Knowledge graph recommendation system fusing depth field information and importance information
By combining deep search layer-by-layer sampling and importance neighborhood sampling strategies, the problem of importance differences in the knowledge graph recommendation model when integrating high-order information is solved, more accurate user preference mining and personalized recommendations are achieved, and the recommendation system is improved intelligence and efficiency.
Patent Information
- Application Number
- CN202411070150.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-19
AI Technical Summary
The existing knowledge graph-based recommendation model fails to fully consider the importance differences between the source node and the target node when integrating higher-order information, resulting in the omission of key information and the introduction of noise, making it difficult to fully capture the potential interests and preferences of users.
The deep search layer-by-layer sampling and importance neighborhood sampling strategy are used to obtain the deep connection and potential preferences behind users and products through the deep search layer-by-layer sampling module, graph embedding module, importance neighborhood sampling module and prediction module. Reduce the uncertainty caused by random sampling and provide clear directional guidance for model training.
It achieves more accurate acquisition of potential preferences between users and products, reduces the impact of noise, accelerates the model training process, improves the intelligence and efficiency of the recommendation system, and provides users with a personalized shopping experience.
Smart Images

Figure CN118981569A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of recommendation methods, and specifically to a knowledge graph recommendation system that integrates deep domain information and importance information. Background Art
[0002] As a powerful assistant for e-commerce websites, the recommendation system aims to accurately push product information and personalized suggestions to customers, simulate and optimize the personalized service experience in the traditional sales process, and help users make efficient decisions and easily purchase their favorite products. With the booming development of e-commerce today, the number and variety of goods are growing explosively. Consumers are often faced with the challenge of an ocean of information and need to spend a lot of time screening. This process is not only cumbersome, but also may lead to user loss due to information overload.
[0003] To cope with this dilemma, personalized recommendation systems have emerged. Relying on powerful data mining technology, they build an intelligent bridge connecting massive data with consumers' personalized needs. However, traditional recommendation systems are limited by the sparsity of historical interaction data between users and products and the cold start challenge, and their effectiveness is difficult to fully exert. In this context, knowledge graphs are introduced into the recommendation system as an auxiliary information, which effectively enhances the accuracy and diversity of recommendations.
[0004] However, when integrating high-order information, existing recommendation models based on knowledge graphs often fail to fully consider the importance difference between source nodes and target nodes, which may lead to the omission of key node information and introduce unnecessary noise, making it difficult to fully capture users' potential interests and preferences. In view of this, the present invention proposes a sampling strategy that subtly combines deep search sampling with importance neighborhood sampling strategies, aiming to more accurately explore the deep connections and potential preferences behind users and products. This strategy not only effectively reduces the uncertainty caused by random sampling, but also provides clear direction guidance for model training, accelerates the learning process, makes the recommendation system more intelligent and efficient, and brings users unprecedented personalized shopping enjoyment. Therefore, in view of this, the existing structure is studied and improved to provide a knowledge graph recommendation system that integrates deep domain information and importance information, in order to achieve a more practical purpose. Summary of the invention
[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:
[0006] The present invention provides a knowledge graph recommendation system integrating deep domain information and importance information, comprising a deep search layer-by-layer sampling module, a graph embedding module, an importance neighborhood sampling module, a neighborhood aggregation module and a prediction module;
[0007] The deep search layer-by-layer sampling module includes a deep search strategy component and a single hidden layer neural network component;
[0008] The graph embedding module adopts the Skip-gram method in the Word2vec model to predict the probability of the context based on the current word;
[0009] The importance neighborhood sampling module quantifies the closeness relationship between each node in the network and the central node;
[0010] The neighborhood aggregation module defines the user's attention factor for the relationship, and uses the attention factor to perform weighted aggregation on the node's neighbor information, and finally merges the node's own information with the aggregated neighbor information;
[0011] The prediction module is configured to obtain recommendation results based on the input user embedding and item embedding vectors.
[0012] As a preferred technical solution of the present invention, a deep search strategy is used in the knowledge graph to sample nodes layer by layer to obtain multiple wandering paths. For the interaction (u, v) of user u on item v, item v is first located in the knowledge graph, and then starting from item node v, the nodes in the graph are sampled from the inside to the outside and wandering layer by layer to obtain multiple node sequences. The depth information of the node is controlled by controlling the number of layers of the node wandering.
