Weight generation in machine learning

• US 9,858,534 B2
• Filed: 08/05/2014
• Issued: 01/02/2018
• Est. Priority Date: 11/22/2013
• Status: Active Grant

• Current Assignee: California Institute of Technology
• Original Assignee: California Institute of Technology
• Inventors: Abu-Mostafa, Yaser Said, Gonzalez, Carlos Roberto
• Primary Examiner(s): SITIRICHE, LUIS A
• Application Number:
  US14/451,899
• Publication Number:
  US 20150206067A1
• Time in Patent Office: 1,246 Days
• Field of Search: None
• CPC Class Codes: G06F 18/211 Selection of the most significant subset of features
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

Technologies are generally described for systems, devices and methods relating to a machine learning environment. In some examples, a processor may identify a training distribution of a training data. The processor may identify information about a test distribution of a test data. The processor may identify a coordinate of the training data and the test data. The processor may determine, for the coordinate, differences between the test distribution and the training distribution. The processor may determine weights based on the differences. The weights may be adapted to cause the training distribution to conform to the test distribution when the weights are applied to the training distribution.

Claims

1. A method to improve predictive capability of a machine learning system, the method comprising:

receiving, by a computer, training data that includes one or more points;

identifying, by the computer, a training distribution of the one or more points of the training data;

receiving, by the computer, test data that includes one or more points;

identifying, by the computer, information about a test distribution of the one or more points of the test data;

identifying, by the computer, one or more coordinates for the one or more points of the training data and the one or more points of the test data;

determining, for each identified coordinate and by the computer differences between the one or more points of the test data and the one or more points of the training data;

determining, by the computer, weights for the one or more points of the training data based on the determined differences, wherein the weights are adapted to cause the training distribution to conform to the test distribution in response to the weights being applied to the training distribution;

generating, by the computer, a weighted function based on the determined weights and the training data; and

generating, by the computer, a first output based on an application of an input to the generated weighted function, wherein the first output is different than a second output generated by an application of the input to a non-weighted function, wherein the first output and the second output respectively correspond to a first predictive capability and a second predictive capability of the machine learning system, and wherein the first predictive capability is greater than the second predictive capability.

2. The method of claim 1, wherein generating the first output includes generating at least one of a recommendation, a classification, a prediction, and a determination.

3. The method of claim 1, wherein:

the training data is generated at a first instance in time; and

the test data is generated at a second instance in time, wherein the second instance in time is later than the first instance in time.

4. The method of claim 1, wherein determining the weights comprises:

iteratively determining, for each identified coordinate, differences between the one or more points of the training data and the one or more points of the test data, wherein the weights for the one or more points of the training data are determined based on a convergent value of the differences between the one or more points of the training data and the one or more points of the test data.

5. The method of claim 1, wherein identifying the one or more coordinates includes identifying a range of values in a coordinate space, and wherein the method further comprises:

dividing the range of values in the coordinate space into bins,

wherein determining the weights is based on a number of the one or more points in the training data and a number of the bins.

6. The method of claim 1, wherein the one or more points of the test data and the one or more points of the training data include at least one first point and at least one second point, respectively, wherein the one or more coordinates include a range of values in a coordinate space, and wherein the method further comprises:

dividing the range of values in the coordinate space into bins,

wherein determining the weights is based on a number of the at least one first point and a number of the at least one second point which are located in the bins.

7. A method to improve predictive capability of a machine learning system, the method comprising, by a computer:

identifying first points of training data;

identifying information about test data, wherein the test data includes second points;

identifying a coordinate of the first points and the second points, wherein the coordinate includes a range of values in a coordinate space;

dividing the range of values in the coordinate space into bins, wherein the bins define subsets of the range of values;

determining a first frequency, wherein the first frequency relates to a first percentage of the first points being located within a particular bin;

determining a second frequency, wherein the second frequency relates to a second percentage of the second points being located within the particular bin;

comparing the first frequency and the second frequency;

determining a weight for the training data, based at least, in part, on the comparison of the first frequency and the second frequency, and on a number of the bins;

generating a weighted function based on the determined weight and the training data; and

generating a first output based on an application of an input to the generated weighted function, wherein the first output is different than a second output generated by an application of the input to a non-weighted function, wherein the first output and the second output respectively correspond to a first predictive capability and a second predictive capability of the machine learning system, and wherein the first predictive capability is greater than the second predictive capability.

8. The method of claim 7, wherein:

the first points follow a training distribution,

the second points follow a test distribution, and

the weight is effective to conform a particular point in the training distribution to a particular point in the test distribution.

