Intelligent Data Analytics using Deep Learning for Data Science

Authors

Maria E. Presa Reyes, Florida International UniversityFollow

Document Type

Dissertation

Degree

Doctor of Philosophy (PhD)

Major/Program

Computer Science

First Advisor's Name

Shu-Ching Chen

First Advisor's Committee Title

Committee chair

Second Advisor's Name

Xudong He

Second Advisor's Committee Title

Committee member

Third Advisor's Name

Jainendra K Navlakha

Third Advisor's Committee Title

Committee member

Fourth Advisor's Name

Cheng-Xian Lin

Fourth Advisor's Committee Title

Committee member

Fifth Advisor's Name

Sitharama S. Iyengar

Fifth Advisor's Committee Title

Committee member

Keywords

Deep Learning, Convolutional Neural Networks

Date of Defense

5-13-2022

Abstract

Nowadays, data science stimulates the interest of academics and practitioners because it can assist in the extraction of significant insights from massive amounts of data. From the years 2018 through 2025, the Global Datasphere is expected to rise from 33 Zettabytes to 175 Zettabytes, according to the International Data Corporation. This dissertation proposes an intelligent data analytics framework that uses deep learning to tackle several difficulties when implementing a data science application. These difficulties include dealing with high inter-class similarity, the availability and quality of hand-labeled data, and designing a feasible approach for modeling significant correlations in features gathered from various data sources. The proposed intelligent data analytics framework employs a novel strategy for improving data representation learning by incorporating supplemental data from various sources and structures. First, the research presents a multi-source fusion approach that utilizes confident learning techniques to improve the data quality from many noisy sources. Meta-learning methods based on advanced techniques such as the mixture of experts and differential evolution combine the predictive capacity of individual learners with a gating mechanism, ensuring that only the most trustworthy features or predictions are integrated to train the model. Then, a Multi-Level Convolutional Fusion is presented to train a model on the correspondence between local-global deep feature interactions to identify easily confused samples of different classes. The convolutional fusion is further enhanced with the power of Graph Transformers, aggregating the relevant neighboring features in graph-based input data structures and achieving state-of-the-art performance on a large-scale building damage dataset. Finally, weakly-supervised strategies, noise regularization, and label propagation are proposed to train a model on sparse input labeled data, ensuring the model's robustness to errors and supporting the automatic expansion of the training set. The suggested approaches outperformed competing strategies in effectively training a model on a large-scale dataset of 500k photos, with just about 7% of the images annotated by a human. The proposed framework's capabilities have benefited various data science applications, including fluid dynamics, geometric morphometrics, building damage classification from satellite pictures, disaster scene description, and storm-surge visualization.

Identifier

FIDC010775

Previously Published In

Maria Presa-Reyes, Shu-Ching Chen, “Multi-Source Weak Supervision Fusion for
Disaster Scene Recognition in Videos,” accepted for publication, IEEE International Conference on Multimedia Information Processing and Retrieval, 2022

Maria Presa-Reyes, Yudong Tao, Shu-Ching Chen, and Mei-Ling Shyu. “Deep
Learning With Weak Supervision for Disaster Scene Description in Low-Altitude
Imagery.” IEEE Transactions on Geoscience and Remote Sensing 60 (2021): 1-10.

Maria Presa-Reyes, Shu-Ching Chen, “Weakly-Supervised Damaged Building Lo-
calization and Assessment with Noise Regularization,” International Conference on
Multimedia Information Processing and Retrieval, Tokyo, Japan, September 8-10,
2021.

Maria Presa-Reyes, Shu-Ching Chen, “Assessing Building Damage by Learning the Deep Feature Correspondence of Before and After Aerial Images,” IEEE 3rd International Conference on Multimedia Information Processing and Retrieval, Shenzhen, Guangdong, China, pp. 43-48, August 6-8, 2020

Maria E. Presa-Reyes, Samira Pouyanfar, Hector Cen Zheng, Hsin-Yu Ha, Shu-
Ching Chen, “Multimedia Data Management for Disaster Situation Awareness,”
International Symposium on Sensor Networks, Systems and Security, Lakeland, FL, USA, pp. 137-146, 8/31-9/2, 2017.

Maria Presa-Reyes, Shu-Ching Chen, “A 3D Virtual Environment for Storm Surge
Flooding Animation,” The Third IEEE International Conference on Multimedia Big
Data, Laguna Hills, California, USA, pp. 244-245, April 19-21, 2017. (Demo Paper)

Recommended Citation

Presa Reyes, Maria E., "Intelligent Data Analytics using Deep Learning for Data Science" (2022). FIU Electronic Theses and Dissertations. 5001.
https://digitalcommons.fiu.edu/etd/5001