Course recommendation method and system based on heterogeneous graph and collaborative attenuation attention mechanism

Abstract

The present invention discloses a course recommendation method and system based on a heterogeneous graph and a collaborative decaying attention mechanism. The method comprises: in a data preprocessing layer, obtaining historical course record information of a target user and candidate recommended course information, constructing a heterogeneous graph of the course, and then constructing input features of an input model; in a knowledge extraction layer, according to the heterogeneous graph between the user and the course, using a graph convolution-based knowledge extraction layer to obtain high-order features of the user and the course, and enriching the input features; in a knowledge evolution layer, using a collaborative decaying attention mechanism to model the learning process of the target user, so that the input features can be combined with the learning process of the target user; in a prediction layer, inputting the output of the knowledge evolution layer into a fully connected network to obtain a predicted course registration rate; and based on the predicted course registration rate, pushing a target recommended course to the target user. The present invention has high recommendation accuracy and comprehensiveness, and can be widely used in the field of artificial intelligence technology.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a course recommendation method and system based on heterogeneous graphs and a collaborative attenuation attention mechanism. Background Art

[0002] With the widespread adoption of internet technology, online platforms like Coursera, edX, and Udacity are now widely known. Many users have access to courses from renowned universities at minimal cost through these platforms. However, the surge in the number of online courses has created a demand for more accurate and targeted recommendations.

[0003] Unlike other recommendation tasks, course recommendation faces three important challenges because it is highly related to learning input and the learning process. (1) According to the input hypothesis in the field of education, the input for learners should be "i+1", where "i" refers to the knowledge of the current stage and "i+1" refers to the learning of the next stage. In other words, the knowledge that learners are exposed to should just exceed their current ability, so that they can understand most of the content but still be challenged to make progress. Given that most courses require some prerequisite knowledge, users also tend to take courses for which they have already mastered the prerequisite knowledge to ensure understandable input. For example, if a user has mastered the knowledge of the courses "Linear Algebra" and "Probability Theory", then the course "Machine Learning" would be a more suitable course to recommend. At the same time, courses generally contain multiple knowledge points. Therefore, course recommendation should consider the order of courses and knowledge points. (2) According to the forgetting curve, forgetting begins immediately after learning and the degree of memory decreases exponentially. At the same time, the courses that most users enroll in during the learning process show an increasing trend in difficulty. This indicates that when the present invention recommends courses, courses learned in the distant past should not be as informative as courses learned recently. (3) Explainable recommendations can guide students to conduct targeted review and consolidation, increasing users' recognition of recommended courses and their enthusiasm for subsequent learning. (4) Popular evaluation methods for recommendation systems include click-through rate, NDCG (normalized discounted cumulative gain), etc., but they all have a common problem: there is no way to comprehensively measure the relevance of the recommended course set as a whole to the target course.

[0004] Traditional course recommendation methods do not focus on modeling students' learning process, nor do they capture the relationship between courses and knowledge points, and are unable to comprehensively evaluate the overall recommendation performance of the course set. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides a course recommendation method and system based on heterogeneous graphs and a collaborative attenuated attention mechanism with high and comprehensive recommendation accuracy.

[0006] One aspect of an embodiment of the present invention provides a course recommendation method based on a heterogeneous graph and a collaborative attenuated attention mechanism, including:

[0007] In the data preprocessing layer, the historical course record information and candidate recommended course information of the target user are obtained to construct a heterogeneous graph of courses, and then the input features of the input model are constructed;

[0008] In the knowledge extraction layer, according to the heterogeneous graph between the user and the course, a graph convolution-based knowledge extraction layer is used to obtain high-level features of the user and the course, and enrich the input features;

[0009] In the knowledge evolution layer, a cooperative decay attention mechanism is used to model the learning process of the target user, so that the input features can be combined with the learning process of the target user;

[0010] In the prediction layer, the output of the knowledge evolution layer is input into the fully connected network to obtain the predicted course enrollment rate;

[0011] Push target recommended courses to the target user based on the predicted course registration rate.

