Methods and systems for quantifying uncertainty in neural link predictors for knowledge graphs

1 . A computing device for quantifying certainty for a prediction based on a knowledge graph, the computing device comprising:

a reception circuitry configured to receive a target triple and a knowledge graph comprising a set of structured data and a set of certainty scores for the structured data;

a knowledge graph embedding generation circuitry configured to convert the target triple to an embeddings space according to neighborhood sampling, wherein the embeddings space includes a set of point coordinates representing the set of structured data in the embeddings space;

a scoring circuitry configured to generate a plausibility prediction for the target triple using a scoring function;

a control circuitry configured to repeat the acts of the knowledge graph embedding generation circuitry and the scoring circuitry N times with dropouts to obtain N plausibility scores for the target triple, wherein N is an integer larger than one; and

an output circuitry configured to generate a predicted plausibility score and a certainty score for the target triple based on the N plausibility scores, and output the predicted plausibility score and the certainty score, wherein the output circuitry is configured to generate the predicted plausibility score and the certainty score for the target triple based on the N plausibility scores and output the predicted plausibility score and the certainty score is carried out by:

calculating a mean and a standard deviation based on the obtained N plausibility scores for the target triple;

determining the mean to represent the predicted plausibility score and the certainty score based on the standard deviation; and

outputting the predicted plausibility score and the certainty score for the target triple.

2 . The computing device according to claim 1 , wherein the knowledge graph embedding generation circuitry is configured to convert the target triple to the embeddings space according to the neighborhood sampling, by:

selecting K neighboring nodes of the target triple based on certainty scores of the neighboring nodes, wherein K is an integer larger than one;

generating an embedding vector for each of the K neighboring nodes by an encoder circuitry; and

aggregating the K embedding vectors to obtain an aggregated embedding vector for the target triple.

3 . The computing device according to claim 2 , wherein selecting the K neighboring nodes of the target triple based on certainty scores of the neighboring nodes, is carried out by:

sampling the neighboring nodes of the target triple according to a sampling with replacement algorithm to obtain the K neighboring nodes with highest certainty scores.

4 . The computing device according to claim 2 , wherein:

the encoder circuitry in the knowledge graph embedding generation circuitry comprises a neural network with dropouts after every weight layer.

5 . The computing device according to claim 2 , wherein aggregating the K embedding vectors to obtain the aggregated embedding vector for the target triple is carried out by summarizing, averaging or taking the weighted average of the corresponding point coordinate of the K embedding vectors to determine each point of the aggregate embedding vector.

6 . The computing device according to claim 1 , wherein:

the computer device further comprises, for each triple in a subset of the knowledge graph:

a corruption generation circuitry configured to generate synthetic negatives corresponding to the triple, and

a loss circuitry configured to calculate a loss score according to a loss function based on scoring the triple and the synthetic negatives; and

the control circuitry is further configured to train the knowledge graph embedding generation circuitry based on the loss score.

7 . The computing device according to claim 6 , wherein:

the scoring circuitry is configured to generate a positive score for the triple and to generate negative scores for the synthetic negatives using the scoring function; and

when the loss circuitry is configured to calculate the loss score, the loss circuitry is configured to calculate the loss score based on the positive score for the triple and the negative scores for the synthetic negatives.

8 . A method for quantifying certainty for a prediction based on a knowledge graph, the method comprising:

receiving, by a device comprising a memory storing instructions and a processing circuitry in communication with the memory, a target triple and a knowledge graph comprising a set of structured data and a set of certainty scores for the structured data;

converting, by the device, the target triple to an embeddings space according to neighborhood sampling by a neural network, wherein the embeddings space includes a set of point coordinates representing the set of structured data in the embeddings space;

generating, by the device, a plausibility prediction for the target triple using a scoring function;

repeating, by the device, converting the target triple to the embedding space and generating another plausibility prediction for the target triple N times with dropouts to obtain N plausibility scores for the target triple, wherein Nis an integer larger than one; and

generating, by the device, a predicted plausibility score and a certainty score for the target triple based on the N plausibility scores, and outputting for display the predicted plausibility score and the certainty score, wherein generating the predicted plausibility score and the certainty score for the target triple based on the N plausibility scores and outputting the predicted plausibility score and the certainty score comprises:

calculating a mean and a standard deviation based on the obtained N plausibility scores for the target triple;

determining the mean to represent the predicted plausibility score and the certainty score based on the standard deviation; and

outputting the predicted plausibility score and the certainty score for the target triple.

