Link prediction method based on multi-path network representation learning

Abstract

The application discloses a link prediction method based on multi-path network representation learning, and comprises the following steps: decomposing a multi-path heterogeneous attribute network into multiple single-path networks by using single paths; constructing a basic network to integrate the structural relationship of each single-path network; using a global-aware graph convolution module with single-path network embedding and global-aware message passing strategy to perceive the global signal of the multi-path heterogeneous attribute network; using an aggregation module to obtain node embedding by weighted sum of the hidden layer embedding of the single-path network, and introducing a final convolution operator; learning a low-dimensional vector for each node, using the learned representation for link learning task, and outputting the result. The application automatically perceives the global and gives the local topology signal of the node embedded in the multi-path heterogeneous attribute network, and fuses network structure and node attribute network representation learning.

Description

Technical Field

[0001] This invention belongs to the field of network representation learning technology, and particularly relates to a link prediction method based on multi-path network representation learning. Background Technology

[0002] Network representation learning (NRL) is an effective method for encoding information networks into a low-dimensional continuous space while preserving the similarity of nodes in network topology and intrinsic properties. Through NRL, a wider range of downstream tasks can benefit from networks in the form of embeddings, such as link prediction, recommender systems, and community detection. In particular, DeepWalk, LINE, and node2vec are pioneering works applying deep learning to NRL. These methods study NRL based on homogeneous graphs.

[0003] In recent years, various models have been introduced to handle heterogeneous NRLs, including metapath-based methods such as metapath2vec and graph neural network-based methods such as HAN. However, it is worth noting that they are only designed for heterogeneous networks with a single view, ignoring the existence of multiple types of nodes and / or edges.

[0004] In the real world, most networks not only contain multiple types of nodes and / or edges, but also possess rich attribute information. Such networks, with multi-channel structures and containing multiple types of nodes and their attribute information, are called multi-path heterogeneous attribute networks. A ttributed M ultiplex He terogeneous N etwork,AMHEN).

[0005] As a simplified academic AMHEN, this network is built from DBLP. Taking paper P4 as an example, it was co-authored by two authors, A2 and A3, and was published at conference C2. Furthermore, paper P2 cites paper P4. In short, there are three types of edges associated with paper P4. Additionally, multiple types of nodes with their own attribute vectors are provided.

[0006] Subsequently, recent research trends have focused on exploring multi-path heterogeneous attribute networks, with GATNE being a significant work that was the first to introduce AMHEN into NRL.

[0007] To further improve the performance of NRL in multi-path heterogeneous attribute networks, DMGI introduces a self-supervised method and designs a consistency regularization scheme to consider the relationships between type-specific node embeddings. GTN leverages the advantages of the Transformer architecture to automatically generate meta-paths without domain knowledge and can generate various meta-paths based on soft selection and combination of edge types.

[0008] As a recent advancement, MHGCN was proposed, which presents a graph convolutional network for AMHEN representation learning. It is designed to aggregate the adjacency matrix through a simplified graph convolutional module and obtain the final embedding via a learnable parameter β.

[0009] Although these studies have made some progress in representation learning for multi-path heterogeneous attribute networks, there are two limitations due to the diversity of the networks.

[0010] Heterogeneity Modeling: To recap, AMHEN contains multiple types of nodes and edges. Current graph neural network methods model AMHEN by manually constructing meta-paths, making it difficult to account for the diversity of the network. Modeling heterogeneity is crucial for NRL in AMHEN.

