Systems and methods for robust feature selection
A device may generate an objective function for determining weights for potential features corresponding to training data. The objective function may be generated using a robust loss function such that the objective function is at least continuously twice differentiable. The objective function may comprise a neighborhood component analysis objective function that includes the robust loss function. The device may determine the weights for the potential features using the objective function. The determining may comprise optimizing a value of the objective function for each potential feature. The weights may represent predictive powers of corresponding potential features. The device may provide the weights for the potential features.
1. A computer-implemented method, comprising:
receiving a loss function, wherein the loss function is selected from the group consisting of a first loss function that is at least continuously twice differentiable and a second loss function that is not at least continuously twice differentiable;
generating an objective function for determining weights for potential features corresponding to training data,
the objective function being generated using the loss function such that the objective function is at least continuously twice differentiable,
the objective function comprising a neighborhood component analysis objective function that includes the loss function as a first component of a term of the neighborhood component analysis objective function, wherein:
the loss function is a function of a set of responses of the training data and not of the weights of the potential features; and
the term further comprises a second component that is a function of the weights of the potential features;
determining the weights for the potential features using the objective function,
the determining comprising optimizing a value of the objective function for each potential feature,
the weights representing predictive powers of corresponding potential features;
providing the weights for the potential features;
selecting, based on the weights for the potential features, a subset of the potential features of the training data as a set of selected features of the training data;
training, using the training data and the set of selected features, a model of a system to perform a prediction; and
providing the trained model.
2. The method of claim 1 , further comprising:
determining a training data set,
the training data set including a matrix of data relating to the potential features,
the training data set including the set of responses,
the subset of the potential features being related to the set of responses; and
where determining the weights for the potential features comprises:
determining the subset of the potential features that are related to the set of responses using neighborhood component analysis; and
selecting the subset of the potential features as the set of selected features.
3. The method of claim 1 , wherein training the model further comprises:
generating, based on the weights, a classification model or a regression model.
4. The method of claim 1 , where the loss function is one of:
a squared loss function,
a bisquare loss function,
an absolute value loss function,
an arbitrary loss function, or
a general loss function.
5. The method of claim 1 , further comprising:
receiving the second loss function that is not at least continuously twice differentiable to be the loss function.
6. The method of claim 1 , further comprising:
generating the objective function based on detecting an interaction with a user interface.
7. The method of claim 1 , further comprising:
providing, via a user interface, a representation relating to the weights for the potential features,
the representation including at least one of:
a graphical representation, or
a textual representation.
8. A non-transitory computer-readable medium storing instructions, the instructions comprising:
one or more instructions that, when executed by one or more processors, cause the one or more processors to:
receive information associated with a data set as input;
receive loss function, wherein the loss function is selected from the group consisting of a first loss function that is at least continuously twice differentiable and a second loss function that is not at least continuously twice differentiable;
generate an objective function for determining weights for features corresponding to the data set using the loss function,
the objective function being generated using the loss function such that the objective function is at least twice continuously differentiable,
the objective function comprising a neighborhood component analysis objective function that includes the loss function as a first component of a term of the neighborhood component analysis objective function, wherein:
the loss function is a function of a set of responses of the data set and not of the weights of the potential features; and
the term further comprises a second component that is a function of the weights of the features;
determine a set of weights for a set of potential features of the data set using the objective function;
provide information identifying the set of weights;
selecting, based on the set of weights for the set of potential features, a subset of the set of potential features of the data set as a set of selected features of the data set;
training, using the data set and the set of selected features, a model of a system to perform a prediction; and
providing the trained model.
9. The non-transitory computer-readable medium of claim 8 , where the one or more instructions, when executed by the one or more processors, cause the one or more processors to:
receive another input identifying at least one of:
a regularization parameter, λ, for the data set,
a fit method for the data set,
an optimizer for the data set, or
a set of user preferences relating to performing feature selection using neighborhood component analysis.
10. The non-transitory computer-readable medium of claim 8 , where the one or more instructions, when executed by the one or more processors, cause the one or more processors to:
determine, based on the data set, at least one of:
an N-by-P matrix of the set of features of the data set,
an N-by-1 response vector of the data set,
the response vector including the set of responses corresponding to the set of features of the data set.
11. The non-transitory computer-readable medium of claim 8 , where the one or more instructions cause the one or more processors to select the loss function, wherein the loss function comprises a least absolute deviations (L 1 ) loss or a ε-insensitive loss.
