Matrix completion method and system of heterogeneous graph neural network in academic recommendation

Abstract

The application discloses a matrix completion method and system of a heterogeneous graph neural network in academic recommendation, and the method comprises the following steps: obtaining interaction information of users and items in an academic network; constructing a heterogeneous graph of user and item relationships based on the interaction information; respectively performing model modeling on the users and the items based on the heterogeneous graph, so as to determine user representation vectors and item embedding information according to the constructed user model and item model; inputting the user representation vectors and the item embedding information into a graph neural network model to calculate the similarity of the users and the items, and performing similarity score ranking on the items in an item list based on the similarity, so as to predict the items clicked by the users according to the score ranking result. The application can recommend academic papers by using auxiliary information, and effectively enhance academic paper recommendation by using a relationship-aware heterogeneous GNN.

Description

Technical Field

[0001] This invention relates to the field of computing network information technology, and in particular to a matrix completion method and system for heterogeneous graph neural networks in academic recommendation. Background Technology

[0002] Graph Neural Networks (GNNs) are widely used in recommender systems and have demonstrated good performance. These networks use message passing and higher-order propagation to learn representations of users and items. However, existing GNN-based models recommend popular items and capture relationships between nodes of the same type, but cannot utilize auxiliary information to recommend academic papers.

[0003] Statistics show that over 68 million researchers have obtained and published more than 295 million academic papers. With the massive growth of research paper databases, recommendation has become one of the most important functions in academic search systems such as Google Scholar and AMiner. Traditional recommendation methods include content-based recommendation or collaborative filtering (CF). These methods rely on analyzing users' queries or historical behavior to model their preferences. Graph Neural Networks (GNNs) have achieved success in recommendation systems. In graph-based recommendation, the relationship between users and items (such as papers) is represented as a user-item bipartite graph. GNNs are used to analyze the graph structure and recommend academic documents to each user. GNNs learn graph structure information through message passing and generate representations of users and items. By integrating the embeddings of each node, GNNs can perform link prediction tasks on two nodes, meaning they can predict the probability of interaction between users and items. GNNs have been widely applied in various fields, such as online social networks and e-commerce. The effectiveness of GNNs has promoted the rapid development of GNN-based recommendation, including academic paper recommendation. GNN-based recommendation models, such as LightGCN, LR-GCCF, and ApeGNN, have enhanced representation learning capabilities and achieved excellent results. However, significant problems still need to be addressed in existing recommendation methods.

[0004] Limitation 1: When it comes to GNN-based recommendations, is the user or the item more important? Many existing studies focus only on user interests, recommending only the most popular items while ignoring others. However, with the increase in mobile users, the goal of this invention is to collect more information and recommend a wider variety of products. To test this, this invention compares its method, MCAP, with the representative baseline LightGCN on the AMiner-PC dataset. This invention finds that when recommending only the top 5 items, its method recommends 6141 items, while LightGCN recommends 5857 items.

[0005] Limitation 2: Does academic recommendation based on GNNs require higher-order propagation? Current GNN-based models typically construct a user-item adjacency matrix during the aggregation process, building user-item interactions into a user-item bipartite graph. Therefore, user-user relationships are obtained through higher-order propagation, e.g., from user u1 to item v1 and then to user u2. Similarly, item-item relationships are generated by propagating multi-hop neighbors, e.g., from item v2 to user u2 and then to item v3. However, long-distance propagation can lead to suboptimal representation learning, meaning that capturing user-user and item-item relationships through first-order propagation is crucial in GNNs. This helps reduce training costs and improve recommendation performance.

[0006] Limitation 3: In academic GNN recommendation, what relationships are most important? In the past, researchers mainly emphasized user-user relationships or item-item relationships, which may lead to missing important relevant information. Take LightGCN as an example; it's a standard recommendation graph convolutional network that constructs user-item interactions as a bipartite graph. To obtain user-user and item-item relationships, it uses 2-hop propagation. GraphRec and GES are social and sequential recommendation models, respectively. GraphRec can enhance user representation learning by constructing user-user relationships, improving social recommendation performance, while GES constructs item similarity in the item relationship graph to smooth item embeddings. However, as one existing study pointed out, representation learning performance can be poor due to the sparse interactions of traditional bipartite graphs. Summary of the Invention

[0007] The present invention aims to at least partially solve one of the technical problems in the related art.

