Methods to Improve the Prediction Accuracy and Performance of Ensemble Models

MPhil Thesis


Adegoke, V. (2022). Methods to Improve the Prediction Accuracy and Performance of Ensemble Models. MPhil Thesis London South Bank University School of Engineering https://doi.org/10.18744/lsbu.92856
AuthorsAdegoke, V.
TypeMPhil Thesis
Abstract

The application of ensemble predictive models has been an important research area in predicting medical diagnostics, engineering diagnostics, and other related smart devices and
related technologies. Most of the current predictive models are complex and not reliable despite numerous efforts in the past by the research community. The performance accuracy of the predictive models have not always been realised due to many factors such as complexity and class imbalance. Therefore there is a need to improve the predictive accuracy of current ensemble models and to enhance their applications and reliability and non-visual predictive tools.
The research work presented in this thesis has adopted a pragmatic phased approach to propose and develop new ensemble models using multiple methods and validated the methods through rigorous testing and implementation in different phases. The first phase comprises of empirical investigations on standalone and ensemble algorithms that were carried out to ascertain their performance effects on complexity and simplicity of the classifiers. The second phase comprises of an improved ensemble model based on the integration of Extended Kalman Filter (EKF), Radial Basis Function Network (RBFN) and AdaBoost algorithms. The third phase comprises of an extended model based on early stop concepts, AdaBoost algorithm, and statistical performance of the training samples to minimize overfitting performance of the proposed model. The fourth phase comprises of an enhanced analytical multivariate logistic regression predictive model developed to minimize the complexity and improve prediction accuracy of logistic regression model.
To facilitate the practical application of the proposed models; an ensemble non-invasive analytical tool is proposed and developed. The tool links the gap between theoretical concepts and practical application of theories to predict breast cancer survivability. The empirical findings suggested that: (1) increasing the complexity and topology of algorithms does not necessarily lead to a better algorithmic performance, (2) boosting by resampling performs slightly better than boosting by reweighting, (3) the prediction accuracy of the proposed ensemble EKF-RBFN-AdaBoost model performed better than several established ensemble models, (4) the proposed early stopped model converges faster and minimizes overfitting better compare with other models, (5) the proposed multivariate logistic regression concept minimizes the complexity models (6) the performance of the proposed analytical non-invasive tool performed comparatively better than many of the benchmark analytical tools used in predicting breast cancers and diabetics ailments.
The research contributions to ensemble practice are: (1) the integration and development of EKF, RBFN and AdaBoost algorithms as an ensemble model, (2) the development and validation of ensemble model based on early stop concepts, AdaBoost, and statistical concepts of the training samples, (3) the development and validation of predictive logistic regression model based on breast cancer, and (4) the development and validation of a non-invasive breast cancer analytic tools based on the proposed and developed predictive models in this thesis.
To validate prediction accuracy of ensemble models, in this thesis the proposed models were applied in modelling breast cancer survivability and diabetics’ diagnostic tasks. In comparison with other established models the simulation results of the models showed improved predictive accuracy.
The research outlines the benefits of the proposed models, whilst proposes new directions for future work that could further extend and improve the proposed models discussed in this
thesis.

Year2022
PublisherLondon South Bank University
Digital Object Identifier (DOI)https://doi.org/10.18744/lsbu.92856
File
License
File Access Level
Open
Publication dates
Print04 Mar 2022
Publication process dates
Deposited05 Dec 2022
Permalink -

https://openresearch.lsbu.ac.uk/item/92856

Download files

  • 15
    total views
  • 2
    total downloads
  • 3
    views this month
  • 1
    downloads this month

Export as