12. The non-transitory computer-readable medium of claim 8 , where the one or more instructions, when executed by the one or more processors, cause the one or more processors to:
partition the data set into a first set of portions to train the model and a second set of portions to validate the model; and
determine a regularization parameter, A, based on partitioning the data.
13. The non-transitory computer-readable medium of claim 8 , where the one or more instructions, that cause the one or more processors to generate the objective function, cause the one or more processors to:
determine a regularization parameter for the objective function based on a range of candidate regularization parameters.
14. A device, comprising:
one or more memories; and
one or more processors, communicatively associated with the one or more memories, to:
generate an object,
the object representing a set of weights for a set of potential features of a data set,
each potential feature, of the data set, being associated with a weight of the set of weights,
each weight corresponding to a predicted relationship between a corresponding potential feature, of the data set, and a response;
create the object using a feature selection technique comprising executing an objective function,
the feature selection technique being associated with a loss function, wherein the loss function is selected from the group consisting of a first loss function that is at least continuously twice differentiable and a second loss function that is not at least continuously twice differentiable,
the objective function being generated using the loss function such that the objective function is at least twice continuously differentiable,
the objective function comprising a neighborhood component analysis objective function that includes the loss function as a first component of a term of the neighborhood component analysis objective function, wherein:
the loss function is a function of a set of responses of the data set and not of the weights of the potential features; and
the term further comprises a second component that is a function of the set of weights of the set of potential features;
store the object;
selecting, based on the set of weights for the set of potential features of the object, a subset of the set of potential features of the data set as a set of selected features of the data set;
training, using the data set and the set of selected features, a model of a system to perform a prediction; and
providing the trained model.
15. The device of claim 14 , where the one or more processors are further configured to:
provide the object.
16. The device of claim 14 , where the one or more processors are further configured to:
perform a machine learning technique to train the model using the object.
17. The device of claim 16 , where the one or more processors are further configured to:
perform classification or regression using the model.
18. A computer-implemented method, comprising:
receiving a loss function, wherein the loss function is not at least continuously twice differentiable;
generating an objective function for determining weights for potential features corresponding to training data,
the objective function being generated using the loss function such that the objective function is at least continuously twice differentiable,
the objective function comprising a neighborhood component analysis objective function that includes the loss function as a first component of a term of the neighborhood component analysis objective function, wherein:
the loss function is a function of a set of responses of the training data and not of the weights of the potential features; and
the term further comprises a second component that is a function of the weights of the potential features;
determining the weights for the potential features using the objective function,
the determining comprising optimizing a value of the objective function for each potential feature,
the weights representing predictive powers of corresponding potential features; and
providing the weights for the potential features;
selecting, based on the weights for the potential features, a subset of the potential features of the training data as a set of selected features of the training data;
training, using the training data and the set of selected features, a model of a system to perform a prediction; and
providing the trained model.
19. The method of claim 18 , further comprising:
determining a training data set,
the training data set including a matrix of data relating to the potential features
the training data set including the set of responses,
the subset of the potential features being related to the set of responses; and
where determining the weights for the potential features comprises:
determining the subset of the potential features that are related to the set of responses using neighborhood component analysis; and
selecting the subset of the potential features as the set of selected features.
20. The method of claim 18 , wherein training the model further comprises:
generating, based on the weights, a classification model or a regression model.
21. The method of claim 18 , further comprising:
generating the objective function based on detecting an interaction with a user interface.
22. A non-transitory computer-readable medium storing instructions, the instructions comprising:
one or more instructions that, when executed by one or more processors, cause the one or more processors to perform:
receiving a loss function, wherein the loss function is not at least continuously twice differentiable;
generating an objective function for determining weights for potential features corresponding to training data,
the objective function being generated using the loss function such that the objective function is at least continuously twice differentiable,
the objective function comprising a neighborhood component analysis objective function that includes the loss function as a first component of a term of the neighborhood component analysis objective function, wherein:
the loss function is a function of a set of responses of the training data and not of the weights of the potential features; and
the term further comprises a second component that is a function of the weights of the potential features;
determining the weights for the potential features using the objective function,
the determining comprising optimizing a value of the objective function for each potential feature,
the weights representing predictive powers of corresponding potential features; and
providing the weights for the potential features;
selecting, based on the weights for the potential features, a subset of the potential features of the training data as a set of selected features of the training data;
training, using the training data and the set of selected features, a model of a system to perform a prediction; and
providing the trained model.