[0008] To address this, this invention proposes a matrix completion method for heterogeneous graph neural networks in academic recommendation, introducing a matrix completion framework MCAP specifically designed for relationship-aware heterogeneous GNNs for academic paper recommendation. This framework establishes a relationship-aware user-user and item-item relationship graph, achieving efficient first-order propagation.

[0009] Another objective of this invention is to propose a matrix completion system for academic recommendation using heterogeneous graph neural networks.

[0010] To achieve the above objectives, this invention proposes a matrix completion method for academic recommendation using heterogeneous graph neural networks, comprising:

[0011] To obtain information about user interactions with items within academic networks;

[0012] A heterogeneous graph of the relationship between users and items is constructed based on the interaction information;

[0013] Based on the heterogeneous graph, user and item models are built respectively to determine user representation vectors and item embedding information according to the constructed user model and item model;

[0014] The user representation vector and item embedding information are input into a graph neural network model to calculate the similarity between the user and the item. Based on the similarity, the items in the item list are sorted by similarity score to predict the item that the user will click.

[0015] The matrix completion method of heterogeneous graph neural networks in academic recommendation according to embodiments of the present invention may also have the following additional technical features:

[0016] In one embodiment of the present invention, the heterogeneous graph includes a user-item interaction graph, a user-user relationship graph, and an item-item relationship graph; a user model is constructed based on the user-item interaction graph and the user-user relationship graph; and an item model is constructed based on the user-user relationship graph and the item-item relationship graph.

[0017] In one embodiment of the present invention, determining the user representation vector based on the constructed user model includes:

[0018] The first user space embedding is obtained by aggregating user information based on the user and user relationship graph.

[0019] The first item space embedding is obtained by aggregating item information based on the user and item interaction graph.

[0020] The user representation vector is learned based on the first user space embedding and the first item space embedding.

[0021] In one embodiment of the present invention, determining item embedding information based on the constructed item model includes:

[0022] The second item space embedding is obtained by aggregating item information based on the items and item relationship graph.

[0023] The second user space is embedded by aggregating user information based on the user and item interaction graph.

[0024] Item embedding information is learned based on the second item space embedding and the second user space embedding.

[0025] In one embodiment of the present invention, the second item space embedding is obtained by aggregating item information based on the items and item relationship graph, including:

[0026] Calculate the similarity between the current item and other items in the item relationship graph;

[0027] The next item spatial embedding is calculated by selecting the items with the highest similarity scores to form a neighbor list and summarizing the embeddings of multiple similar items.

[0028] To achieve the above objectives, another aspect of the present invention proposes a matrix completion system for academic recommendation using heterogeneous graph neural networks, comprising:

[0029] The interactive information acquisition module is used to acquire interactive information between users and items in the academic network;

[0030] A heterogeneous graph construction module is used to construct a heterogeneous graph of the relationship between users and items based on the interaction information;

[0031] The data modeling and representation module is used to model users and items based on the heterogeneous graph, so as to determine user representation vectors and item embedding information according to the constructed user model and item model.

[0032] The item click prediction module is used to input the user representation vector and item embedding information into a graph neural network model to calculate the similarity between the user and the item, and to sort the items in the item list according to the similarity score based on the similarity score, so as to predict the item to be clicked by the user based on the score sorting result.

[0033] This invention presents a matrix completion method and system for academic recommendation using heterogeneous graph neural networks (MCAP). The proposed MCAP can be easily integrated into any GNN-based model, and experimental results on three academic datasets demonstrate that MCAP outperforms state-of-the-art methods across all performance metrics. Online A / B testing was also conducted on multiple platforms, including PCs and WeChat, to verify the effectiveness of MCAP in providing high-quality recommendations. These tests show that MCAP is indeed effective in providing valuable suggestions. The method of this invention utilizes auxiliary information to recommend academic papers and predict items that users are likely to click on next.

[0034] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0035] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0036] Figure 1 This is a flowchart of a matrix completion method for academic recommendation using heterogeneous graph neural networks according to an embodiment of the present invention;

[0037] Figure 2 This is a framework diagram of a matrix completion method for academic recommendation using heterogeneous graph neural networks according to an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of a matrix completion method for academic recommendation using a heterogeneous graph neural network according to an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the overall experimental results according to an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of the overall experimental results compared with the knowledge graph perception method according to an embodiment of the present invention;

[0041] Figure 6 These are schematic diagrams of ablation experimental results on the AMiner-PC and CiteULIke datasets, respectively, according to embodiments of the present invention.

