A Learning Resource Recommendation Method and Related Device Integrating Spatiotemporal Multi-Granularity Interests

Abstract

The present invention discloses a learning resource recommendation method and related device integrating spatio-temporal multi-granularity interests, belonging to the field of learning resource recommendation. The present invention includes the following steps: 1) Knowledge concept space multi-granularity interest modeling; 2) Time dimension multi-granularity interest modeling; 3) Learning resource recommendation with multi-granularity interest adaptive fusion. The learning resource recommendation provided by the present invention can extract the interest preferences of learners at different granularities such as knowledge points, learning resources, courses, etc. from the knowledge concept space, and extract the interest preferences of learners at different granularities within recent courses, across recent courses and globally from the time dimension, so as to realize more accurate modeling of learners' interest preferences. On the basis of adaptively fusing the single-point multi-granularity interests in the learners' knowledge concept space and the dynamic multi-granularity interests in the time dimension, the present invention provides personalized recommendations of learning resources that meet the interest preferences of learners.

Description

Technical Field

[0001] The present invention relates to a learning resource recommendation method, and in particular to a learning resource recommendation method and related device that fuse spatio-temporal multi-granularity interests. Background Art

[0002] Learning resource recommendation is mainly divided into recommendation algorithms based on traditional machine learning and those based on deep learning. Pang et al. located the unpassed and passed learning resources in the learner's historical learning sequence. For the former, they recommended the prerequisites for the learner; for the latter, they used a collaborative filtering algorithm to recommend the subsequent learning resources for the learner. Fauzan et al. proposed a learning resource recommendation method based on association rules. This method sequentially performed data standardization, data cleaning and preprocessing, grouped users by K-Models, and finally used the Apriori algorithm to form association rules. Chen et al. proposed a hybrid recommendation algorithm based on collaborative filtering and association rule mining and incorporated the learning styles of learners into the algorithm. With the continuous development of deep learning, learning resource recommendation algorithms based on deep learning have received extensive attention. Trirat et al. used technologies such as deep neural networks and confusion classifiers to build a recommendation system, which recommended learning video segments for learners according to their questions. Gong et al. and Wang et al. constructed a learner-course-learning video-knowledge point heterogeneous information network, used a meta-path-based graph representation method to obtain node representations, and input the node representations into a matrix factorization module to recommend knowledge points for learners in a personalized manner. Chen et al. introduced a knowledge graph into learning video recommendation to improve the recommendation effect. Ren et al. proposed a multi-modal course recommendation model based on LSTM and Attention. This model took different modal data such as course pictures, course introductions, audio, and the demographic information of learners as inputs, and combined the explicit and implicit feedback of learners to recommend courses for learners.

[0003] As can be seen from the above literature, in learning resource recommendation, the modeling of learners' interest preferences is an important aspect. However, there are the following two difficulties in modeling learners' interest preferences: First, learners' interests are complex and diverse and are affected by different granularity factors such as knowledge points, learning resources, and courses, resulting in difficulty in characterizing learners' interests; Second, learners' interests change dynamically over time, resulting in difficulty in capturing learners' interest patterns. Existing research often focuses on the single interest of learners in learning resources themselves, lacks modeling of learners' multi-granularity interests, ignores the multi-granularity interests shown by learners in the knowledge concept space, and in the time dimension, does not fully solve the problems of lack of supervision signals for multi-granularity interests and difficulty in interest extraction due to the short length after sequence segmentation. Therefore, how to overcome the influence of the above factors, capture learners' dynamic interest preferences at the same time, and more accurately recommend learning resources that match their interests for learners is an urgent problem to be solved. Summary of the Invention

[0004] The object of the present invention is to overcome the shortcomings of the above-mentioned existing technologies, and to provide a learning resource recommendation method and related device that integrate spatio-temporal multi-granularity interests.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A learning resource recommendation method that integrates spatio-temporal multi-granularity interests includes the following steps:

[0007] 1) Construct a knowledge concept - multi-granularity interest - learner hierarchical heterogeneous graph to express the knowledge concept space information, design nodes and edges related to aggregating interest granularities, and extract different granularity interest vectors of learners in the knowledge concept space;

[0008] From the time dimension, the learner's historical behavior sequence is segmented into the recent course sequence and the recent cross-course sequence. Then, combined with the complete learner's historical behavior sequence, the enhanced and non-enhanced interest vectors of the learner at different granularities in the recent course, recent cross-course, and global time dimensions are extracted through the GRU layer and the learner interest representation method enhanced by the learning resource relationship respectively. The interest vectors of each granularity before and after enhancement are aggregated through average pooling operation to obtain the final learner's recent course interest vector e u2l' , recent cross-course interest vector e u2l” and global interest vector e u2g ;