[0012] Optionally, the heterogeneous graph includes user interaction data, course precedence and follow-up relationships, and course knowledge points;

[0013] The process of constructing a heterogeneous graph of courses and then constructing input features of an input model includes the following steps:

[0014] Based on the registration history of users in the training set, a bipartite graph of user-corresponding courses is constructed;

[0015] Construct a course precedence and follow-up relationship diagram;

[0016] Determine the knowledge points included in the course based on the course catalog and syllabus;

[0017] Connect the knowledge points included in the course with each course, and merge the bipartite graph and the precedence-successor relationship graph to obtain a heterogeneous graph between users and courses;

[0018] According to the heterogeneous graph, a graph convolutional network is used to obtain high-order features of each node in the heterogeneous graph.

[0019] Optionally, the method of obtaining high-level features of users and courses based on the heterogeneous graph between the users and courses by using a graph convolution-based knowledge extraction layer to enrich the input features includes:

[0020] Obtain the initial features of users and courses through the embedding layer;

[0021] For knowledge points, the semantic features of the knowledge points are obtained through the pre-trained w2v model;

[0022] Each time through a knowledge extraction layer, by aggregating the node features of the neighbors corresponding to users and courses on the graph, the features of the neighboring nodes are summed and normalized to obtain the high-level features of each node. The neighbors of the user aggregation are courses, and the neighbors of the course aggregation include: semantic information of the knowledge points contained in the course, the courses before and after the current course, and the user;

[0023] After multiple layers of knowledge extraction, the features of different layers of each user node and course node are averaged, and the features of the course are added to the category features corresponding to the course to obtain the final features of the node.

[0024] Optionally, the adopting of a collaborative attenuation attention mechanism to model the target user's learning process so that the input features can be combined with the target user's learning process includes:

[0025] For each course, first calculate the monotonic attention weight of each course; wherein, the monotonic is used to characterize the query feature of the course and the key features of the course learned before the course to calculate the attention score;

[0026] Calculating the co-decayed attention weight according to the exponential decay rate and the monotonic attention weight;

[0027] According to the attention weight after collaborative attenuation, the features of the current course are obtained by weighted summing the attention weight and each feature;

[0028] The final features of each course are obtained as the input features.

[0029] Optionally, the method further includes: determining evaluation indicators for course recommendation results in combination with knowledge points.

[0030] Optionally, the evaluation indicators for determining course recommendation results based on knowledge points include:

[0031] According to the coverage of course knowledge points, the course similarity between any two courses is calculated by counting the number of common knowledge points contained in the courses;

[0032] For each user, we calculate the set of courses that the user is most likely to register for. Then, based on the course similarity, we calculate the mean similarity between the courses actually selected by the user and the recommended target courses. We obtain the overall similarity between the set of recommended courses and the target courses, and then obtain the evaluation index of the course recommendation results.

[0033] Another aspect of the present invention provides a course recommendation system based on heterogeneous graphs and a collaborative attenuated attention mechanism, including:

[0034] The data preprocessing layer is used to obtain the target user's historical course record information and candidate recommended course information, build a heterogeneous graph of courses, and then construct the input features of the input model;

[0035] A knowledge extraction layer is used to obtain high-level features of users and courses based on the heterogeneous graph between the users and courses, and enrich the input features by using a graph convolution-based knowledge extraction layer;

[0036] A knowledge evolution layer, configured to adopt a cooperative decay attention mechanism to model the learning process of the target user, so that the input features can be combined with the learning process of the target user;

[0037] A prediction layer, configured to input the output of the knowledge evolution layer into a fully connected network to obtain a predicted course enrollment rate;

[0038] A push module is used to push target recommended courses to the target user based on the predicted course registration rate.

