CN113672735B - Link prediction method based on theme-aware heterogeneous graph neural network - Google Patents
Link prediction method based on theme-aware heterogeneous graph neural network Download PDFInfo
- Publication number
- CN113672735B CN113672735B CN202111048789.4A CN202111048789A CN113672735B CN 113672735 B CN113672735 B CN 113672735B CN 202111048789 A CN202111048789 A CN 202111048789A CN 113672735 B CN113672735 B CN 113672735B
- Authority
- CN
- China
- Prior art keywords
- topic
- meta
- path
- aware
- perception
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 31
- 230000008447 perception Effects 0.000 claims abstract description 41
- 230000002776 aggregation Effects 0.000 claims abstract description 20
- 238000004220 aggregation Methods 0.000 claims abstract description 20
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 19
- 239000013598 vector Substances 0.000 claims abstract description 19
- 239000011159 matrix material Substances 0.000 claims abstract description 11
- 230000009466 transformation Effects 0.000 claims abstract description 5
- 238000013507 mapping Methods 0.000 claims description 22
- 238000009826 distribution Methods 0.000 claims description 21
- 238000012549 training Methods 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 11
- 238000005070 sampling Methods 0.000 claims description 11
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000004913 activation Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000002360 preparation method Methods 0.000 claims description 2
- 230000007246 mechanism Effects 0.000 abstract description 8
- 238000010586 diagram Methods 0.000 description 4
- 238000005065 mining Methods 0.000 description 3
- 239000002355 dual-layer Substances 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000005295 random walk Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Databases & Information Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a link prediction method based on a topic perception heterogeneous graph neural network, which aims to further mine fine-granularity topic perception semantics based on structural semantics and is used for multi-aspect topic perception representation learning in HGs. Specifically, the present invention first maps vector representations of different types of nodes into a plurality of topic sensing subspaces using a multi-aspect transformation matrix; then an alternate two-step aggregation mechanism is applied, including intra-meta-path decomposition and inter-meta-path merging, to learn the multi-aspect topic perceptual representation of each target node; in addition, a topic prior guiding module is introduced, global statistical knowledge is acquired from unstructured text content by using topic modeling, the aggregation of guide contexts is facilitated, the module serves as a regularization device, the inferred topic perception subspace is more orthogonalized, and the interpretability of the topic perception representation in multiple aspects is improved.
Description
Technical Field
The invention relates to the technical field of networks, in particular to a link prediction method based on a theme-aware heterogeneous graph neural network.
Background
Heterograms (HGs), which consist of different types of nodes and relationships, also known as heterogeneous information networks, are ubiquitous in a variety of real-world scenarios, such as academic networks, social networks, and recommendation systems. As shown in fig. 1 (a), an academic network has multiple types of nodes (authors, papers, meetings, and terms) and edges defined by their relationships (e.g., authors-papers, papers-meetings). HGs typically bear extremely rich and diverse semantics due to the heterogeneity of graph structures and node properties. Accordingly, much research effort is devoted to heterograph representation to map HGs to low-dimensional vector space and for downstream tasks. Among tasks related to HGs, link prediction is a fundamental and important task that can evaluate the probability of a link being present between two nodes, which is the basis of many data mining tasks (e.g., recommendations).
Recently, graph Neural Networks (GNNs) have been used as a powerful deep representation learning method, and have been combined with the structure and node characteristics of graph data with considerable success. Heterogeneous Graphic Neural Networks (HGNNs) have also attracted much attention in recent years inspired by the well-designed mechanisms for homogeneous graphics in GNNs. One mainline of HGNNs is to preserve semantics and model heterostructures by defining and utilizing meta-paths, as different meta-paths can reveal different aspects of the target node from a global perspective. For example, in the academic network illustrated in FIG. 1 (a), author-Paper-Author (APA) and Author-Paper-Conference-Paper-Author (APCPA) are meta-paths that describe two different relationships between authors. The APA meta-path associates two co-authors, while the APCPA meta-path associates two authors who published papers on the same meeting. Specifically, the classical paradigm of HGNNs is to employ hierarchical aggregation at the node level and semantic level to encode and fuse information from different meta-paths, as shown in fig. 1 (b).
Although pel paths such heterostructures contain much semantic information, rich unstructured text content carried by nodes, such as paper summaries, descriptions, or comments, are also common in HGs. Furthermore, text content is often a semantic mix produced by multiple topic perceptions, which fundamentally explains why different types of nodes can link and form specific heterostructures. This topic-aware semantics is more fine-grained for link prediction than structural semantics. For example, author node a in FIG. 1 (c) 1 There may be multiple research interests for different topics, which may be reflected in locally heterogeneous contexts. In the APA-based context, she is in contact with author a 3 Co-operate together because they have a common interest in the "graph mining" topic, while at the same time being directed to author a 4 The same interests on the "NLP application" establish a partnership. Similarly, from a 1 From APCPA-based context, a can be inferred 1 And a 5 Is accepted by the same conference and is also relevant to the "graph mining" topic. If we do not consider such fine-grained semantics by identifying potential multi-aspect topic-aware factors, but simply fuse the resulting confusing features, the performance of node representation in link prediction will inevitably be limited.
