FDR2-BD: A Fast Data Reduction Recommendation Tool for Tabular Big Data Classification Problems

In this paper, a methodological data condensation approach for reducing tabular big datasets in classification problems is presented, named FDR2-BD. The key of our proposal is to analyze data in a dual way (vertical and horizontal), so as to provide a smart combination between feature selection to generate dense clusters of data and uniform sampling reduction to keep only a few representative samples from each problem area. Its main advantage is allowing the model's predictive quality to be kept in a range determined by a user's threshold. Its robustness is built on a hyper-parametrization process, in which all data are taken into consideration by following a k-fold procedure. Another significant capability is being fast and scalable by using fully optimized parallel operations provided by Apache Spark. An extensive experimental study is performed over 25 big datasets with different characteristics. In most cases, the obtained reduction percentages are above 95%, thus outperforming state-of-the-art solutions such as FCNN_MR that barely reach 70%. The most promising outcome is maintaining the representativeness of the original data information, with quality prediction values around 1% of the baseline.