[0013] As a preferred technical solution of the present invention, after obtaining multiple wandering paths, a single hidden layer neural network is used to train the node sequence set obtained after sampling. The input layer of the single hidden layer neural network inputs the node sequence, where "1" represents the node at the corresponding position of the input, and "0" represents the predicted node. That is, in the training process, the current node is used to predict the information of other nodes connected to it. After convergence through similar multiple path training, the weight matrix of the hidden layer is used as the node embedding matrix vD.
[0014] As a preferred technical solution of the present invention, the importance domain adopts a module, which is characterized by quantifying the closeness relationship between each node and the central node in the network, and specifically includes the following steps:
[0015] Step S1: Calculate the closeness between all nodes in the network and the central node. The higher the value, the closer the relationship between the node and the central node.
[0016] Step S2: Sort the compactness and select the first k values. The nodes corresponding to these k value pairs are the target nodes for sampling.
[0017] The domain aggregation module includes the following steps:
[0018] Step S1: Define the attention factor of user u to relationship r. The attention factor of user u to relationship r indicates the degree of attention of the user to a specific relationship. We use function f to represent the attention score of user u to relationship r. The specific formula is as follows:
[0019]
[0020] Step S2: Normalization. In order to make the attention factors comparable between different relationships, we need to normalize the attention factors. The specific formula is as follows:
[0021]
[0022] Step S3: Use the normalized attention factor to perform weighted aggregation on the neighbor information of node v, using Indicates, as follows:
[0023]
[0024] Step S4: Fusion of node v's own information and neighbor information. Fusion of node v's own information and aggregated neighbor information to obtain a new node information representation. The specific formula is as follows:
[0025]
[0026] The beneficial effects of the present invention are:
[0027] 1. The present invention adopts a knowledge graph network recommendation method based on a combination of embedding and path to optimize the existing model;
[0028] 2. The present invention uses two strategies, deep search node sampling and importance sampling, to learn the knowledge graph in order to more accurately obtain the potential preferences of users and projects. This can not only reduce the uncertainty caused by random sampling, but also provide a guide for model training, accelerate the training process, and has high practical value. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0030] Figure 1 A simplified schematic diagram of a system structure in one embodiment of the present invention;
[0031] Figure 2 A schematic diagram of traversing sampling nodes layer by layer in an embodiment of the present invention;
[0032] Figure 3 A schematic diagram of a graph embedding model in an embodiment of the present invention;
[0033] Figure 4 A comparison chart of prediction results in a Top-k task based on the Last.FM dataset in an embodiment of the present invention;
[0034] Figure 5 A comparison chart of prediction results in the Top-k task based on the Book-Crossing dataset in one embodiment of the present invention;
[0035] Figure 6 A comparison chart of prediction results in Top-k tasks based on the MoiveLens dataset in one embodiment of the present invention; DETAILED DESCRIPTION
[0036] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0037] Example: Figure 1-6 As shown, the present invention is a knowledge graph recommendation system that integrates deep domain information and importance information, including the deep search layer-by-layer sampling strategy of the present invention for sampling nodes layer by layer to obtain a variety of wandering paths, such as Figure 2 As shown in Figure 1. For example, for the interaction (u, v) between user u and item v, we first locate item v in the knowledge graph, and then take item node v as the starting point, sample the nodes in the graph by walking from the inside out and layer by layer to obtain multiple node sequences. The depth information of the node is controlled by controlling the number of layers of the node walk. Unlike the traditional random walk method, we specify the global direction during the walk, rather than limiting it to the local direction. For example, in the following Figure 2 In , we start walking from node v (layer 0), the next node can be any random node in layer 1, and the next object to walk is in layer 2 instead of walking inside layer 1. We continue walking in this way until we reach the specified exploration depth N (here 4).
[0038] After obtaining multiple walking paths, a single hidden layer neural network is used to train the node sequence set obtained after sampling, such as Figure 3 As shown in the figure, the input layer inputs a node sequence, where "1" represents the node at the corresponding position of the input, and "0" represents the predicted node. That is, during the training process, the current node is used to predict the information of other nodes connected to it. After similar multi-path training converges, the weight matrix of the hidden layer is used as the node embedding matrix vD.