Recently, there have been some attempts to identify potential interpretations behind the data using decoupled representation learning, with a considerable result. Most of the previous work with decoupled representation learning is mainly directed to the field of image representation learning. In order to process non-euclidean graph data, some work has explored the potential factors for forming edges between a pair of nodes of a homogeneous graph. While decoupling representation learning has been used by work in HGNNs, it focuses only on coarse granularity and local level, aims to automatically distinguish structural semantics and avoid selecting element paths only from neighboring nodes, and it cannot further identify and reveal fine granularity semantics under bare node connections.
In view of the above limitations of current methods, the applicant explores and leverages heterostructures and unstructured text content in HGs. In particular, we have studied more deeply to identify potential but fundamental topic-aware factors based on the rich structural semantics in HGs in order to learn the multi-aspect topic-aware representation of nodes while preserving such hierarchical semantics for link prediction. However, this presents some challenges and therefore cannot be directly extended by existing solutions. First, HGs typically contain complex interactions between nodes and diverse attribute information, but no explicit label indicates potential and subtle topic perceptions. This presents difficulties in distinguishing the blended information and decomposing the feature vectors into multi-aspect topic aware components. Second, after identifying potential topic awareness factors, an appropriate mechanism is needed to combine structure and topic awareness semantics. Third, to synchronize structural semantics at the global level, it is also important to maintain global features of topic-aware semantics and to maintain the quality of multi-aspect topic-aware representations, thereby improving both model performance and interpretability.
Disclosure of Invention
In order to solve the problems, the invention provides a link prediction method based on a topic perception heterogeneous graph neural network (THGNN), which aims to further mine fine-granularity topic perception semantics based on structural semantics and is used for multi-aspect topic perception representation learning in HGs. Specifically, the present invention first maps vector representations of different types of nodes into a plurality of topic sensing subspaces using a multi-aspect transformation matrix; then an alternate two-step aggregation mechanism is applied, including intra-meta-path decomposition and inter-meta-path merging, to learn the multi-aspect topic perceptual representation of each target node; the main objective of the decomposition step in the meta-path is to infer the topic perception distribution based on the context of the meta-path, and aggregate the context information according to the distribution to form multiple aspects of representation, thereby capturing topic perception semantics with fine granularity; the merging step between the element paths adopts a multi-aspect attention mechanism to merge different element paths for final multi-aspect representation, so that the structure and the theme perception semantics for link prediction are reserved; in addition, a topic prior guiding module is introduced, global statistical knowledge is acquired from unstructured text content by using topic modeling, the aggregation of guide contexts is facilitated, the module serves as a regularization device, the inferred topic perception subspace is more orthogonalized, and the interpretability of the topic perception representation in multiple aspects is improved.
In order to achieve the above object, the present invention provides the following technical solutions:
the invention provides a link prediction method based on a theme-aware heterogeneous graph neural network, which comprises the steps of multi-aspect mapping, multi-aspect heterogeneous graph neural network establishment and theme priori guidance; the multi-aspect mapping step is used as a preprocessing stage, different types of node characteristic representations are projected into the same shared potential vector subspace indicating multi-aspect theme perception semantics, and preparation is made for a theme perception heterogeneous graph neural network aggregation process; the multi-aspect heterogeneous graph neural network building step comprises two steps of decomposition in a meta-path and merging between meta-paths, wherein the decomposition in the meta-path is used for capturing theme perception semantics, the merging between the meta-paths is used for retaining structural semantics, and the two steps are used for iteratively learning multi-aspect theme perception representations with multi-aspect factor nodes; the topic prior guiding step is used as a regularizer, global statistical knowledge is obtained from unstructured text content by using a topic model, context aggregation is guided, and the quality of multi-aspect topic perception representation is further maintained.
Further, in the multi-aspect mapping step, it is assumed that there are K potential topic sensing subspaces in HG, for each type of node, the feature vector is projected to the K potential topic sensing subspaces by applying K type-specific linear transformation matrices, and after multi-aspect mapping, all node features share the same D dimension and are composed of K parts.
Further, for types ofThe multi-aspect mapping process is represented as follows:
where k=1, 2, …, K,is the original eigenvector of node u, +.>Is a mapping vector under the kth underlying topic-aware subspace, < ->Is directed to type +.>A kth trainable weight matrix of nodes of (c).