[0039] Another aspect of an embodiment of the present invention further provides an electronic device, including a processor and a memory;

[0040] The memory is used to store programs;

[0041] The processor executes the program to implement the method described above.

[0042] Another aspect of the embodiment of the present invention further provides a computer-readable storage medium, the storage medium stores a program, the program is executed by a processor to implement the method as described above.

[0043] The present invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the above method.

[0044] In the embodiment of the present invention, at the data preprocessing layer, historical course record information and candidate recommended course information of the target user are obtained to construct a heterogeneous graph of the course, thereby constructing the input features of the input model. At the knowledge extraction layer, based on the heterogeneous graph between the user and the course, a graph convolution-based knowledge extraction layer is used to obtain high-order features of the user and the course, and the input features are enriched. At the knowledge evolution layer, a collaborative decay attention mechanism is used to model the learning process of the target user, so that the input features can be combined with the learning process of the target user. At the prediction layer, the output result of the knowledge evolution layer is input into a fully connected network to obtain a predicted course registration rate. Based on the predicted course registration rate, the target recommended course is pushed to the target user. The present invention has high recommendation accuracy and comprehensiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0046] Figure 1 An overall step flow chart provided for an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the structure of a heterogeneous graph provided by an embodiment of the present invention;

[0048] Figure 3 A flowchart of the knowledge extraction process based on graph convolution provided by an embodiment of the present invention;

[0049] Figure 4 A processing flow chart of the knowledge evolution layer provided by an embodiment of the present invention;

[0050] Figure 5 This is a heat map of the attention weights of the knowledge evolution layer provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0052] In view of the problems existing in the prior art, an embodiment of the present invention provides a course recommendation method based on heterogeneous graphs and a collaborative attenuation attention mechanism. Figure 1 As shown, the method of the present invention comprises the following steps:

[0053] In the data preprocessing layer, the historical course record information and candidate recommended course information of the target user are obtained to construct a heterogeneous graph of courses, and then the input features of the input model are constructed;

[0054] In the knowledge extraction layer, according to the heterogeneous graph between the user and the course, a graph convolution-based knowledge extraction layer is used to obtain high-level features of the user and the course, and enrich the input features;

[0055] In the knowledge evolution layer, a cooperative decay attention mechanism is used to model the learning process of the target user, so that the input features can be combined with the learning process of the target user;

[0056] In the prediction layer, the output of the knowledge evolution layer is input into the fully connected network to obtain the predicted course enrollment rate;

[0057] Push target recommended courses to the target user based on the predicted course registration rate.

[0058] Optionally, the heterogeneous graph includes user interaction data, course precedence and follow-up relationships, and course knowledge points;

[0059] The process of constructing a heterogeneous graph of courses and then constructing input features of an input model includes the following steps:

[0060] Based on the registration history of users in the training set, a bipartite graph of user-corresponding courses is constructed;

[0061] Construct a course precedence and follow-up relationship diagram;

[0062] Determine the knowledge points included in the course based on the course catalog and syllabus;

[0063] Connect the knowledge points included in the course with each course, and merge the bipartite graph and the precedence-successor relationship graph to obtain a heterogeneous graph between users and courses;

[0064] According to the heterogeneous graph, a graph convolutional network is used to obtain high-order features of each node in the heterogeneous graph.

[0065] Optionally, the method of obtaining high-level features of users and courses based on the heterogeneous graph between the users and courses by using a graph convolution-based knowledge extraction layer to enrich the input features includes:

[0066] Obtain the initial features of users and courses through the embedding layer;

[0067] For knowledge points, the semantic features of the knowledge points are obtained through the pre-trained w2v model;

[0068] Each time through a knowledge extraction layer, by aggregating the node features of the neighbors corresponding to users and courses on the graph, the features of the neighboring nodes are summed and normalized to obtain the high-level features of each node. The neighbors of the user aggregation are courses, and the neighbors of the course aggregation include: semantic information of the knowledge points contained in the course, the courses before and after the current course, and the user;

[0069] After multiple layers of knowledge extraction, the features of different layers of each user node and course node are averaged, and the features of the course are added to the category features corresponding to the course to obtain the final features of the node.