[0042] Figure 7 This is a schematic diagram of the matrix completion system structure of a heterogeneous graph neural network in academic recommendation according to an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0045] The following describes, with reference to the accompanying drawings, a matrix completion method and system for academic recommendation using heterogeneous graph neural networks proposed according to embodiments of the present invention.

[0046] This invention proposes a framework called MCAP, which uses a relation-aware heterogeneous GNN to enhance academic paper recommendation. The framework employs an attention mechanism to learn top-U relationships by constructing a user-user relationship graph based on co-authors and interacting papers. It utilizes the Sentence-BERT method to fuse auxiliary information from papers (titles and abstracts) to generate similarity between items, thereby constructing an item-item relationship graph. Finally, the relation-aware user-user and paper-paper graphs are integrated into an existing GNN-based model to generate user and paper representations, thereby enhancing academic paper recommendation. Figure 2 As shown.

[0047] Figure 1 This is a flowchart of a matrix completion method for heterogeneous graph neural networks in academic recommendation, according to an embodiment of the present invention.

[0048] like Figure 1 As shown, the method includes, but is not limited to, the following steps:

[0049] S1, obtain interaction information between users and items in the academic network;

[0050] S2, constructing a heterogeneous graph of the relationship between users and items based on interactive information;

[0051] S3, based on heterogeneous graphs, performs user and item modeling respectively, in order to determine user representation vectors and item embedding information according to the constructed user model and item model;

[0052] S4 inputs the user representation vector and item embedding information into the graph neural network model to calculate the similarity between the user and the item, and sorts the items in the item list according to the similarity score to predict the item that the user will click based on the score sorting result.

[0053] Understandably, the problem this invention aims to solve can be formally defined as follows: the input is a bipartite graph G = (U, I, R), where U represents the set of user nodes, I represents the set of item nodes, and R represents the interaction records between users and items. Generally, the goal of a collaborative filtering encoding model f is to first compute the representations of users and items, then rank all items by calculating the similarity between the user and item representations, and finally predict the items the user might click next based on the ranked item list.

[0054] Specifically, to address the aforementioned problems, this invention proposes a simple and effective framework that utilizes relation-aware heterogeneous items to perform the behavioral matrix of the Massive Exploration and Recommendation of Academic Papers (MCAP). The main idea behind MCAP is to directly capture user-user and item-item relationships by constructing a correlation graph between users and items. To achieve this, a user-user relationship-aware graph is first constructed using co-authors and locations in papers interacting with two users. Then, an attention mechanism is used to capture the top-U user relationships for each user. Next, an item-item relationship-aware graph is constructed by mining elements to identify co-authors and co-locations of two items. This invention also uses Sentence-BERT to calculate representational similarity of auxiliary information such as titles and abstracts, and selects top-I item relationships for each item. Finally, a bipartite graph is combined with the user-user and item-item relationship graphs as input to a general GNN model to learn representations of users and papers. By doing so, this invention effectively solves the problem.

[0055] In one embodiment of the present invention, the framework diagram and flowchart of MCAP are as follows: Figure 2 and Figure 3 As shown:

[0056] The MCAP framework consists of five parts: graph construction, user modeling, item modeling, propagation, and pooling. Each component will be explained in detail below.

[0057] Preferably, a graph is constructed. This invention constructs a heterogeneous graph that combines user-item interactions, user-user relationships, and item-item relationships. This graph has three parts: a bipartite graph for user-item interactions, a user-user relationship graph, and an item-item relationship graph. This invention constructs the user-user relationship graph and the item-item relationship graph based on content similarity.

[0058] Preferably, user modeling. The second component of the proposed model learns user representations through two independent subgraphs—a user-item graph and a relationship-aware user-user graph. The former captures user-item interactions, while the latter considers relationships between users based on co-authorship, location, and other paper-related factors. This invention uses two aggregations to comprehensively and accurately represent users in academic recommendations.

