Regression for metric dataset

Patent metadata

Field	Value
Publication number	US-11423325-B2
Application number	US-201715793214-A
Country	US
Kind code	B2
Filing date	Oct 25, 2017
Priority date	Oct 25, 2017
Publication date	Aug 23, 2022
Grant date	Aug 23, 2022

Abstract

A method, a system, and a computer program product for predicting an outcome expected for a particular positional value is provided. In the method, an input set of data records, each having a label and a positional value, and a target positional value are obtained. The label of each data record is one in a label set. A learning model that includes an output layer, an input layer corresponding to the label set and a network structure provided therebetween is read. In the learning model, the network structure has a plurality of functions trained so as to evaluate influence from each label in the label set depending on a relationship between the target positional value and a representative positional value associated with the label in the label set. A target outcome is estimated for the target positional value from the input set using the learning model.

First claim

What is claimed is:

1. A computer-implemented method for training a learning model to predict an outcome expected for a particular positional value, the method comprising:
obtaining an input set of data records and a target positional value, each data record having a label and a positional value, the label of each data record being one in a label set; and
training a logistic regression model including an output layer, an input layer corresponding to the label sets and a network structure provided therebetween, the network structure having a plurality of functions trained so as to evaluate influence from each label in the label set depending on a relationship between the target positional value and a representative positional value associated with the label in the label set, each of the plurality of functions including a weight function that takes the representative positional value, the target positional value, and a positional parameter as an input, and the output layer having an output function that receives a sum of outputs from the plurality of functions to estimate a target output.

2. The method of claim 1 , wherein the method further comprises:
generating an input vector for the logistic regression model from the input set, the input vector including a plurality of elements each having a value representing at least whether a corresponding label in the label set is observed in the input set or not, each element being associated with a positional value obtained from one or more data records having the corresponding label in the input set as the representative positional value.

3. The method of claim 1 , wherein each function is parameterized by a positional parameter and a weight parameter for a corresponding label in the label set, the positional parameter representing a range of influence from the corresponding label on a target outcome, the weight parameter representing a magnitude of the influence from the corresponding label on the target outcome.

4. The method of claim 3 , wherein each function is monotonic to the positional parameter.

5. The method of claim 3 , wherein the relationship is a difference or distance between the representative positional value and the target positional value, and each function is monotonic to the difference or the distance.

6. The method of claim 4 , wherein the polarity of monotonicity of each positional parameter does not depends on the corresponding weight parameter.

7. The method of claim 1 , wherein each positional value and the target positional value represent a time and a target time, respectively, each label represents an event, and a target outcome is estimated as a probability that a target event is observed at the target time.

8. The method of claim 1 , wherein each positional value and the target positional value represent a location and a target location, respectively, each label represents an object, and a target outcome is estimated as a probability that a target result is obtained at the target location.

9. The method of claim 1 , wherein each positional value and the target positional value represent a location and a target location, respectively, each label represents an object, and a target outcome is estimated as an estimated value of an evaluation item for the target location.

10. The method of claim 3 , wherein the method further comprises:
preparing a collection of training data, each training data including a set of data records each having a label and a positional value, a given positional value, and an answer given for the given positional value, the positional parameter and the weight parameter being trained by using the collection of the training data.

11. The method of claim 10 , wherein the method further comprises:
outputting the trained positional parameter as a range of the corresponding label to affect a target outcome.

12. The method of claim 10 , wherein the method further comprises:
estimating an outcome for the given positional value from the set of data records in one training data using the logistic regression model; and
updating the positional parameter and the weight parameter by comparing the answer given for the given positional value with the outcome estimated for the given positional value.

13. The method of claim 12 , wherein the weight parameter is updated under a regularization constraint having a strength different from the positional parameter.

14. The method of claim 1 , wherein the output function is an inverse function of a link function and the link function is a logit function.

15. The method of claim 1 , wherein the weight functions are expressed as:

k
n

⁡

(

s
n

)

=

w
n

1
+

e

t
*

-

s
n

-

a
n

,

where the k n , is the weight function, s n is the representative position, w n is a magnitude of influence from a label on the target outcome, t* is a target timestamp, and a n is a positional parameter.

16. The method of claim 1 , wherein the weight functions are expressed as:

k
n

⁡

(

s
n

)

=

w

n
+

1
+

e

t
*

-

s
n

-

a

n
+

Assignees

• IBM

Inventors

• Kabeya Yoshinori
• Takeuchi Emiko
• Takuma Daisuke
• Toyoshima Hirobumi

Classifications

• G06T2207/20081

  Training; Learning · CPC title

• G06N5/02

  Knowledge representation; Symbolic representation · CPC title

• G06F3/013

  Eye tracking input arrangements (G06F3/015 takes precedence) · CPC title

• G16H10/60 Primary

  for patient-specific data, e.g. for electronic patient records · CPC title

• G16H50/20

  for computer-aided diagnosis, e.g. based on medical expert systems · CPC title