogeneous relationship network for weighted fusion to obtain the static learning demand embedding expression vector of the target user; Obtaining the dynamic learning demand embedding expression vector of the target user according to the knowledge point embedding expression vector, and splicing the static learning demand embedding expression vector of the target user and the dynamic learning demand embedding expression vector of the target user to obtain the learning demand embedding expression vector of the target user; wherein, the obtaining of the dynamic learning demand embedding expression vector of the target user according to the knowledge point embedding expression vector includes: taking course learning as a unit, accumulating and aggregating the knowledge point embedding expression vectors included in the courses learned at different times to obtain the knowledge level embedding expression vectors of the user at different times, and sorting them in time order into a knowledge level embedding expression vector sequence; according to the knowledge level embedding expression vector sequence, using a gated recurrent unit grid for sequence modeling, mining and conversion to obtain the dynamic learning demand embedding expression vector of the target user; Obtaining the knowledge point learning times according to the learning demand embedding expression vector of the target user and the knowledge point embedding expression vector, performing matrix decomposition on the knowledge point learning times to obtain a knowledge point recommendation list, and obtaining a recommendation result from the knowledge point recommendation list.

2. The deep learning knowledge point recommendation method based on learning demand perception according to claim 1, characterized in that The obtaining process of the knowledge point embedding expression vector includes: Processing according to the initial embedding expression vector of the knowledge point using a second meta-path guided graph convolutional network to obtain the initial embedding expression vector of the knowledge point; According to the initial embedding expression vector of the knowledge point, combining with the initial feature matrices corresponding to different heterogeneous association relationships, using a second attention heterogeneous relationship network for weighted fusion to obtain the knowledge point embedding expression vector.

3. The deep learning knowledge point recommendation method based on learning demand perception according to claim 2, characterized in that, The expression of the initial embedding expression vector of the user is: f user = D -1 · A uc ; D = diag(A uc 1); The expression of the initial embedding expression vector of the knowledge point is: f concept = Word2vec(concept); Among them, f user is the content feature embedding expression vector of the user, A uc is the user-course interaction adjacency matrix. If the user has studied the course, the corresponding element value is 1; if the user has not studied the course, the corresponding element is 0. D is the degree matrix of the adjacency matrix, and (·) -1 is the inverse, f concept is the content feature embedding expression vector of the knowledge point. Word2vec(·) is to convert the name of the knowledge point into a vector by using the Word2vector word vectorization method. concept is the name of the knowledge point, diag(·) is the matrix diagonalization operation, and 1 is the all-ones vector.

4. The deep learning knowledge point recommendation method based on learning demand perception according to claim 1, wherein, Initializing the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network includes initializing the meta-paths of the heterogeneous association relationships and initializing the meta-paths of prerequisite and successor knowledge points. Among them, The initialization of the heterogeneous association relationship meta-path includes initializing the representation matrix of the association relationship between different types of entities to obtain the initial feature matrix P of different types of entity pairs uku , and its expression is: Among them, A uku is the relational matrix corresponding to the specific meta-path uku, I is the identity matrix, and D1 is the degree matrix of A uku +I; The initialization of the prerequisite-successor knowledge point meta-path includes initializing the prerequisite-successor relationship between knowledge points to obtain the feature matrix P of knowledge point pairs kk , and its expression is: Among them, A kk is the first-order prerequisite and successor adjacency matrix of knowledge points, I is the identity matrix, and D2 is the degree matrix of A kk +I.

5. The deep learning knowledge point recommendation method based on learning demand perception according to claim 1, wherein, The expression of the knowledge level embedding expression vector of the user at different times is: Among them, is the improved knowledge level of user u after learning course , is the number of knowledge points included in the course , is the i-th knowledge point embedding expression vector in the course , is the t-th course in all course sequences learned by user u, where t is the order of the course sequence arranged in time; The dynamic learning requirements of the target user are embedded in the expression vector v dynamic , and its expression is: v dynamic = h m ; Among them, h m is the hidden state output at the last moment of the gated recurrent unit network, and m is the length of the learning sequence of the target user's course set.

6. The deep learning knowledge point recommendation method based on learning demand perception according to claim 1, characterized in that, The expression vector v embedding the learning needs of the target user user is expressed as: v user = concat(v static , v dynamic ); Among them, concat(i,j) is an operation of concatenating the embedded expression vectors i and j, and v static is the embedded expression vector of the static learning needs of the target user, and v dynamic is the embedded expression vector of the dynamic learning needs of the target user.

