Matching Items (2)
Filtering by
- All Subjects: Feature Selection
- Creators: Dinu, Valentin
Description
Random forest (RF) is a popular and powerful technique nowadays. It can be used for classification, regression and unsupervised clustering. In its original form introduced by Leo Breiman, RF is used as a predictive model to generate predictions for new observations. Recent researches have proposed several methods based on RF for feature selection and for generating prediction intervals. However, they are limited in their applicability and accuracy. In this dissertation, RF is applied to build a predictive model for a complex dataset, and used as the basis for two novel methods for biomarker discovery and generating prediction interval.
Firstly, a biodosimetry is developed using RF to determine absorbed radiation dose from gene expression measured from blood samples of potentially exposed individuals. To improve the prediction accuracy of the biodosimetry, day-specific models were built to deal with day interaction effect and a technique of nested modeling was proposed. The nested models can fit this complex data of large variability and non-linear relationships.
Secondly, a panel of biomarkers was selected using a data-driven feature selection method as well as handpick, considering prior knowledge and other constraints. To incorporate domain knowledge, a method called Know-GRRF was developed based on guided regularized RF. This method can incorporate domain knowledge as a penalized term to regulate selection of candidate features in RF. It adds more flexibility to data-driven feature selection and can improve the interpretability of models. Know-GRRF showed significant improvement in cross-species prediction when cross-species correlation was used to guide selection of biomarkers. The method can also compete with existing methods using intrinsic data characteristics as alternative of domain knowledge in simulated datasets.
Lastly, a novel non-parametric method, RFerr, was developed to generate prediction interval using RF regression. This method is widely applicable to any predictive models and was shown to have better coverage and precision than existing methods on the real-world radiation dataset, as well as benchmark and simulated datasets.
Firstly, a biodosimetry is developed using RF to determine absorbed radiation dose from gene expression measured from blood samples of potentially exposed individuals. To improve the prediction accuracy of the biodosimetry, day-specific models were built to deal with day interaction effect and a technique of nested modeling was proposed. The nested models can fit this complex data of large variability and non-linear relationships.
Secondly, a panel of biomarkers was selected using a data-driven feature selection method as well as handpick, considering prior knowledge and other constraints. To incorporate domain knowledge, a method called Know-GRRF was developed based on guided regularized RF. This method can incorporate domain knowledge as a penalized term to regulate selection of candidate features in RF. It adds more flexibility to data-driven feature selection and can improve the interpretability of models. Know-GRRF showed significant improvement in cross-species prediction when cross-species correlation was used to guide selection of biomarkers. The method can also compete with existing methods using intrinsic data characteristics as alternative of domain knowledge in simulated datasets.
Lastly, a novel non-parametric method, RFerr, was developed to generate prediction interval using RF regression. This method is widely applicable to any predictive models and was shown to have better coverage and precision than existing methods on the real-world radiation dataset, as well as benchmark and simulated datasets.
ContributorsGuan, Xin (Author) / Liu, Li (Thesis advisor) / Runger, George C. (Thesis advisor) / Dinu, Valentin (Committee member) / Arizona State University (Publisher)
Created2017
Description
This dissertation presents three novel algorithms with real-world applications to genomic oncology. While the methodologies presented here were all developed to overcome various challenges associated with the adoption of high throughput genomic data in clinical oncology, they can be used in other domains as well. First, a network informed feature ranking algorithm is presented, which shows a significant increase in ability to select true predictive features from simulated data sets when compared to other state of the art graphical feature ranking methods. The methodology also shows an increased ability to predict pathological complete response to preoperative chemotherapy from genomic sequencing data of breast cancer patients utilizing domain knowledge from protein-protein interaction networks.
Second, an algorithm that overcomes population biases inherent in the use of a human reference genome developed primarily from European populations is presented to classify microsatellite instability (MSI) status from next-generation-sequencing (NGS) data. The methodology significantly increases the accuracy of MSI status prediction in African and African American ancestries.
Finally, a single variable model is presented to capture the bimodality inherent in genomic data stemming from heterogeneous diseases. This model shows improvements over other parametric models in the measurements of receiver-operator characteristic (ROC) curves for bimodal data. The model is used to estimate ROC curves for heterogeneous biomarkers in a dataset containing breast cancer and cancer-free specimen.
ContributorsSaul, Michelle (Author) / Dinu, Valentin (Thesis advisor) / Liu, Li (Committee member) / Wang, Junwen (Committee member) / Arizona State University (Publisher)
Created2021