Neural Network Data Preparation

Article MQL5 articles

Summary

This article surveys data preparation for neural network classifiers, focusing on filtering mislabeled training examples and reducing input dimensions. It compares data-level noise filtering with methods that try to make the classifier itself more robust, then illustrates use of a filter on raw, normalized, and transformed versions of a financial dataset. The examples report that the filter removes roughly 30% of training samples across those variants. Visual inspection and an information-value analysis suggest several predictors may be weak or irrelevant, but the author says model experiments are needed to confirm those judgments.

The dimensionality-reduction section introduces PCA, ICA, probabilistic PCA, autoencoders, and inverse nonlinear PCA, though much of its detail is omitted in the supplied text. It notes that the shown PPCA representation separates classes no better than an earlier autoencoder, while offering speed and simplicity. The article also discusses train, validation, and test splits, sliding or growing windows, bootstrap, and cross-validation. It emphasizes that the proper size of these sets remains unresolved and cautions that preprocessing and model fitting across partitions require careful evaluation.

Key ideas

Filtering mislabeled examples before classifier training lets the same cleaned data support different models.
The illustrated noise filter removes about 30% of samples across raw and transformed dataset versions.
Plots and predictor-importance measures can flag weak features, but model-based experiments are needed to validate them.
Dimensionality-reduction methods trade representation size against retained information and should be judged with a classification model.
The article treats partition sizing and preprocessing across train, validation, and test data as open practical questions.

Neural Network Data Preparation: Noise Filtering and Dimensionality Reduction

Summary

Key ideas

Tags