A user learning interest label prediction method based on a graph neural network

Abstract

The application relates to a user learning interest label prediction method based on a graph neural network, which comprises the following steps: obtaining user historical learning behavior data of a user platform within a set time range; mapping the user platform into a graph structure; pre-constructing and training a learning behavior interest feature extraction model based on closeness and learning behavior similarity; the model is a double-layer graph neural network, the first layer extracts user learning behavior interest features based on the closeness between users, and the second layer extracts user learning behavior interest features based on the historical learning behavior similarity of the user; taking the structure of the user platform and the obtained user historical learning behavior data as the input of the trained learning behavior interest feature extraction model, extracting long-term and short-term learning behavior interest features of the user; and inputting the long-term and short-term learning behavior interest features of the user into a pre-constructed and trained prediction model to obtain accurate learning interest labels of the user. Compared with the prior art, the application has the advantages of improved learning interest label prediction accuracy and good practicability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for predicting user learning interest tags based on graph neural networks. Background Technology

[0002] With the rapid development of information technology, research on educational issues has expanded from traditional educational disciplines to how to integrate information technology, especially artificial intelligence, into educational practice. The development of modern network communication technology enables learners to acquire knowledge through online knowledge-sharing communities such as Zhihu. Whether user learning behaviors and interactions occur in such online public learning spaces and whether they promote user learning performance has become a quality evaluation standard for online knowledge-sharing communities. High-quality online knowledge-sharing communities can foster effective interaction among users, thereby promoting user learning performance.

[0003] A user's historical learning behavior can reflect their learning interests to some extent, and the social relationships within online knowledge-sharing communities also influence these interests. As an education-related social network, online knowledge-sharing communities can map these social relationships into a graph structure, mapping non-Euclidean network distances to edges connecting nodes. Currently, research on learner learning characteristics mainly relies on data analysis based on questionnaires, along with human judgment from professional educators. Deep learning methods have achieved breakthroughs in various fields in recent years, such as image recognition, natural language processing, bioengineering, and feature learning and analysis of spatiotemporal structures. Deep learning-based embedding models have also gained attention. Through effective training on large amounts of data, deep learning can effectively extract the temporal and spatial correlations between data, a capability that traditional prediction methods lack.

[0004] Traditional research on the analysis of educational and learning characteristics still relies on data analysis based on questionnaire surveys. While this method can yield good results, it is susceptible to subjective factors and cannot handle the large amounts of user interaction information and learning behavior sequence information stored in online knowledge-sharing communities. Common representation learning models, when using graph neural networks to process graph data and treating users in online knowledge-sharing communities as user-related features, still face several problems: Traditional graph neural network models such as GNN, GCN, and GAT can significantly improve the performance of embedding models, but require the participation of all nodes, making them unsuitable for real-world complex networks; Efficient graph embedding models based on GNN, such as GraphSAGE and PinSAGE, can overcome these shortcomings by extracting subgraph structures from the global graph for partial sampling, greatly reducing computation. However, the neighbor nodes selected based on the random walk algorithm during sampling differ from the user's choice of learning interaction objects (people or learning content) in the online knowledge-sharing community. Using random selection results affects the accuracy of the aggregated representation during feature extraction, reducing the accuracy of user label prediction. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a user learning interest tag prediction method based on graph neural networks, thereby improving the accuracy of user learning interest tag prediction.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A user learning interest label prediction method based on graph neural networks includes the following steps:

[0008] Acquire historical learning behavior data of users on the user platform within a set time range;

[0009] The user platform is mapped as a graph structure. A learning behavior interest feature extraction model based on intimacy and similarity of learning behavior is pre-built and trained. The learning behavior interest feature extraction model is a two-layer graph neural network. The first layer of the two-layer graph neural network extracts user learning behavior interest features based on the intimacy relationship between users, and the second layer extracts user learning behavior interest features based on the similarity of users' historical learning behaviors.

[0010] The structure of the user platform and the acquired historical learning behavior data of users are used as inputs to the trained learning behavior interest feature extraction model to extract users' long-term and short-term learning behavior interest features.

[0011] The extracted user long-term and short-term learning behavior interest features are input into a pre-built and trained prediction model to obtain accurate user learning interest labels.

