Job description
In this paper, a new inductive reasoning method, namely TACT , is proposed, which can effectively exploit the topology-aware correlation between relations in knowledge graphs .
Specifically, TACT models semantic correlations from two aspects of correlation patterns and correlation coefficients .
Divide all relationship pairs into 7 categories. Different topological structures correspond to different association patterns. Then, the original knowledge graph is transformed into a Relational Correlation Graph (RCG) , where nodes represent relations and edges represent association patterns between any two relations in the original knowledge graph . Based on RCG , a Relational Network (RCN) is proposed to learn the correlation coefficients of different modalities for inductive link prediction.
TACT can effectively absorb the information of neighbor relations, thereby improving the performance of link prediction in an inductive setting.
Article contribution:
(a) Innovatively classify all relation pairs into seven topological patterns, and propose a new relation-dependent relation network to model topology-aware correlation .
(b) Considers the inductive link prediction task , while the aforementioned knowledge graph embedding methods have difficulties in handling it. (?)
(c) outperforms existing state-of-the-art inductive inference methods on benchmark datasets.
Job Introduction - Introduction to TACT Model Structure
TACT aims to score a given triple (u , rt , v) in an entity-independent manner, where rt is the target relation between entities u and v . Specifically, TACT consists of two modules : a relational association module and a graph structure module .
The semantic association between any two relations is highly related to its topology.
A graph structure module is designed based on the idea of GraIL .
Job Introduction - Relationship Association Module - Association Mode
To model the semantic relatedness between relations, two aspects of relatedness are considered :
(a) Association pattern : The association between any two relations is highly correlated with its topological structure in the knowledge graph.
Based on the definitions of different correlation patterns, we can transform the original knowledge graph into a relational correlation graph ( RCG) , where nodes represent relations and edges represent the correlation patterns between any two relations in the original knowledge graph.
Classify all relation pairs into 7 topological patterns. These topological patterns are "head-to-tail", "tail-to-tail", "head-to-head", "tail-to-head", "parallel", "ring" and "disconnected".
Job Introduction - Relational Association Module - Correlation Coefficient
(b) Correlation coefficient : We use correlation coefficient to denote the degree of semantic relatedness between any two relations.
For an edge with relation rt , we can divide all its adjacent edges in RCG into six groups according to topological patterns " HT" , " TT" , " HH" , " TH" , " PARA" and " LOOP" , respectively . For these 6 groups, 6 linear transformations are used to learn different semantic associations corresponding to topological patterns.
To distinguish different degrees of relevance rt of relations , an attention network is further used to learn the correlation coefficients of all the relationships. Specifically, all correlation coefficients of the relation rt are aggregated to obtain the neighborhood embedding in the local graph , denoted by rNt :
W is a weight matrix, R is the relational embedding, Npt is the "indicator vector", and Λpt represents the weight parameters of different degrees of association in the association pattern.
The final relation expresses:
Job Introduction - Graph Structure Modeling
Graph embeddings follow ( grail ) the node labeling of the enclosing subgraph, and the nodes in the subgraph have initial embeddings.
Use R-GCN to learn an embedding layer on the extracted closed subgraph G(u , rt , v) :
The embedding of subgraph G(u , rt , v) is:
The final output of this module:
Design a scoring network to combine the outputs of the two modules and get a score for a given triplet (u , rt , v) :
Experiment -AUC-PR value
TACT-base：
Experiment -H@1 and MRR , Link Prediction and Running Time
How does inductive link prediction work? (There are doubts, it seems that it is not displayed in the code)