[0009] 2) Design a multi-granularity interest adaptive fusion layer to adaptively fuse the different granularity interest vectors of the learner in the knowledge concept space and the time dimension, and obtain the interest vector of the learner after fusion in the knowledge concept space and the interest vector of the learner after fusion in the time dimension;

[0010] 3) Input the ID features of the learner, the learner's last interacted learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u , the learner's last interacted learning resource representation vector e last and the target learning resource representation vector e v ;

[0011] Input the learner representation vector, the learner's last interacted learning resource representation vector, the interest vector of the learner after fusion in the knowledge concept space, and the interest vector of the learner after fusion in the time dimension into a deep neural network to obtain the final learner representation vector

[0012] Input the final learner representation vector and the target learning resource representation vector into the prediction layer, and the prediction layer outputs a list of recommended learning resources.

[0013] Furthermore, in the knowledge concept layer of the knowledge concept-multi-granularity interest-learner hierarchical heterogeneous graph, it includes knowledge point nodes, learning resource nodes, course nodes, teacher nodes, and school nodes, and also includes virtual type nodes corresponding to each node;

[0014] In the multi-granularity interest layer, it includes knowledge point interest nodes, learning resource interest nodes, course interest nodes, overall granularity interest nodes, and other interest nodes; the other interest nodes are the influence of the teaching teacher of the learning resource or the school where the teacher teaches on the selection of learning resources;

[0015] In the learner layer, it includes learner nodes.

[0016] Furthermore, in step 1), extracting the interest vectors of the learner at different granularities in the knowledge concept space includes extracting the interest vector e u2k at the knowledge point granularity, the interest vector e u2v at the learning resource granularity, and the interest vector e u2c at the course granularity. The extraction process is as follows:

[0017] Using the attention mechanism to calculate the importance score of the relevant edges of the granularity and the virtual type nodes, and then aggregating the vectors of the granularity-related edges and the virtual type nodes according to the corresponding attention weights to obtain the interest vector of the granularity;

[0018] The overall interest vector e ov of the learner is to aggregate the relevant edges of the knowledge point, learning resource, course, teacher, and school nodes, as well as the vectors of the corresponding virtual type nodes;

[0019] The other interest vector e ot is to aggregate the vectors of the teacher and school-related edges and the virtual type nodes.

[0020] For the learning resource recommendation that fuses spatio-temporal multi-granularity interests, further, in step 1), the specific process of extracting the interest vectors of the learner within the recent courses, across recent courses, and at different global granularities before enhancement is as follows:

[0021] Based on the historical behavior sequences within the recent courses, across recent courses, and the learner's historical behavior sequences, use GRU to initially model the multi-granularity interests of the learner in the time dimension to obtain the interest vector of the learner within the recent courses before enhancement the interest vector across recent courses and the global interest vector and / or

[0022] In step 1), use the learning resource relationship to enhance the learner's interest representation to obtain the enhanced interest vector of the learner within the recent courses the interest vector across recent courses and the global interest vector The specific process is as follows:

[0023] First, use the learner's historical interaction sequence to construct a learner behavior graph;

[0024] Secondly, use the learning resources interacted by the learner to construct a learning resource relationship graph, and integrate the learning resource relationship graph into the learner behavior graph to obtain a learning resource - learning behavior graph;

[0025] Finally, through gated graph neural network representation learning, obtain the representation vectors of each node in the learning resource - learning behavior graph, and obtain the enhanced learner interest vector through the Readout layer.

[0026] Furthermore, design multi - granularity interest self - supervised learning for the multi - granularity interest extraction process in the time dimension. Using the fact that the similarity of interest vectors before and after enhancement at the same granularity is higher than that of interest vectors at different granularities as a constraint condition, use the Bayesian personalized ranking loss function to implement the constraint condition.

[0027] Furthermore, step 2) adaptively fuses the multi - granularity interests of the learner in the knowledge concept space and the time dimension. Specifically:

[0028] Use the attention mechanism to calculate the weights of each granularity of the learner in the knowledge concept space, and aggregate the interest vectors of each granularity in the knowledge concept space according to the weights to obtain the fused interest vector e of the learner in the knowledge concept space u2kc ;

[0029] Use the attention mechanism to calculate the weights of each granularity of the learner in the time dimension, and aggregate the interest vectors of each granularity in the time dimension according to the weights to obtain the fused interest vector e of the learner in the time dimension u2ts 。

[0030] Furthermore, in step 3), the final representation vector of the learner The calculation formula is as shown in Equation (15):

[0031]

[0032] By inputting the learner's final representation vector and the target learning resource representation vector into the prediction layer, obtain the preference score of the learner for the learning resource, and select the top N learning resources with higher scores to recommend to the learner.