Further, the sampling method of the multi-aspect mapping step is as follows: given one of HG's having multifaceted factor node u, a meta-path based context is collected according to equation (2) that includes a plurality of text-related nodes with high subject-matter consistency:
wherein,is context c based on meta-path M u Sampling probability of +.>Is context c pre-computed by topic model LDA u Middle node v s The topic probability distribution of the contained text content.
Further, the multi-aspect heterogeneous graph neural network builds the equation:
where g (·) is the final multi-aspect topic awareness representation y of the learning node u u Is used as a function of the aggregation function of (a),refers to the context c given to the meta-path in the kth topic aware subspace u Is represented by y u Composed of K parts y u ＝[z u，1 ，z u，2 ，…，z u,K ]Wherein->The kth topic perceptron property of node u can be characterized.
Further, the specific method of the decomposing step in the meta-path is as follows: given the multi-aspect projected feature vector of the target node u, and a set of selected meta-paths m= (M 1 ，M 2 ，…，M P ) For u, decomposing into u within the meta-path generates a P-component path specific multi-aspect topic aware representation, denoted as
Further, the specific method of the merging step between the meta paths is as follows: converting the single meta-path specific representation with a weight matrix W and then passing the converted single meta-path specific representation for the target node in each batch of dataAveraging the total element path M in the kth topic aware subspace i ：
Where k=1, 2, …, K,is a learnable parameter, B is the size of the batch;
and then using the multifaceted attention vectorThe meta-path M in each topic-aware subspace is measured according to formulas (8) and (9) i Is of importance to (a):
wherein the method comprises the steps of For element path M i Contribution of (2);
using the calculated coefficientsMerging all meta-path specific multi-faceted representations, obtaining a final multi-faceted topic perceptual representation in the current iteration as:
where k=1, 2, …, K.
Further, the two steps of decomposition in the meta-path and merging between the meta-paths are performed iteratively, and the current output is outputtedSequentially taking the three parameters as guidance of the next theme perception distribution inference, and after T rounds of iteration, the theme perception heterogeneous graph neural network passes through an activation function +.> The final post-output K-segment final representation, the final inferred topic perceptual distribution of the meta-path based context for all samples is noted +.>Is based on meta-path M i Soft cluster allocation matrix of context of +.>Comprises K elements calculated at the last iteration by equation (5):
where k=1, 2, …, K anda set of meta-based paths M representing u i The context of the sampling is such that,the formula is as follows:
wherein,f (·) select average pooling operation.
Further, the form of the subject priori guidance is shown in formulas (11) and (12):
wherein,represents the meta-path M-based from equation (5) i Inferred topic awareness distribution of the context of (2) and +.>Representing a first document node d in a single meta-path based context c 1 Theme distribution pre-computed with LDA, +.>Meta-path M representing all samples of a target node in a current small batch i In the context of (a).
Further, the method further comprises the step of model training, specifically: similarity of training pairs (u, v) is estimated by inner products in the multi-aspect topic perceptual subspace, added as a final matching score for the link prediction according to equation (13):
the loss function of graph reconstruction is constructed in the following two modes:
(1) Only one node in the training pair (u, v) has multiple factors, assuming u, then:
wherein b= { B + ∪B - The training pair set is represented by a visible edge set B + And invisible edge set B - ；
(2) The training pair (u, v) has multiple factors for both nodes, and then:
wherein b= { B + ∪B - }；
The overall training loss function combines graph reconstruction and regularization term loss, written as:
compared with the prior art, the invention has the beneficial effects that:
the link prediction method based on the topic perception heterogeneous graph neural network, which is provided by the invention, is named as THGNN, and is used for hierarchically mining topic perception semantics and learning multi-aspect node representation of link prediction in HGs. Specifically, the invention mainly applies an alternate two-step aggregation mechanism comprising decomposition in meta-paths and merging among meta-paths, and can differentially aggregate abundant heterogeneous information according to inferred theme perception factors so as to preserve hierarchical semantics. In addition, a topic a priori guidance module is designed to rely on global knowledge from unstructured text content in HGs to maintain the quality of a multi-aspect topic-aware representation, which helps to improve both performance and interpretability. Experimental results on three real world HGs indicate that our proposed model can be effectively superior to the most advanced methods in linking prediction tasks and demonstrate the potential interpretability of learned multi-aspect topic-aware representations.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present invention, and other drawings may be obtained according to these drawings for a person having ordinary skill in the art.
