Utils¶
This module contains methods needed for the construction of the constrained objective space and to analyze and interpret the results obtained by BFS.
Methods¶
circle_points
circle_points(r,n)
This function, first presented by [Lin2019], is used to divide the objective space into k regions. The idea is to create regions that have the same angle. Since we normalize our objective functions before applying BFS, we can use a balanced space where all regions are equal.
Parameters
- r: int
Radius of the circle to be created. To get a space between 0 and 1, r should be equal to 1.
- n: int
Number of preference vectors to be generated. n should be 1 value higher than the number of regions desired. For instance, if we want a space split into 10 regions, n should be equal to 11.
Returns
List of the preference vectors’ cootdinates.
s_score
s_score(network_weights, sparsity_level, bfs_bool)
This function defines a sparsity metric used to measure the number of weights that are null or almost null in a network.
Parameters
- network_weights: list
List of a network’s weights (weights of all layers, biases and other initialized parameters).
- sparsity_level: float
Threshold value below which a connection’s weight is considered equal to 0.
- bfs_bool: boolean
Boolean to indicate whether the method from which the weights were extracted is BFS. If we are working with BFS results, then bfs_bool should be set to True.
Returns
Float indicating the proprotion of weights that were pruned. This value will always be between 0 and 1.
average_fs_score
average_fs_score(network_weights, features_name, relevant_features)
This method defines a feature selection metric that can only be used when the important feature in a dataset are known. Its value can be between 0 and 1.
Parameters
- network_weights: list
List of a network’s weights (weights of all layers, biases and other initialized parameters).
- features_name: list
List of the name of all features.
- relevant_features: list
List of the name of each relevant feature.
Returns
A float that represents the feature selection score after a model has been training.
individual_fs_score
individual_fs_score(network_weights, feature_name)
This method is used for feature ranking. It can be used whether the dataset has already known relevant features or non, as it will return the features’ contributions’ to prediction task sorted in decreasing order from highest to lowest.
Parameters
- network_weights: list
List of a network’s weights (weights of all layers, biases and other initialized parameters).
- feature_name: list
List of the name of all features.
Returns
Dataframe containing all features with their feature importance value ranked from highest to lowest.
Lin, X., Zhen, H. L., Li, Z., Zhang, Q. F., & Kwong, S. (2019). Pareto multi-task learning. Advances in neural information processing systems, 32.