[0033] A learning resource recommendation device that fuses spatio - temporal multi - granularity interests, including an extraction module for spatio - temporal different - granularity interest vectors, an adaptive fusion module, and a learning resource recommendation module;

[0034] The extraction module of spatio-temporal different granularity interest vectors is used to construct a knowledge concept - multi-granularity interest - learner hierarchical heterogeneous graph to express the knowledge concept space information, design nodes and edges related to aggregating interest granularity, and extract different granularity interest vectors of learners in the knowledge concept space;

[0035] It is also used to slice the learner's historical behavior sequence from the time dimension into in-recent-course and in-recent-cross-course sequences. Then, combined with the complete learner's historical behavior sequence, the enhanced and non-enhanced interest vectors of learners at different granularities in the in-recent-course, in-recent-cross-course, and global time dimensions are extracted respectively through the GRU layer and the learner interest representation method enhanced by learning resource relationships. The interest vectors of each granularity before and after enhancement are aggregated through average pooling operation to obtain the final in-recent-course interest vector e of the learner u2l' , in-recent-cross-course interest vector e u2l” and global interest vector e u2g ;

[0036] The adaptive fusion module is used to adaptively fuse the different granularity interest vectors of learners in the knowledge concept space and the time dimension to obtain the interest vector of the learner after fusion in the knowledge concept space and the interest vector of the learner after fusion in the time dimension;

[0037] The learning resource recommendation module inputs the ID features of the learner, the learner's last interacted learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u , the learner's last interacted learning resource representation vector e last and the target learning resource representation vector e v ;

[0038] Input the learner representation vector, the learner's last interacted learning resource representation vector, the interest vector of the learner after fusion in the knowledge concept space, and the interest vector of the learner after fusion in the time dimension into a deep neural network to obtain the final learner representation vector

[0039] Input the final learner representation vector and the target learning resource representation vector into the prediction layer, and the prediction layer outputs a list of recommended learning resources.

[0040] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above learning resource recommendation method for fusing spatio-temporal multi-granularity interests.

[0041] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the steps of the above learning resource recommendation method for fusing spatio-temporal multi-granularity interests.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The learning resource recommendation method integrating spatio-temporal multi-granularity interests of the present invention can take into account the interest preferences of learners at different granularities compared with the existing learning resource recommendation methods; the model proposed by the present invention can extract the interests of learners at different granularities such as knowledge points, learning resources, and courses from the knowledge concept space, overcoming the problems that learners' interests are complex and diverse and are affected by factors such as knowledge points, learning resources, and courses; extracting the interests of learners at different granularities within the recent courses, across recent courses, and globally from the time dimension, capturing the dynamic interest preferences of learners, the model proposed by the present invention can combine the single-point interest preferences of learners in the knowledge concept space and the interest preferences that dynamically change in the time dimension to achieve more accurate modeling of learners' interest preferences and learning resource recommendation.

[0044] The learning resource recommendation device integrating spatio-temporal multi-granularity interests provided by the present invention includes specific modules for completing the above working method.

[0045] The present invention provides a computer device and a storage medium for a learning resource recommendation method integrating spatio-temporal multi-granularity interests, which are used to implement the specific steps of the above working method. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the flowchart of the learning resource recommendation of the present invention;

[0047] Figure 2 is the framework diagram of the learning resource recommendation model for integrating multi-granularity interest modeling in the time dimension;

[0048] Figure 3 is the internal structure diagram of the computer device. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0050] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or devices.

[0051] Different from the existing learning resource recommendation methods that often focus on the single interest of learners in learning resources themselves and lack the modeling of the multi-granularity interests of learners. The present invention starts from the knowledge concept space and time dimension, models the multi-granularity interests of learners, and on this basis recommends appropriate learning resources for learners. The present invention extracts the interest vectors of learners at different granularities such as knowledge points, learning resources, and courses from the knowledge concept space, and extracts the interest vectors of learners at different granularities within the recent courses, across recent courses, and globally from the time dimension. After adaptively fusing the multi-granularity interest vectors in the knowledge concept space of learners and the multi-granularity interest vectors in the time dimension, personalized learning resources are recommended for learners.

