Processing graphs using graph patterns
In an approach, a processor identifies subgraphs of predefined patterns in a first graph, the first graph: (i) representing a specific ontology and (ii) comprising nodes representing first entities and first edges representing relationships between the first entities. A processor represents the identified subgraphs by respective second graphs, thereby forming multi-relational graphs, each second graph comprising nodes representing second entities and second edges representing relationships between the second entities, where: the second entities are the respective nodes of the first graph; and each second edge indicates that the two second entities linked by the second edge are part of a pattern of the predefined patterns. A processor inputs the multi-relational graphs to a multi-relational graph neural network for generating output in accordance with the specific ontology. A processor provides the output.
1 . A computer-implemented method comprising:
training, using a full-batch gradient descent technique, a multi-relational graph neural network using training data comprising a set of multi-relational graphs with labeled nodes, wherein:
the multi-relational graph neural network is a relational graph convolutional network (RGCN), for generating embeddings for input entities and input edges of multi-relational graphs, that takes a set of vectors representing nodes as input, together with a structure of a graph, and generates a new collection of representations for nodes in the graph; and
multiple layers of the multi-relational graph neural network are stacked such that output of a preceding layer is used as input for a current layer;
identifying subgraphs of predefined patterns in a first graph, the first graph: (i) representing a specific ontology and (ii) comprising nodes representing first entities and first edges representing relationships between the first entities, wherein:
the specific ontology comprises concepts of a specific domain; and
each edge is connected to two nodes;
generating, by one or more processors, multi-relational graphs by representing the identified subgraphs by respective second graphs, each second graph comprising a matrix with nodes representing second entities and second edges representing relationships between the second entities, wherein:
for each respective matrix, each axis of the matrix corresponds to edges present within first graph cells set to a predefined value for each edge associated with the respective subgraph pattern;
the second entities are the respective nodes of the first graph;
each second edge indicates that the two second entities linked by the second edge are part of a pattern of the predefined patterns; and
the multi-relational graphs are represented as a three-dimensional matrix obtained by concatenating each individual matrix;
inputting the multi-relational graphs to the multi-relational graph neural network for generating output in accordance with the specific ontology; and
providing the output of the multi-relational graph neural network, the output including an indication of a class of the ontology for each node in the multi-relational graphs.
2 . The computer-implemented method of claim 1 , further comprising prior to inputting the multi-relational graphs to the multi-relational graph neural network, adding an additional graph to the multi-relational graphs.
3 . The computer-implemented method of claim 1 , wherein representing the identified subgraphs by the respective second graphs further comprises adding a node to the respective second graph that represents the pattern associated with a subgraph of the identified subgraphs and linking the second entities that belong to the pattern with a created entity.
4 . The computer-implemented method of claim 1 , wherein responsive to the second entities being the first entities of the first graph, the second edges comprise first edges of the first graph.
5 . The computer-implemented method of claim 1 , further comprising:
creating a table having a record per second entity of the second graph, the table comprising fields representing a predicted embedding and one or more attributes descriptive of the second entities; and
providing records of the table as input to a Light Gradient Boosting Model for classification of the second entities into classes of the specific ontology.
6 . The computer-implemented method of claim 1 , wherein the predefined patterns are defined according to rules of the specific ontology.
7 . The computer-implemented method of claim 1 , wherein the predefined patterns are selected from the group consisting of: fan-out pattern, bipartite pattern, cycle pattern, gather-scatter pattern, fan-in pattern, stacked bipartite pattern, and scatter-gather pattern.
8 . The computer-implemented method of claim 1 , further comprising creating a table having a record per second entity of the second graph, the table comprising fields representing a predicted embedding and one or more attributes descriptive of the second entities and storing the table.
9 . The computer-implemented method of claim 1 , wherein:
for each pattern, multiple subgraphs are identified in the first graph; and
for each pattern, the second graph represents the multiple subgraphs representing the pattern.
10 . The computer-implemented method of claim 2 , wherein, responsive to the second entities of the second graph representing the first entities, the additional graph is the first graph.
11 . The computer-implemented method of claim 2 , wherein, responsive to the second entities of the second graph representing the first edges, the additional graph is a line graph representing the first graph.
12 . The computer-implemented method of claim 1 , further comprising:
minimizing cross-entropy loss on labeled nodes.
13 . The computer-implemented method of claim 1 , wherein:
the specific ontology comprises a biology ontology; and
the relationships between the first entities for each of the first edges comprise associative relationships.