[0011] Multipath information fusion: Obtaining each embedding of multiple subnetworks of a multipath heterogeneous network with attributes. Most previous work aimed to aggregate metapaths or adjacency matrices; however, these information fusion strategies may be simple. Summary of the Invention

[0012] In view of this, this invention proposes a novel method based on the concept of a single path and conceives a globally perceptive neural network (UAN) based on a single path to better handle the representation learning problem of multi-path heterogeneous attribute networks. A single path is a channel in a composite structure, and a single-path network is a sub-network of a multi-path heterogeneous attribute network, which aggregates by connecting meta-paths with single-sided attributes. Therefore, the single-path network aims to capture local characteristics. Based on these concepts, the multi-path heterogeneous attribute network is first decomposed into multiple single-path networks by utilizing some representative single paths. Simultaneously, this invention constructs a base network that integrates the structural relationships of each single-path network. To address the heterogeneous model problem and perceive the global signals of the multi-path heterogeneous attribute network, this invention designs a globally perceptive graph convolution module with single-path network embedding and a globally perceptive message passing strategy. The message passing strategy effectively perceives the correlation between the base network and each single-path network. To effectively fuse multi-path information and improve model performance, this invention proposes an aggregation module that obtains node embeddings by weighted summation of the hidden layer embeddings of the single-path networks and introduces the final convolution operator. Experiments have demonstrated the effectiveness of UAN compared to existing NRL models.

[0013] The link prediction method based on multi-path network representation learning disclosed in this invention includes the following steps:

[0014] Obtaining a multi-channel heterogeneous attribute network containing information on multiple types of nodes and their attributes. in, Let be the set of nodes and ε be the set of edges between nodes. They are all associated with an attribute vector Among them, node set The dimensions of the attribute vectors are represented as n and m, respectively;

[0015] A multi-path heterogeneous attribute network is decomposed into multiple single-path networks using single-path methods.

[0016] A base network is constructed to integrate the structural relationships of each single-path network. To address the heterogeneous model problem, a globally aware graph convolutional module with single-path network embedding and a globally aware message passing strategy is used to perceive the global signals of the multi-path heterogeneous attribute network. The message passing strategy perceives the correlation between the base network and each single-path network.

[0017] To effectively fuse multipath information and improve model performance, an aggregation module is used to obtain node embeddings by weighted summation of the hidden layer embeddings of the single-path network, and the final convolution operator is introduced.

[0018] For each node Learning a low-dimensional vector in Let G be the dimension of the attribute vector in the network. The learned representation is input into the link predictor to perform link prediction and output the link prediction result.

[0019] Furthermore, each edge e of the multi-path heterogeneous attribute network ij =(v i v j )∈ε means v i and v j The connections between nodes and edges, due to the consistency of nodes and edges, have a mapping function φ for node types and edge types: and ψ: in and These represent the number of node types and the number of edge types, respectively; each node Both edges e∈ε are associated with a point in ε. and Specific types within; in addition, Where ε i Let i ∈ k represent the edge type, and k represent the number of edge types in network G. This means the same pair of nodes Connect them using multiple types of edges.

[0020] Furthermore, a single-path network is defined as follows: given a defined multi-path heterogeneous attribute network... Single-path network This means that the single-path network is a single channel of a multi-path heterogeneous attribute network, which aggregates meta-paths with a single edge attribute connection.

[0021] Furthermore, a combined single-path network and a basic network are constructed as follows:

[0022] G s =(G1′,G2′,…,G k ′,G base ),

[0023] G1′,G2′,…,G k '' represents the 1st, 2nd, and kth single-path networks, respectively; all single-path networks share the same set of nodes. Special Network G base ∈G s , ε i G represents the type of the i-th edge in a multi-path heterogeneous attribute network; base It possesses structural information about multi-path heterogeneous attribute networks.

[0024] Furthermore, a multi-channel global awareness graph convolutional module is used to quickly learn all data passing through G... base The metapath, for a single channel layer:

[0025]

[0026] in It represents the hidden state of the network, and σ is the non-linear activation function. Let be the adjacency matrix of the i-th single-path network. For the attribute matrix in a multi-path heterogeneous attribute network, b is a learnable parameter of the first layer. 1 It is the bias term of the first level.