[0039] The graph embedding module described in the present invention uses Word2vec, which mainly has two models: CBOW and Skip-gram. Relatively speaking, Skip-gram is more commonly used, and it predicts the probability of the context based on the current word. Therefore, it will perform more predictions and adjustments than CBOW during the training process, and obtain more accurate word vectors.
[0040] When adopting the above-mentioned deep search layer-by-layer sampling strategy to sample node information, some important node information may be lost while introducing a lot of noise, and the user's potential preferences cannot be fully understood. In order to solve this problem, the present invention proposes a method of combining the deep search sampling strategy with the importance neighborhood sampling strategy during the sampling process, so as to more accurately obtain the potential preferences of users and projects. This can not only reduce the uncertainty caused by random sampling, but also provide a guide for model training and accelerate the training process.
[0041] The importance neighbor sampling algorithm of the present invention comprises the following specific steps:
[0042] S1: Calculate the closeness between all nodes in the network and the central node. The higher the value, the closer the relationship between the node and the central node.
[0043] S2: Sort the compactness and select the first k values. The nodes corresponding to these k value pairs are the target nodes for sampling.
[0044] The relationship closeness calculation includes the following indicators: C number (common neighbors), Adamic-Adar coefficient, RA coefficient (resource allocation coefficient), etc.
[0045] The CN coefficient measures the closeness of the relationship between nodes by comparing the number of common neighbors. The more common neighbors there are, the closer the relationship between nodes is. The specific calculation formula is as follows:
[0046] sim(x,y)=|Γ(x)∩Γ(y)|
[0047] The Resource Allocation Index Algorithm (RA) introduces the consideration of degree value based on the calculation of the number of common neighbors to distinguish the different effects of common neighbors on the strength of the relationship, and believes that the higher the degree of common neighbors, the more valuable they are. The specific calculation formula is as follows:
[0048]
[0049] The Adamic-Adar Index (AA) algorithm performs logarithmic processing on the degree calculation in the RA coefficient. This is to take into account that the high degree of some nodes may interfere with the calculation of the closeness of the relationship. For example, if a node is connected to most nodes, it will become a common neighbor of many nodes, and its degree value is large, which will make the influence of other common neighbors very small. In this case, the size of the degree is not conducive to calculating the closeness of the relationship. Therefore, by taking the logarithm of the degree, the absolute value of the degree can be reduced, thereby effectively considering the influence of the degree factor without making its influence too large. The specific calculation formula is as follows:
[0050]
[0051] The Jaccard similarity index was proposed by Jaccard himself. The index raises the question of whether there are common neighbors between scientists, because each scientist has many neighbors. However, these common neighbors are not necessarily closely related to each other. The specific definition is as follows.
[0052]
[0053] In the formula, Sxy represents the closeness of the relationship between the neighborhood node y and the central node x, Γ(x) represents the set of first-order neighbor nodes of node x, and k(i) is the node degree value.
[0054] The neighborhood aggregation module of the present invention is configured to learn the embedded representation of nodes in the local network of project v through a graph convolutional network.
[0055] Define the attention factor of user u to relationship r, and the function f represents the attention score of user u to relationship r:
[0056]
[0057] The user's attention factor on the relationship is normalized to obtain:
[0058]
[0059] The attention factor is used to aggregate the neighbor information of the node. The combined results are as follows:
[0060]
[0061] Finally, the information of node v itself is fused with its neighbor information to obtain a new node information representation:
[0062]
[0063] The prediction module of the present invention is configured to generate recommendation results based on the input vector data. For each user u and each item v, the model will output a prediction score, denoted as At the same time, the true rating we know is marked as yuv. In order to more accurately evaluate the model prediction performance, we introduced the cross entropy loss function Able to keenly capture the subtle differences between predicted scores and actual scores.
[0064]
[0065] The loss function Loss is calculated as shown in the formula. Where P+ is a positive sample and P- is a negative sample.
