IP Library Granted Patent US 11,080,618
Granted Patent B2
US 11,080,618 · App. 15/806,161 · Granted Aug 3, 2021

Augmenting incomplete training datasets for use in a machine learning system

Inventors: Brandon Ray Kam (Los Altos, CA); Viraj Navkal (San Bruno, CA); Grant Schneider (Foster City, CA); Paul Gu (Redwood City, CA); Alec M. Zimmer (San Francisco, CA)
Assignee: Upstart Network, Inc.
G06N20/00G06N7/005
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 11,080,618
App. No.
15/806,161
Granted
Aug 3, 2021
Kind
B2
Abstract

Systems and methods for augmenting incomplete training dataset for use in a machine learning system are described herein. In an embodiment, a server computer receives a plurality of input training datasets including one or more incomplete input training datasets and one or more complete datasets which contain one or more failure training datasets, the incomplete input training datasets comprising a plurality of parameters. Using the one or more failure training datasets, the server computer generates temporal failure data describing a likelihood of failure of an item as a function of time. Using the one or more complete training datasets, the server computer generates parameter specific likelihoods of failure of an item. The server computer augments the one or more incomplete input training datasets using the temporal failure data and/or the parameter specific likelihoods of failure to create one or more augmented training datasets. The server computer uses the one or more augmented training datasets as input for training a machine learning model that is programmed to generate a probability of failure of a particular item represented by an input dataset.

Claims (60)

1. A data processing system comprising:

one or more processors;

a memory coupled to the one or more processors and storing instructions which, when executed by the one or more processors, cause performance of:

receiving a plurality of input training datasets including one or more incomplete input training datasets and one or more complete training datasets which comprise one or more success training datasets and one or more failure training datasets, the one or more complete training datasets and the one or more incomplete input training datasets comprising a plurality of parameters;

augmenting the one or more incomplete input training datasets to create one or more augmented training datasets, wherein augmenting the one or more incomplete input training datasets comprises:

generating, using the one or more success training datasets and failure training datasets, a temporal failure curve that describes a likelihood of success or failure of an item as a function of time;

generating temporal failure data by extracting, from the temporal failure curve, a value describing the likelihood of failure of the item at a particular time; and

using the temporal failure data to augment the incomplete input training datasets;

training a machine learning model using the one or more augmented training datasets as an input; and

generating, using the trained machine learning model, a probability of success or failure of a particular item represented by an input dataset.

2. The system of claim 1 , wherein generating the temporal failure data curve that describes a likelihood of success or failure of an item as a function of time comprises:

identifying, in each of the failure training datasets, temporal data indicating when a failure occurred for the item; and

using the temporal data, generating a cumulative distribution function that describes the likelihood of success or failure of the item over time.

3. The system of claim 1 , wherein the instructions, when executed by the one or more processors, further cause performance of:

receiving the input dataset comprising a second plurality of parameters relating to the particular item;

using the second plurality of parameters as inputs in the machine learning model to compute the probability of success or failure for the input dataset.

4. The system of claim 1 , wherein the machine learning model comprises a set of program instructions that implement a logistic regression and/or a gradient boosted decision tree.

5. The system of claim 1 , wherein augmenting the one or more incomplete input training datasets further comprises generating parameter specific likelihoods of success or failure of an item and using the parameter specific likelihoods of success or failure to augment the incomplete input datasets.

6. The system of claim 5 wherein the generated parameter specific likelihoods of success or failure of an item are computed for each of a plurality of values of a particular parameter, wherein the particular parameter relates to a status of the incomplete training dataset at a computation time.

7. The system of claim 1 , wherein the one or more incomplete training datasets comprise data regarding loan transactions that have neither been fully paid nor defaulted upon.

8. The system of claim 7 , wherein the plurality of parameters includes a current status of an associated loan transaction, a number of days since closing the associated loan transaction, and one or more parameters related to a recipient of the associated loan transaction.

9. A method comprising, by a computing system:

receiving a plurality of input training datasets including one or more incomplete input training datasets and one or more complete training datasets which comprise one or more success training datasets and one or more failure training datasets, the one or more complete training datasets and the one or more incomplete input training datasets comprising a plurality of parameters;

augmenting the one or more incomplete input training datasets to create one or more augmented training datasets, wherein augmenting the one or more incomplete input training datasets comprises:

generating, using the one or more success training datasets and failure training datasets, a temporal failure curve that describes a likelihood of success or failure of an item as a function of time;

generating temporal failure data by extracting, from the temporal failure curve, a value describing the likelihood of failure of the item at a particular time; and

using the temporal failure data to augment the incomplete input training datasets;

training a machine learning model using the one or more augmented training datasets as an input; and

generating, using the trained machine learning model, a probability of success or failure of a particular item represented by an input dataset.

10. The method of claim 9 , wherein generating the temporal failure data curve that describes a likelihood of success or failure of an item as a function of time comprises:

identifying, in each of the failure training datasets, temporal data indicating when a failure occurred for the item; and

using the temporal data, generating a cumulative distribution function that describes the likelihood of success or failure of the item over time.

11. The method of claim 9 , further comprising:

receiving the input dataset comprising a second plurality of parameters relating to the particular item;

using the second plurality of parameters as inputs in the machine learning model to compute the probability of success or failure for the input dataset.