Fig. 1 is a schematic diagram of the architecture of three academic networks. FIG. 1 (a) is a heterogeneous diagram and a predefined meta-path thereof, FIG. 1 (b) is a dual-layer aggregation mode of the conventional HGNN method, and FIG. 1 (c) is a multi-aspect theme perception factor in the heterogeneous diagram of the present invention.
Fig. 2 is an overall frame diagram of THGNN provided in an embodiment of the present invention. Where fig. 2 (a) converts the original features to potential topic-aware subspaces through multiple aspect mapping. FIG. 2 (b) updates the multifaceted node representation by multifaceted heterogeneous GNNs, comprising two iterative steps: intra-meta-path decomposition and inter-meta-path merging. Fig. 2 (c) regularizes the decomposition process within the meta-path through topic a priori guidance.
Detailed Description
For a better understanding of the present technical solution, the method of the present invention is described in detail below with reference to the accompanying drawings.
Definition 1: heterogeneous information networks. A heterogeneous graph may be defined as a graph g= (V, E) containing a set of multi-type nodes V and a set of edges E. Each node and each edge are respectively associated with a type mapping functionAnd psi E.fwdarw.R, wherein A and R represent the node and the set of relationship types, respectively. When |A|+|R| > 2.
Definition 2: a meta-path. Meta-path M may be defined as a path in the form of a bar:target object A 1 And A l+1 The compound relationship between is defined as r=r 1 R 2 …R l+1 。
Definition 3: meta-path instance. Given the meta-path M of the heterogeneous graph, the meta-path instance M of M is defined as the sequence of nodes in the graph that match the sequence of types in M. The meta-path instance connecting nodes u and v (where node u is the target node) is denoted m u 。
Definition 4: based on the context of the meta-path. Given a meta-path instance m u The meta-path based context c of the target node u is defined as the sequence of nodes in the meta-path instance other than node u, denoted as c u ＝m u { u }, where there is a node containing text content. More specifically, the set of meta-path M-based contexts with target node u is represented asc u The text-related node sequence in (a) is denoted +.>
Examples: as shown in FIG. 1 (a), we constructed an HG to model DBLP. It consists of multiple types of objects (author (A), paper (P), term (T), meeting (C)) and relationships (P-A, P-T and P-C). Two authors may be connected based on multiple meta paths, e.g., author-paper-author (APA) and author-paper-meeting-paper-author (APCPA). Given element path APA, a 3 -p 2 -a 1 Is with target node a 1 An associated meta-path instance. Author a 1 By connecting with other authors through different meta-paths, the node sequence (a 3 ,p 2 ) (APA-based) and (a) 5 ,p 3 ,c 1 ,p 4 ) (APCPA-based) constitutes a 1 Different meta-path based contexts may reveal different topic-aware semantics because the paper nodes in the meta-path based context carry rich text content.
To guide the information aggregation of nodes with multifaceted factors according to hidden topic-aware semantics behind HGs, we propose a topic-aware heterographic neural network (THGNN). The architecture of THGNN comprises three basic components, respectively: multiple aspect mapping, multiple aspect heterogeneous graph neural networks, topic prior guidance. Wherein, the multi-aspect mapping module is used as a preprocessing stage to prepare for the aggregation process of THGNN. The key module 'multi-aspect heterogeneous graph neural network' comprises two steps: the intra-meta-path decomposition and inter-meta-path merging iteratively learn a multi-faceted topic-aware representation with multifaceted factor nodes. In addition, topic a priori guidance is introduced to guide context aggregation, further preserving the quality of the multi-aspect topic-aware representation. FIG. 2 illustrates an overall multi-aspect topic aware representation generation process. Each of which is described in detail below.
1. Multiple aspect mapping
Because of the heterogeneity of nodes in HGs, different types of nodes and edges have different properties, often in disparate feature spaces. To mine the underlying topic-aware subspaces in HGs, we need to project different types of node feature representations into the same shared underlying vector subspace that is indicative of multi-aspect topic-aware semantics.
Thus, assuming there are K potential topic-aware subspaces in HG, for each type of node, we apply K type-specific linear transformation matrices to project feature vectors into the K potential topic-aware subspaces. As shown in fig. 2 (a), for a type ofThe multifaceted mapping process can be expressed as follows:
where k=1, 2, …, K.Is the original eigenvector of node u, +.>Is the mapping vector in the kth potential topic aware subspace. />Is directed to type +.>A kth trainable weight matrix of nodes of (c).
After multi-aspect mapping, all node features share the same D dimension, grouped by K parts.
Sampling mechanism: given one of the HGs with multifaceted factor node u, we first need to sample some of the meta-path based context according to the different meta-paths. To identify multi-aspect topic perceptions, we employ a sampling strategy to focus more on meta-path based contexts that contain multiple text-related nodes with high topic consistency. The sampling process is as follows:
wherein the method comprises the steps ofIs context c based on meta-path M u Sampling probability of +.>Is context c pre-computed by topic model LDA u Middle node v s The topic probability distribution of the contained text content.
