Rocío L. Cecchini

Discretization of gene expression data revised

Gene expression measurements represent the most important source of biological data used to unveil the interaction and functionality of genes. In this regard, several data mining and machine learning algorithms have been proposed that require, in a number of cases, some kind of data discretization to perform the inference. Selection of an appropriate discretization process has a major impact on the design and outcome of the inference algorithms, as there are a number of relevant issues that need to be considered. This study presents a revision of the current state-of-the-art discretization techniques, together with the key subjects that need to be considered when designing or selecting a discretization approach for gene expression data.

A wrapper-based feature selection method for ADMET prediction using evolutionary computing

Wrapper methods look for the selection of a subset of features or variables in a data set, in such a way that these features are the most relevant for predicting a target value. In chemoinformatics context, the determination of the most significant set of descriptors is of great importance due to their contribution for improving ADMET prediction models. In this paper, a comprehensive analysis of descriptor selection aimed to physicochemical property prediction is presented. In addition, we propose an evolutionary approach where different fitness functions are compared. The comparison consists in establishing which method selects the subset of descriptors that best predicts a given property, as well as maintaining the cardinality of the subset to a minimum. The performance of the proposal was assessed for predicting hydrophobicity, using an ensemble of neural networks for the prediction task. The results showed that the evolutionary approach using a non linear fitness function constitutes a novel and a promising technique for this bioinformatic application.

Multi-objective evolutionary approaches for intelligent design of sensor networks in the petrochemical industry

Highlights? Pose of semi-automation of the initialization of instrumentation design procedures. ? Implementation of different evolutionary alternatives for instrumentation design. ? Use of MOEAs to support experts with alternative high quality sensor networks. ? Contrast of traditional methods apropos closeness to the Pareto-front and dispersion. The design of optimal sensor networks for an industrial process is a complex problem that requires the resolution of several tasks with a high level of expertise. The first of these subproblems consists in selecting an initial sensor network as the starting point for the instrumentation design. This particular task constitutes a combinatorial optimization problem, where several goals are prosecuted by the designer. Therefore, the initialization procedure can be defined as a multi-objective optimization problem. In this paper, the use of multi-objective evolutionary approaches to assist experts in the design of an initial sensor network is proposed and analyzed. The aim is to contrast the advantages and limitations of Pareto and non-Pareto techniques in the context of this industrial application. The algorithms consider objectives related to cost, reliability and level of information associated with a sensor network. The techniques were evaluated by means of a comparative analysis for a strongly non-linear mathematical model that represents an ammonia synthesis plant. Results have been contrasted in terms of the set coverage and spacing metrics. As a final conclusion, the non-Pareto strategy converged closer to the Pareto front than the Pareto-based algorithms. In contrast, the Pareto-based algorithms achieved better relative distance among solutions than the non-Pareto method. In all cases, the use of evolutionary computation is useful for the expert to take the final decision on the preferred initial sensor network.

An Evolutionary Algorithm for Automatic Recommendation of Clustering Methods and its Parameters

Abstract One of the main problems being faced at the time of performing data clustering consists in the deteremination of the best clustering method together with defining the ideal amount (k) of groups in which these data should be separated. In this paper, a preliminary approximation of a clustering recommender method is presented which, starting from a set of standardized data, suggests the best clustering strategy and also proposes an advisable k value. For this aim, the algorithm considers four indices for evaluating the final structure of clusters: Dunn, Silhouette, Widest Gap and Entropy. The prototype is implemented as a Genetic Algorithm in which individuals are possible configurations of the methods and their parameters. In this first prototype, the algorithm suggests between four partitioning methods namely K-means, PAM, CLARA and, Fanny. Also, the best set of parameters to execute the suggested method is obtained. The prototype was developed in an R environment, and its findings could be corroborated as consistent when compared with a combination of results provided by other methods with similar objectives. The idea of this prototype is to serve as the initial basis for a more complex framework that also incorporates the reduction of matrices with vast numbers of rows.

Topic relevance and diversity in information retrieval from large datasets: A multi-objective evolutionary algorithm approach

Abstract Enabling effective information search is an increasing problem, as technology enhances the ability to publish information rapidly, and large quantities of information are instantly available for retrieval. In this scenario, topical search is the process of searching for material that is relevant to a given topic. Multi-objective Evolutionary Algorithms have demonstrated great potential for addressing the topical search problem in very large datasets. In an evolutionary approach to topical search, a population of queries is automatically generated from a given topic, and the population of queries then evolves towards successively better candidate queries. Despite the promise of this approach, previous studies have revealed a common genotypic phenomenon: throughout evolution, the population tends to converge to almost identical sets of terms. This situation reduces the solution set to a few queries and leads to the exploration of a very limited region of the search space, which constitutes a limitation when users require different options from a topical search tool. This paper proposes and evaluates strategies to favor diversity in evolutionary topical search. These strategies rely on novel fitness functions, different parameterization for the crossover and mutation rates, and the use of multiple populations to favor diversity preservation. Experimental results conducted using these strategies in combination with the NSGA-II algorithm on a dataset consisting of more than 350,000 labeled web pages indicate that the proposed strategies show great promise for searching very large datasets, by helping to achieve query and search result diversity without giving up precision.

Multi-objective Query Optimization Using Topic Ontologies

Formulating search queries based on a thematic context is a challenging problem due to the large number of combinations in which terms can be used to reflect the topic of interest. This paper presents a novel approach to learn topical queries that simultaneously satisfy multiple retrieval objectives. The proposed method consists in using a topic ontology to train an Evolutionary Algorithm that incrementally moves a population of queries towards the proposed objectives. We present an analysis of different single- and multi-objective strategies, discuss their strengths and limitations and test the most promising strategies on a large set of labeled Web pages. Our evaluations indicate that the tested strategies that apply multi-objective Evolutionary Algorithms are significantly superior to a baseline approach that attempts to generate queries directly from a topic description.