[0052] The following further describes the present invention in detail with reference to the drawings:

[0053] See Figure 1 , Figure 1 which is the flowchart of the present invention. The learning resource recommendation method for fusing spatio-temporal multi-granularity interests of the present invention includes the following steps:

[0054] Step 1: Modeling multi-granularity interests in the knowledge concept space

[0055] Use concept entities such as knowledge points, learning resources, courses and the relationships between entities to construct a heterogeneous graph, which constitutes the knowledge concept space. Through graph neural network representation learning, design nodes and edges related to aggregating interest granularity, and extract the interest vectors of learners in the knowledge concept space for different granularity concept entities such as knowledge points, learning resources, courses, etc. Specifically:

[0056] 101) Use the entities such as knowledge points, learning resources, courses, teachers, learners, etc. and the relationships between entities in the scenario of using learning resources to construct a knowledge concept - multi - granularity interest - learner hierarchical heterogeneous graph to express the knowledge concept space information. Among them, in the knowledge concept layer, it includes knowledge point nodes, learning resource nodes, course nodes, teacher nodes, and school nodes. In addition, considering the large number of nodes in the knowledge concept space, directly aggregating them will introduce noise, so virtual type nodes are designed for each type of node. In the multi - granularity interest layer, knowledge point, learning resource, course, and overall granularity interest nodes are designed. In addition, considering that learners often also consider the teaching teachers of learning resources or the schools where teachers teach when choosing learning resources, other interest nodes are introduced. In the learner layer, it includes learner nodes.

[0057] 102) Design the nodes and edges related to aggregating interest granularity. The specific process of extracting the multi - granularity interest vector of learners in the knowledge concept space is as follows:

[0058] The process of extracting the knowledge point granularity interest vector is as follows: Use the edges related to the knowledge point granularity and virtual type nodes to form a vector set e knowledge , considering that different edges or virtual type nodes have different importance degrees for interest aggregation, use the attention mechanism to calculate the importance degree scores of different edges and virtual type nodes, and aggregate the vectors of the edges and virtual type nodes related to the knowledge point according to the attention weights to obtain the interest vector e of the learner at the knowledge point granularity u2k .

[0059]

[0060]

[0061]

[0062] In formula (1), e knowledge is the set of the representation vectors of the edges related to the knowledge point and the representation vectors of the virtual type nodes of the knowledge point, is the representation vector of the edge related to the knowledge point, e k' is the representation vector of the virtual type node of the knowledge point. In formula (2), w k,i is the learnable weight parameter of the i - th representation vector in the set e knowledge for the knowledge point granularity interest vector, w k,j is the learnable weight parameter of the j - th representation vector in the set e knowledge for the knowledge point granularity interest vector, K is the number of elements in the set e knowledge , and α(k,i) is the importance degree score of the i - th representation vector in the set e knowledge for the knowledge point granularity interest vector. In formula (3), e k,i is the set e knowledgeThe i-th representation vector in, e u2k is the interest vector of the learner at the knowledge point granularity.

[0063] Similarly, the interest vectors e u2v and e u2c of the learner at the learning resource and course granularities can be obtained. The difference is that the global interest vector e u2ov of the learner is the vector that aggregates all types of edges and all virtual type nodes in the knowledge concept layer, and the other interest vectors e u2ot are the vectors that aggregate the edges and virtual type nodes related to teachers and schools.

[0064] Step 2: Multi-granularity interest modeling in the time dimension

[0065] From the time dimension, the learner's historical behavior sequence is segmented into the recent in-course and recent cross-course sequences. Combining the learner's complete historical behavior sequence, through the GRU layer and the learner interest representation method enhanced by learning resource relationships, the learner's interest vectors in the recent in-course, recent cross-course, and global in the time dimension are extracted, specifically as follows:

[0066] 201) Segment the learner's historical behavior sequence from the time dimension, specifically as follows: Given the learner's historical behavior sequence seq = [v1, v2,..., v n , the learner's historical behavior sequence is segmented into the recent in-course historical behavior sequence seq_l' = [v p , v2,..., v n , the recent cross-course historical behavior sequence seq_l” = [v q , v2,..., v n . Combining the learner's complete historical behavior sequence, the learner's interest vectors at different granularities in the recent in-course, recent cross-course, and global are extracted.