[0059] Specifically, the user modeling component of the model proposed in this invention is used to learn the user's representation vector, where user u uses h u The user representation is obtained from two subgraphs: a user-item graph and a user-user graph. The main challenge is how to aggregate each part individually and combine them correctly. To address this, this invention uses two types of embeddings to form the user representation from the two graphs. One type of aggregation, called "user aggregation," is performed on the user-user relationship graph to learn the user embeddings in the user space. The other type of aggregation, called "item aggregation," is performed on the user-item interaction graph to learn the user's item space embeddings. User modeling may include the following steps:

[0060] Step 1: User Aggregation. Social recommendations can infer user interests from their social relationships. In contrast, academic paper recommendations face the challenge of representing user preferences in the absence of direct social relationships. In academic paper recommendation scenarios, users lack direct social relationships, such as friendships, to obtain user preferences. However, each paper contains rich and substantial supplementary information, such as the author and the conference or journal in which it was published, which can effectively influence user behavior. Therefore, this invention can leverage internal user relationships to learn user embeddings in the user space.

[0061] Step 2: Item Aggregation. In MCAP, this invention follows the basic idea of ​​LightGCN, learning the user's representation in the item space within the user-item graph. Specifically, the item space embedding of the user interface represents the user interface by capturing the interactions between the user and items. The purpose of item aggregation is to model the user embedding by utilizing the items that user u interacts with.

[0062] Step 3: Joint Representation. To better learn user representations, this invention combines user space embedding in the user-user relationship graph and item space embedding in the user-item interaction graph to learn rich information from different perspectives.

[0063] Preferably, item modeling. Moving to the third part, it aims to learn item representations in an academic paper recommendation system. The data for item modeling includes a relationship-aware paper-to-paper graph. Similar to user modeling, this invention obtains item representations from two subgraphs—a user-item bipartite graph and an item-item relationship graph. The former captures user-item interactions, while the latter captures relationships and similarities between documents. Compared to general GNN-based recommendation models, the aggregation of MCAP in this invention can learn item representations from user-item interactions and paper relationships, thereby enabling the invention to capture the intrinsic attributes of papers, such as topic similarity. This result further improves the accuracy and effectiveness of academic paper recommendations.

[0064] Specifically, the purpose of item modeling is to learn the embeddings of items, where item v is represented by h. v The item representation can be obtained from two parts: one part comes from the user-item graph, which contains the interactions between users and items; the other part comes from the item-item graph, which contains the similarities between items. Therefore, this invention aggregates the embeddings from these two parts from the user space and the item space to form an embedding. The specific process of item modeling may include:

[0065] Step 1: Item Aggregation. In the model proposed in this invention, item-item graphs are used to learn item embeddings in the item space. Specifically, the item space embedding of item v is used to aggregate related items. To learn and calculate the similarity between items, this invention selects specific information content including title, abstract, author, and the conference or journal of the item. Relationships are captured for v. i and v j This invention uses three relationships: co-author (co-a), co-location (co-v), and similarity coefficient sim. If a paper has co-authors, then co-a is 1. To improve training efficiency, this invention selects top-I similar items to construct the item-item interaction graph. Specifically, this invention calculates item v... iThe similarity score between the item and all other items in the item-item graph is used to select the top-I items with the highest similarity scores to form a neighbor list. Item spatial embedding is calculated by aggregating the embeddings of the top-I similar items. By selecting the top-I similar items for each item, the model of this invention can reduce the computational complexity of item aggregation, improve training efficiency, and still capture the most relevant and informative relationships between items.

[0066] Step 2: User Aggregation. Similar to item aggregation, user aggregation aims to learn item embeddings in the user space from the user-item graph. Specifically, this invention uses a simple weighted sum aggregator to aggregate items v. j Interact with each neighbor. Item v j User space embedding. By aggregating the embeddings of neighbors in the user space, the model of this invention can effectively capture the characteristics of users interacting with the item, thereby representing the item in the user space more accurately and effectively.

[0067] Step 3: Joint representation. To learn item v j The present invention combines item space embedding and user space embedding to obtain item v. j The final representation. By jointly representing the item embeddings from the item space and the user space, the model of the present invention can capture complementary information from two perspectives.

[0068] Preferably, propagation and pooling are employed. The final component of MCAP obtains the final representations of the user and items. These representations are then used to predict which items the user is likely to click.

[0069] Therefore, this invention inputs user interaction records with items, then learns vector representations of users and items through an encoder, and finally sorts the item list by calculating the similarity between users and items, and predicts the list of items that the user may click.

[0070] Furthermore, this invention conducted comprehensive experiments on three datasets (AMiner-WeChat, AMiner-PC, and CiteULike), with Recall@20 and NDCG@20 used as evaluation metrics.