9 . The method according to claim 8 , wherein converting the target triple to the embeddings space according to the neighborhood sampling comprises:

selecting K neighboring nodes of the target triple based on certainty scores of the neighboring nodes, wherein K is an integer larger than one;

generating an embedding vector for each of the K neighboring nodes by an encoder; and

aggregating the K embedding vectors to obtain an aggregated embedding vector for the target triple.

10 . The method according to claim 9 , wherein selecting the K neighboring nodes of the target triple based on certainty scores of the neighboring nodes comprises:

sampling the neighboring nodes of the target triple according to a sampling with replacement algorithm to obtain the K neighboring nodes with highest certainty scores.

11 . The method according to claim 9 , wherein the encoder in the knowledge graph embedding generation circuitry comprises a neural network with dropouts after every weight layer.

12 . The method according to claim 9 , wherein aggregating the K embedding vectors to obtain the aggregated embedding vector for the target triple comprises:

summarizing, averaging or taking the weighted average of the corresponding point coordinate of the K embedding vectors to determine each point of the aggregate embedding vector.

13 . The method according to claim 8 , further comprising:

for each triple in a subset of the knowledge graph:

generating synthetic negatives corresponding to the triple, and

calculating a loss score according to a loss function based on scoring the triple and the synthetic negatives; and

training the neural network based on the loss score.

14 . The method according to claim 13 , further comprising:

generating a positive score for the triple and to generate negative scores for the synthetic negatives using the scoring function; and

wherein calculating the loss score comprises calculating the loss score based on the positive score for the triple and the negative scores for the synthetic negatives.

15 . A non-transitory computer-readable storage medium storing computer-readable instructions, wherein, the computer-readable instructions, when executed by a processing circuitry, are configured to cause the processing circuitry to perform:

receiving a target triple and a knowledge graph comprising a set of structured data and a set of certainty scores for the structured data;

converting the target triple to an embeddings space according to neighborhood sampling by a neural network, wherein the embeddings space includes a set of point coordinates representing the set of structured data in the embeddings space;

generating a plausibility prediction for the target triple using a scoring function;

repeating converting the target triple to the embedding space and generating another plausibility prediction for the target triple N times with dropouts to obtain N plausibility scores for the target triple, wherein N is an integer larger than one; and

generating a predicted plausibility score and a certainty score for the target triple based on the N plausibility scores, and outputting for display the predicted plausibility score and the certainty score, wherein generating the predicted plausibility score and the certainty score for the target triple based on the N plausibility scores and outputting the predicted plausibility score and the certainty score comprises:

calculating a mean and a standard deviation based on the obtained N plausibility scores for the target triple;

determining the mean to represent the predicted plausibility score and the certainty score based on the standard deviation; and

outputting the predicted plausibility score and the certainty score for the target triple.

16 . The non-transitory computer-readable storage medium according to claim 15 , wherein, when the computer-readable instructions are configured to cause the processor to perform converting the target triple to the embeddings space according to the neighborhood sampling, the computer-readable instructions are configured to cause the processing circuitry to perform:

selecting K neighboring nodes of the target triple based on certainty scores of the neighboring nodes, wherein K is an integer larger than one;

generating an embedding vector for each of the K neighboring nodes by an encoder; and

aggregating the K embedding vectors to obtain an aggregated embedding vector for the target triple.

17 . The non-transitory computer-readable storage medium according to claim 16 , wherein, when the computer-readable instructions are configured to cause the processor to perform selecting the K neighboring nodes of the target triple based on certainty scores of the neighboring nodes, the computer-readable instructions are configured to cause the processing circuitry to perform:

sampling the neighboring nodes of the target triple according to a sampling with replacement algorithm to obtain the K neighboring nodes with highest certainty scores.