[0067] 202) Use GRU to initially model the multi-granularity interest of the learner in the time dimension to obtain the learner's recent in-course interest vector before enhancement recent cross-course interest vector and global interest vector Aiming at the problem that the sequence length is short after sequence segmentation, resulting in great difficulty in interest extraction, a method for enhancing learners' interest representation based on learning resource relationships is proposed. The overall idea of this method is to use the association information between learning resources to assist in describing the relationship between the learning resources interacted by learners. The specific process is as follows: First, use the learner's historical interaction sequence to construct a learner behavior graph. Second, use the learning resources interacted by the learner to construct a learning resource relationship graph and integrate it into the learner behavior graph to obtain a learning resource - learning behavior graph. Finally, through gated graph neural network representation learning, obtain the representation vectors of each node in the graph, and obtain the enhanced learner interest vector through the Readout layer. The calculation process of gated graph neural network representation learning is shown in equations (4)-(8), and the Readout calculation process is shown in equations (9)-(10).

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] In equation (4), N v is the set of neighbor nodes of node v, u is a neighbor node of node v, is the representation vector of node u at the l-th layer, b is the bias vector, is the representation vector of node v at the (l + 1)-th layer. In equation (5), is the representation vector of node v at the l-th layer, W z and U z are learnable parameter matrices, and σ is a non-linear activation function. In equation (6), W r and U r are learnable parameter matrices. In equation (7), W and U are learnable parameter matrices, and ⊙ is the Hadamard product. In equation (9), is the representation vector of the i-th node in the graph at the l-th layer, α i is the weight of the i-th node, V is the set of nodes in the graph. In equation (10), q T 、W1, W2, r are learnable parameter matrices, is the representation vector of the learning resource that the learner interacted with last at the l-th layer.

[0076] A method for enhancing learners' interest representation through learning resource relationships can obtain the enhanced interest vector of learners in recent courses Interest vector across recent courses and global interest vector

[0077] Step 3: Learning resource recommendation with multi-granularity interest adaptive fusion

[0078] Input the ID features of the learner, the learner's last interactive learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u , the representation vector e last of the learner's last interactive learning resource and the representation vector e v of the target learning resource. Input the learner representation vector, the representation vector of the learner's last interactive learning resource, the interest vector of the learner after fusion in the knowledge concept space, and the interest vector of the learner after fusion in the time dimension into the neural network to obtain the final learner representation vector Input the final learner representation vector and the target learning resource representation vector into the prediction layer, and the prediction layer outputs a list of recommended learning resources. Specifically:

[0079] 301) Adaptively fuse the multi-granularity interests of the learner in the knowledge concept space and the time dimension. Taking the multi-granularity interests of the learner in the knowledge concept space as an example, use the method in Equation (12) to aggregate the interest vectors of each granularity in the knowledge concept space according to the weights. The weights are calculated by the attention mechanism, and the calculation formula is as shown in Equations (13)-(14).

[0080]

[0081]

[0082] ω i = τ1(e u ‖e kc_i ‖(e u - e kc_i )‖(e u · e kc_i )||e last ), (14)

[0083] In Equation (12), e kc_i is the i-th interest vector of the learner in the knowledge concept space, α i is the weight of the i-th interest vector, and e u2kc is the interest vector of the learner after fusion in the knowledge concept space. In Equation (13), ω i is the learnable weight parameter of the i-th interest vector, ω jis the learnable weight parameter of the j-th interest vector. In Equation (14), τ1 is a deep neural network, and || is the vector concatenation operation.

[0084] Similarly, the interest vector e after the learner's time dimension fusion can be obtained. u2ts .

[0085] 302) Fuse the learner's representation vector, the learner's interest vectors in the knowledge concept space and the spatio-temporal dimension, and the representation vector of the learning resource that the learner interacted with last to obtain the learner's final representation vector that more comprehensively represents the learner's interest. The calculation formula is as shown in Equation (15).

[0086]

[0087] In Equation (15), τ2 is a deep neural network.

[0088] By inputting the learner's final representation vector and the learning resource representation vector into the prediction layer, the predicted preference score of the learner for the learning resource is obtained, and the top N learning resources with higher scores are selected and recommended to the learner.

[0089] Another embodiment of the present invention provides a learning resource recommendation device that fuses spatio-temporal multi-granularity interests, including an extraction module for spatio-temporal different granularity interest vectors, an adaptive fusion module, and a learning resource recommendation module;

[0090] The extraction module for spatio-temporal different granularity interest vectors is used to construct a knowledge concept - multi-granularity interest - learner hierarchical heterogeneous graph to express the knowledge concept space information, design nodes and edges related to aggregating interest granularity, and extract different granularity interest vectors of the learner in the knowledge concept space;