[0070] Optionally, the adopting of a collaborative attenuation attention mechanism to model the target user's learning process so that the input features can be combined with the target user's learning process includes:

[0071] For each course, first calculate the monotonic attention weight of each course; wherein, the monotonic is used to characterize the query feature of the course and the key features of the course learned before the course to calculate the attention score;

[0072] Calculating the co-decayed attention weight according to the exponential decay rate and the monotonic attention weight;

[0073] According to the attention weight after collaborative attenuation, the features of the current course are obtained by weighted summing the attention weight and each feature;

[0074] The final features of each course are obtained as the input features.

[0075] Optionally, the method further includes: determining evaluation indicators for course recommendation results in combination with knowledge points.

[0076] Optionally, the evaluation indicators for determining course recommendation results based on knowledge points include:

[0077] According to the coverage of course knowledge points, the course similarity between any two courses is calculated by counting the number of common knowledge points contained in the courses;

[0078] For each user, we calculate the set of courses that the user is most likely to register for. Then, based on the course similarity, we calculate the mean similarity between the courses actually selected by the user and the recommended target courses. We obtain the overall similarity between the set of recommended courses and the target courses, and then obtain the evaluation index of the course recommendation results.

[0079] Another aspect of the present invention provides a course recommendation system based on heterogeneous graphs and a collaborative attenuated attention mechanism, including:

[0080] The data preprocessing layer is used to obtain the target user's historical course record information and candidate recommended course information, build a heterogeneous graph of courses, and then construct the input features of the input model;

[0081] A knowledge extraction layer is used to obtain high-level features of users and courses based on the heterogeneous graph between the users and courses, and enrich the input features by using a graph convolution-based knowledge extraction layer;

[0082] A knowledge evolution layer, configured to adopt a cooperative decay attention mechanism to model the learning process of the target user, so that the input features can be combined with the learning process of the target user;

[0083] A prediction layer, configured to input the output of the knowledge evolution layer into a fully connected network to obtain a predicted course enrollment rate;

[0084] A push module is used to push target recommended courses to the target user based on the predicted course registration rate.

[0085] Another aspect of an embodiment of the present invention further provides an electronic device, including a processor and a memory;

[0086] The memory is used to store programs;

[0087] The processor executes the program to implement the method described above.

[0088] Another aspect of the embodiment of the present invention further provides a computer-readable storage medium, the storage medium stores a program, the program is executed by a processor to implement the method as described above.

[0089] The present invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the above method.

[0090] The specific implementation principle of the present invention is described in detail below with reference to the accompanying drawings:

[0091] This invention introduces three key factors: first, a heterogeneous graph, which includes user-course interaction records, the order of succession between courses, and the knowledge points included in the courses. A knowledge extraction layer is used to capture high-order signals and semantic information of knowledge points in the heterogeneous graph to enrich the representation of users and courses. Second, a knowledge evolution layer uses a context-dependent attention mechanism to model the user's learning process, better capturing the user's knowledge level. Third, the model's visualization provides guidance for students to review courses. This not only achieves better recommendation results but also provides review guidance for students' continued learning.

[0092] Traditional evaluation indicators generally focus on whether the course set recommended by the recommendation system includes the user's target course, and do not comprehensively evaluate the overall recommendation performance of the set. Therefore, the present invention uses the knowledge point-related course similarity evaluation indicator to evaluate the recommendation system, helping the recommendation system achieve better recommendation results.

[0093] For the dataset used in this paper, we first dedupe and cleanse user interaction data, then extract the last course enrolled by 20% of users as the test set. Each instance in the training and test sets represents a user and their corresponding historical course registration. This paper selects the last course enrolled in the historical record as the target course recommendation. For each target course in the training set, four courses are randomly sampled as negative examples, while for the test set, 99 courses are randomly sampled as negative examples.