[0027] Furthermore, to better aggregate global signals, construct a basic network, and use a globally aware message passing strategy to enhance model representation capabilities and maintain the input order of magnitude for each channel in the model, the globally aware message passing strategy includes:

[0028] The base network is embedded through an additional graph convolutional layer, and information G is absorbed from the learned data. base For each single-path network, the embedding of the first-layer base network is represented as follows:

[0029]

[0030] It is the adjacency matrix of the basic network;

[0031] Therefore, the global awareness information transmission strategy follows the following equation:

[0032]

[0033] Where l represents the l-th layer of the model. It is the embedded representation of the (l-1)th layer of the basic network.

[0034] Furthermore, combining the aforementioned globally aware message passing strategy, a two-layer graph convolutional network (GCN) is used as follows:

[0035]

[0036] in The learnable weights of the second layer, b 2 It is the bias term of the second layer;

[0037] Furthermore, the model uses different numbers of graph convolutional layers to capture meta-path single-path networks and base networks of different lengths; after multiple convolutional operations, the model acquires multiple pieces of information and has a better representation of multi-path heterogeneous attribute networks.

[0038] Furthermore, after obtaining the embeddings of each single-path network, the aggregation module is used to aggregate these embeddings together, the steps of which include:

[0039] Considering the different connections between single-path networks, the embeddings are first weighted and summed, and then the adjacency matrix of the base network is fused into the final graph convolutional layer.

[0040] The aggregation module performs nonlinear transformations and feature extraction on the node embeddings to extract more representative and distinguishable node representations, as shown in the following formula:

[0041]

[0042] in,

[0043]

[0044] It is a weighted sum of single-path network embeddings. Let p be the adjacency matrix of the basic network, p be the number of graph convolutional modules in the last layer, and β be a learnable parameter used to measure the graph-controlled convolutional modules for each channel. It is the learnable weight of the last layer, b final This is the bias term of the last layer; thus, the final embedding is obtained.

[0045] Furthermore, unsupervised learning is used for training in link prediction:

[0046] The model parameters are optimized by minimizing the following binary cross-entropy loss function through negative sampling:

[0047]

[0048] in, For node v i The expression, For node v j The expression, For node v′ i The expression, For node v′ j The representation is as follows: T is the matrix transpose, σ(·) is the LogSigmoid function, and Ω is the set of positive node pairs. - It is a set of negative node pairs sampled from all unobserved node pairs. Therefore, a loss function is used to increase the similarity between node representations in positive samples, while decreasing the similarity between node representations in negative samples.

[0049] Furthermore, the aforementioned online datasets are real-world public datasets from the academic, e-commerce, and film industries, including at least one of the following: Alibaba, DBLP, IMDB, and Amazon datasets.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] This invention proposes a single-path global perception neural network, or UAN, which can automatically perceive the global network and provide local topology signals of nodes embedded in a multi-path heterogeneous attribute network.

[0052] This invention integrates network structure and node attribute network representation learning. Through the aggregation module, the model can effectively fuse multiple pieces of information.

[0053] Extensive experiments were conducted on several public datasets, and the results demonstrate the superiority of the proposed method in link prediction, with UAN outperforming state-of-the-art competitors. Attached Figure Description

[0054] Figure 1 A framework diagram of the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0056] Define network in Let ε be the set of nodes and ε be the set of edges between nodes. Each edge eij =(v i v j )∈ε means v i and v j The connections between them are shown in Table 1.

[0057] Table 1: Symbols and Descriptions. The three blocks from top to bottom list the symbols used to describe variables in the AMHEN, single-path network construction, and graph convolution modules.

[0058]

[0059] Definition 1 (Multi-path Heterogeneous Attribute Networks). Given a network... Each node They are all associated with an attribute vector Among them, node set Let the dimensions of the attribute vectors be n and m, respectively. Due to the consistency between nodes and edges, the mapping function between node types and edge types is defined as φ: and ψ: in and These represent the number of node types and the number of edge types, respectively. Each node Both edges e∈ε are associated with a point in ε. and Specific types within. Furthermore... Where ε i Let the edge type be i∈k, and This means the same pair of nodes It can be connected by multiple types of edges.

[0060] Definition 2 (Metapath). The metapath Φ is defined as... The form describes objects U1 and U l+1 The composite relationship between them in Represents a composite operator on a relation. Definition 3 (Single-path network). Given a defined multi-path heterogeneous attribute network... Single-path network This means that the single-path network is a single channel of a multi-path heterogeneous attribute network. In different instances of the single-path network, the nodes and attributes remain the same, but the edges differ. From another perspective, the single-path network is a subgraph of the multi-path heterogeneous attribute network that aggregates meta-paths connected by a single edge attribute.

[0061] Problem 1 (Representation Learning for Multi-path Heterogeneous Attribute Networks). Given a multi-path heterogeneous attribute network with attributes... The task of the network representation in this invention is to represent each node Learning a low-dimensional vector in These learned representations can be used for downstream tasks, including but not limited to node classification and link prediction.

[0062] In this section, using Figure 1 The overall architecture shown describes the UAN framework of this invention.

[0063] Single-path network construction

[0064] Unlike ordinary heterogeneous networks, multi-path heterogeneous attribute networks have the following characteristics: a multi-path heterogeneous attribute network can be viewed as a set of single-path networks, where the nodes representing each single-path network are fixed, but each node can interoperate on different single-path networks. To better capture the interaction relationships between these nodes, this paper now demonstrates how to convert a multi-path heterogeneous attribute network into a single-path network and how to construct the base network. This invention constructs a combination of a single-path network and a base network:

[0065] G s =(G1,G2,…,G k G base ),

[0066] Refer to definition 3. ε i This represents the type of the i-th edge in a multi-path heterogeneous attribute network. Note that all single-path networks share the same set of nodes. A special network G base ∈G s Defined as G base Having structural information about multi-path heterogeneous attribute networks, they are easy to construct. Therefore, this construction method easily decomposes multi-path heterogeneous attribute networks and extends to large inputs.

[0067] Global Attention Graph Convolution Module

[0068] Multi-path heterogeneous attribute networks contain different types of nodes and edges. Furthermore, the same node may have potential connections in different single-path networks of AMHEN. Edges and nodes of specified types form meta-paths, which have a significant impact on the learning of heterogeneous network representations. Existing techniques require manually defining meta-paths and learning node representations on sampled heterogeneous meta-paths. However, setting up manually sampled meta-paths is a complex task. Sampling a large number of meta-paths in a large-scale network takes a long time. Simultaneously, aggregating meta-paths into the meta-path network also requires significant memory overhead. This invention effectively solves this problem through a multi-channel global-aware graph convolutional module, which can easily learn all nodes in a set G of single-path networks. base Metapath.

[0069] like Figure 1 As shown, multi-channel global awareness graph convolutional layers are used for network search for information AMHEN. For a single-channel layer:

[0070]

[0071] in The hidden state of the network is σ, which is a non-linear activation function (i.e., the ReLU function).

[0072] Let be the adjacency matrix of the i-th single-path network. This is the attribute matrix in AMHEN. These are the learnable parameters of the first layer. To ensure that the model of this invention focuses on the structural information and properties of the AMHEN itself, note that this invention uses a simplified graph convolutional layer, i.e., without using a degree matrix.

[0073] While single-graph embeddings can be obtained through convolutional layers, the loss of inter-layer interaction information can lead to suboptimal results. Therefore, simple modeling of single paths is insufficient; the interactions between nodes across different single-path networks should be considered. Note that to better aggregate global signals, this invention constructs a base network and designs a message-passing strategy step in the model: global information representation and integrated message passing.

[0074] Global awareness. While graph convolution operators can model single-path networks, they cannot handle topological imbalance problems. Therefore, this invention proposes a globally aware message-passing strategy to aggregate all information, enhancing model representation capabilities and maintaining the order of magnitude of the input for each channel in the model. This invention embeds the underlying network through an additional graph convolutional layer and absorbs information G from learning. base To each single-path network. Taking the embedding representation of the first-layer base network as an example:

[0075]

[0076] Therefore, the first step is that the global perception information transmission strategy follows the following equation:

[0077]

[0078] Where l represents the l-th layer of the model.

[0079] A comprehensive strategy. Combining a globally-aware message passing strategy, a two-layer GCN can be used as an example, as follows:

[0080]

[0081] in These are the learnable weights of the second layer. Furthermore, the model can use different numbers of graph convolutional layers to capture meta-paths of varying lengths in single-path networks and the base network. After multiple convolutional operations, the model can acquire more information and has a better representation of multi-path heterogeneous attribute networks.

[0082] Aggregation Module

[0083] When the embeddings for each single-path network are obtained, an aggregation module is used to combine these embeddings. Considering the different connections between the single-path networks, the embeddings are first weighted and summed, and then the adjacency matrices of the base networks are fused into the final graph convolutional layer. The aggregation module allows for non-linear transformations and feature extraction of the node embeddings to extract more representative and distinguishable node representations. This is illustrated in the following formula:

[0084]

[0085] in,

[0086]

[0087] It is a weighted sum of single-path network embeddings. is the adjacency matrix of the basic network. k is the number of graph convolutional modules in the last layer. β is a learnable parameter used to measure the graph-controlled convolutional modules for each channel. This allows us to obtain the final embedding. Model optimization

[0088] This module introduces a defined objective function to train the model of this invention to learn the final node representation. Depending on the requirements of different downstream tasks and the ability to utilize node labels, two learning modes can be used to train UAN: unsupervised learning and semi-supervised learning. Unsupervised learning is used in link prediction, and semi-supervised learning is used in node classification.

[0089] Unsupervised learning. In unsupervised learning, this invention optimizes model parameters by minimizing the following binary cross-entropy loss function through negative sampling:

[0090]

[0091] in For node v i The representation is as follows: T is the matrix transpose, σ(·) is the LogSigmoid function, and Ω is the set of positive node pairs. - It is a set of negative node pairs sampled from all unobserved node pairs. In other words, this invention uses a loss function to increase the similarity between node representations in positive samples, while decreasing the similarity between node representations in negative samples.

[0092] Semi-supervised learning. In semi-supervised learning, the backpropagation and gradient descent of the UAN parameters are optimized by minimizing the cross-entropy. The cross-entropy loss is applied across all labeled nodes between the ground truth and the prediction, and is expressed as:

[0093]

[0094] in It is a set of labeled node metrics, Y i ∈R n Let C be the label of the i-th node, and C be the node classifier parameters. It is the embedding of the i-th node. Guided by a subset of labeled nodes, UAN can be optimized, and then the node embeddings can be learned for classification.

[0095] experiment

[0096] Table 2: Statistical data of the network in the experiment. Wherein, "n-type" (representing "e-type") indicates the node type (representing the edge type).

[0097] Alibaba 21,318 41,676 2 4 19 DBLP 26,128 119,783 4 3 4,635 IMDB 12,772 18,644 3 2 1,256 Amazon 7,621 1,384,799 4 3 2,000

[0098] Dataset and Experiment Introduction

[0099] This invention utilizes four real-world public datasets from different fields (academic, e-commerce, and film): Alibaba, DBLP, IMDB, and Amazon. Detailed statistical descriptions of the data can be found in Table 2. Furthermore, to evaluate the effectiveness of this invention, a link prediction task was performed. Two datasets (Alibaba and DBLP) were used in this experiment for the link prediction task.

[0100] Baseline model

[0101] This invention is compared with the following 19 network learning baselines, which are divided into three categories.

[0102] Homogeneous network embedding method:

[0103] node2vec is a representative method that learns network representations by generating a sequence of nodes on a graph using random walks.

[0104] The projection process is performed iteratively to capture the high-order graph structure, and a matrix factorization objective function is used. This method generates a similarity matrix to model the transitive relationships between nodes. Then, sparse random projection is used to reduce dimensionality.

[0105] SGC proposed a method to simplify graph convolutional networks by removing nonlinear projections during inter-layer information propagation. AM-GNN is an advanced graph convolutional network, an adaptive multi-channel graph convolutional network used for node classification tasks.

[0106] Heterogeneous network embedding methods:

[0107] R-GCN considers the impact of different edge types and applies weight sharing and coefficient constraints to heterogeneous networks.

[0108] HAN applies graph attention networks to heterogeneous networks, taking into account the interactions within and between networks, and learns node embeddings using manually selected meta-paths.

[0109] NARS decouples heterogeneous networks according to edge type and aggregates features of neighbors on the decoupled subgraph.

[0110] MAGNN is a graph neural network based on aggregate metapaths, used to process heterogeneous graphs.

[0111] HPN designs a semantic propagation mechanism to alleviate semantic confusion, and then integrates rich semantic information. Multi-path heterogeneous network embedding method:

[0112] PMNE includes three different models for merging multiple networks to generate a holistic embedding for each node, referred to as PMNE-n, PMNE-r, and PMNE-c.

[0113] MNE obtains network embeddings by combining high-dimensional general embeddings with low-dimensional hierarchical embeddings. GATNE proposes using basic embeddings, edge embeddings, and attribute representations to generate holistic node embeddings.

[0114] GTN learns network representation tasks by identifying effective meta-paths with high-order relations.

[0115] DMGI developed a consensus regularization scheme that considers the relationships between type-specific node embeddings. Furthermore, it aggregates the contextual signals for each relationship through an attention mechanism.

[0116] FAME uses spectral transformation to capture metapaths and employs random projection to improve model efficiency.

[0117] HGSL is a state-of-the-art heterogeneous graph neural network that simultaneously learns heterogeneous graph structure and graph neural network parameters for classification.

[0118] DualHGCN uses a dual hypergraph convolutional network to learn node embeddings for a multi-bipart network.

[0119] HDMI proposes a supervisory signal based on high-order mutual information and an attention-based fusion module for fusing node embeddings from different channels of multiple networks.

[0120] MHGCN uses hyperparameters to sum the adjacency matrix of the single-path network in AMHEN, and then uses a simplified graph convolutional network for node representation.

[0121] Experimental setup

[0122] In terms of experimental setup, this invention maintains... Compared with other methods, and with other parameter settings following the original implementation in published papers, the number of convolutional layers l in this invention is set to 2. For fair comparison, a uniform 500-round training run is performed on the link prediction task. Optimal performance is obtained by adjusting the learning rate and weight decay from 5e-2 to 5e-4. According to Table 2, the number of single-path networks for Alibaba, DBLP, IMDB, and Amazon are 4, 3, 2, and 3, respectively. Other results are derived from previously published work.

[0123] All experiments were conducted on a Linux server equipped with one GPU (Tesla A100) and one CPU (Intel(R) Xeon(R) Gold6240 CPU @ 2.60GHz), running Ubuntu 18.04.6LTS. The proposed UAN model was implemented using Python and the deep learning library PyTorch. Python and PyTorch versions 3.9.13 and 1.13.0 were used, respectively.

[0124] Furthermore, the statistical significance was reported using a two-tailed t-test, and... Use The UAN tag represents a significant improvement over MHGCN.

[0125] Link prediction is widely used to evaluate the quality of network embeddings obtained by different methods. For link prediction, connected nodes in the network are considered positive node pairs, and all unconnected nodes are considered negative node pairs. For each edge type, positive node pairs are divided into training, validation, and test sets at ratios of 80%, 10%, and 10%. First, the model's performance on the link prediction task is evaluated in an unsupervised manner against sixteen baseline methods. Experimental evaluation is conducted using one of the commonly used evaluation metrics, the area under the ROC curve (ROC-AUC) and the area under the PR curve (PR-AUC). The F1 score is also used as another evaluation metric. The results are shown in Table 3, with the best results shown in bold. The first seven baseline methods are homogeneous or heterogeneous network representation learning methods, and the latter nine are multi-path network representation learning methods.

[0126] Table 3: Comparison of model performance for link prediction on different datasets.

[0127]

[0128] As can be seen, UAN significantly outperforms all baseline methods across all evaluation metrics on both datasets. The UAN model of this invention achieves high accuracy on all datasets (Alibaba and DBLP), particularly achieving 100% prediction performance on the e-commerce network (Alibaba). This is because UAN captures effective multi-single-path topologies through graph convolutions built on single-path networks and a globally aware propagation strategy. Compared to GATNE and MAGNN, the model of this invention achieves better results, demonstrating that manually setting meta-paths is less effective than the model's natural awareness of meta-paths.

[0129] MHGCN can automatically capture meta-paths, but UAN outperforms MHGCN on both datasets. This is due to the difference in information aggregation methods between the two models. In MHGCN, it uses a learnable hyperparameter β to sum the adjacency matrix. However, in UAN, additional embeddings are computed, and a final aggregation layer is used to enable the model to better fuse cross-layer information.

[0130] The beneficial effects of this invention are as follows:

[0131] This invention proposes a network representation learning model (UAN) for multi-path heterogeneous attribute networks. The model mainly consists of two key components: a globally aware graph convolutional module and an aggregation module. Through the globally aware graph convolutional module, this invention can distinguish the importance of relationships between different nodes in a multi-path heterogeneous attribute network and capture arbitrary-length meta-path interactions across multiple relationships. Meaningful node embeddings are learned during the training phase through model parameter learning. Experimental results on four real-world heterogeneous networks demonstrate the superiority of UAN.

[0132] As used herein, the term "preferred" is meant as an example, illustration, or illustration. Any aspect or design described herein as "preferred" need not be construed as being more advantageous than other aspects or designs. Rather, the use of the term "preferred" is intended to present the concept in a specific manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusionary "or." That is, unless otherwise specified or clear from the context, "X uses A or B" naturally includes either of the permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0133] Furthermore, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art based on a reading and understanding of this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the aforementioned components (e.g., elements, etc.), the terminology used to describe such components is intended to correspond to any component (unless otherwise indicated) that performs the specified function of said component (e.g., is functionally equivalent to it), even if structurally not equivalent to the disclosed structure performing the functions in the exemplary implementations of this disclosure shown herein. Moreover, although specific features of this disclosure have been disclosed with respect to only one of several implementations, such features may be combined with one or more features of other implementations that may be desirable and advantageous for a given or particular application. Furthermore, with regard to the use of the terms “comprising,” “having,” “containing,” or variations thereof in the Detailed Description or claims, such terms are intended to be included in a manner similar to the term “including.”

[0134] The functional units in this invention embodiment can be integrated into a processing module, or each unit can exist physically separately, or multiple units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. The aforementioned devices or systems can execute the storage methods in the corresponding method embodiments.

[0135] In summary, the above embodiments are one implementation of the present invention, but the implementation of the present invention is not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made that deviate from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A link prediction method based on multi-path network representation learning, characterized in that, Includes the following steps: Obtain a network dataset containing a multi-channel heterogeneous attribute network with multiple types of nodes and their attribute information. ,in, For a set of nodes, It is the set of edges between nodes, each node They are all associated with an attribute vector , where the node set The dimensions of the attribute vectors are respectively represented as and This paper describes a method for decomposing a multi-path heterogeneous attribute network into multiple single-path networks using single-path methods. A base network is constructed to integrate the structural relationships of each single-path network. To address the heterogeneous model problem, a globally aware graph convolutional module with single-path network embedding and a globally aware message-passing strategy is used to perceive the global signals of the multi-path heterogeneous attribute network. The message-passing strategy perceives the correlation between the base network and each single-path network. To effectively fuse multi-path information and improve model performance, an aggregation module is used to obtain node embeddings by weighted summation of the hidden layer embeddings of the single-path networks, and a final convolution operator is introduced. For each node... Learning a low-dimensional vector ,in , For the network The attribute vector dimension in the model is used to input the learned representation into the link predictor for link prediction and output the link prediction result. The following is a combination of a single-path network and a basic network: These represent the 1st, 2nd, and kth single-path networks, respectively; all single-path networks share the same set of nodes. Special Network , , This indicates the type of the i-th edge in a multi-path heterogeneous attribute network; It possesses structural information about multi-path heterogeneous attribute networks; Rapidly learn all data through a multi-channel global awareness graph convolutional module. The metapath, for a single channel layer: in It is the hidden state of the network. It is a non-linear activation function. For the first Adjacency matrix of a single-path network For the attribute matrix in a multi-path heterogeneous attribute network, These are the learnable parameters of layer 1. It is the bias term of layer 1; To better aggregate global signals, a base network is constructed, and a globally aware message passing strategy is used to enhance model representation capabilities and maintain the input order of magnitude for each channel in the model. This strategy includes embedding the base network through an additional graph convolutional layer and absorbing information from the learned data. For each single-path network, the embedding of the first-layer base network is represented as follows: The adjacency matrix of the base network is used; therefore, the global awareness information transmission strategy follows the following equation: in For the model's first layer, It is the first Embedded representation of layer-based networks; Combining the aforementioned globally aware message passing strategy, a two-layer graph convolutional network (GCN) is used as follows: = in , are the learnable weights of the second layer. It is the bias term of the second layer; in addition, the model uses different numbers of graph convolutional layers to capture meta-path single-path networks and base networks of different lengths; after multiple convolutional operations, the model acquires multiple pieces of information and has a better representation of multi-path heterogeneous attribute networks.

2. The link prediction method based on multi-path network representation learning according to claim 1, characterized in that, Each edge of the multi-path heterogeneous attribute network express and The connections between nodes and edges, due to the consistency of nodes and edges, have a mapping function for node types and edge types as follows: , ,in and These represent the number of node types and the number of edge types, respectively; each node and edge Each is associated with one in and Specific types within; in addition, , Representing edge type , Represents network The number of edge types in the, and The same pair of nodes Connect them using multiple types of edges.

3. The link prediction method based on multi-path network representation learning according to claim 2, characterized in that, A single-path network is defined as follows: given a predefined multi-path heterogeneous attribute network... Then a single-path network This means that the single-path network is a single channel of a multi-path heterogeneous attribute network, which aggregates meta-paths with a single edge attribute connection.

4. The link prediction method based on multi-path network representation learning according to claim 3, characterized in that, After obtaining the embeddings for each single-path network, the aggregation module is used to aggregate these embeddings together. The steps include: considering the different connections between the single-path networks, the embeddings are first weighted and summed, and then the adjacency matrix of the base network is fused into the final graph convolutional layer; the aggregation module performs nonlinear transformation and feature extraction on the node embeddings to extract more representative and distinguishable node representations, as shown in the following formula: in, It is a weighted sum of single-path network embeddings. The adjacency matrix of the basic network. It is the number of layers in the last layer of the graph convolutional module. These are learnable parameters used to measure the graphical control of the convolutional module for each channel. These are the learnable weights of the last layer. This is the bias term of the last layer; thus, the final embedding is obtained. .

5. The link prediction method based on multi-path network representation learning according to claim 1, characterized in that, The network datasets are real-world public datasets from the academic, e-commerce, and film industries, including at least one of Alibaba, DBLP, IMDB, and Amazon datasets.

Images