A productivity metric based on statistical pattern recognition

Abstract

The generally accepted calculation to measure the productivity of a software engineer is based on economic theory and is borrowed from traditional product manufacturing environments. Managers often measure the productivity of a worker to determine merit-based raises or to provide feedback to workers with poor productivity. The assumption is that this calculation of a worker's productivity is directly proportional to a worker's value to the company. The motivation for the approach proposed here is that such relationship may not be algebraically captured with respect to the productivity of software engineers. To better capture the productivity of a software engineer and his value to a company, the productivity problem is reformulated here as a pattern recognition problem and solved using clustering. By defining a general productivity operator, clustering has been used to map the domain of the productivity operator to a range of productivity classes. This new approach has been successfully applied to randomly generated project data and actual project data from the NASA SATC software metric database.