2. Multi-aspect heterogeneous graphic neural network
In the following discussion, we will amplify the key module of THGNN, which consists of two steps: the meta-path inner decomposition captures topic awareness semantics and the meta-path merge to preserve structural semantics. The purpose of the decomposition step within the meta-path is to initially inferThe topic aware distribution of the context based meta path M and aggregate the context information according to the inferred distribution to form a multi-aspect representation. The inter-meta-path merging step aims at merging the different meta-paths to produce a representation that retains structural and topic-aware semantics in many ways in the current iteration step. The above two steps are alternately performed and summarized as follows: a step of
Where g (·) is the final multi-aspect topic awareness representation y of the learning node u u Is used as a function of the aggregation function of (a),refers to the context c given to the meta-path in the kth topic aware subspace u Is a representation of (c). Just as there are K potential theme-aware subspaces previously opened, we want to let y u Composed of K parts y u ＝[z u,1 ,z u,2 ,…,z u,K ]Wherein->The kth topic perceptron property of node u can be characterized.
2.1 in-path decomposition
As in FIG. 2 (b) -1), a meta-path M is given i E M, the goal of this step is to infer to which topic each sampled meta-path based context belongs. We expectThe current meta-path based context can be utilized to capture the k-aspect topic awareness factors of the target node u. The context encoder is as follows:
wherein the method comprises the steps ofWe choose the mean pooling operation in the experiment f (·) in view of simplicity and efficiency.
After encoding the meta-path based context as a multifaceted representation in the topic-aware subspace, we employ cosine similarity between the multifaceted representation of the target node and its current meta-path based context to infer the most relevant topic-aware subspace they share. The reasoning topic distribution process is given by:
where k=1, 2, …, K anda set of meta-based paths M representing u i Context of sampling. Naturally, if the target node u and the kth topic perceive the current meta-path based context c in the subspace u Correlated, thenShould be high. At the same time, it also serves as an importance weight when the current meta-path based context information propagates to the target node in the kth topic aware subspace. Thus, we can obtain a meta-path specific multi-aspect topic aware representation of the target node u:
where k=1, 2, …, K, M i ∈M。
In summary, the feature vector after the multi-aspect projection of a given target node u, and a set of selected meta-paths m= (M) 1 ,M 2 ,…,M P ) For u, decomposing into u in meta-path to generate P-component meta-pathDiameter-specific multi-aspect topic aware representations, expressed as
The topic-aware distribution and the multi-aspect topic-aware representation derived from the meta-path decomposition step are intended to be as diverse as possible, while the inter-meta-path merging step is intended to find the overall importance of the multi-aspect information under specific structural semantics.
2.2-element inter-path merging
Since each node in HG contains multiple types of structural semantic information that can be revealed by meta-paths, we propose multiple aspects of attention to learn the importance of the different meta-paths and combine them to generate the final multiple aspects of topic aware representation of target node u, as in FIG. 2 (b) -2).
To learn a given meta-path M i First we transform the meta-path specific single representation with a weight matrix W and then summarize the meta-path M in the kth topic aware subspace by averaging the transformed meta-path specific single representation of the target node in each batch of data i ：
Where k=1, 2, …, K,is a learnable parameter and B is the size of the batch.
And then using the multifaceted attention vectorTo measure meta-path M in each topic-aware subspace i Is of importance. Since the different meta-paths describe the various semantics behind the HG at a structural level, the overall importance of the different meta-paths should be shared in all topic-aware subspaces, so we aggregate the multi-aspect relevance importance into a meta-pathM i Is of overall importance:
wherein the method comprises the steps ofCan be interpreted as meta-path M i Is a contribution of (a).
Using the calculated coefficientsWe can combine all meta-path specific multi-faceted representations to obtain the final multi-faceted topic perceptual representation in the current iteration:
where k=1, 2, …, K.
The above two steps are carried out as iteration, and the current outputThis can also be interpreted as updated K topic-aware cluster centers, in turn as guidance for the next topic-aware distribution inference. After iteration of the T round THGNN is subjected to an activation function +.>1,2, …, K } outputs a final representation of the K segment (let z for text-dependent node v v,k ＝σ(h v，k ) Without any iteration), and the final inferred topic perceptual distribution of the meta-path based context for all samples, noted +.>Here->Can be interpreted as being based on meta-path M i Soft cluster allocation matrix of context of +.>Comprises K elements calculated by equation (5) at the last iteration.