[0066]
[0067] Example:
[0068] This paper innovatively proposes a knowledge graph recommendation system named DSINS (Knowledge graph network thatintegrates depth search layer by layer sampling and importance neighbors sampling) that integrates deep domain information and importance information. The system combines the advantages of deep domain exploration and importance neighbor sampling, such as Figure 2 To verify its superiority, this example specifically compares DSINS with the current mainstream recommendation algorithms based on knowledge graphs, and comprehensively evaluates the performance of the model through three representative real data sets of music, books, and movies.
[0069] This example focuses on three widely recognized public datasets (Last.FM (https: / / grouplens.org / datasets / hetrec-2011 / ), Book-Crossing (http: / / www2.in-formatik.uni-freiburg.de / cziegler / BX / ), and MovieLens-20M) to conduct a detailed performance evaluation of CRT and Top-K recommendation tasks. These three datasets are not only publicly accessible, but also have their own characteristics, showing significant differences in scale and data sparsity, providing a solid experimental basis for comprehensively verifying the effectiveness and robustness of the model. Table 1 gives the relevant statistical information of the dataset.
[0070] The Last.FM dataset is derived from its online music system, which records in detail the listening behavior of 1,872 users on 3,846 pieces of music. In this dataset, we regard users who listen more than once as positive interactions, i.e., positive samples, and those below this threshold as negative samples, in order to distinguish the degree of user preference for music.
[0071] The Book-Crossing dataset originates from a reader rating activity in a book community. It brings together rich evaluation information of readers on books, providing valuable resources for studying book recommendations.
[0072] The MovieLens-20M dataset is a popular open-source movie rating dataset that covers detailed feedback from users on movie websites. The ratings of this dataset range from 1 to 5 points. In order to convert it into a format suitable for implicit feedback analysis, we set a rating threshold - ratings of 3 points and above are considered positive samples, representing the user's positive attitude towards the movie, while ratings below this threshold are considered negative samples.
[0073]
[0074] Table 1 Experimental dataset statistics
[0075] The experimental setup of this example is as follows:
[0076] In the critical stage of data preprocessing, we implemented a conversion strategy from explicit feedback to implicit feedback to optimize subsequent analysis. Specifically, positive user ratings (marked as 1) were directly regarded as samples of positive feedback, while in order to construct a balanced and challenging training set, we randomly extracted negative samples (marked as 0) from the set of samples where no interaction occurred. Given the sparsity of interaction data in the Last.FM and Book-Crossing datasets, we decided not to apply rating threshold screening to these negative samples to maintain the original sparse characteristics of the data. However, for the relatively dense MovieLens-20M dataset with a wider distribution of ratings, we set a positive rating threshold (4 points) to ensure that the ratings selected as positive samples do reflect the user's positive preferences.
[0077] In the design of the deep search layer-by-layer sampling strategy, we set the exploration depth and the number of searches to adapt to the characteristics of different datasets. The exploration depth is uniformly set to 8, which is designed to balance the expressiveness and computational efficiency of the model. However, in setting the number of searches, we adopted a differentiated strategy: for the Last.FM dataset, considering the sparsity of its interaction data, we set a lower number of searches (8 times) to reduce the computational burden and avoid overfitting; for the Book-Crossing and MovieLens-20M datasets, given their relatively rich interaction information and higher information density, we increased the number of searches to 32 times to more fully explore potential user-item relationships and improve the model's recommendation accuracy.
[0078] This example is evaluated in the following two recommendation scenarios: (1) Click-through rate prediction (CTR), predicting the interaction probability between a specific user and an item in a trained recommendation model. (2) Top-k recommendation, using the recommendation model learned from the training set to select the k items with the highest predicted probability for a specified user in the test set. To verify the effectiveness of these methods, this example applies the following evaluation metrics:
[0079] Precision: The accuracy of the model's recommended items. R(u) represents the list of recommended items generated by the system for user u based on the training set, and T(u) represents the list of recommended items generated for user u based on the test set.
[0080]
[0081] Recall: The hit rate of the candidate recommendation list.
[0082]
[0083] F1: A weighted combination of Precision and Recall. The F1 value can better reflect the performance of the model.