[0071] Specifically, in terms of overall performance, this invention was compared with several representative methods (traditional methods, MF-based, GNNs-based, and KB-based) in a top-20 performance comparison. Experimental results show that this method outperforms all compared methods in both Recall@20 and NDCG@20. Empirical statistical results (p-value) indicate that this method significantly and consistently outperforms all compared baseline models. Figure 4 and Figure 5As shown.

[0072] Specifically, in ablation experiments, taking AMiner-PC as an example, this invention uses LightGCN as the basic encoding model and performs comprehensive ablation experiments to verify the necessity of each component in MCAP. The results show that removing one or two of the three components leads to performance degradation. This phenomenon indicates that MCAP's user modeling and item modeling design is effective in heterogeneous graph collaborative filtering. Figure 6 (a) and Figure 6 As shown in (b) of the diagram.

[0073] In summary, this invention validated the effectiveness of MCAP using three academic datasets (AMiner-WeChat, AMiner-PC, and CiteULike), which include user-item interactions and paper-supporting information. Comprehensive experiments show that on the AMiner-WeChat dataset, MCAP outperforms state-of-the-art models in Recall@10, NDCG@10, and HR@10, achieving performance of 69.12%, 70.70%, and 77.42%, respectively.

[0074] In summary, the proposed method MCAP can be easily integrated into any existing GNN-based model without requiring any changes to the model architecture. This makes it a convenient plugin for recommending academic items. Extensive experiments were conducted to evaluate the effectiveness of MCAP using three academic datasets (AMiner-WeChat, AMiner-PC, and CiteULike) containing user-item interactions and supplementary information from papers. MCAP was compared with traditional methods and representative GNN-based models. The comprehensive experiments demonstrate that MCAP outperforms state-of-the-art models in recall, NDCG, and click-through rate across all datasets. These results highlight the benefits of relation-aware heterogeneous GNNs for recommending academic papers.

[0075] The matrix completion method of heterogeneous graph neural networks in academic recommendation, as described in this invention, is specifically designed for relationship-aware heterogeneous GNNs for academic paper recommendations. This framework establishes a relationship-aware user-user and item-item relationship graph, achieving efficient first-order propagation. The proposed MCAP can be easily integrated into any GNN-based model to further improve the accuracy and effectiveness of academic paper recommendations.

[0076] To achieve the above embodiments, such as Figure 7As shown, this embodiment also provides a matrix completion system 10 for heterogeneous graph neural networks in academic recommendation. The system 10 includes an interactive information acquisition module 100, a heterogeneous graph construction module 200, a data modeling and representation module 300, and an item click prediction module 400.

[0077] The interactive information acquisition module 100 is used to acquire interactive information between users and items in the academic network;

[0078] Heterogeneous graph construction module 200 is used to construct heterogeneous graphs of user-item relationships based on interaction information;

[0079] The data modeling and representation module 300 is used to model users and items based on heterogeneous graphs, so as to determine user representation vectors and item embedding information according to the constructed user model and item model.

[0080] The item click prediction module 400 is used to input user representation vectors and item embedding information into a graph neural network model to calculate the similarity between users and items, and to sort the items in the item list based on similarity scores, so as to predict the item to be clicked by the user according to the score sorting results.

[0081] Furthermore, the heterogeneous graph includes a user-item interaction graph, a user-user relationship graph, and an item-item relationship graph; a user model is constructed based on the user-item interaction graph and the user-user relationship graph; and an item model is constructed based on the user-user relationship graph and the item-item relationship graph.

[0082] Furthermore, the data modeling and representation module 300 is also used for:

[0083] The first user space embedding is obtained by aggregating user information based on the user and user relationship graph.

[0084] The first item space embedding is obtained by aggregating item information based on the user and item interaction graph.

[0085] The user representation vector is learned based on the first user space embedding and the first item space embedding.

[0086] Furthermore, the data modeling and representation module 300 also includes:

[0087] An item information aggregation unit is used to aggregate item information based on the item and item relationship graph to obtain a second item space embedding.

[0088] The user information aggregation unit is used to aggregate user information based on the user and item interaction graph to obtain the second user space embedding.

[0089] The item embedding learning unit is used to learn item embedding information based on the second item space embedding and the second user space embedding.