3. Topic a priori guidance
While iterative processes in multi-aspect heterogeneous graph neural networks can infer topic-aware distributions based on the context of the meta-path and generate segmented representations for target nodes, there are two more points to further explore:
(1) Due to the heterogeneity of meta-path based contexts, there may still be confusing topic-aware subspaces, even resulting in extreme collapse scenarios;
(2) The meta-path decomposition step is essentially performed as a local cluster, which makes it difficult to obtain global knowledge from the K potential topic-aware subspaces we hypothesize.
The first point is to inspire that the topic-aware subspaces we infer need to be more independent, and each subspace remains of a certain size. Another point means that global a priori knowledge is necessary to preserve the global nature of the topic-aware subspace. To meet these two requirements, we introduced another module named topic a priori guidance to encourage the inferred topic-aware subspaces to be more orthogonal and improve interpretability.
As shown in fig. 2 (c), the added module will act as a regularizer, skillfully using the topic model to obtain global statistical knowledge from unstructured text content, and is used to guide context aggregation in HG, which is in the form of:
wherein the method comprises the steps ofRepresents the meta-path M based from equation 5 i Inferred topic awareness distribution of the context of (2) and +.>Representing a first document node d in a single meta-path based context c 1 The topic distribution is pre-computed using LDA. Note that we will sample all meta-paths M based on the target node in the current batch i The contexts are spliced to form a matrix->Rather than handling a single target node. Similarly, a->Meta-path M, which also represents all samples of the target node in the current small lot i In the context of (a). Thus (S)>The underlying topic-aware subspace under HG plays a priori guiding role for context aggregation.
4. Model training
We estimate the similarity of training pairs (u, v) by inner products in the multi-aspect topic perceptual subspace, adding as the final matching score for the link prediction:
we construct a loss function for graph reconstruction with the following two paradigms:
(1) Only one node in the training pair (u, v) has multiple factors, assuming u, then:
wherein b= { B + ∪B - The training pair set is represented by a visible edge set B + And invisible edge set B - 。
(2) The training pair (u, v) has multiple factors for both nodes, and then:
wherein b= { B + ∪B - The final matching score is averaged here because THGNN is subjected to the l2 regularization operation before the final multifaceted representation is output.
The overall training loss function combines graph reconstruction and regularization term loss, and can be written as:
the invention proposes, from a new perspective, recognition of multiple aspects of theme perception factors behind bare links between associated nodes, and fully utilizes heterostructure and unstructured text content to conduct HGs' link prediction. The rich heterogeneous information can be differentially aggregated according to the inferred multi-aspect topic awareness factors, thereby generating multi-aspect topic awareness representations that preserve structural and topic awareness semantics, and preserving dual-layer semantics of structural and topic awareness. In addition, the designed topic prior guiding module further uses global knowledge from unstructured text content to guide the blocking aggregation of heterogeneous neighborhood information, meanwhile, the inferred topic perception subspace is more orthogonalized, the interpretability of the learned multi-aspect topic perception representation is improved, and the quality of multi-aspect topic perception embedding is further maintained.
In addition, experiments on three real heterogeneous graph datasets of DBLP, YELP and Amazon show that the method provided by the invention is obviously superior to the most advanced method in the link prediction task, and the potential interpretability of the multi-aspect theme perception representation learned by us is also shown.
These comparative graph submerging methods can be summarized roughly into three categories: random walk-based (deep, metaath 2vec, HERec), GNN-based (graphpage, GAT, distnegcn), HGNN-based (HAN, disenHAN, MAGNN). In addition, THGNN \MA Is a variant of the THGNN model, i.e. removes the multifaceted attention mechanism while simply imparting an averaging operation to the meta-path.
Table 1. Performance comparisons of linked prediction tasks.
