Adrian L. Arnaud

The Power of Sampling and Stacking for the PAKDD-2007 Cross-Selling Problem

This article presents an efficient solution for the PAKDD-2007 Competition cross-selling problem. The solution is based on a thorough approach which involves the creation of new input variables, efficient data preparation and transformation, adequate data sampling strategy and a combination of two of the most robust modeling techniques. Due to the complexity imposed by the very small amount of examples in the target class, the approach for model robustness was to produce the median score of the 11 models developed with an adapted version of the 11-fold cross-validation process and the use of a combination of two robust techniques via stacking, the MLP neural network and the n-tuple classifier. Despite the problem complexity, the performance on the prediction data set (unlabeled samples), measured through KS2 and ROC curves was shown to be very effective and finished as the first runner-up solution of the competition.

Neural Networks vs Logistic Regression: a Comparative Study on a Large Data Set

Neural networks and logistic regression have been among the most widely used AI technique in applications of pattern classification.Much has been discussed about if there is any significant difference in between them but much less has been actually done with real-world applications data (large scale) to help settle this matter, with a few exceptions.This paper presents a performance comparison between these two techniques on the market application of credit risk assessment, making use of a large database from an outstanding credit bureau and financial institution (a sample of 180,000 examples).The comparison was carried out through a 30-fold stratified cross-validation process to define the confidence intervals for the performance evaluation. Several metrics were applied both on the optimal decision point and along the continuous output domain.The statistical tests showed that multilayer perceptrons perform better than logistic regression at 95% confidence level, for all the metrics used.

Continuous Dynamical Combination of Short and Long-Term Forecasts for Nonstationary Time Series

This brief generalizes the forecasting method that has been awarded first-place winner in the International Competition of Time Series Forecasting (ICTSF 2012). It is based on a short-term forecasting approach of multilayer perceptrons (MLP) ensembles, combined dynamically with a long-term forecasting. The main feature of this general approach is the original concept of continuous dynamical combination of forecasts, in which the weights of the forecasting combination are a function of forecast horizon. Experiments in ICTSFs and NN5s nonstationary time series show that this new combination method improves the performance in multistep forecasting of MLP ensembles when compared to the MLP ensembles alone.

The role of temporal feature extraction and bagging of MLP neural networks for solving the WCCI 2008 Ford Classification Challenge

This paper presents an approach for solving WCCI 2008s Ford Classification Challenge Problem. The solution is based on the creation of new input variables through temporal feature extraction and on the combination via bagging of an ensemble of 30 multi-layer perceptrons trained on sets divided by multiple random sampling of the labeled data. Signal power, signal to noise ratio and signal frequency were some of the meaningful features extracted for improving the systems performance. The data sampling strategy produced a robust median MLP response and allowed for the definition of the appropriate decision threshold. The performance measured on the 30 test samples (statistically independent from the training data) reached an average of Max_KS2 = 0.91, AUC_ROC = 0.99 and accuracy of 95.6% for Ford_A and Max_KS2 = 0.88, AUC_ROC = 0.98 and accuracy of 94.1% for Ford_B. These results have been confirmed on the competition for the noiseless data and have degraded around 15% for the noisy data.

A systematic solution for the NN3 Forecasting Competition problem based on an ensemble of MLP neural networks

This work presents an award winning approach for solving the NN3 forecasting competition problem. It consisted of predicting 18 future values of 111 monthly short time series. This approach consists of applying the median value of a 15-MLP ensemble for predicting each time series. The system performed very well on test data, finishing as the second best solution of the competition with a SMAPE=16.17%.

Data transformations and seasonality adjustments improve forecasts of MLP ensembles

This work describes the first place winner forecasting method for solving the 1st International Competition on Time Series Forecasting (ICTSF 2012). It is based on an already award winning approach of MLP ensembles [1]. The ICTSF 2012 consisted on predicting 8 time series of different time frequency and different forecasting horizons. The main feature of the present method was applying different data pre-processing and seasonality adjustments to a combined forecast of 225 MLPs predicting each time series. Experimental comparison and the competitions result shows that this new predictive system increases its performance in multi-step forecasting when compared to ensembles of MLP.

Hybrid Approach for Time Series Prediction using Neural Networks and Simulated Annealing

This paper proposes a new hybrid approach which combines simulated annealing and standard backpropagation for optimizing Multi Layer Perceptron Neural Networks (MLP) for time series prediction. Experimental tests were carried out on four simulated series with known features and on the Sunspot series. The results have shown that this approach selects the appropriate time series lags and builds an MLP with the minimum number of hidden neurons required for achieving good performance on the task. The performance attained was better than some results recently reported for hybrid systems combining Genetic Algorithms (GA) and MLPs for the same purpose

Sistema de Apoio à Decisão para Estimação do Sucesso do Aluno no Programa de Mestrado em Ciência da Computação da UFPE

The candidates for the Center for Computer Science postgraduate programme of UFPE, Brazil, are selected based on their CVs. Five attributes are linearly weighed to produce the candidate’s score, which serves as ranking criterion. The postgraduate collegiate was concerned this process could be restricting access to the programme. This paper analyzes the profiles both of the 148 students and of the other 666 candidates who did not enter the programme in the last 4 years. This research concludes that the 5 attributes used could not discriminate between good and weak students. The 4 additional attributes made it feasible. Moreover, this paper has proposed an effective system and has shown that the previous system was indeed being restrictive.

DECISION SUPPORT SYSTEM FOR PREVENTING NO-SHOW TO MEDICALAPPOINTMENTS

The failure of customers to attend booked appointments, known as no-show, is a problem faced by companies worldwide, operating in medical assistance and transportation businesses. This work focused on the former, more particularly, on the medical appointments made via call-center. This paper presents a decision support system based on data mining for identifying, at booking time, the medical appointments with high risk of no-show, for helping online re-scheduling. Preprocessing and data transformation yielded embedding experts knowledge and behavioral information. The a priori algorithm explicited the knowledge contained on the data and an MLP neural network estimated the risk of no-show. The system has been developed on a data set of 30,000 and tested on other 10,000 appointments from a healthcare company operating in Brazil. Both the risk estimation and the rules extracted attained high quality in the metrics defined and were considered very relevant by the companys specialist.