[0094] Specifically, the method of the present invention comprises the following steps:

[0095] Step 1: Data Preprocessing

[0096] Regarding user features, the present invention fixes the number of user history course records to 20, and uses them together with user and target recommended courses as feature input models.

[0097] The present invention constructs a heterogeneous graph about courses, which includes user interaction data, course precedence and follow-up relationships, and course knowledge points. Specifically, there are user registration history records in the training set. The present invention constructs a bipartite graph of user-corresponding courses through these history records. That is to say, if a user has studied a certain course, there is a relationship between the user and the course in the interaction graph. Then the present invention constructs a precedence and follow-up relationship graph of courses. For example, if the prerequisite course required by a course is another course, then there is an edge between the two courses in the precedence and follow-up relationship graph of courses. The present invention then determines which knowledge points are included in each course based on the course catalog and outline, and connects the relevant knowledge points with the courses. Finally, the above graphs are merged, and the heterogeneous graph schematic diagram is shown as follows. Figure 2 shown.

[0098] For heterogeneous graphs, the present invention can then use graph convolutional networks to obtain high-order features of each node in the graph.

[0099] Step 2: Knowledge extraction layer based on graph convolution

[0100] Users and courses are positioned differently in a heterogeneous graph, so the graph's structure inevitably influences its features. Furthermore, the graph contains information about the precedence and subsequent stages of courses, knowledge points, and collaborative signals related to user course selection. Therefore, this paper applies graph convolution to heterogeneous graphs to capture high-level features of users and courses, thereby enriching their inherent features.

[0101] like Figure 3 As shown, the present invention first obtains the initial features of users and courses through the embedding layer. For knowledge points, the present invention uses the pre-trained model w2v (a pre-trained model that maps words to embeddings) to obtain the semantic features of knowledge points. Each time after passing through a layer of knowledge extraction layer, the present invention obtains the high-level features of the node by aggregating the node features of their corresponding neighbors on the graph, that is, summing and normalizing the features of the neighbor nodes. It should be noted here that the aggregated neighbors of the user are courses, and the aggregated neighbors of the course include three things: the semantic information of the knowledge points contained in itself, its preceding and subsequent courses, and the user. Finally, after passing through multiple layers of knowledge extraction layers, the present invention averages the features of different layers of each node (user and course) obtained to obtain the final features of the node. In addition, the course features need to be added to the category features corresponding to the course. This not only captures the structural information and collaborative signals of the heterogeneous graph, but also captures the knowledge points and preceding and subsequent information of the course, greatly enriching the features of users and courses.

[0102] Step 3: Knowledge Evolution Layer - User Learning Process Modeling

[0103] History learning courses contribute to a user's knowledge level. However, according to the forgetting curve, a user's history courses should provide different amounts of information for their target recommended courses. The relevance of the history learning course to the target recommended course, as well as the time interval between them, should jointly determine the contribution of the history learning course. Therefore, for the user, history learning course, and target recommended course, this paper uses a monotonic co-attenuated attention mechanism to model the user's learning process. Co-attenuation refers to the trend of the forgetting curve.

[0104] Among them, each course will pass through 3 different fully connected layers to obtain the corresponding query, key and value features with the same dimensions. Figure 4As shown, taking the calculation of the characteristics of a course as an example, the present invention first obtains the monotonic attention weight. Monotonic means that the query characteristics of the course will be multiplied by the key characteristics of the course and the courses learned before the course to obtain the calculated attention score. Co-attenuation refers to multiplying the obtained attention weight by an exponential decay rate. The decay rate is determined by the time interval between the past course and the registration of the course. The longer the time interval, the more decay, which simulates the process of user learning and forgetting. After obtaining the attenuated attention weight, the present invention obtains the characteristics of the current course by weighted summation of the attention weight and each feature. Similarly, this method can obtain the characteristics of each course and splice them for output.