7. The deep learning knowledge point recommendation method based on learning demand perception according to claim 1, characterized in that Obtaining the knowledge point learning times according to the target user learning demand embedding expression vector and the knowledge point embedding expression vector, performing matrix factorization on the knowledge point learning times, and obtaining a recommendation result from the knowledge point recommendation list, including: Multiplying the target user learning demand embedding expression vector by the knowledge point embedding expression vector to obtain the knowledge point learning times; Performing matrix factorization on the knowledge point learning times to obtain the knowledge point recommendation list, and obtaining a recommendation result from the knowledge point recommendation list, the expression of which is: Among them, R u,k is the number of times user u has learned knowledge point k in the knowledge point recommendation list, [·] T is for transpose, x u is the implicit feature embedding expression vector of user u, y k is the implicit feature embedding expression vector of knowledge point k, d is the dimension size of the implicit feature embedding expression vector x of user u u and the implicit feature embedding expression vector y of knowledge point k k .

8. The deep learning knowledge point recommendation method based on learning demand perception according to claim 7, wherein The implicit feature embedding expression vector x of the user u u and the implicit feature embedding expression vector y of the knowledge point k k are obtained by optimizing and training the objective function of the matrix factorization model based on the user's historical learning behavior using the gradient descent algorithm; wherein, the expression of the objective function of the matrix factorization model is: where r u,k is the actual learning times of user u for knowledge point k, is the predicted learning times of user u for knowledge point k, U is the set of users, K is the set of knowledge points, |U| is the total number of users, |K| is the total number of knowledge points, λ is the regularization parameter, λ(||x u ||2 + ||y k ||2 + ||W u ||2 + ||W k ||2) is for regularization processing, v user is the embedded expression vector of the learning needs of the target user, e k ∈E is the embedded expression vector of the knowledge point, T is for transpose, W u and W k are trainable weight parameters, β u and β k are adjustment parameters.

9. A deep learning knowledge point recommendation device based on learning demand perception, characterized in that, Including: A heterogeneous knowledge network construction module for converting a plurality of different types of entity interaction data in the user's historical learning behavior into a heterogeneous knowledge network; wherein, the heterogeneous knowledge network includes multiple entity nodes, namely users, knowledge points, courses, videos or teachers, and heterogeneous association relationships are formed by connecting different entity nodes; A content feature initialization module for initializing the features of the user and the knowledge points in the heterogeneous knowledge network to obtain the initial embedding expression vectors of the corresponding user and the knowledge points; A meta-path feature initialization module for initializing the meta-path features of the heterogeneous association relationships in the heterogeneous knowledge network to obtain initial feature matrices corresponding to different heterogeneous association relationships; A static learning demand perception module for respectively obtaining the static learning demand embedding expression vector and the knowledge point embedding expression vector of the target user according to the initial embedding expression vector of the user, the initial embedding expression vector of the knowledge points, and the initial feature matrices corresponding to the different heterogeneous association relationships; among them, the process of obtaining the static learning demand embedding expression vector of the target user includes: processing the features of the user in the heterogeneous knowledge network using a first meta-path guided graph convolutional network to obtain the initial embedding expression vector of the user; according to the initial embedding expression vector of the user, combining the initial feature matrices corresponding to the different heterogeneous association relationships, and using a first attention heterogeneous relationship network for weighted fusion to obtain the static learning demand embedding expression vector of the target user; A dynamic learning requirement perception module, which is used to obtain the dynamic learning requirement embedding expression vector of a target user according to the knowledge point embedding expression vector, splice the static learning requirement embedding expression vector and the dynamic learning requirement embedding expression vector of the target user, and obtain the learning requirement embedding expression vector of the target user; wherein, obtaining the dynamic learning requirement embedding expression vector of the target user according to the knowledge point embedding expression vector includes: taking course learning as a unit, accumulating and aggregating the knowledge point embedding expression vectors included in the courses learned at different times to obtain the knowledge level embedding expression vectors of the user at different times, and sorting them in time order into a knowledge level embedding expression vector sequence; according to the knowledge level embedding expression vector sequence, using a gated recurrent unit grid for sequential modeling, mining and transformation to obtain the dynamic learning requirement embedding expression vector of the target user; A knowledge point recommendation module, which is used to obtain the knowledge point learning times according to the learning requirement embedding expression vector of the target user and the knowledge point embedding expression vector, perform matrix decomposition on the knowledge point learning times to obtain a knowledge point recommendation list, and obtain a recommendation result from the knowledge point recommendation list.

Images