[0084]
[0085] AUC: It is used to evaluate the performance of the recommendation system in distinguishing between items that the user likes and dislikes. a is the item that the user likes, b is the item that the user dislikes, and the recommendation system scores a and b each time. m is the total number of comparisons, m′ is the number of times a's score is greater than b's score, and m″ is the number of times a's score is equal to b's score.
[0086]
[0087] In order to verify the effectiveness of the DSINS model proposed in this paper in the CTR task, this example selects the following five benchmark models and conducts an in-depth comparison of their performance in the CTR prediction task based on the AUC evaluation index. The specific comparison models are as follows:
[0088] BPRMF: Bayesian Matrix Factorization uses the Bayesian pairwise loss function to optimize the matrix factorization algorithm.
[0089] CKE (Collaborative Knowledge base Embedding): is a model based on entity embedding, which jointly trains knowledge graph and collaborative filtering and utilizes feature embedding of multimodal information.
[0090] PER: Utilizing the relational heterogeneity in the project knowledge graph, meta-paths are introduced to represent the connectivity between users and projects in different relational paths, and projects are recommended based on path similarity.
[0091] Ripple Net: The water wave network model is the first model based on preference propagation. This model compares the high-order transmission of user preference information in the knowledge graph to water wave diffusion, thereby expanding the user's high-order representation.
[0092] KGCN: Introduces the traditional graph convolutional network model into the knowledge graph, and takes the combination of users and relationships in the knowledge graph as an important factor in aggregating neighborhood information in the knowledge graph, thereby expanding the high-level representation of project entities in the knowledge graph.
[0093] The prediction results in top-k are as follows Figure 4-6 . The experimental results shown in Table 2 clearly verify the excellent performance of the DSINS model in the CTR prediction task, and its effect is significantly better than the four comparative models of BPRMF, CKE, PER and RippleNet. . Specifically, the BPRMF model is limited by the use of only the information of the dataset itself and fails to effectively integrate external auxiliary data; although the PER model attempts to introduce additional information through the knowledge graph, it is highly dependent on the manually preset meta-path, which limits its generalization ability; the CKE model fails to fully realize its potential because it ignores the high-order information of the knowledge graph and fails to effectively integrate key information such as text and vision; the RippleNet model shows asymmetry in the generation of user and project embeddings, which affects its performance; the KGCN model focuses on the knowledge graph information on the project side and ignores the equally important information dimension on the user side, which also limits its performance improvement.
[0094] It is worth noting that the experimental results show that movie datasets (such as Movielens-20M) exhibit superior performance among all methods. This is most likely due to the rich interactive behaviors and complex relationship networks in the movie datasets, which provide a more diverse source of information for the recommendation model, thereby promoting more accurate latent feature learning.
[0095] In particular, the DSINS model proposed in this paper has achieved the best performance on the three datasets, especially on the Last.FM and Book-Crossing datasets, where its performance is significantly improved compared to the mainstream algorithms RippleNet and KGCN. On the Movielens-20M dataset, DSINS not only improves the AUC index by about 1.77% compared to the BPRMF model, but also achieves a 1.64% increase in F1 score; compared to the KGCN model, its AUC and F1 scores have also increased by about 0.31% and 0.22%, respectively.
[0096]
[0097] Table 2 Prediction results based on AUC indicator in CTR task
[0098] In order to deeply analyze the respective contributions of the two core components DSINS_1 and DSINS_2 in the DSINS model, a series of ablation experiments were conducted. Specifically, the DSINS_1 component focuses on accurately capturing and learning the complex structural information in the knowledge graph through advanced graph embedding technology combined with deep sampling methods, providing a solid skeleton support for the model. The DSINS_2 component focuses on using efficient important domain sampling strategies, focusing on mining and integrating the information of key neighbor nodes in the knowledge graph to further enhance the representation ability of the model. By evaluating the performance of these two components in the model separately, we can more clearly understand the role and importance of each of them in improving the accuracy of CTR prediction.
