It has become apparent that in many problems that require machine learning based solutiuons, the complexity of the data's underlying patterns is not always directly captured by the sets of signals directly measurable from the process. Moreover, increasing the number of instances or data samples captured does not lead to any improvement in the predictive quality of the systems. The expansion of the representation space of machine learning algorithms has beed widely acknowledged as one of the crucial factors in improving their predictive performance. Through the design of new features, new information is readily made available to the machine learning algorithm allowing for dramatic performance increases should such features be relevant to the problem being solved.
New features can be constructued in a variety of ways, from applying simple fixed mathematical operations using single features of single instances to applying complex functions involving many samples spread acrosss the existing variable space. The Features Analytics process of designing features relies on a multitude of approaches not only to engineer and efficiently calculate features, but also to test their relevance within the context of the problem at hand. The ever-increasing amount of data available as well as the speed with which is generated does pose a huge challenge for organizations, not only from an analysis and modeling standpoint but also from a pure computational one as well: benefitting from Features Analytics' fast and efficient calculation of multiple complex mathematical operations using data spread across both time and existing feature space, organizations can improve not only the speed with which business insights are generated but the quality of these insights as well, as more features can be calculated and evaluated during the feature design process.
While generating many complex features can certainly be beneficial to the final performance of predictive models, their employment alone does not necessarily guarantee this. Hence, the whole design process is organized in an iterative manner such that in each iteration features are being engineered, their quality assesed and, depending of their predictive qualities either discarded or kept and passed in the next iteration, until the desired performance is achieved. Exploiting both computational efficiency as well as rigorous testing and business knowledge, Features Analytics' feature design process yields up to hundreds of new relevant predictive features to support organizations to use in their analytical endeavours.