[0012] Furthermore, the process of acquiring the user's historical learning behavior data is as follows:

[0013] Acquire users' historical learning behavior data, and perform classification sampling and encoding to obtain classification results for users' answer collection sequences, answer browsing sequences, and search sequences.

[0014] Furthermore, mapping the user platform as a graph structure specifically involves:

[0015] Each user in the user platform is considered as a node in the graph, and the follow relationships between users are considered as the edges between nodes.

[0016] Furthermore, the first layer of the dual-layer graph neural network constructs a probability transition matrix based on close relationships and obtains the user neighborhood structure through probabilistic walks. The computational expression for the first layer of the dual-layer graph neural network is:

[0017]

[0018]

[0019] In the formula, A uu Let N represent the adjacency matrix of the network graph structure, where p(v|u) is the probability of traversing from node u to node v in the next time step. (u) N (v) Let u and v represent the number of direct neighbor nodes of nodes u and v, respectively, and α be a trainable parameter. The transition probability matrix based on proximity is obtained when traversing the first layer of the network. The network is traversed starting from node u based on this transition matrix. The convergence result N is obtained when traversing back to node u or reaching a J-hop neighbor. DFS N DFS Filter the neighbors of node u and aggregate the neighbor information.

[0020] Furthermore, the second layer of the dual-layer graph neural network constructs a probability transition matrix based on the similarity of users' historical learning behaviors and obtains the user's neighborhood structure through probabilistic walks. The calculation expression for the second layer of the dual-layer graph neural network is as follows:

[0021] d uv =ω1D+ω2I

[0022]

[0023]

[0024] In the formula, the distance between the historical learning behavior features of users u and v is obtained by aligning the historical learning behavior features of users u and v. u,v Let D and I represent the deletion and insertion operands during the alignment of the learning action sequence, respectively, and wi |u represents the cost of the corresponding operation. k-1 | represents the embedding vector obtained by node u after passing through the previous GCN layer, |v k-1 Similarly, β is a learnable parameter; we obtain the transition probability matrix based on behavioral similarity:

[0025]

[0026] The high-frequency nodes in the traversal sequence form the first-order neighbors of node u in the virtual neighborhood, and the neighbor information is aggregated.

[0027] Furthermore, the different historical learning behavior sequence features of each layer of the dual-layer graph neural network based on the node's neighbor structure are extracted by the trained model. The calculation expression for this feature extraction process is as follows:

[0028]

[0029]

[0030] In the formula, This represents the hidden long-term learning behavior and interest features mined after aggregating the features of nodes in the neighborhood of user u at layer k. The long-term learning behavior and interest features of node u at layer k are obtained through feature concatenation and linear transformation. Similarly This indicates that the extracted short-term learning behavior interest features are concatenated with the extracted long-term and short-term learning behavior interest features to obtain the user's long-term and short-term learning behavior interest features. W represents the final short-term and long-term learning behavior and interest features extracted from user u after training a k-layer learning behavior and interest feature extraction model. k represents the model weights for the k-th layer.

[0031] Furthermore, the prediction model performs accurate classification prediction of user learning interest tags based on the extracted long-term and short-term learning behavior interest features. The expression for this classification prediction is:

[0032]

[0033] In the formula, This represents the label prediction result after user node classification, where σ is the model parameter, W is the model weight, and m is the model weight. u This identifies users' short-term and long-term learning behaviors and interest characteristics.

[0034] Furthermore, the prediction model is trained by minimizing the cross-entropy between the predicted and true values ​​and by backpropagation;

[0035] The loss function expression of the prediction model is:

[0036]

[0037] In the formula, L is the loss function of the prediction model, l u The actual results of labeling user nodes after classification.

[0038] Furthermore, the method is used to predict user learning interest tags on online knowledge-sharing communities.

[0039] Furthermore, the user's learning behavior interest features are embedded from the node features extracted from the learner's learning behavior and social relationships in online knowledge-sharing communities over a period of time, in order to reflect the possible learning behavior sequence that the learner may generate under the influence of close relationships and historical behavior.

[0040] The learning behavior refers to a series of actions that learners take on online knowledge-sharing communities.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] (1) Improve the accuracy of learning interest tag prediction: This invention focuses on the currently widespread online knowledge sharing community, classifies and processes users' historical learning behavior data to represent the learning behavior interest trends of users at different times, and uses the graph neural network-based tag prediction model proposed in this invention to deeply extract the influencing factors of users' learning behavior interest trends: close relationship and learning behavior similarity dependence, and finally obtains the user's learning behavior interest features and user education-related tag information. Compared with traditional methods, it has the advantages of good feature extraction effect and high tag prediction accuracy.

