Granted Patent B2

US 12,288,143 · App. 17/930,046 · Granted Apr 29, 2025

Inventors: Daniel Erenrich (Mountain View, CA); Matthew Elkherj (Menlo Park, CA)

Assignee: Palantir Technologies Inc.

G06N20/00G06N5/04

Abstract

Claims (53)

1. A system comprising:

one or more processors and a memory storing instructions that, when executed by the one or more processors, cause the system to:

obtain a dataset, the dataset including examples, each example of the examples including a pair of records comprising a first record from a first storage construct and a second record from a second storage construct;

selectively extract a subset of the examples from the dataset, wherein the selectively extracting is based on a degree of uncertainty of each pair, the degree of uncertainty indicating whether the first record matches the second record; and

for the subset of the examples:

request labels indicating whether each of the first records matches the corresponding second record;

receive the labels;

annotate the subset of data according to the labels;

determine that the first record from the first storage construct matches or potentially matches the second record from the second storage construct according to a first criteria;

determine that the first record from the first storage construct matches or potentially matches a third record from the second storage construct according to a second criteria;

and

generate a graph representing the first record matching or potentially matching the second record or representing the first record matching or potentially matching the third record, wherein the generating of the graph comprises generating record representations corresponding to the first record, the second record, the third record, and one or more other records, determining potential edges to indicate matches or potential matches among any of the first record, the second record, the third record, and the one or more other records, and pruning one or more of the determined potential edges according to one or more probabilities of the matches or potential matches.

2. The system of claim 1 , wherein the selectively extracting is based on degrees of diversity among the examples.

3. The system of claim 2 , wherein the degrees of diversity are indicative of

a first degree of difference between a first example and a second example; and

a second degree of difference between the second example and one or more remaining examples, of the examples.

4. The system of claim 1 , wherein the selectively extracting comprises extracting pairs of records having among highest degrees of uncertainty.

5. The system of claim 1 , wherein the selectively extracting comprises determining a size of the subset based on a size of available memory.

6. The system of claim 1 , wherein the selectively extracting comprises determining a size of the subset based on a processing speed of evaluating the subset and a time consumed during the labelling.

7. The system of claim 1 , wherein the selectively extracting comprises determining a size of the subset based on degrees of uncertainty within the subset.

8. The system of claim 7 , wherein the size of the subset is inversely correlated with the degrees of certainty within the subset.

9. A method comprising:

obtaining a dataset, the dataset including examples, each example of the examples including a pair of records comprising a first record from a first storage construct and a second record from a second storage construct;

selectively extracting a subset of the examples from the dataset, wherein the selectively extracting is based on a degree of uncertainty of each pair, the degree of uncertainty indicating whether the first record matches the second record; and

for the subset of the examples:

requesting labels indicating whether each of the first records matches the corresponding second record;

receiving the labels;

annotating the subset of data according to the labels;

determining that the first record from the first storage construct matches or potentially matches the second record from the second storage construct according to a first criteria;

determining that the first record from the first storage construct matches or potentially matches a third record from the second storage construct according to a second criteria;

associating the first record with the second record and associate the first record with the third record; and

generating a graph representing the first record matching or potentially matching the second record or representing the first record matching or potentially matching the third record, wherein the generating of the graph comprises generating record representations corresponding to the first record, the second record, the third record, and one or more other records, determining potential edges to indicate matches or potential matches among any of the first record, the second record, the third record, and the one or more other records, and pruning one or more of the determined potential edges according to one or more probabilities of the matches or potential matches.

10. The method of claim 9 , wherein the selectively extracting is based on degrees of diversity among the examples.

11. The method of claim 10 , wherein the degrees of diversity are indicative of:

a first degree of difference between a first example and a second example; and

a second degree of difference between the second example and one or more remaining examples, of the examples.

12. The method of claim 9 , wherein the selectively extracting comprises extracting pairs of records having among highest degrees of uncertainty.

13. The method of claim 9 , wherein the selectively extracting comprises determining a size of the subset based on a size of available memory.

14. The method of claim 9 , wherein the selectively extracting comprises determining a size of the subset based on a processing speed of evaluating the subset and a time consumed during the labelling.

15. The method of claim 9 , wherein the selectively extracting comprises determining a size of the subset based on degrees of uncertainty within the subset.

16. The method of claim 15 , wherein the size of the subset is inversely correlated with the degrees of certainty within the subset.

17. A non-transitory computer readable medium comprising instructions that, when executed, cause one or more processors to perform:

obtaining a dataset, the dataset including examples, each example of the examples including a pair of records comprising a first record from a first storage construct and a second record from a second storage construct;

selectively extracting a subset of the examples from the dataset, wherein the selectively extracting is based on a degree of uncertainty of each pair, the degree of uncertainty indicating whether the first record matches the second record; and

for the subset of the examples:

requesting labels indicating whether each of the first records matches the corresponding second record;

receiving the labels;

annotating the subset of data according to the labels;

determining that the first record from the first storage construct matches or potentially matches the second record from the second storage construct according to a first criteria;

determining that the first record from the first storage construct matches or potentially matches a third record from the second storage construct according to a second criteria;

associating the first record with the second record and associate the first record with the third record; and

generating a graph representing the first record matching or potentially matching the second record or representing the first record matching or potentially matching the third record, wherein the generating of the graph comprises generating record representations corresponding to the first record, the second record, the third record, and one or more other records, determining potential edges to indicate matches or potential matches among any of the first record, the second record, the third record, and the one or more other records, and pruning one or more of the determined potential edges according to one or more probabilities of the matches or potential matches.

18. The non-transitory computer readable medium of claim 17 , wherein the selectively extracting is based on degrees of diversity among the examples.

Assignments (1)

Continuity (4)

References Cited (236)