The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may be modified or some technical features may be replaced with others, which may not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (7)
1. The link prediction method based on the topic perception heterogeneous graph neural network is characterized by comprising the steps of multi-aspect mapping, multi-aspect heterogeneous graph neural network establishment and topic priori guidance; the multi-aspect mapping step is used as a preprocessing stage, different types of node characteristic representations are projected into the same shared potential vector subspace indicating multi-aspect theme perception semantics, and preparation is made for a theme perception heterogeneous graph neural network aggregation process; the multi-aspect heterogeneous graph neural network building step comprises two steps of decomposition in a meta-path and merging between meta-paths, wherein the decomposition in the meta-path is used for capturing theme perception semantics, the merging between the meta-paths is used for retaining structural semantics, and the two steps are used for iteratively learning multi-aspect theme perception representations with multi-aspect factor nodes; the topic prior guiding step is used as a regularizer, global statistical knowledge is obtained from unstructured text content by using a topic model, context aggregation is guided, and the quality of multi-aspect topic perception representation is further maintained;
the multi-aspect heterogeneous graph neural network builds the equation:
where g (·) is the final multi-aspect topic awareness representation y of the learning node u u Is used as a function of the aggregation function of (a),refers to the context c given to the meta-path in the kth topic aware subspace u Is represented by y u Composed of K parts y u ＝[z u,1 ,z u,2 ,…,z u,K ]Wherein->The kth theme perception factor property of the node u can be described;
the specific method of the decomposition step in the meta-path is as follows: given the multi-aspect projected feature vector of the target node u, and a set of selected meta-paths m= (M 1 ,M 2 ,…,M P ) For u, decomposing into u within the meta-path generates a P-component path specific multi-aspect topic aware representation, denoted as
The specific method of the merging step between the element paths is as follows: converting the single meta-path specific representation with a weight matrix W, and then summarizing the meta-path M in the kth topic aware subspace by averaging the converted single meta-path specific representation of the target node in each batch of data i ：
Where k=1, 2, …, K,is a learnable parameter, B is the size of the batch;
and then using the multifaceted attention vectorThe meta-path M in each topic-aware subspace is measured according to formulas (8) and (9) i Is of importance to (a):
wherein the method comprises the steps of For element path M i Contribution of (2);
using the calculated coefficientsMerging all meta-path specific multi-faceted representations, obtaining a final multi-faceted topic perceptual representation in the current iteration as:
where k=1, 2, …, K.
2. The link prediction method based on a topic-aware heterogeneous graph neural network according to claim 1, wherein in the multi-aspect mapping step, assuming that K potential topic-aware subspaces exist in HG, for each type of node, applying K type-specific linear transformation matrices to project feature vectors to the K potential topic-aware subspaces, and after multi-aspect mapping, all node features share the same D-dimension and consist of K parts.
3. The link prediction method based on the topic-aware heterogeneous graphic neural network according to claim 2, wherein for a type ofThe multi-aspect mapping process is represented as follows:
where k=1, 2, …, K,is the original eigenvector of node u, +.>Is a mapping vector under the kth underlying topic-aware subspace, < ->Is directed to type +.>A kth trainable weight matrix of nodes of (c).
4. The link prediction method based on the topic-aware heterogeneous graph neural network according to claim 2, wherein the sampling method of the multi-aspect mapping step is as follows: given one of HG's having multifaceted factor node u, a meta-path based context is collected according to equation (2) that includes a plurality of text-related nodes with high subject-matter consistency:
wherein,is context c based on meta-path M u Sampling probability of +.>Is context c pre-computed by topic model LDA u Middle node v s The topic probability distribution of the contained text content.
5. The link prediction method based on the topic-aware heterogeneous graph neural network according to claim 1, wherein the two steps of intra-path decomposition and inter-path merging are iteratively performed, and the current outputSequentially taking the three parameters as guidance of the next theme perception distribution inference, and after T rounds of iteration, the theme perception heterogeneous graph neural network passes through an activation function +.> Post-output K-segment final representation, final inferred topic awareness for all sampled meta-path based contextsDistribution is marked as-> Is based on meta-path M i Soft cluster allocation matrix of context of +.>Comprises K elements calculated at the last iteration by equation (5):
where k=1, 2, …, K and a set of meta-based paths M representing u i Context of sampling, ++>The formula is as follows:
wherein,f (·) select average pooling operation.
6. The link prediction method based on the topic-aware heterogeneous graph neural network according to claim 5, wherein the topic a priori guidance is in the form of formulas (11) and (12):
wherein,represents the meta-path M-based from equation (5) i Inferred topic awareness distribution of the context of (2) and +.>Representing a first document node d in a single meta-path based context c 1 Theme distribution pre-computed with LDA, +.>Meta-path M representing all samples of a target node in a current small batch i In the context of (a).