[0043] (2) High Practicality: This invention leverages the advantages of deep learning algorithms to process large-scale data in online knowledge-sharing communities without requiring data analysis based on questionnaires or manual annotation. Furthermore, considering the unique nature of online knowledge-sharing communities—their social network attributes—the data processing is mapped to graph-based data processing, resulting in fast processing speed and high practicality. This invention also provides reliable data for downstream tasks such as learning content recommendation and learning partner recommendation, thus possessing practical application value. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the main steps of a user learning interest tag prediction method based on graph neural networks provided in an embodiment of the present invention;

[0045] Figure 2 This is a neighborhood construction graph provided in an embodiment of the present invention;

[0046] Figure 3This is a schematic diagram of the structure of a learning behavior interest feature extraction model provided in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0048] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0049] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0050] In the description of this invention, it should be noted that the definitions of learning behavior and learning behavior interest characteristics are as follows:

[0051] Definition 1: Learning behavior: Learners in online knowledge-sharing communities have both educational and social attributes. Therefore, a series of behaviors generated by learners in online knowledge-sharing communities can be considered their learning behaviors.

[0052] Definition 2 Learning Behavior Interest Features: In this embodiment, learning behavior interest features refer to the node feature embedding extracted from the learner's learning behavior and social relationships in online knowledge sharing communities over a period of time, which is used to reflect the possible learning behavior sequence that the learner may generate under the influence of close relationships and historical behavior.

[0053] Example 1

[0054] This embodiment provides a user learning interest tag prediction method based on graph neural networks, including the following steps:

[0055] Acquire historical learning behavior data of users on the user platform within a set time range;

[0056] The user platform is mapped as a graph structure. A learning behavior interest feature extraction model based on intimacy and similarity of learning behavior is pre-built and trained. The learning behavior interest feature extraction model based on intimacy and similarity of learning behavior is a two-layer graph neural network. The first layer of the two-layer graph neural network extracts user learning behavior interest features based on the intimacy relationship between users, and the second layer extracts user learning behavior interest features based on the similarity of users' historical learning behavior.

[0057] The structure of the user platform and the acquired historical learning behavior data are used as inputs to a trained learning behavior interest feature extraction model to extract users' short-term and long-term learning behavior interest features.

[0058] The extracted user long-term and short-term learning behavior interest features are input into a pre-built and trained prediction model to obtain accurate user learning interest labels.

[0059] The aforementioned user platform can be a platform with both educational and social attributes. In this embodiment, the above method is used to predict user learning interest tags on an online knowledge-sharing community, such as... Figure 1-3 As shown, the specific implementation process of this method is as follows:

[0060] Step S1: Obtain historical learning behavior data of users within a time range T on the online knowledge sharing community, and classify and sample the data;

[0061] Data on users' historical learning behavior within a time frame T in an online knowledge-sharing community is collected. The data is then categorized, sampled, and coded. Users' historical learning behavior includes their answer collection sequence, answer browsing sequence, and search sequence. These three types of learning behavior are categorized, with the answer collection sequence used to reflect users' long-term learning interests, while the answer browsing and search sequences are used to reflect users' short-term learning interests because these sequences overwrite older records. This represents the sequence of collection actions of user i. Let X represent the m1-th favorite answer by user i, where X = (x1, x2, ..., x...). n ) T This represents the feature vector matrix of the answer-collection behavior sequence of n users in the network; similarly... and Y = (y1, y2, ..., y3) represents the sequence of user i's answer browsing behavior and the sequence of search keywords, respectively. n ) T and Z = (z1, z2, ..., z n ) T These represent the feature vector matrix of the answer browsing behavior sequence and the feature vector matrix of the search keyword sequence of n users in the network, respectively, which together characterize the user's short-term learning interest.

[0062] Step S2: Map the online knowledge-sharing community as a graph structure and construct a learning behavior interest feature extraction model based on intimacy and similarity in learning behavior;

[0063] Specifically, based on deep learning principles, this paper utilizes a graph neural network-based method for predicting user learning interest tags in online knowledge-sharing communities to more accurately extract user learning behavior interest features, thereby improving the accuracy of learning interest tag prediction. This also enhances the effectiveness of downstream tasks such as learning content recommendation and user clustering in online learning communities. Each user in the online knowledge-sharing community is considered a node in a graph, with connections between nodes representing follower relationships. A user's historical learning behavior data over time period T is used as the node feature vector. A two-layer GNN network is constructed: the first layer extracts user learning behavior interest features based on the closeness of relationships between users, and the second layer extracts user learning behavior interest features based on the similarity of users' historical learning behaviors.

[0064] Based on the close relationships between users, user learning behavior and interest features are extracted. A user neighborhood structure is constructed based on these close relationships.

[0065]

[0066]

[0067] Where A uv Let N represent the adjacency matrix of the network graph structure, where p(v|u) is the probability of traversing from node u to node v in the next time step. (u) N (v) Let u and v represent the number of direct neighbor nodes of nodes u and v, respectively, and α be a trainable parameter. The transition probability matrix based on proximity is obtained when traversing the first layer of the network. The network is traversed starting from node u based on this transition matrix. The convergence result N is obtained when traversing back to node u or reaching a J-hop neighbor. DFS N DFS Filter the neighbors of node u and aggregate the neighbor information.

[0068] Based on the similarity of users' historical learning behaviors, user learning behavior interest features are extracted, and user neighborhood nodes are constructed using the similarity of historical learning behaviors.

[0069] d uv =w1D+w2I

[0070]

[0071]

[0072] Aligning the historical learning behavior features of users u and v, we obtain the distance between the historical learning behaviors of users u and v, denoted by d. u,v Let ω represent the number of deletion and insertion operands when aligning learning action sequences, where D and I represent the number of deletion and insertion operands, respectively. i |u represents the cost of the corresponding operation. k-1 | represents the embedding vector obtained by node u after passing through the previous GCN layer, |v k-1 Similarly, β is a learnable parameter; we obtain the transition probability matrix based on behavioral similarity:

[0073]

[0074] The high-frequency nodes in the traversal sequence form the first-order neighbors of node u in the virtual neighborhood, and the neighbor information is aggregated.

[0075] Step S3: Using community structure and user historical learning behavior data as input to the trained learning behavior interest feature extraction model, obtain user short-term and long-term learning behavior interest features. The features of different historical learning behavior sequences of users in a single-layer network based on the node neighbor structure are extracted by the trained model:

[0076]

[0077]

[0078] This represents the hidden long-term learning behavior and interest features mined after aggregating the features of nodes in the neighborhood of user u at layer k. The long-term learning behavior and interest features of node u at layer k are obtained through feature concatenation and linear transformation. Similarly This indicates that the extracted short-term learning behavior interest features are concatenated with the extracted long-term and short-term learning behavior interest features to obtain the user's long-term and short-term learning behavior interest features. This represents the final learning behavior and interest features extracted from user u after training a k-layer network model.

[0079] Step S4: Use the user learning behavior and interest features extracted by the model to classify and predict users.

[0080]

[0081] in This represents the label prediction result after user node classification. By minimizing the cross-entropy between the predicted and true values ​​and training the model through backpropagation, the accuracy of learning interest label prediction can be improved. This can also enhance the effectiveness of downstream tasks such as learning content recommendation and user clustering in online learning communities.

[0082]

[0083] In the formula, L is the loss function of the prediction model, l u The actual results of labeling user nodes after classification.

[0084] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting user learning interest tags based on graph neural networks, characterized in that, Includes the following steps: Acquire historical learning behavior data of users on the user platform within a set time range; The user platform is mapped as a graph structure. A learning behavior interest feature extraction model based on intimacy and similarity of learning behavior is pre-built and trained. The learning behavior interest feature extraction model is a two-layer graph neural network. The first layer of the two-layer graph neural network extracts user learning behavior interest features based on the intimacy relationship between users, and the second layer extracts user learning behavior interest features based on the similarity of users' historical learning behaviors. The structure of the user platform and the acquired historical learning behavior data of users are used as inputs to the trained learning behavior interest feature extraction model to extract users' long-term and short-term learning behavior interest features. The extracted user long-term and short-term learning behavior interest features are input into a pre-built and trained prediction model to obtain accurate user learning interest labels. The prediction model accurately classifies and predicts users' learning interest tags based on the extracted long-term and short-term learning behavior interest features. The prediction model is trained by minimizing the cross-entropy between the predicted and the true values ​​and by backpropagation. The first layer of the dual-layer graph neural network constructs a probability transition matrix based on close relationships and obtains the user neighborhood structure through probabilistic walks. The computational expression for the first layer of the dual-layer graph neural network is: In the formula, Represents the adjacency matrix of the network graph structure. p(v|u) To start from node in the next time step u Wander to the node v The probability, N (u) 、N (v) Representing the nodes respectively u、v The number of direct neighbor nodes, α For a trainable parameter, It is the transition probability matrix based on affinity obtained when passing through the first layer of the network. Based on the transition matrix, the nodes... u Start traversing the network, and when you reach a node again... u Or traverse to J Obtain the convergence result of the traversal when jumping to the neighbor. N DFS , N DFS Filtering nodes u It aggregates neighbor information; The second layer of the dual-layer graph neural network constructs a probability transition matrix based on the similarity of users' historical learning behaviors and obtains the user's neighborhood structure through probabilistic walks. The calculation expression for the second layer of the dual-layer graph neural network is as follows: In the formula, align users u、v Historical learning behavior characteristics were obtained u, v User historical learning behavior distance d u,v It means that, among them D and I These represent the deletion and insertion operands, respectively, when aligning the learning behavior sequence. Indicates the cost of the corresponding operation, | u k-1 | Represents a node u The embedding vector obtained after the previous GCN layer, | v k-1 Similarly, β Given a learnable parameter; obtain the transition probability matrix based on behavioral similarity: High-frequency elements in the traversal sequence are represented by nodes. u First-order neighbors in the virtual neighborhood, and aggregate neighbor information.

2. The user learning interest tag prediction method based on graph neural networks according to claim 1, characterized in that, The process of obtaining the user's historical learning behavior data is as follows: Acquire users' historical learning behavior data, and perform classification sampling and encoding to obtain classification results for users' answer collection sequences, answer browsing sequences, and search sequences.

3. The user learning interest tag prediction method based on graph neural networks according to claim 1, characterized in that, Mapping the user platform as a graph structure specifically involves: Each user in the user platform is considered as a node in the graph, and the follow relationships between users are considered as the edges between nodes.

4. The user learning interest tag prediction method based on graph neural networks according to claim 1, characterized in that, Each layer of the dual-layer graph neural network, based on the node's neighbor structure, extracts features from the user's different historical learning behavior sequences through the trained model. The calculation expression for this feature extraction process is as follows: In the formula, Indicates user u No. k Hidden long-term learning behavioral interest features are mined after aggregating the features of nodes within the neighborhood of the layer. The 1st layer is obtained by feature concatenation and linear transformation. k Layer nodes u Long-term learning behavior and interest characteristics are represented as Similarly , This indicates that the extracted short-term learning behavior interest features are concatenated with the extracted long-term and short-term learning behavior interest features to obtain the user's long-term and short-term learning behavior interest features. , indicating user u After training k The final user long-term and short-term learning behavior interest features extracted after the layered learning behavior interest feature extraction model. For the first k Model weights for each layer.

5. The user learning interest tag prediction method based on graph neural networks according to claim 1, characterized in that, The expression for the classification prediction is: In the formula, This represents the label prediction result after user node classification. For model parameters, For model weights, This identifies users' short-term and long-term learning behaviors and interest characteristics.

6. The user learning interest tag prediction method based on graph neural networks according to claim 5, characterized in that, The loss function expression of the prediction model is: In the formula, The loss function of the prediction model, The actual results of labeling user nodes after classification.

7. The user learning interest tag prediction method based on graph neural networks according to claim 1, characterized in that, The method described is used to predict user learning interest tags in online knowledge-sharing communities.

8. The user learning interest tag prediction method based on graph neural networks according to claim 7, characterized in that, The user learning behavior interest features are embedded from node features extracted from the learner’s learning behavior and social relationships in online knowledge sharing communities over a period of time, in order to reflect the possible learning behavior sequence that learners may generate under the influence of close relationships and historical behavior. The learning behavior refers to a series of actions that learners take on online knowledge-sharing communities.

Images