Oral Defense of Doctoral Dissertation: Sze Wing Wong

Oral Defense of Doctoral Dissertation

Doctor of Philosophy in Computational Sciences and Informatics
Department of Computational and Data Sciences
College of Science
George Mason University  

Sze Wing Wong
Bachelor of Science, New Mexico Institute of Mining and Technology, 2008
Master of Science, University of New Mexico, 2009
 

CLASSIFYING THREE TIERS OF SUCCESS IN CROWDFUNDING WITH MACHINE LEARNING AND NATURAL LANGUAGE PROCESSING 

Thursday, November 18, 2021, 9 to 11 a.m. Eastern Time (US and Canada)

Join Zoom Meeting
https://gmu.zoom.us/j/92693763551
Meeting ID: 926 9376 3551

One tap mobile
+13017158592,,92693763551# US (Washington DC)
+12678310333,,92693763551# US (Philadelphia)
Dial by your location
+1 301 715 8592 US (Washington DC)
+1 267 831 0333 US (Philadelphia)

Meeting ID: 926 9376 3551

Find your local number: https://gmu.zoom.us/u/aRlw9Jps  

All are invited to attend.

Committee

Dr. Robert L. Axtell, Committee Chair
Dr. William G. Kennedy
Dr. Igor Griva
Dr. Olga Gkountouna

ABSTRACT:  With the continued growing adoption of crowdfunding for raising capital, much research and numerous studies have been done using econometric analyses and traditional machine learning models to identify key indicators and to classify success in campaigns. These findings are informative and hold beneficial insights for both entrepreneurs and investors. However, in recent years, the increasing number of successful campaigns far exceed the number of failed campaigns. Therefore, previous studies focusing on binary classification is no longer sufficient to capture different levels of emergent success in crowdfunding. This dissertation examines three tiers of success in reward-based campaigns of the 2019 Kickstarter data to gain insights on the evolved crowdfunding phenomenon. How I demonstrate new key indicators are identified by exploiting campaign information relates to “people” versus “products” with hierarchical multiple and ordinal logistic regressions. In conjunction, I adopt Binary Particle Swarm Optimization (BPSO) for feature selection to classify campaign success. The BPSO improved Extreme Gradient Boosting (XGBoost) classifier shows favorable model performance in multiclass classification.  Most importantly, the interest in this research extends beyond using categorical and numeric features, but also fuses multiple textual information with natural language processing (NLP) and deep neural networks (DNNs) for multiclass classification. The proposed multimodal Long Short-Term Memory (LSTM) concatenates BPSO selected meta data with project’s pitch and creator’s biography text yields the best performing multiclass classification accuracy of 71.04% after tuning. However, the BPSO improved Extreme Gradient Boosting (XGBoost) classifier achieves the highest accuracy of 74.61% overall. These impactful findings allow entrepreneurs and researchers to gain further insights and to optimize the cybermarketing space effectively.