[0091] It is also used to segment the learner's historical behavior sequence from the time dimension into recent in-course and recent cross-course sequences, and then combine the complete learner's historical behavior sequence. Respectively, through the GRU layer and the learner interest representation method enhanced by learning resource relationships, the enhanced and un-enhanced interest vectors of the learner at different granularities in the recent in-course, recent cross-course, and global time dimensions are extracted. Through the average pooling operation, the interest vectors of each granularity before and after enhancement are aggregated to obtain the final learner's recent in-course interest vector e u2l' , recent cross-course interest vector e u2l” and global interest vector e u2g ;

[0092] The adaptive fusion module is used to adaptively fuse the different granularity interest vectors of the learner in the knowledge concept space and the time dimension to obtain the interest vector after the learner's fusion in the knowledge concept space and the interest vector after the learner's fusion in the time dimension;

[0093] The learning resource recommendation module inputs the ID features of the learner, the learner's last interactive learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u and the learner's last interactive learning resource representation vector e last and the target learning resource representation vector e v ;

[0094] The learner representation vector, the learner's last interactive learning resource representation vector, the learner's interest vector after fusion in the knowledge concept space, and the learner's interest vector after fusion in the time dimension are input into the deep neural network to obtain the learner's final representation vector

[0095] The learner's final representation vector and the target learning resource representation vector are input into the prediction layer, and the prediction layer outputs a list of recommended learning resources.

[0096] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structural diagram may be as shown in Figure 3 The figure. The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the computer program is executed by the processor, it implements a learning resource recommendation method that fuses spatio-temporal multi-granularity interests.

[0097] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the following steps are implemented: 1) Construct a knowledge concept-multi-granularity interest-learner hierarchical heterogeneous graph to express the knowledge concept space information, design nodes and edges related to aggregating interest granularity, and extract different granularity interest vectors of the learner in the knowledge concept space; Cut the learner's historical behavior sequence from the time dimension into recent in-course and recent cross-course sequences, and then combine the complete learner's historical behavior sequence. Respectively, through the GRU layer and the learner interest representation method enhanced by the learning resource relationship, the enhanced and unenhanced interest vectors of the learner in different granularities in the recent in-course, recent cross-course, and global time dimensions are extracted. By performing average pooling operations on the enhanced and unenhanced interest vectors of each granularity, the final learner's recent in-course interest vector e u2l' and the recent cross-course interest vector e u2l” and the global interest vector e u2g; 2) Design a multi-granularity interest adaptive fusion layer to adaptively fuse the different granularity interest vectors of the learner in the knowledge concept space and time dimension, obtaining the fused interest vector of the learner in the knowledge concept space and the fused interest vector of the learner in the time dimension; 3) Input the ID features of the learner, the learner's last interacted learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u , the representation vector e last of the learner's last interacted learning resource, and the representation vector e v of the target learning resource; Input the learner representation vector, the representation vector of the learner's last interacted learning resource, the fused interest vector of the learner in the knowledge concept space, and the fused interest vector of the learner in the time dimension into the deep neural network to obtain the final learner representation vector Input the final learner representation vector and the target learning resource representation vector into the prediction layer, and the prediction layer outputs a list of recommended learning resources.

[0098] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: 1) Construct a knowledge concept - multi-granularity interest - learner hierarchical heterogeneous graph to express the knowledge concept space information, design nodes and edges related to aggregating interest granularity, and extract different granularity interest vectors of the learner in the knowledge concept space; Segment the learner's historical behavior sequence from the time dimension into the recent course and recent cross-course sequences, and then combine the complete learner's historical behavior sequence. Respectively, through the GRU layer and the learner interest representation method enhanced by learning resource relationships, extract the enhanced and unenhanced interest vectors of different granularities of the learner within the recent course, recent cross-courses, and globally in the time dimension. Aggregate the enhanced and unenhanced interest vectors of each granularity through average pooling operation to obtain the final learner's recent course interest vector e u2l' , recent cross-course interest vector e u2l” , and global interest vector e u2g ; 2) Design a multi-granularity interest adaptive fusion layer to adaptively fuse the different granularity interest vectors of the learner in the knowledge concept space and time dimension, obtaining the fused interest vector of the learner in the knowledge concept space and the fused interest vector of the learner in the time dimension; 3) Input the ID features of the learner, the learner's last interacted learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u , the representation vector e last of the learner's last interacted learning resource, and the representation vector e v of the target learning resource; Input the learner representation vector, the representation vector of the learner's last interacted learning resource, the fused interest vector of the learner in the knowledge concept space, and the fused interest vector of the learner in the time dimension into the deep neural network to obtain the final learner representation vector The final representation vector of the learner and the representation vector of the target learning resource are input into the prediction layer, and the prediction layer outputs a list of recommended learning resources.

[0099] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0100] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the above division of each functional unit and module is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0101] Embodiment

[0102] The method proposed in the present invention was experimentally tested on the publicly available dataset MOOCCubeQ&A of the XuetangX platform and the publicly available dataset MOOPer of the TopCode practical teaching platform. The MOOCCubeQ&A dataset contains 126,131 interaction records of 1,713 learners; the MOOPer dataset contains 641,744 interaction records of 44,523 learners. The recommended effects of the method proposed in the present invention and classical sequential recommendation methods including NextItNet, Caser, GRU4Rec, STAMP, SRGNN, BERT4Rec, CORE, GCSAN, and LightSANs were compared in the experiment. The experimental evaluation metrics were Recall@10, Recall@20, NDCG@10, and NDCG@20. The experimental results are shown in Table 1. The results indicate that the method proposed in the present invention can achieve the optimal recommendation results.

[0103] Table 1 Evaluation Metrics of the Embodiment

[0104]

[0105]

[0106] The above content is only to illustrate the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any modification made on the basis of the technical solution according to the technical idea proposed in the present invention falls within the protection scope of the claims of the present invention.

Claims

1. A learning resource recommendation method integrating spatio-temporal multi-granularity interests, characterized in that, Including the following steps: 1) Construct a knowledge concept - multi - granularity interest - learner hierarchical heterogeneous graph to express the knowledge concept space information, design nodes and edges related to aggregating interest granularity, and extract different - granularity interest vectors of learners in the knowledge concept space; The learner's historical behavior sequence is sliced into the recent in-course and recent cross-course sequences from the time dimension. Then, combined with the complete learner's historical behavior sequence, the learner's interest vectors before and after enhancement at different granularities in the recent in-course, recent cross-course, and global time dimensions are extracted through the GRU layer and the learner interest representation method enhanced by learning resource relationships. By performing average pooling operations on the interest vectors at each granularity before and after enhancement, the final learner's recent in-course interest vector e u2l' , recent cross-course interest vector e u2l” and global interest vector e u2g ; 2) Design a multi - granularity interest adaptive fusion layer to adaptively fuse different - granularity interest vectors of learners in the knowledge concept space and time dimension, obtaining the fused interest vector of learners in the knowledge concept space and the fused interest vector of learners in the time dimension; 3) Input the ID features of the learner, the learner's last interacted learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u , the learner's last interacted learning resource representation vector e last and the target learning resource representation vector e v ; Input the learner representation vector, the learner's last interactive learning resource representation vector, the learner's interest vector after fusion in the knowledge concept space, and the learner's interest vector after fusion in the time dimension into the deep neural network to obtain the learner's final representation vector Input the final representation vector of the learner and the representation vector of the target learning resource into the prediction layer, and the prediction layer outputs a list of recommended learning resources.

2. The learning resource recommendation method integrating spatio-temporal multi-granularity interests according to claim 1, wherein In the knowledge concept layer of the knowledge concept - multi - granularity interest - learner hierarchical heterogeneous graph, it includes knowledge point nodes, learning resource nodes, course nodes, teacher nodes, and school nodes, and also includes virtual type nodes corresponding to each node; In the multi - granularity interest layer, it includes knowledge point interest nodes, learning resource interest nodes, course interest nodes, overall granularity interest nodes, and other interest nodes; the other interest nodes are the influence of the teaching teacher of the learning resource or the school where the teacher teaches on the selection of the learning resource; In the learner layer, it includes learner nodes.

3. The learning resource recommendation method integrating spatio-temporal multi-granularity interests according to claim 2, wherein, Step 1) Extract the interest vectors of the learner in different granularities of the knowledge concept space, including the interest vector e at the knowledge point granularity u2k extraction, the interest vector e at the learning resource granularity u2v extraction, and the interest vector e at the course granularity u2c extraction. The extraction process is as follows: Use the attention mechanism to calculate the importance score of the relevant edges of the granularity and the virtual type nodes, and then aggregate the vectors of the granularity - related edges and virtual type nodes according to the corresponding attention weights to obtain the interest vector of the granularity; The overall interest vector e of the learner ov is a vector that aggregates the relevant edges of knowledge points, learning resources, courses, teachers, and school nodes, as well as the corresponding virtual type nodes; Other interest vector e ot is a vector that aggregates teacher- and school-related edges and virtual type nodes.

