A transfer learning-based feature reduction method to improve classification accuracy

UNCG Author/Contributor (non-UNCG co-authors, if there are any, appear on document)
Maha Mohammed Asiri (Creator)
The University of North Carolina at Greensboro (UNCG )
Web Site: http://library.uncg.edu/
Fereidoon Sadri

Abstract: The need for efficient data use grows in machine learning algorithm for dataset with larger feature sets. Feature selection is the process of selecting minimum set of features that fully represent the learning problem. Transfer learning can motivate in scenario where we train model with the common problem and use it to identify important features needed to build model for target problem. In this thesis, we propose transfer learning algorithm combined with or without suggested features from experts, to learn from the source dataset and recognize important feature sets needed to train models in target dataset. Also, we compared this algorithm with classical machine learning algorithm with or without using the suggested features recommended by the experts. In series of experiment, it shows that our method is adequate to find the minimum feature sets which also outperformed then using only the suggested features by the experts. Furthermore, it also shows that the subsequent reduce in number of features in transfer learning method have better or almost same performance then using all the features of the dataset. We performed our experiments using heart disease, readmission dataset and BMI dataset.

Additional Information

Language: English
Date: 2017
Decision Tree, KNN, Machine Learning, MLP, Random Forest, Transfer Learning
Machine learning
Random graphs
Decision trees

Email this document to