Review and comparison of methods to study the contribution of variables in artificial neural network models

Gevrey, Muriel; Dimopoulos, Ioannis; Lek, Sovan

doi:10.1016/s0304-3800(02)00257-0

Cited by 1,073 publications

(706 citation statements)

References 40 publications

Supporting

Mentioning

646

Contrasting

Unclassified

Order By: Relevance

“…A novel sensitivity analysis method was used for the models (Gevrey et al, 2003;Olden et al, 2004), which is simple and effective in identifying key variables. The method is introduced as follows:…”

Section: Predicting Important Variables Via Sensitivity Analysis and mentioning

confidence: 99%

Ensemble artificial neural networks applied to predict the key risk factors of hip bone fracture for elders

Liu

Cui

Chou

et al. 2015

Biomedical Signal Processing and Control

View full text Add to dashboard Cite

Hip bone fracture is one of the most important causes of morbidity and mortality in the elder adults. It is necessary to establish a prediction model to provide suggestions for elders. A total of 725 subjects were involved, including 228 patients with first low-trauma hip fracture and 497 ages-, sex-, and living area-matched controls (215 from the same hospital and 282 from community). All the subjects were interviewed with the same questionnaire, and the answers of the interviewees were recorded to be the database. Three-layer back-propagation Artificial Neural Networks (ANN) models were applied for females and males separately in this study to predict the risk of hip bone fracture for elders. Furthermore, to improve the accuracies and the generalizations of the models, the ensemble ANNs method was applied. To understand variables contributions and find the important variables for predicting hip fracture, sensitivity analysis and connection weights approach were applied. In this study, three ANNs prediction models were tested with different architectures. With the fivefold crossvalidation method evaluating the performances, one of the three models turned out to be the best prediction model and achieved a big success of prediction. The best area under the receiver operating characteristic (ROC) curve and the accuracy of the prediction model are 0.91 ± 0.028 (mean ± SD) and 0.85 ± 0.029 for females, while for males are 0.99 ± 0.015 and 0.93 ± 0.020. With the method of sensitivity analysis and connection weights, input variables were ranked according to contributions/importance, and the top 10 variables show great proportion of contribution to predict hip fracture. The top 10 important variables causing hip fracture for both females and males are similar to our previous results got from logistic regression model and other related researches. In conclusion, ANNs has successfully been to establish prediction models for predicting the risk of hip bone fracture for both female and male elder adults respectively and identified the top 10 important variables from 74 input variables to predict hip bone fracture of elders. This study verified the performance of ANNs to be a highly complex prediction model. Sept 2014Dear Prof. R. Allen, I use the electronic version to send this manuscript to you. The manuscript title is: "Ensemble back-propagation neural networks for predicting the risk of hip bone fracture for elders in Taiwan". We are submitting this material for possible publication in "Biomedical Signal Processing and Control". This material has not been submitted for publication or published elsewhere in whole or part. We believe this manuscript represents an original and significant contribution to the field of "Neural networks for predicting the risk of hip bone fracture" and therefore would like to be considered for publication in "Original Articles". AbstractHip bone fracture is one of the most important causes of morbidity and mortality in the elder adults. It is necessary to establish a prediction mod...

show abstract

Section: Predicting Important Variables Via Sensitivity Analysis and mentioning

confidence: 99%

Ensemble artificial neural networks applied to predict the key risk factors of hip bone fracture for elders

Liu

Cui

Chou

et al. 2015

Biomedical Signal Processing and Control

View full text Add to dashboard Cite

show abstract

“…Unfortunately, the "very black-box" nature of feed-forward neural networks ob- scures the meaning of the parameters and, despite the several attempts done in the literature, it is not easy to evaluate the importance of the different variables; furthermore, the different methodologies suitable to this end may lead to inconsistent conclusions (Gevrey et al, 2003).…”

Section: Input Variables Significancementioning

confidence: 99%

Air quality prediction in Milan: feed-forward neural networks, pruned neural networks and lazy learning

Corani

2005

Ecological Modelling

216

View full text Add to dashboard Cite

Ozone and PM 10 constitute the major concern for air quality of Milan. This paper addresses the problem of the prediction of such two pollutants, using to this end several statistical approaches. In particular, feed-forward neural networks (FFNNs), currently recognized as state-of-the-art approach for statistical prediction of air quality, are compared with two alternative approaches derived from machine learning: pruned neural networks (PNNs) and lazy learning (LL). PNNs constitute a parameterparsimonious approach, based on the removal of redundant parameters from fully connected neural networks; LL, on the other hand, is a local linear prediction algorithm, which performs a local learning procedure each time a prediction is required. All the three approaches are tested in the prediction of ozone and PM 10 ; predictors are trained to return at 9 a.m. the concentration estimated for the current day.No strong differences are found between the forecast accuracies of the different models; nevertheless, LL provides the best performances on indicators related to average goodness of the prediction (correlation, mean absolute error, etc.), while PNNs are superior to the other approaches in detecting of the exceedances of alarm and attention thresholds. In some cases, datadeseasonalization is found to improve the prediction accuracy of the models.Finally, some striking features of lazy learning deserve consideration: the LL predictor can be quickly designed, and, thanks to the simplicity of the local linear regressors, it both gets rid of overfitting problems and can be readily interpreted; moreover, it can be also easily kept up-to-date.

show abstract

“…Shojaeefard et al [47] have published the report where there was a similarity between results obtained using the PaD method and the profile method, as well as the classical stepwise method, but the results obtained using the connection weights method were significantly different. Gevrey et al [49] employed several methods of testing the variables contribution to study a brown trout reproduction phenomenon. The results obtained by authors show that more than one method should be used to analyze the contribution of the inputs, and results should be compared because for each method they are not always the same.…”

Section: Resultsmentioning

confidence: 99%

“…In this work, the absolute values of connection weights were used, the same as in work of Gevrey et al [49].…”

Section: Connection Weights Methodsmentioning

confidence: 99%

The identification of relationships between selected honey parameters by extracting the contribution of independent variables in a neural network model

Pentoś

Łuczycka

Kapłon

2015

Eur Food Res Technol

View full text Add to dashboard Cite

show abstract

Review and comparison of methods to study the contribution of variables in artificial neural network models

Cited by 1,073 publications

References 40 publications

Ensemble artificial neural networks applied to predict the key risk factors of hip bone fracture for elders

Ensemble artificial neural networks applied to predict the key risk factors of hip bone fracture for elders

Air quality prediction in Milan: feed-forward neural networks, pruned neural networks and lazy learning

The identification of relationships between selected honey parameters by extracting the contribution of independent variables in a neural network model

Contact Info

Product

Resources

About