[0105] The course features obtained in this way model the user's learning process. Ultimately, when recommending courses, historical courses similar to the recommended course provide more information, while courses older than the recommended course provide less information.

[0106] Step 4: Prediction Layer

[0107] The output of knowledge evolution is a large-scale feature, incorporating contextual information and high-level features of the collaborative sequence graph. However, the present invention aims to predict the user enrollment rate for a target course. Therefore, the present invention uses a fully connected network to predict the user enrollment rate for the target course. By inputting the output of the knowledge evolution layer into a fully connected network, the predicted course enrollment rate can be obtained.

[0108] Step 5: Recommendation system evaluation indicators combined with knowledge points

[0109] In addition to popular evaluation indicators such as AUC (area under the sensitivity curve), HR (click-through rate), NDCG (normalized discounted cumulative gain) and MRR (mean relative ranking), the present invention also proposes a course recommendation evaluation indicator based on knowledge points. Specifically: according to the coverage of the knowledge points of the course, the present invention calculates the similarity between the two courses by calculating the number of identical knowledge points contained in the courses. For each user, the present invention can calculate the set of courses that the user is most likely to register for through the model, and then calculate the mean of the course similarity of these courses with the recommended target course based on the course similarity. The present invention can obtain the overall similarity between the set of recommended courses and the target course. This evaluation can more comprehensively assess whether the recommended course set is in line with the user's intentions and interests, and can help improve the overall effect of the recommendation.

[0110] Step 6: Comparative experiment

[0111] In order to prove the effectiveness of the model proposed in the present invention, the present invention conducted a series of ablation experiments and comparative experiments.

[0112] In the comparative experiment, the comparison method of the present invention includes:

[0113] MLP uses multi-layer perception to represent users and corresponding courses and predict the probability of recommending courses to users.

[0114] NeuMF: It combines the traditional matrix decomposition method and multi-layer perceptron, which can extract low-dimensional and high-dimensional features at the same time and has good recommendation effect.

[0115] CKE. This is a knowledge graph-based representation regularization method that leverages semantic embeddings to enhance matrix factorization.

[0116] LightGCN. It learns user and item embeddings by linearly propagating on the user-item interaction graph. The weighted sum of the user and item embeddings obtained at each level is then calculated as the final prediction score.

[0117] KGAT: It integrates the knowledge graph and interaction graph into a unified graph space; explicitly models high-order relations via recursive neighbor propagation of target node embeddings, and uses knowledge-dependent attention to distinguish the importance of neighbor embeddings.

[0118] HRL: It uses a hierarchical reinforcement model to modify the user's profile, which is helpful for predicting scores.

[0119] DIN: It uses attention to capture the user's interest in historical behavior. The user's embedding changes with the change of candidate products, which effectively improves the performance of the model.

[0120] BST: It introduces the Transformer layer to model the user's historical behavior sequence to capture the user's dynamic interests.

[0121] Table 1 Comparative experimental results

[0122] Models AUC HR@5 HR@10 NDCG@5 NDCG@10 MRR MLP 0.8507 0.4427 0.618 0.3059 0.3624 0.3026 NeuMF 0.882 0.4944 0.6534 0.3518 0.4033 0.3429 CKE 0.8643 0.4472 0.6055 0.3126 0.3623 0.305 lightgcn 0.8621 0.4834 0.6434 0.3457 0.3973 0.3385 KGAT 0.8888 0.4871 0.6533 0.3477 0.4013 0.3403 HRL 0.8853 0.4854 0.6567 0.343 0.3984 0.3367 BST 0.9026 0.5376 0.699 0.3853 0.4374 0.3718 DIN 0.8958 0.5331 0.6923 0.3795 0.431 0.3654 Our Model 0.901 0.5444 0.7094 0.391 0.4446 0.3773