4. The learning resource recommendation method integrating spatio-temporal multi-granularity interests according to claim 1, wherein, In step 1), the specific process of extracting the learner's recent in - course, recent cross - course, and global different - granularity interest vectors before enhancement is: Based on the recent in-course historical behavior sequence, the recent cross-course historical behavior sequence, and the learner's historical behavior sequence, use GRU to initially model the multi-granularity interests of the learner in the time dimension, and obtain the learner's recent in-course interest vector before enhancement Recent cross-course interest vector And the global interest vector And / or In step 1), the interest representation of the learner is enhanced by using the learning resource relationship, and the enhanced interest vector of the learner in the recent courses is obtained. Recent cross-course interest vector and the global interest vector The specific process is as follows: First, use the learner's historical interaction sequence to construct a learner behavior graph; Secondly, use the learning resources interacted by the learner to construct a learning resource relationship graph, and integrate the learning resource relationship graph into the learner behavior graph to obtain a learning resource - learning behavior graph; Finally, through gated graph neural network representation learning, obtain the representation vectors of each node in the learning resource - learning behavior graph, and obtain the enhanced learner interest vector through the Readout layer.

5. The learning resource recommendation method integrating spatio-temporal multi-granularity interests according to any one of claims 1-4, characterized in that Design multi - granularity interest self - supervised learning for the time - dimension multi - granularity interest extraction process. Using the fact that the similarity of interest vectors before and after enhancement at the same granularity is higher than that of different - granularity interest vectors as a constraint condition, use the Bayesian personalized ranking loss function to achieve the constraint condition.

6. The learning resource recommendation method integrating spatio-temporal multi-granularity interests according to claim 1, wherein Step 2) adaptively fuses the multi - granularity interests of learners in the knowledge concept space and time dimension. Specifically: The weights of the learner in each granularity of the knowledge concept space are calculated using the attention mechanism, and the interest vectors of each granularity in the knowledge concept space are aggregated according to the weights to obtain the fused interest vector e of the learner in the knowledge concept space u2kc ; The weights of the learner at each granularity in the time dimension are calculated using the attention mechanism, and the interest vectors of each granularity within the time dimension are aggregated according to the weights to obtain the fused interest vector \(e\) of the learner in the time dimension u2ts .

7. The learning resource recommendation method integrating spatio-temporal multi-granularity interests according to claim 6, wherein The learner's final representation vector in step 3) The calculation formula is as shown in Equation (15): By inputting the final representation vector of the learner and the representation vector of the target learning resource into the prediction layer, obtain the preference score of the learner for the learning resource, and select the top N learning resources with higher scores to recommend to the learner.

8. A learning resource recommendation device that integrates spatio-temporal multi-granularity interests, characterized in that, Including an extraction module for spatio - temporal different - granularity interest vectors, an adaptive fusion module, and a learning resource recommendation module; The extraction module for spatio - temporal different - granularity interest vectors is used to construct a knowledge concept - multi - granularity interest - learner hierarchical heterogeneous graph to express the knowledge concept space information, design nodes and edges related to aggregating interest granularity, and extract different - granularity interest vectors of learners in the knowledge concept space; It is also used to slice the learner's historical behavior sequence from the time dimension into the recent in-course and recent cross-course sequences. Then, combined with the complete learner's historical behavior sequence, the interest vectors of the learner before and after enhancement at different granularities in the recent in-course, recent cross-course, and global time dimensions are extracted through the GRU layer and the learner interest representation method enhanced by learning resource relationships. By performing average pooling operations on the interest vectors at each granularity before and after enhancement, the final learner's recent in-course interest vector e u2l' , recent cross-course interest vector e u2l” and global interest vector e u2g ; The adaptive fusion module is used to adaptively fuse different granularity interest vectors of the learner in the knowledge concept space and the time dimension, so as to obtain the interest vector of the learner after fusion in the knowledge concept space and the interest vector of the learner after fusion in the time dimension; The learning resource recommendation module inputs the ID features of the learner, the learner's last interacted learning resource, and the target learning resource into the embedding layer to obtain the learner representation vector e u , the learner's last interacted learning resource representation vector e last and the target learning resource representation vector e v ; Input the learner representation vector, the learner's last interactive learning resource representation vector, the learner's interest vector after fusion in the knowledge concept space, and the learner's interest vector after fusion in the time dimension into the deep neural network to obtain the learner's final representation vector Input the final representation vector of the learner and the representation vector of the target learning resource into the prediction layer, and the prediction layer outputs a list of recommended learning resources.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, the steps of the learning resource recommendation method for fusing spatio-temporal multi-granularity interests described in any one of claims 1-7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the steps of the learning resource recommendation method for fusing spatio-temporal multi-granularity interests described in any one of claims 1-7 are implemented.

Images