Training a machine learning model
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a machine learning model. One of the methods includes receiving a complete set of training data; receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data; training the predictive model on the initial subset; storing data representing a first state of the predictive model after training the predictive model on the initial subset; receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data; and training the predictive model on the new subset.
1. A method performed by a computer-implemented node, the method comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing completed financial product transactions including transaction prices;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model predicting a next transaction price or a next transaction price direction for one or more financial products;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
2. The method of claim 1 , wherein the model training process is a limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
3. The method of claim 1 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
4. The method of claim 1 , further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
5. A non-transitory computer storage medium encoded with a computer program, the program comprising instructions that when executed by a computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing completed financial product transactions including transaction prices;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model predicting a next transaction price or a next transaction price direction for one or more financial products;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
6. The computer storage medium of claim 5 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
7. The computer storage medium of claim 5 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
8. The computer storage medium of claim 5 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
9. A system comprising:
a computer-implemented node storing instructions that when executed by the computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing completed financial product transactions including transaction prices;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model predicting a next transaction price or a next transaction price direction for one or more financial products;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
10. The system of claim 9 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
11. The system of claim 9 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
12. The system of claim 9 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
13. A system comprising:
a computer-implemented node storing instructions that when executed by the computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing completed credit card transactions or debit card transactions or both;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model classifying particular transactions as likely being anomalous or not;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
14. The system of claim 13 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
15. The system of claim 13 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
16. The system of claim 13 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
17. A non-transitory computer storage medium encoded with a computer program, the program comprising instructions that when executed by a computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing completed credit card transactions or debit card transactions or both;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model classifying particular transactions as likely being anomalous or not;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
18. The computer storage medium of claim 17 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
19. The computer storage medium of claim 17 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
20. The computer storage medium of claim 17 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
21. A method performed by a computer-implemented node, the method comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing completed credit card transactions or debit card transactions or both;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model classifying particular transactions as likely being anomalous or not;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
22. The method of claim 21 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
23. The method of claim 21 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
24. The method of claim 21 , further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
25. A system comprising:
a computer-implemented node storing instructions that when executed by the computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing purchase transactions representing purchases of products or services or both and including respective prices paid for the products or services or both;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is further a model classifying particular prices for particular products or services as likely being fraudulent or not;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
26. The system of claim 25 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
27. The system of claim 25 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
28. The system of claim 25 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
29. A non-transitory computer storage medium encoded with a computer program, the program comprising instructions that when executed by a computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing purchase transactions representing purchases of products or services or both and including respective prices paid for the products or services or both;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is further a model classifying particular prices for particular products or services as likely being fraudulent or not;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
30. The computer storage medium of claim 29 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
31. The computer storage medium of claim 29 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
32. The computer storage medium of claim 29 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
33. A method performed by a computer-implemented node, the method comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing purchase transactions representing purchases of products or services or both and including respective prices paid for the products or services or both;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is further a model classifying particular prices for particular products or services as likely being fraudulent or not;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
34. The method of claim 33 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
35. The method of claim 33 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
36. The method of claim 33 , further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
37. A system comprising:
a computer-implemented node storing instructions that when executed by the computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing user actions on an interactive computer-based system;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model predicting user actions on the interactive computer-based system;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
38. The system of claim 37 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
39. The system of claim 37 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
40. The system of claim 37 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
41. A non-transitory computer storage medium encoded with a computer program, the program comprising instructions that when executed by a computer-implemented node cause the computer-implemented node to perform operations comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing user actions on an interactive computer-based system;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model predicting user actions on the interactive computer-based system;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
42. The computer storage medium of claim 41 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
43. The computer storage medium of claim 41 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
44. The computer storage medium of claim 41 , the operations further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.
45. A method performed by a computer-implemented node, the method comprising:
receiving a complete set of training data, wherein the complete set of training data comprises a plurality of feature vectors, and wherein the plurality of feature vectors comprises vectors representing user actions on an interactive computer-based system;
receiving instructions to train a predictive model having a plurality of parameters on an initial subset of the complete set of training data, wherein the predictive model is a model predicting user actions on the interactive computer-based system;
training the predictive model on the initial subset using a model training process to generate a respective first parameter value for each of the model parameters;
storing data representing a first state of the predictive model after training the predictive model on the initial subset, wherein the first state includes the first parameter values and components of the model training process used to generate the first parameter values;
receiving updated parameter values and instructions to train the predictive model on a new subset of the complete set of training data, wherein the new subset contains feature vectors from the initial subset and new feature vectors from the complete set of training data that have not been used to train the predictive model; and
training the predictive model on the new subset beginning from the first state of the predictive model with the updated parameter values in place of the first parameter values using the model training process to generate a respective second parameter value for each of the model parameters.
46. The method of claim 45 , wherein the model training process is a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) process, and wherein the first state of the predictive model comprises the first parameter values and L-BFGS components used to generate the first parameter values during a last iteration of the L-BFGS process performed on the initial subset.
47. The method of claim 45 , wherein training the predictive model on the new subset comprises:
updating the first state of the model by replacing the first parameter values with the updated parameter values; and
using the updated state as an initial state for a first iteration of the model training process performed on the new subset.
48. The method of claim 45 , further comprising:
operating under control of a controller, including receiving the instructions to train the predictive model on the initial subset and receiving updated parameter values and instructions to train the predictive model on the new subset;
providing the first parameter values to the controller; and
after training the predictive model on the new subset, providing the second parameter values to the controller.