[0123] As can be seen from the results in Table 1, the method proposed in this paper outperforms the baseline in all cases. MLP and NeuMF cannot effectively extract users' potential interests from their historical behaviors, which has a negative impact on their performance. Regarding methods based on graph convolutional networks or knowledge graphs, this paper organizes collaborative sequence graphs into a knowledge graph for KGAT and CKE. It turns out that they cannot achieve good recommendation results due to the same defects as MLP and NeuMF. Sequence-based methods excel in the comparison of all methods because they overcome the problem of user preference bias. However, HRL performs poorly because there is less user interaction data in the dataset, and it is more difficult to capture users' interests after filtering their historical courses. As two industrialized methods, DIN and BST also perform poorly because they do not capture additional information (such as sequential information between courses) or do not model the user's learning process.

[0124] Step 7: Explainable Recommendations

[0125] Figure 5 An explanation of the monotonically co-decreasing attention mechanism is provided, which shows a heat map of a user's attention weight. The vertical axis represents the current course index, and the horizontal axis represents the course index of attention focus. The present invention uses attention weights to calculate the contribution of his 20 registered courses to the recommended courses. This observation shows that related courses in the past contain highly predictive information. The user's preference for the current course can be effectively captured by attention with a certain decay rate. At the same time, the attention weights of courses learned in the distant past are not as informative as those learned recently. Similarity and time interval are key factors in controlling attention weights.

[0126] These observations suggest that the monotonic attention decay mechanism of the present invention can provide user feedback by linking the user's target course with the courses they have studied in the past. This information can enable users to select specific courses to review and consolidate, and then continue studying.

[0127] In summary, compared with the prior art, the present invention has the following advantages:

[0128] 1. This paper constructs a heterogeneous graph of courses, which contains user interaction data, course precedence and follow-up information, and course knowledge points. For each user and each course, this paper obtains their feature embeddings through the embedding layer, then feeds these into the knowledge extraction layer based on a graph convolutional network to obtain high-level collaborative signals between users and courses in the heterogeneous graph, as well as course precedence and follow-up information and knowledge points.

[0129] 2. This invention combines the knowledge evolution layer with a co-decaying attention mechanism to model the user's learning process. Co-decay refers to the fact that the weight of attention decays as the user's memory curve decays. In general, in the knowledge evolution layer, in addition to the similarity between courses, this invention also uses exponential decay and absolute distance metrics to calculate attention weight.

[0130] 3. The method of the present invention can provide explanations through visual attention weights, enabling users to select specific relevant courses for review and consolidation before continuing learning.

[0131] 4. The knowledge point-related recommendation system evaluation indicators can comprehensively capture the similarities between courses based on knowledge points. Applying this indicator to measure the recommendation system can more comprehensively evaluate the recommendation system, thereby improving the recommendation effect.

[0132] This paper demonstrates the effectiveness of the proposed recommendation algorithm through a series of ablation experiments and comparative experiments with real-world data collected from XuetangX. The algorithm can be deployed on online course platforms to provide learners with personalized course recommendations and review guidance.