[0090] Furthermore, the item information aggregation unit is also used for:

[0091] Calculate the similarity between the current item and other items in the item relationship graph;

[0092] The next item spatial embedding is calculated by selecting the items with the highest similarity scores to form a neighbor list and summarizing the embeddings of multiple similar items.

[0093] This invention presents a matrix completion system for heterogeneous graph neural networks in academic recommendation, specifically designed for relationship-aware heterogeneous GNNs for academic paper recommendations. This framework establishes a relationship-aware user-user and item-item relationship graph, achieving efficient first-order propagation. The proposed MCAP can be easily integrated into any GNN-based model to further improve the accuracy and effectiveness of academic paper recommendations.

[0094] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions 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 one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0095] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A matrix completion method for academic recommendation using heterogeneous graph neural networks, characterized in that, The method includes the following steps: To obtain information about user interactions with items within academic networks; Construct a user-item interaction graph based on the aforementioned interaction information; Based on the interaction information, the co-author and interactive paper relationships between users are extracted, and a user-user relationship graph is constructed. The Sentence-BERT method is used to fuse the titles and summaries of the items to generate similarity between items. Based on the similarity, the items with the highest similarity are selected to form a neighbor list, and an item-item relationship graph is constructed. User modeling is performed based on the user-item interaction graph and the user-user relationship graph, and item modeling is performed based on the user-item interaction graph and the item-item relationship graph to determine user representation vectors and item embedding information; wherein, the user modeling and the item modeling capture user-user and item-item relationships through first-order propagation. The user representation vector and item embedding information are input into a graph neural network model to calculate the similarity between the user and the item. Based on the similarity, the items in the item list are sorted by similarity score to predict the item that the user will click.

2. The method according to claim 1, characterized in that, The determination of the user representation vector includes: The first user space embedding is obtained by aggregating user information based on the user-user relationship graph. The first item space embedding is obtained by aggregating item information based on the user-item interaction graph. The user representation vector is learned based on the first user space embedding and the first item space embedding.

3. The method according to claim 1, characterized in that, The determination of the embedded information of the item includes: The second item space embedding is obtained by aggregating item information based on the item-item relationship graph. The second user space is embedded by aggregating user information based on the user-item interaction graph. Item embedding information is learned based on the second item space embedding and the second user space embedding.

4. The method according to claim 3, characterized in that, The step of aggregating item information based on the item-item relationship graph to obtain the second item space embedding includes: The spatial embedding of the second item is calculated by aggregating the embeddings of multiple similar items in the neighbor list.

5. A matrix completion system for academic recommendation using heterogeneous graph neural networks, characterized in that, include: The interactive information acquisition module is used to acquire interactive information between users and items in the academic network; The heterogeneous graph construction module is used to construct a user-item interaction graph based on the interaction information; and to extract co-author and interactive paper relationships between users based on the interaction information to construct a user-user relationship graph. Furthermore, the Sentence-BERT method is used to fuse the titles and summaries of the items, generate similarity between items, and select the items with the highest similarity to form a neighbor list based on the similarity, thus constructing an item-item relationship graph; The data modeling and representation module is used to perform user modeling based on the user-item interaction graph and the user-user relationship graph, and to perform item modeling based on the user-item interaction graph and the item-item relationship graph, so as to determine the user representation vector and item embedding information; wherein, the user modeling and the item modeling capture user-user and item-item relationships through first-order propagation; The item click prediction module is used to input the user representation vector and item embedding information into a graph neural network model to calculate the similarity between the user and the item, and to sort the items in the item list according to the similarity score based on the similarity score, so as to predict the item to be clicked by the user based on the score sorting result.

6. The system according to claim 5, characterized in that, The data modeling and representation module is also used for: The first user space embedding is obtained by aggregating user information based on the user-user relationship graph. The first item space embedding is obtained by aggregating item information based on the user-item interaction graph. The user representation vector is learned based on the first user space embedding and the first item space embedding.

7. The system according to claim 5, characterized in that, The data modeling and representation module also includes: The item information aggregation unit is used to aggregate item information based on the item-item relationship graph to obtain a second item space embedding. The user information aggregation unit is used to aggregate user information based on the user-item interaction graph to obtain the second user space embedding. The item embedding learning unit is used to learn item embedding information based on the second item space embedding and the second user space embedding.

8. The system according to claim 7, characterized in that, The item information aggregation unit is also used for: The spatial embedding of the second item is calculated by aggregating the embeddings of multiple similar items in the neighbor list.

Images