Identifying optimal weights to improve prediction accuracy in machine learning techniques

1 . A computer-implemented method for improving prediction accuracy in machine learning techniques, the computer-implemented method comprising:

training a candidate student model using training data and weights generated by a teacher model;

comparing a current student model with the candidate student model using the training data and testing data to determine which is better at predicting an observed target;

determining whether the candidate student model is better at predicting the observed target than the current student model based on how close predictions generated by the candidate student model are to the observed target; and

in response to determining that the candidate student model is better at predicting the observed target than the current student model, updating the current student model with the candidate student model and updating current weights with the weights.

2 . The computer-implemented method of claim 1 , further comprising:

training the current student model using the training data and a current weight for each data case generated by the teacher model;

generating a reward based on the comparing the current student model with the candidate student model; and

updating the teacher model with the reward in response to a stopping rule not being satisfied.

3 . The computer-implemented method of claim 2 , further comprising:

prior to the updating the teacher model, generating the weights by the teacher model using state features; and

subsequent to the updating the teacher model, generating a set of weights by the updated teacher model using the state features in response to determining that the candidate student model is not better at predicting the observed target than the current student model.

4 . The computer-implemented method of claim 3 , further comprising:

training the candidate student model using the training data and the set of weights.

5 . The computer-implemented method of claim 3 , wherein the state features comprise one or more of data features, case weights, student model features, or features representing a combination of data and the current student model.

6 . The computer-implemented method of claim 1 , further comprising:

returning the weights and the current student model to a user in response to a stopping rule being satisfied, wherein the returned current student model provides a prediction of the observed target, and wherein the stopping rule comprises one or more of: reaching a specified number of trials, reaching a specified training time, converging of a prediction accuracy, or a user-initiated termination.

7 . The computer-implemented method of claim 1 , wherein the weights are generated by the teacher model using reinforcement learning techniques.

8 . A computer program product for improving prediction accuracy in machine learning techniques, the computer program product comprising one or more computer readable storage mediums having program code embodied therewith, the program code comprising programming instructions for:

training a candidate student model using training data and weights generated by a teacher model;

comparing a current student model with the candidate student model using the training data and testing data to determine which is better at predicting an observed target;

determining whether the candidate student model is better at predicting the observed target than the current student model based on how close predictions generated by the candidate student model are to the observed target; and

in response to determining that the candidate student model is better at predicting the observed target than the current student model, updating the current student model with the candidate student model and updating current weights with the weights.

9 . The computer program product of claim 8 , wherein the programming instructions are further for:

training the current student model using the training data and a current weight for each data case generated by the teacher model;

generating a reward based on the comparing the current student model with the candidate student model; and

updating the teacher model with the reward in response to a stopping rule not being satisfied.

10 . The computer program product of claim 9 , wherein the programming instructions are further for:

prior to the updating the teacher model, generating the weights by the teacher model using state features; and

subsequent to the updating the teacher model, generating a set of weights by the updated teacher model using the state features in response to determining that the candidate student model is not better at predicting the observed target than the current student model.

11 . The computer program product of claim 10 , wherein the programming instructions are further for:

training the candidate student model using the training data and the set of weights.

12 . The computer program product of claim 10 , wherein the state features comprise one or more of data features, case weights, student model features, or features representing a combination of data and the current student model.

13 . The computer program product of claim 8 , wherein the programming instructions are further for:

returning the weights and the current student model to a user in response to a stopping rule being satisfied, wherein the returned current student model provides a prediction of the observed target, and wherein the stopping rule comprises one or more of the following: reaching a specified number of trials, reaching a specified training time, converging of a prediction accuracy, and a user-initiated termination.

14 . The computer program product of claim 8 , wherein the weights are generated by the teacher model using reinforcement learning techniques.

15 . A system, comprising:

a memory for storing a computer program for improving prediction accuracy in machine learning techniques; and

a processor connected to the memory, wherein the processor is configured to execute program instructions of the computer program comprising:

training a candidate student model using training data and weights generated by a teacher model;

comparing the current student model with the candidate student model using the training data and testing data to determine which is better at predicting an observed target;

determining whether the candidate student model is better at predicting the observed target than the current student model based on how close predictions generated by the candidate student model are to the observed target; and

in response to determining that the candidate student model is better at predicting the observed target than the current student model, updating the current student model with the candidate student model and updating current weights with the weights.

16 . The system as recited in claim 15 , wherein the program instructions of the computer program further comprise:

training the current student model using the training data and a current weight for each data case generated by the teacher model;

generating a reward based on the comparing the current student model with the candidate student model; and

updating the teacher model with the reward in response to a stopping rule not being satisfied.

17 . The system of claim 16 , wherein the program instructions of the computer program further comprise:

prior to the updating the teacher model, generating the weights by the teacher model using state features; and

subsequent to the updating the teacher model, generating a set of weights by the updated teacher model using the state features in response to determining that the candidate student model is not better at predicting the observed target than the current student model.

18 . The system of claim 17 , wherein the program instructions of the computer program further comprise:

training the candidate student model using the training data and the set of weights.

19 . The system of claim 15 , wherein the program instructions of the computer program further comprise:

returning the weights and the current student model to a user in response to a stopping rule being satisfied, wherein the returned current student model provides a prediction of the observed target, and wherein the stopping rule comprises one or more of: reaching a specified number of trials, reaching a specified training time, converging of a prediction accuracy, or a user-initiated termination.

20 . The system of claim 17 , wherein the state features comprise one or more of data features, case weights, student model features, or features representing a combination of data and the current student model.