[0099]
[0100] Table 3 Ablation experiment results
[0101] According to the experimental results shown in Table 3, we can draw the following conclusions. First, the performance of the DSINS1 component on all datasets is relatively poor, which strongly proves that it is far from enough to rely solely on the deep sampling method of meta-path embedding and only use the structural information of the knowledge graph for recommendation. Although this method can capture the skeleton information of the graph, it ignores the richer and more detailed semantic and relational data, thus limiting the improvement of the recommendation effect. Secondly, the various indicators of the DSINS2 component are significantly better than those of DSINS1, which fully demonstrates the superiority of the graph convolutional network based on importance sampling in the recommendation system. This method effectively integrates these important information by accurately identifying and focusing on the key neighbor nodes in the knowledge graph, which has a profound impact on the optimization of the recommendation results and proves its significant effect in improving the accuracy and personalization of recommendations. Further comparing DSINS1 and DSINS2, we can clearly see the significant improvement in the performance of the latter. This result not only verifies the effectiveness of combining wide-area semantic information with graph neural networks, but also reveals that the fusion of multiple information sources can significantly enhance the limitations of a single method and achieve a leap in performance. Finally, the DSINS model proposed in this paper, with its best performance, fully demonstrates the advantages of the fusion strategy of multiple methods based on graph neural network in making recommendations using knowledge graphs. This method can not only capture the structure, semantics and relationship information in the knowledge graph more comprehensively, but also realize the utilization of information through an efficient fusion mechanism, bringing higher accuracy and personalized experience to the recommendation system.
[0102] Finally, it should be noted that in the description of the present invention, it should be noted that the terms "vertical", "up", "down", "horizontal", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention.
[0103] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A knowledge graph recommendation system that integrates deep domain information and importance information, including a deep search layer-by-layer sampling module, a graph embedding module, an importance neighborhood sampling module, a neighborhood aggregation module, and a prediction module; The deep search layer-by-layer sampling module includes a deep search strategy component and a single hidden layer neural network component; The graph embedding module adopts the Skip-gram method in the Word2vec model to predict the probability of the context based on the current word; The importance neighborhood sampling module quantifies the closeness relationship between each node in the network and the central node; The neighborhood aggregation module defines the user's attention factor for the relationship, and uses the attention factor to perform weighted aggregation on the node's neighbor information, and finally merges the node's own information with the aggregated neighbor information; The prediction module is configured to obtain recommendation results based on the input user embedding and item embedding vectors.
2. A knowledge graph recommendation system integrating deep domain information and importance information according to claim 1, characterized in that: The knowledge graph adopts a deep search strategy to sample nodes layer by layer to obtain multiple walking paths. For the interaction (u, v) of user u on project v, project v is first located in the knowledge graph, and then starting from the project node v, the nodes in the graph are sampled from the inside to the outside and walking layer by layer to obtain multiple node sequences. The depth information of the node is controlled by controlling the number of layers of the node walking.
3. According to claim 1, a knowledge graph recommendation system integrating deep domain information and importance information is characterized in that: After obtaining multiple wandering paths, a single hidden layer neural network is used to train the node sequence set obtained after sampling. The input layer of the single hidden layer neural network inputs the node sequence, where "1" represents the node at the corresponding position of the input, and "0" represents the predicted node. That is, in the training process, the current node is used to predict the information of other nodes connected to it. After convergence through similar multiple path training, the weight matrix of the hidden layer is used as the embedding matrix vD of the node.
4. The importance domain adoption module according to claim 1, characterized in that The closeness relationship between each node and the central node in the quantified network also includes the following steps: Step S1: Calculate the closeness between all nodes in the network and the central node. The higher the value, the closer the relationship between the node and the central node. Step S2: Sort the compactness and select the first k values. The nodes corresponding to these k value pairs are the target nodes for sampling. The domain aggregation module includes the following steps: Step S1: Define the attention factor of user u to relationship r. The attention factor of user u to relationship r indicates the degree of attention of the user to a specific relationship. We use function f to represent the attention score of user u to relationship r. The specific formula is as follows: Step S2: Normalization. In order to make the attention factors comparable between different relationships, we need to normalize the attention factors. The specific formula is as follows: Step S3: Use the normalized attention factor to perform weighted aggregation on the neighbor information of node v, using Indicates, as follows: Step S4: Fusion of node v's own information and neighbor information. Fusion of node v's own information and aggregated neighbor information to obtain a new node information representation. The specific formula is as follows:
