Grzegorz Trawiński

An attempt to employ genetic fuzzy systems to predict from a data stream of premises transactions

An approach to apply ensembles of genetic fuzzy systems, built over the chunks of a data stream, to aid in residential premises valuation was proposed. The approach consists in incremental expanding an ensemble by systematically generated models in the course of time. The output of aged component models produced for current data is updated according to a trend function reflecting the changes of premises prices since the moment of individual model generation. An experimental evaluation of the proposed method using real-world data taken from a dynamically changing real estate market revealed its advantage in terms of predictive accuracy.

Evaluation of Neural Network Ensemble Approach to Predict from a Data Stream

We have recently worked out a method for building reliable predictive models from a data stream of real estate transactions which applies the ensembles of genetic fuzzy systems and neural networks. The method consists in building models over the chunks of a data stream determined by a sliding time window and enlarging gradually an ensemble by models generated in the course of time. The aged models are utilized to compose ensembles and their output is updated with trend functions reflecting the changes of prices in the market. In the paper we present the next series of extensive experiments to evaluate our method with the ensembles of artificial neural networks. We examine the impact of the number of aged models used to compose an ensemble on the accuracy and the influence of the degree of polynomial trend functions employed to modify the results on the performance of neural network ensembles. The experimental results were analysed using statistical approach embracing nonparametric tests followed by post-hoc procedures designed for multiple N×N comparisons.

Evaluation of Fuzzy System Ensemble Approach to Predict from a Data Stream

In the paper we present extensive experiments to evaluate our recently proposed method applying the ensembles of genetic fuzzy systems to build reliable predictive models from a data stream of real estate transactions. The method relies on building models over the chunks of a data stream determined by a sliding time window and incrementally expanding an ensemble by systematically generated models in the course of time. The aged models are utilized to compose ensembles and their output is updated with trend functions reflecting the changes of prices in the market. The experiments aimed at examining the impact of the number of aged models used to compose an ensemble on the accuracy and the influence of degree of polynomial trend functions applied to modify the results on the performance of single models and ensembles. The analysis of experimental results was made employing statistical approach including nonparametric tests followed by post-hoc procedures devised for multiple N×N comparisons.

Weighting Component Models by Predicting from Data Streams Using Ensembles of Genetic Fuzzy Systems

Our recently proposed method to predict from a data stream of real estate sales transactions based on ensembles of genetic fuzzy systems was extended to include weighting component models. The method consists in incremental expanding an ensemble by models built over successive chunks of a data stream. The predicted prices of residential premises computed by aged component models for current data are updated according to a trend function reflecting the changes of the market. The impact of different techniques of weighting component models on the accuracy of an ensemble was compared in the paper. Three techniques of weighting component models were proposed: proportional to their estimated accuracy, time of ageing, and dependent on property market fluctuations.

An analysis of change trends by predicting from a data stream using genetic fuzzy systems

A method to predict from a data stream of real estate sales transactions based on ensembles of genetic fuzzy systems was proposed. The approach consists in incremental expanding an ensemble by models built over successive chunks of a data stream. The predicted prices of residential premises computed by aged component models for current data are updated according to a trend function reflecting the changes of the market. The impact of different trend functions on the accuracy of single and ensemble fuzzy models was investigated in the paper. The results proved the usefulness of ensemble approach incorporating the correction of individual component model output.

An Analysis of Change Trends by Predicting from a Data Stream Using Neural Networks

A method to predict from a data stream of real estate sales transactions based on ensembles of artificial neural networks was proposed. The approach consists in incremental expanding an ensemble by models built over successive chunks of a data stream. The predicted prices of residential premises computed by aged component models for current data are updated according to a trend function reflecting the changes of the market. The impact of different trend functions on the accuracy of ensemble neural models was investigated in the paper. The results indicate it is necessary to make selection of correcting functions appropriate to the nature of market changes.

Investigation of random subspace and random forest regression models using data with injected noise

The ensemble machine learning methods incorporating random subspace and random forest employing genetic fuzzy rule-based systems as base learning algorithms were developed in Matlab environment. The methods were applied to the real-world regression problem of predicting the prices of residential premises based on historical data of sales/purchase transactions. The accuracy of ensembles generated by the proposed methods was compared with bagging, repeated holdout, and repeated cross-validation models. The tests were made for four levels of noise injected into the benchmark datasets. The analysis of the results was performed using statistical methodology including nonparametric tests followed by post-hoc procedures designed especially for multiple N×N comparisons.

Investigation of rotation forest ensemble method using genetic fuzzy systems for a regression problem

The rotation forest ensemble method using a genetic fuzzy rule-based system as a base learning algorithm was developed in Matlab environment. The method was applied to the real-world regression problem of predicting the prices of residential premises based on historical data of sales/purchase transactions. The computationally intensive experiments were conducted aimed to compare the accuracy of ensembles generated by our proposed method with bagging, repeated holdout, and repeated cross-validation models. The statistical analysis of results was made employing nonparametric Friedman and Wilcoxon statistical tests.

An Attempt to Use Self-Adapting Genetic Algorithms to Optimize Fuzzy Systems for Predicting from a Data Stream

In this chapter we present the continuation of our research into prediction from a data stream of real estate sales transactions using ensembles of regression models. The method consists in building models over the chunks of a data stream determined by a sliding time window and incrementally expanding an ensemble by systematically generated models in the course of time. The aged models are utilized to compose ensembles and their output is updated with trend functions reflecting the changes of prices in the market. In the study reported we attempted to incorporate self-adapting techniques into genetic fuzzy systems aimed to construct base models for property valuation. Six self-adapting genetic algorithms with varying mutation, crossover, and selection were developed and tested using real-world datasets. The analysis of experimental results was made employing non-parametric statistical techniques devised for multiple N×N comparisons.

Application of Self-adapting Genetic Algorithms to Generate Fuzzy Systems for a Regression Problem

Six variants of self-adapting genetic algorithms with varying mutation, crossover, and selection were developed. To implement self-adaptation the main part of a chromosome which comprised the solution was extended to include mutation rates, crossover rates, and/or tournament size. The solution part comprised the representation of a fuzzy system and was real-coded whereas to implement the proposed self-adapting mechanisms binary coding was employed. The resulting self-adaptive genetic fuzzy systems were evaluated using real-world datasets derived from a cadastral system and included records referring to residential premises transactions. They were also compared in respect of prediction accuracy with genetic fuzzy systems optimized by a classical genetic algorithm, multilayer perceptron and radial basis function neural network. The analysis of the results was performed using statistical methodology including nonparametric tests followed by post-hoc procedures designed especially for multiple N×N comparisons.