Adina Cocu

A hybrid feature selection algorithm for the QSAR problem

In this paper we discuss a hybrid feature selection algorithm for the Quantitative Structure Activity Relationship (QSAR) modelling. This is one of the goals in Predictive Toxicology domain, aiming to describe the relations between the chemical structure of a molecule and its biological or toxicological effects, in order to predict the behaviour of new, unknown chemical compounds. We propose a hybridization of the ReliefF algorithm based on a simple fuzzy extension of the value difference metric. The experimental results both on benchmark and real world applications suggest more stability in dealing with noisy data and our preliminary tests give a promising starting point for future research.

Improved occupancy-based solutions for energy saving in buildings

In the past decades, despite the huge interest for the topics related to energy conservation and reducing the CO2 emissions, it became obvious that there is an economic reason for the weak progress in these fields: the visible correlation between the economic growth and the energy consumption. As a result, it is very likely that the demand for energy will continue to grow. Since buildings are responsible for 40% of the total energy demand, in this paper we review the vast literature dedicated to energy saving in buildings from the perspective of the feasibility on a large scale. We emphasize the solutions based on detecting and forecasting the building occupancy in order to control the HVAC and lighting systems for energy saving without affecting the comfort of the users. Several improvements of these solutions are proposed.

A Hybrid Approach for Data Preprocessing in the QSAR Problem

One of the approaches in the Knowledge Discovery in Databases (KDD) domain is Predictive Toxicology (PT). Its aim is to discover and represent the relationships between the chemical structure of chemical compounds and biological and toxicological processes. The challenges in real toxicology problems are big amount of the chemical descriptors and imperfect data (means noisy, redundant, incomplete, and irrelevant). The main goals in knowledge discovery field are to detect these undesirable proprieties and to eliminate or correct them. This supposes noise reduction, data cleaning and feature selection because the performance of the applied Machine Learning algorithms is strongly related with the quality of the used data. In this paper, we present some of the issues that can be performed for preparing data before the knowledge discovery process begin.