Compound discovery via information divergence with knowledge graphs
Complex computer system architectures are described for utilizing a knowledge data graph comprised of elements, and selecting a discovery element to replace an existing element of a formulation depicted in the knowledge data graph. The substitution process takes advantage of the knowledge data graph structure to improve the computing capabilities of a computing device executing a substitution calculation by translating the knowledge data graph into an embedding space, and determining a discovery element from within the embedding space.
1. A system comprising:
a knowledge graph reception circuitry configured to receive an original knowledge graph including a set of structured data;
a knowledge graph embedding circuitry configured to convert the original knowledge graph to an embedding space comprising nodes by converting the set of structured data into sets of vector triples and modelling the vector triples with a neural network architecture to learn the representations of the knowledge graph;
a region slicing circuitry configured to:
identify a selected node within the embedding space;
determine a region of interest within the embedding space containing the selected node, and calculate a similarity score for each node within the region of interest, wherein the similarity score represents a similarity between each node within the region of interest and the selected node;
calculate a weighted look up score for nodes within the region of interest, and select a predetermined number of candidate nodes having a highest weighted look up score; and
calculate an information divergence score for the predetermined number of candidate nodes having the highest weighted look up score, wherein the information divergence score represents a divergence of a discovery candidate node from the selected node; and
computation circuitry configured to:
calculate a discovery score for at least one candidate node within the predetermined number of candidate nodes having the highest weighted look up score; and
select a discovery node from the at least one candidate node according to a respective discovery score.
2. The system of claim 1 , wherein the at least one candidate node represents a compound included in a formulation.
3. The system of claim 1 , wherein the region slicing circuitry is configured to identify the selected node within the embedding space based on at least part of a received user query.
4. The system of claim 1 , wherein the region slicing circuitry is configured to determine the region of interest based on at least part of a received user query.
5. The system of claim 1 , wherein the computation circuitry is configured to calculate the discovery score by a weighted sum of the similarity score, the weighted look up score, and the information divergence score for a respective candidate node.
6. The system of claim 1 , wherein the region slicing circuitry is configured to determine the region of interest as including nodes within a predetermined vector distance from the selected node.
7. The system of claim 1 , wherein the computation circuitry is configured to select a candidate node having a highest discovery score as the discovery node.
8. The system of claim 1 , wherein the region slicing circuitry is configured to calculate the information divergence score by:
determining a first link prediction probability for each node linked to the selected node based on at least part of a received user query;
determining a second link prediction probability for each node linked to the nodes within the region of interest based on at least part of the received user query; and
determining the information divergence score based on the first link prediction probability and the second link prediction probability.
9. The system of claim 1 , wherein the region slicing circuitry is configured to calculate the information divergence score using a Kullback-Leibler (KL) divergence technique.
10. A method comprising:
receiving, by a knowledge graph reception circuitry, an original knowledge graph including a set of structured data;
converting, by a knowledge graph embedding circuitry, the original knowledge graph to an embedding space comprising nodes by converting the set of structured data into sets of vector triples and modelling the vector triples with a neural network architecture to learn the representations of the knowledge graph;
identifying, by a region slicing circuitry, a selected node within the embedding space;
determining, by the region slicing circuitry, a region of interest within the embedding space containing the selected node, and calculating a similarity score for each node within the region of interest, wherein the similarity score represents that depicts a similarity between each node within the region of interest and the selected node;
calculating, by the region slicing circuitry, a weighted look up score for nodes within the region of interest, and selecting a predetermined number of candidate nodes having a highest weighted look up score; and
calculating, by the region slicing circuitry, an information divergence score for the predetermined number of candidate nodes having the highest weighted look up score, wherein the information divergence score represents a divergence of a discovery candidate node from the selected node;
calculating, by a computation circuitry, a discovery score for at least one candidate node within the predetermined number of candidate nodes having the highest weighted look up score; and
selecting, by the computation circuitry, a discovery node from the at least one candidate node according to a respective discovery score.
11. The method of claim 10 , wherein identifying, by the region slicing circuitry, the selected node within the embedding space is based on at least part of a received user query.
12. The method of claim 10 , wherein determining the region of interest comprises including nodes within a predetermined vector distance from the selected node, wherein the predetermined vector distance is included in a received user query.
13. The method of claim 10 , wherein calculating, by the computation circuitry, the discovery score comprises calculating a weighted sum of the similarity score, the weighted look up score, and the information divergence score for a respective candidate node.
14. The method of claim 10 , wherein calculating the information divergence score comprises:
determining a first link prediction probability for each node linked to the selected node based on at least part of a received user query;
determining a second link prediction probability for each node linked to the nodes within the region of interest based on at least part of the received user query; and
determining the information divergence score based on the first link prediction probability and the second link prediction probability.
15. The method of claim 10 , wherein selecting, by the computation circuitry, the discovery node comprises selecting a candidate node having a highest substitution score as the discovery node.
16. A product comprising:
a machine-readable medium, other than a transitory signal; and
instructions stored on the machine-readable medium, the instructions configured to, when executed, cause processing circuitry to:
receive an original knowledge graph including a set of structured data;
convert the original knowledge graph to an embedding space comprising nodes by converting the set of structured data into sets of vector triples and modelling the vector triples with a neural network architecture to learn the representations of the knowledge graph;
identify a selected node within the embedding space for substitution;
determine a region of interest within the embedding space containing the selected node, and calculating a similarity score for each node within the region of interest wherein the similarity score represents a similarity between each node within the region of interest and the selected node;
calculate a weighted look up score for nodes within the region of interest, and selecting a predetermined number of candidate nodes having a highest weighted look up score; and
calculate an information divergence score for the predetermined number of candidate nodes having the highest weighted look up score, wherein the information divergence score represents a divergence of a discovery candidate node from the selected node;
calculate a discovery score for at least one candidate node within the predetermined number of candidate nodes having the highest weighted look up score; and
select a discovery node from the at least one candidate node according to a respective discovery score.
17. The product of claim 16 , wherein the instructions, when executed, cause the processing circuitry to:
identify the selected node within the embedding space based on at least part of a received user query.
18. The product of claim 16 , wherein the instructions, when executed, cause the processing circuitry to calculate the discovery score by a weighted sum of the similarity score, the weighted look up score, and the information divergence score for a respective substitute candidate node.
19. The product of claim 16 , wherein the instructions, when executed, cause the processing circuitry to calculate the information divergence score by:
determining a first link prediction probability for each node linked to the selected node based on at least part of a received user query;
determining a second link prediction probability for each node linked to the nodes within the region of interest based on at least part of the received user query; and
determining the information divergence score based on the first link prediction probability and the second link prediction probability.