Use Disclaimer, Please Read

Disclaimer

This is a stable release of a work-in-progress under active development. We make no guarantees about the quality of this software. To the best of our knowledge, we have followed the machine learning “best practices” when developing this software, but if you know better than us, please let us know! File any and all issues at GitHub. Please read all of the documentation carefully before using this package. In some cases, you need to know about the implicit assumptions that make for a smooth user experience. In other cases, you need to know about the key terminology that could prevent gross negligence. Otherwise, happy learning!

Warnings against improper use

• plGrid, plGridMulti, plMonteCarlo, plNested: For a high-throughput classification pipeline, if you supply an \(x\) number of top features to the top argument greater than the number of total number of features available in a training set, exprso will automatically use all features instead.
• pipeFilter, buildEnsemble: For an ExprsPipeline model extraction, if you supply an \(x\) number of top models to the top argument greater than the total number of models available in a filtered cut of models, exprso will automatically use all models instead. If you are concerned about this default behavior, call pipeFilter first, then call buildEnsemble on the pipeFilter results after inspecting them manually.
• plCV: This internal function calculates a simple metric of cross-validation during high-throughput classification. When the function receives data that have already undergone feature selection (which is in most cases), plCV provides an invalid and overly-optimistic metric of classifier performance. However, the results of plCV appear to have at least some relative validity. Therefore, it is reasonable to let the results posted by plCV guide your choice of classifier parameters. However, you must never report any performance statistic derived from a validation set that contains subjects who have undergone feature selection with the training set.
• splitSample: The splitSample method builds the training and validation sets by randomly sampling all subjects in an ExprsArray object. However, splitSample is not truly random; it iteratively samples until at least one of every class appears in the test set. This rule makes it easier to run analyses and interpret results, but requires caution when articulating in a report how you chose the test set.

Known issues

• fsMrmre: This feature selection method will crash when used with too many (> 46340) features. Unfortunately, this has to do with the implementation of the mRMRe package and we can do nothing about it.
• buildDNN: This classification method will exhaust RAM unless you manually clear old models with h2o::h2o.shutdown(). The ExprsModel objects for deep learning store links to back-end classifiers and do not contain the classifiers themselves. These back-end classifiers will continue to exist long after the R object has been destroyed unless manually cleared. Unfortunately, this has to do with the implementation of the h2o package and we can do nothing about it.
• buildRF: This classification method will crash sometimes when working with very small or unbalanced datasets within a large high-throughput classification pipeline. We are actively working on finding a solution for this problem.