[0133] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0134] Furthermore, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise indicated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It will also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art using ordinary skill will be able to implement the present invention set forth in the claims without undue experimentation. It will also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0135] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0136] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0137] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0138] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0139] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0140] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0141] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A course recommendation method based on heterogeneous graph and cooperative attenuation attention mechanism, characterized by: include: In the data preprocessing layer, the historical course record information and candidate recommended course information of the target user are obtained to construct a heterogeneous graph of courses, and then the input features of the input model are constructed; In the knowledge extraction layer, according to the heterogeneous graph between the user and the course, a graph convolution-based knowledge extraction layer is used to obtain high-level features of the user and the course, and enrich the input features; In the knowledge evolution layer, a cooperative decay attention mechanism is used to model the learning process of the target user, so that the input features can be combined with the learning process of the target user; In the prediction layer, the output of the knowledge evolution layer is input into the fully connected network to obtain the predicted course enrollment rate; Pushing target recommended courses to the target user based on the predicted course enrollment rate; The heterogeneous graph includes user interaction data, course precedence and follow-up relationships, and course knowledge points; The process of constructing a heterogeneous graph of courses and then constructing input features of an input model includes the following steps: Based on the registration history of users in the training set, a bipartite graph of user-corresponding courses is constructed; Construct a course precedence and follow-up relationship diagram; Determine the knowledge points included in the course based on the course catalog and syllabus; Connect the knowledge points included in the course with each course, and merge the bipartite graph and the precedence-successor relationship graph to obtain a heterogeneous graph between users and courses; According to the heterogeneous graph, a graph convolutional network is used to obtain high-order features of each node in the heterogeneous graph; According to the heterogeneous graph between the user and the course, a knowledge extraction layer based on graph convolution is used to obtain high-level features of the user and the course, and the input features are enriched, including: Obtain the initial features of users and courses through the embedding layer; For knowledge points, the semantic features of the knowledge points are obtained through the pre-trained w2v model; Each time through a knowledge extraction layer, by aggregating the node features of the neighbors corresponding to users and courses on the graph, the features of the neighboring nodes are summed and normalized to obtain the high-level features of each node. The neighbors of the user aggregation are courses, and the neighbors of the course aggregation include: semantic information of the knowledge points contained in the course, the courses before and after the current course, and the user; After multiple layers of knowledge extraction, the features of different layers of each user node and course node are averaged, and the features of the course are added to the category features corresponding to the course to obtain the final features of the node.

2. The course recommendation method based on heterogeneous graph and cooperative attenuation attention mechanism according to claim 1 is characterized in that: The adopting of the cooperative attenuation attention mechanism to model the learning process of the target user so that the input features can be combined with the learning process of the target user includes: For each course, first calculate the monotonic attention weight of each course; wherein, the monotonic is used to characterize the query feature of the course and the key features of the course learned before the course to calculate the attention score; Calculating the co-decayed attention weight according to the exponential decay rate and the monotonic attention weight; According to the attention weight after collaborative attenuation, the features of the current course are obtained by weighted summing the attention weight and each feature; The final features of each course are obtained as the input features.

3. The course recommendation method based on heterogeneous graph and cooperative attenuation attention mechanism according to claim 1 is characterized in that: The method further includes: determining evaluation indicators of course recommendation results in combination with knowledge points.

4. The course recommendation method based on heterogeneous graph and cooperative attenuation attention mechanism according to claim 3 is characterized in that: The evaluation indicators for determining course recommendation results based on knowledge points include: According to the coverage of course knowledge points, the course similarity between any two courses is calculated by counting the number of common knowledge points contained in the courses; For each user, we calculate the set of courses that the user is most likely to register for. Then, based on the course similarity, we calculate the mean similarity between the courses actually selected by the user and the recommended target courses. We obtain the overall similarity between the set of recommended courses and the target courses, and then obtain the evaluation index of the course recommendation results.

5. A system for implementing the course recommendation method based on heterogeneous graphs and cooperative attenuation attention mechanism as described in any one of claims 1 to 4, characterized in that: include: The data preprocessing layer is used to obtain the target user's historical course record information and candidate recommended course information, build a heterogeneous graph of courses, and then construct the input features of the input model; A knowledge extraction layer is used to obtain high-level features of users and courses based on the heterogeneous graph between users and courses, and enrich the input features by using a graph convolution-based knowledge extraction layer; a knowledge evolution layer, configured to adopt a cooperative decay attention mechanism to model the learning process of the target user, so that the input features can be combined with the learning process of the target user; A prediction layer, configured to input the output of the knowledge evolution layer into a fully connected network to obtain a predicted course enrollment rate; A push module is used to push target recommended courses to the target user based on the predicted course registration rate.

6. An electronic device, characterized in that: including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that The storage medium stores a program, and the program is executed by a processor to implement the method according to any one of claims 1 to 4.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images