9. The method of claim 7, wherein comparing the first frequency and the second frequency includes:

identifying a first comparison value;

comparing frequency values of the test data and the training data in the bins to produce a difference value;

updating the first comparison value to produce a second comparison value based on the difference value; and

iteratively repeating the identifying the first comparison value, comparing frequency values of the test data and the training data in the bins to produce the difference value, and updating the first comparison value to produce the second comparison value based on the difference value, until the second comparison value converges to a convergent value.

10. The method of claim 9, wherein updating the first comparison value to produce the second comparison value based on the difference value comprises:

adding a fraction of the difference value to the first comparison value to produce the second comparison value.

11. The method of claim 7, wherein determining the weight for the training data is based on a number of the first points.

12. The method of claim 7, wherein determining the weight for the training data is based on a number of the first points and a number of the second points which are located in the bins.

13. A computing device, comprising:

a first processor;

a second processor; and

a memory configured to be in communication with the first processor and the second processor, the memory effective to store training data and test data, wherein the training data comprises first points and the test data comprises second points, and wherein;

the first processor is effective to;

identify a coordinate of the first points and the second points, wherein the coordinate includes a range of values in a coordinate space;

divide the range of values in the coordinate space into bins, wherein the bins define subsets of the range of values;

determine a first frequency, wherein the first frequency relates to a first percentage of the first points being located within a particular bin;

determine a second frequency, wherein the second frequency relates to a second percentage of the second points being located within the particular bin;

compare the first frequency and the second frequency; and

determine a weight for the training data, based at least, in part, on the comparison of the first frequency and the second frequency, and on a number of bins,

the second processor is effective to;

generate a weighted function based on the determined weight and the training data; and

generate a first output based on an application of an input to the generated weighted function, wherein the first output is different than a second output generated by an application of the input to a non-weighted function, wherein the first output and the second output respectively correspond to a first predictive capability and a second predictive capability of the computing device, and wherein the first predictive capability is greater than the second predictive capability, and

the memory is further effective to store the determined weight.

14. The computing device of claim 13, wherein:

the first points follow a training distribution,

the second points follow a test distribution, and

the weight is effective to conform a particular point in the training distribution to a particular point in the test distribution.

15. The computing device of claim 13, wherein the first processor is further effective to:

identify a first comparison value;

compare frequency values of the test data and the training data in the bins to produce a difference value;

update the first comparison value to produce a second comparison value based on the difference value;

iteratively repeat the identification of the first comparison value, the comparison of frequency values of the test data and the training data in the bins to produce the difference value, and the update of the first comparison value to produce the second comparison value based on the difference value, until the second comparison value converges to a convergent value; and

store the converged second comparison value in the memory.

16. The computing device of claim 15, wherein the first processor is further effective to update the first comparison value to produce the second comparison value based on the difference value, by addition of a fraction of the difference value to the first comparison value.

17. The computing device of claim 13, wherein the second processor is further effective to:

store the weighted function in the memory.

18. The computing device of claim 13, wherein the first processor is effective to determine the weight for the training data based on a number of the first points.

19. The computing device of claim 13, wherein the processor is effective to determine the weight for the training data based on a number of the first points and a number of the second points which are located in the bins.

20. A computer-implemented method to improve predictive capability of a machine learning system, the method comprising:

receiving, by a weight generation module of the machine learning system, training data that includes one or more training points;

identifying, by a processor of the machine learning system, a training distribution of the one or more training points of the training data;

retrieving, by the weight generation module of the machine learning system, test data from a memory of the machine learning system, wherein the test data is different from the training data, and wherein the test data includes one or more test points;

identifying, by the processor of the machine learning system, information about a test distribution of the one or more test points of the test data;

identifying, by the weight generation module of the machine learning system, one or more coordinates for the one or more training points of the training data and the one or more test points of the test data;

determining, based on the identified information and by the weight generation module of the machine learning system and for each identified coordinate, differences between the one or more test points of the test data and the one or more training points of the training data;

determining, by the weight generation module of the machine learning system and based on the determined differences, weights for the one or more points of the training data, wherein;

the weights are adapted to cause the training distribution to conform to the test distribution in response to the weights being applied to the training distribution,

the training data includes a number of points, and

each coordinate includes a range of values in a coordinate space;

dividing the range of values in the coordinate space into bins;

calculating a frequency of each bin based on a number of points in each bin and a total number of points included in the training data, wherein determining the weights is based on the calculated frequency of each bin and a number of the bins;

transmitting, by the weight generation module of the machine learning system, the determined weights and the training data to a machine learning module of the machine learning system;

producing, by the machine learning module of the machine learning system, a weighted function based on the determined weights and the training data, wherein the weighted function corresponds to a first predictive capability greater than a second predictive capability that corresponds to a function produced based on the training data; and

operating the machine learning system to use the weighted function to provide the first predictive capability.