A Model for Feature Selection with Binary Particle Swarm Optimisation and Synthetic Features

Recognising patterns and inferring nonlinearities between data that are seemingly random and stochastic in nature is one of the strong suites of machine learning models. Given a set of features, the ability to distinguish between useful features and seemingly useless features, and thereafter extract a subset of features that will result in the best prediction on data that are highly stochastic, remains an open issue. This study presents a model for feature selection by generating synthetic features and applying Binary Particle Swarm Optimisation with a Long Short-Term Memory-based model. The study analyses the correlation between data and makes use of Apple stock market data as a use case. Synthetic features are created from features that have weak/low correlation to the label and analysed how synthetic features that are descriptive of features can enhance the model’s predictive capability. The results obtained show that by expanding the dataset to contain synthetic features before applying feature selection, the objective function was better optimised as compared to when no synthetic features were added.