Parameters Tuning

This page contains parameters tuning guides for different scenarios.

For Faster Speed

Add More Computational Resources

On systems where it is available, OpenFE uses multiprocessing to parallelize the calculation and evaluation of new features. The maximum number of threads used by OpenFE is controlled by the parameter n_jobs. For best performance, set this to the number of real CPU cores available.

Increase n_data_blocks

This parameter is an integer that controls the number of data blocks the data is split into for successive feature-wise halving. The overall complexity of OpenFE is \(O(q2^{-q}\cdot mn^2)\), where \(m\) is the number of features, \(n\) is the number of samples, and \(2^q\) is the number of data blocks. Although increasing the value of n_data_blocks can speed up computation, doing so may have a negative impact on overall performance because important candidate features may be discarded during successive feature-wise halving. This is a trade-off between efficiency and effectiveness. Besides, setting n_data_blocks to a large value will result in few candidate features left for stage2 selection. This can be controlled by setting the parameter min_candidate_features to early-stop the successive feature-wise halving.

Perform Feature Selection Before Feature Generation

Another way to speed up is to perform feature selection and remove redundant features before feature generation. This can reduce the number of candidate features to speed up. However, this method of speeding up may also hurt the performance, because uninformative features may also yield informative candidate features after transformation. In a Diabetes dataset, for example, when the goal is to forecast if a patient will be readmitted to the hospital, the feature ‘patient id’ is useless. However, ‘freq(patient id)’, which is the number of times the patient has been admitted to the hospital, is a strong predictor of whether the patient would be readmitted. An alternative way to perform feature selection is to determine the candidate_features_list yourself by openfe.get_candidate_features().

In a word, there is no free way to accelerate. If there was, we would have been happy to incorporate it into our algorithm.

For Better Performance

Decrease n_data_blocks Or Increase min_candidate_features

This is particularly important if there are many useful candidate features in the pool of candidate features. However, changing these two parameters may not make a difference if there isn’t many useful candidate features.

Perform Feature Selection on Generated Features

In our examples, we directly include the top generated features recommended by OpenFE. Users can also perform more delicate feature selection methods (such as forward feature selection) to achieve better performance.

Set feature_boosting to True or False and See Which Provides Better Results

feature_boosting is an important parameter that determines whether to use feature boosting in OpenFE (see more details in our paper). In general, using feature boosting yields better results on most datasets. However, on some of the datasets, we find that disabling feature boosting can provide better results. We are planning to investigate this issue further in the future.