14 . A computer program product comprising:
one or more computer readable storage media; and
program instructions stored on the one or more computer readable storage media to perform operations comprising:
training, using a full-batch gradient descent technique, a multi-relational graph neural network using training data comprising a set of multi-relational graphs with labeled nodes, wherein:
the multi-relational graph neural network is a relational graph convolutional network (RGCN), for generating embeddings for input entities and input edges of multi-relational graphs, that takes a set of vectors representing nodes as input, together with a structure of a graph, and generates a new collection of representations for nodes in the graph; and
multiple layers of the multi-relational graph neural network are stacked such that output of a preceding layer is used as input for a current layer;
identifying subgraphs of predefined patterns in a first graph, the first graph: (i) representing a specific ontology and (ii) comprising nodes representing first entities and first edges representing relationships between the first entities, wherein:
the specific ontology comprises concepts of a specific domain; and
each edge is connected to two nodes;
generating multi-relational graphs by representing the identified subgraphs by respective second graphs, each second graph comprising a matrix with nodes representing second entities and second edges representing relationships between the second entities, wherein:
for each respective matrix, each axis of the matrix corresponds to edges present within first graph cells set to a predefined value for each edge associated with the respective subgraph pattern;
the second entities are the respective nodes of the first graph;
each second edge indicates that the two second entities linked by the second edge are part of a pattern of the predefined patterns; and
the multi-relational graphs are represented as a three-dimensional matrix obtained by concatenating each individual matrix;
inputting the multi-relational graphs to the multi-relational graph neural network for generating output in accordance with the specific ontology; and
providing the output of the multi-relational graph neural network, the output including an indication of a class of the specific ontology for each node in the multi-relational graphs.
15 . The computer program product of claim 14 , wherein the operations further comprise:
prior to inputting the multi-relational graphs to the multi-relational graph neural network, adding an additional graph to the multi-relational graphs.
16 . The computer program product of claim 14 , wherein representing the identified subgraphs by the respective second graphs further comprises adding a node to the respective second graph that represents the pattern associated with a subgraph of the identified subgraphs and linking the second entities that belong to the pattern with a created entity.
17 . The computer program product of claim 14 , wherein responsive to the second entities being the first entities of the first graph, the second edges comprise first edges of the first graph.
18 . The computer program product of claim 14 , wherein the operations further comprise:
creating a table having a record per second entity of the second graph, the table comprising fields representing a predicted embedding and one or more attributes descriptive of the second entities; and
providing records of the table as input to a Light Gradient Boosting Model for classification of the second entities into classes of the specific ontology.
19 . The computer program product of claim 14 , wherein the predefined patterns are defined according to rules of the specific ontology.
20 . A computer system comprising:
a processor set;
one or more computer readable storage media; and
program instructions stored on the one or more computer readable storage media to cause the processor set to perform operations comprising:
training, using a full-batch gradient descent technique, a multi-relational graph neural network using training data comprising a set of multi-relational graphs with labeled nodes, wherein:
the multi-relational graph neural network is a relational graph convolutional network (RGCN), for generating embeddings for input entities and input edges of multi-relational graphs, that takes a set of vectors representing nodes as input, together with a structure of a graph, and generates a new collection of representations for nodes in the graph; and
multiple layers of the multi-relational graph neural network are stacked such that output of a preceding layer is used as input for a current layer;
identifying subgraphs of predefined patterns in a first graph, the first graph: (i) representing a specific ontology and (ii) comprising nodes representing first entities and first edges representing relationships between the first entities, wherein:
the specific ontology comprises concepts of a specific domain; and
each edge is connected to two nodes;
generating multi-relational graphs by representing the identified subgraphs by respective second graphs, each second graph comprising a matrix with nodes representing second entities and second edges representing relationships between the second entities, wherein:
for each respective matrix, each axis of the matrix corresponds to edges present within first graph cells set to a predefined value for each edge associated with the respective subgraph pattern;
the second entities are the respective nodes of the first graph;
each second edge indicates that the two second entities linked by the second edge are part of a pattern of the predefined patterns; and
the multi-relational graphs are represented as a three-dimensional matrix obtained by concatenating each individual matrix;
inputting the multi-relational graphs to the multi-relational graph neural network for generating output in accordance with the specific ontology; and
providing the output of the multi-relational graph neural network, the output including an indication of a class of the specific ontology for each node in the multi-relational graphs.