7. The link prediction method based on the topic-aware heterogeneous graph neural network according to claim 1, further comprising the step of model training, specifically: similarity of training pairs (u, v) is estimated by inner products in the multi-aspect topic perceptual subspace, added as a final matching score for the link prediction according to equation (13):
the loss function of graph reconstruction is constructed in the following two modes:
(1) Only one node in the training pair (u, v) has multiple factors, assuming u, then:
wherein b= { B + ∪B - The training pair set is represented by a visible edge set B + And invisible edge set B - ；
(2) The training pair (u, v) has multiple factors for both nodes, and then:
wherein b= { B + ∪B - }；
The overall training loss function combines graph reconstruction and regularization term loss, written as:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111048789.4A CN113672735B (en) | 2021-09-08 | 2021-09-08 | Link prediction method based on theme-aware heterogeneous graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111048789.4A CN113672735B (en) | 2021-09-08 | 2021-09-08 | Link prediction method based on theme-aware heterogeneous graph neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113672735A CN113672735A (en) | 2021-11-19 |
| CN113672735B true CN113672735B (en) | 2024-01-19 |
Family
ID=78548799
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111048789.4A Active CN113672735B (en) | 2021-09-08 | 2021-09-08 | Link prediction method based on theme-aware heterogeneous graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113672735B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114528221B (en) * | 2022-02-24 | 2023-04-07 | 北京航空航天大学 | Software defect prediction method based on heterogeneous graph neural network |
| CN114491080B (en) * | 2022-02-28 | 2023-04-18 | 中国人民解放军国防科技大学 | Unknown entity relationship inference method oriented to character relationship network |
| CN115080846B (en) * | 2022-05-30 | 2024-06-11 | 广州大学 | Scene information-fused graphic neural network clothing recommendation method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018188240A1 (en) * | 2017-04-10 | 2018-10-18 | 北京大学深圳研究生院 | Cross-media retrieval method based on deep semantic space |
| CN111737535A (en) * | 2020-06-22 | 2020-10-02 | 复旦大学 | Network characterization learning method based on element structure and graph neural network |
| CN112364161A (en) * | 2020-09-25 | 2021-02-12 | 天津大学 | Microblog theme mining method based on dynamic behaviors of heterogeneous social media users |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150332158A1 (en) * | 2014-05-16 | 2015-11-19 | International Business Machines Corporation | Mining strong relevance between heterogeneous entities from their co-ocurrences |
-
2021
- 2021-09-08 CN CN202111048789.4A patent/CN113672735B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018188240A1 (en) * | 2017-04-10 | 2018-10-18 | 北京大学深圳研究生院 | Cross-media retrieval method based on deep semantic space |
| CN111737535A (en) * | 2020-06-22 | 2020-10-02 | 复旦大学 | Network characterization learning method based on element structure and graph neural network |
| CN112364161A (en) * | 2020-09-25 | 2021-02-12 | 天津大学 | Microblog theme mining method based on dynamic behaviors of heterogeneous social media users |
Non-Patent Citations (1)
| Title |
|---|
| 融合关联规则的学术论文主题学习及表示方法;赵慧茹;林民;;计算机工程与应用(20);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113672735A (en) | 2021-11-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113672735B (en) | Link prediction method based on theme-aware heterogeneous graph neural network | |
| Qin et al. | Semantic communications: Principles and challenges | |
| CN110163258B (en) | Zero sample learning method and system based on semantic attribute attention redistribution mechanism | |
| Suzuki et al. | Joint multimodal learning with deep generative models | |
| Lee et al. | Detecting handcrafted facial image manipulations and GAN-generated facial images using Shallow-FakeFaceNet | |
| CN109919204B (en) | Noise image-oriented deep learning clustering method | |
| US11928957B2 (en) | Audiovisual secondary haptic signal reconstruction method based on cloud-edge collaboration | |
| Patel et al. | Sparsity-motivated automatic target recognition | |
| CN112364161B (en) | Microblog theme mining method based on dynamic behaviors of heterogeneous social media users | |
| Farhadinia et al. | A family of similarity measures for q‐rung orthopair fuzzy sets and their applications to multiple criteria decision making | |
| CN111985538A (en) | Small sample picture classification model and method based on semantic auxiliary attention mechanism | |
| CN112991160A (en) | Image processing method, image processing device, computer equipment and storage medium | |
| CN116304367B (en) | Algorithm and device for obtaining communities based on graph self-encoder self-supervision training | |
| Miyamoto | Different Objective Functions in Fuzzy c-Means Algorithms and Kernel-Based Clustering. | |
| CN116030296A (en) | Social platform data mining method and system for graphic data collaboration | |
| Lai et al. | Faithful vision-language interpretation via concept bottleneck models | |
| CN113222002B (en) | Zero sample classification method based on generative discriminative contrast optimization | |
| Patel et al. | Dictionaries for image and video-based face recognition | |
| Shen et al. | Channel recombination and projection network for blind image quality measurement | |
| CN113822183B (en) | Zero sample expression recognition method and system based on AU-EMO association and graph neural network | |
| Sampaio et al. | Are facial expression recognition algorithms reliable in the context of interactive media? A new metric to analyse their performance | |
| Suzuki et al. | Improving bi-directional generation between different modalities with variational autoencoders | |
| Sarkar et al. | Multi-source transfer learning for facial emotion recognition using multivariate correlation analysis | |
| Gonzaga et al. | Multimodal intent classification with incomplete modalities using text embedding propagation | |
| Giljum et al. | General neural network approach to compressive feature extraction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |