Feature Engineering Procedure for Information Enrichment

Data is essential for an organization's development, maintenance, and building of intelligence systems. Furthermore, AI procedures are key to generating recommendations when data contains multiple features. A notable challenge in developing a descriptive or predictive model is reducing the number of features to increase the model's performance metrics. Selecting a significant set of features for a dataset requires human intervention and analysis due to the diversity of techniques suitable for different models. Ho wever, insignificant features often need to be used in the models. This study presents a novel method for generating relevant features and enriching the dataset for learning. It is based on measuring the shared information between features and dividing the dataset into two subsets according to their similarities, such that one is more informative than the other. The significant advantage of our method is the increased ability to detect patterns in the dataset, leading to dimensionality reduction or enhanced performance metrics (i.e., precision, recall) for each target in the dataset. We tested our procedure using six datasets. The results showed an improvement of up to 43.75% in the recall, 20.77% in precision, and 29.62% in the weighted fl-score.