12. The method of claim 9 , wherein the machine learning model comprises a set of program instructions that implement a logistic regression and/or a gradient boosted decision tree.

13. The method of claim 9 , wherein augmenting the one or more incomplete input training datasets further comprises generating parameter specific likelihoods of success or failure of an item and using the parameter specific likelihoods of success or failure to augment the incomplete input datasets.

14. The method of claim 13 wherein the generated parameter specific likelihoods of success or failure of an item are computed for each of a plurality of values of a particular parameter, wherein the particular parameter relates to a status of the incomplete training dataset at a computation time.

15. The method of claim 9 , wherein incomplete training datasets comprise data regarding loan transactions that have neither been fully paid nor defaulted upon.

16. The method of claim 15 , wherein the plurality of parameters includes a current status of an associated loan transaction, a number of days since closing the associated loan transaction, and one or more parameters related to the recipient of the associated loan transaction.

17. One or more non-transitory computer-readable media storing instructions which, when executed by one or more processors, cause:

receiving a plurality of input training datasets including one or more incomplete input training datasets and one or more complete training datasets which comprise one or more success training datasets and one or more failure training datasets, the one or more complete training datasets and the one or more incomplete input training datasets comprising a plurality of parameters;

augmenting the one or more incomplete input training datasets to create one or more augmented training datasets, wherein augmenting the one or more incomplete input training datasets comprises:

generating, using the one or more success training datasets and failure training datasets, a temporal failure curve that describes a likelihood of success or failure of an item as a function of time;

generating temporal failure data by extracting, from the temporal failure curve, a value describing the likelihood of failure of the item at a particular time; and

using the temporal failure data to augment the incomplete input training datasets;

training a machine learning model using the one or more augmented training datasets as an input; and

generating, using the trained machine learning model, a probability of success or failure of a particular item represented by an input dataset.

18. The one or more non-transitory computer-readable media of claim 17 , wherein generating the temporal failure data curve that describes a likelihood of success or failure of an item as a function of time comprises:

identifying, in each of the failure training datasets, temporal data indicating when a failure occurred for the item; and

using the temporal data, generating a cumulative distribution function that describes the likelihood of success or failure of the item over time.

19. The one or more non-transitory computer-readable media of claim 17 , wherein the instructions, when executed by the one or more processors, further cause performance of:

receiving the input dataset comprising a second plurality of parameters relating to the particular item;

using the second plurality of parameters as inputs in the machine learning model to compute the probability of success or failure for the input dataset.

20. The one or more non-transitory computer-readable media of claim 17 , wherein the machine learning model comprises a set of program instructions that implement a logistic regression and a gradient boosted decision tree.

21. The one or more non-transitory computer-readable media of claim 17 , wherein the instructions, when executed by the one or more processors, further cause performance of:

generating parameter specific likelihoods of success or failure of an item and using the parameter specific likelihoods of success or failure to augment the incomplete input datasets.

22. The one or more non-transitory computer-readable media of claim 21 wherein the generated parameter specific likelihoods of success or failure of an item are computed for each of a plurality of values of a particular parameter, wherein the particular parameter relates to a status of the incomplete training dataset at a computation time.

23. The one or more non-transitory computer-readable media of claim 17 , wherein incomplete training datasets comprise data regarding loan transactions that have neither been fully paid nor defaulted upon.

24. The one or more non-transitory computer-readable media of claim 23 , wherein the plurality of parameters includes a current status of an associated loan transaction, a number of days since closing the associated loan transaction, and one or more parameters related to the recipient of the associated loan transaction.

Assignments (7)
INTELLECTUAL PROPERTY SECURITY AGREEMENT TERMINATION AT REEL/FRAME NO. 52918/0923 Recorded Jun 2, 2021
From: SILICON VALLEY BANK
To: UPSTART NETWORK, INC.
Reel/Frame 056461/0869 →
INTELLECTUAL PROPERTY SECURITY AGREEMENT TERMINATION AT REEL/FRAME NO. 53117/0334 Recorded Jun 2, 2021
From: SILICON VALLEY BANK
To: UPSTART NETWORK, INC.
Reel/Frame 056461/0888 →
INTELLECTUAL PROPERTY SECURITY AGREEMENT TERMINATION AT REEL/FRAME NO. 53117/0360 Recorded Jun 2, 2021
From: SILICON VALLEY BANK
To: UPSTART NETWORK, INC.
Reel/Frame 056461/0861 →
SECURITY INTEREST - MEZZ LOAN Recorded Jul 1, 2020
From: UPSTART NETWORK, INC.
To: SILICON VALLEY BANK
Reel/Frame 053117/0360 →
SECURITY INTEREST - SR. LOAN Recorded Jul 1, 2020
From: UPSTART NETWORK, INC.
To: SILICON VALLEY BANK
Reel/Frame 053117/0334 →
SECURITY INTEREST Recorded Jun 11, 2020
From: UPSTART NETWORK, INC.
To: SILICON VALLEY BANK
Reel/Frame 052918/0923 →
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Nov 7, 2017
From: KAM, BRANDON RAY; NAVKAL, VIRAJ; SCHNEIDER, GRANT; GU, PAUL; ZIMMER, ALEC M.
To: UPSTART NETWORK, INC.
Reel/Frame 044058/0281 →
Continuity (1)
Related Publication 20190138941A1 · May 9, 2019
Cited By (1)
US 12,524,439