Crowdfunding Campaigns. Success Prediction Through Natural Language Processing

Table of Contents

1. Introduction

2. Theoretical Foundation

2.1. Data Mining Process

2.2. Natural Language Processing

2.2.1. Definition and Distinction

2.2.2. Preprocessing and Feature Selection

2.2.3. Vectorization of Words

2.3. Architecture and Types of Artificial Neural Networks

2.3.1. Feed Forward Neural Networks

2.3.2. Convolutional Neural Networks

2.3.3. Graph Neural Networks

2.3.4. Recurrent Neural Networks

2.3.5. Methods for Performance Evaluation

2.4. Applied technologies

3. Methodology

3.1. Business Understanding

3.2. Data Understanding

3.3. Data Preparation

3.4. Modeling

3.4.1. Convolutional Neural Networks

3.4.2. Graph Neural Networks

3.4.3. Bidirectional LSTM

4. Evaluation of Results

4.1. Convolutional Neural Network Results

4.1.1. Category Prediction by Convolutional Neural Network

4.1.2. Success Prediction by Convolutional Neural Network

4.2. Graph Convolutional Neural Network Results

4.2.1. Category Prediction by Graph Convolutional Neural Network

4.2.2. Success Prediction by Convolutional Neural Network

4.3. Bidirectional LSTM Results

4.3.1. Category Prediction by Bidirectional LSTM

4.3.2. Success Prediction by Bidirectional LSTM

5. Discussion

6. Conclusion and Outlook

Objective & Topics

This thesis investigates whether natural language processing (NLP) techniques applied to user-generated campaign text can assist crowdfunding founders by predicting appropriate categories and overall campaign success. The study aims to provide a competitive advantage to founders by leveraging data mining frameworks to turn unstructured textual descriptions into actionable insights.

• Application of the CRISP-DM framework to crowdfunding data.
• Implementation and comparison of Convolutional Neural Networks (CNN), Graph Convolutional Neural Networks (GCN), and Bidirectional LSTM models.
• Analysis of dataset imbalances on predictive model performance.
• Evaluation of "typicality" and "ambiguity" metrics in relation to campaign success.

Excerpt from the Thesis

Summary of Chapters

1. Introduction: Outlines the growth of the crowdfunding market, the problem of selecting the right categories, and sets the research questions regarding NLP's ability to predict categorization and campaign success.

2. Theoretical Foundation: Provides the conceptual background on the CRISP-DM framework, natural language processing techniques, and the architectural details of various artificial neural networks.

3. Methodology: Details the practical implementation, including the business understanding, data sourcing from Kickstarter and Indiegogo, data cleaning processes, and the model architectures used.

4. Evaluation of Results: Presents the findings from the neural network experiments, comparing model performance across categories and success objectives using test accuracy and F1 scores.

5. Discussion: Interprets the results within the research context, addressing the impact of data imbalance and hyperparameter tuning on model reliability.

6. Conclusion and Outlook: Synthesizes the main findings, confirms the effectiveness of BI-LSTM models, and suggests future research directions, such as Hierarchical Neural Networks.

Keywords

Crowdfunding, Natural Language Processing, Artificial Neural Network, Kickstarter, Indiegogo, CRISP-DM, Data Mining, Convolutional Neural Networks, Graph Convolutional Networks, Bidirectional LSTM, Success Prediction, Text Classification, Hyperparameter Tuning, Deep Learning, Feature Selection

Frequently Asked Questions

What is the primary focus of this research?

The research explores the application of natural language processing on user-generated text from crowdfunding campaigns to identify patterns that predict campaign categorization and financial success.

Which crowdfunding platforms were analyzed for this study?

The study utilizes datasets from Kickstarter (KS) and Indiegogo (IG).

What is the core research question addressed by the thesis?

The thesis answers two main questions: Can NLP predict the category for a given campaign, and can it predict whether a campaign will reach its funding target?

Which machine learning architectures were employed?

The author implemented and evaluated three distinct architectures: Convolutional Neural Networks (CNN), Graph Convolutional Neural Networks (GCN), and Bidirectional Long Short-Term Memory networks (BI-LSTM).

How is the methodology structured?

The analysis follows the standardized Cross-Industry Standard Process for Data Mining (CRISP-DM), moving through phases of business understanding, data collection, preparation, and modeling.

Which performance metrics were used to determine model quality?

Model success was primarily evaluated using test accuracy and F1 scores, with secondary analysis of typicality and entropy metrics to assess prediction confidence.

Which model provided the best overall performance?

The BI-LSTM architecture consistently achieved the highest test accuracy and F1 scores across both objectives and datasets.

How does the author propose dealing with dataset imbalance?

The author acknowledges that dataset imbalance affects results, particularly for rare categories, and discusses weight balancing as a potential mitigation strategy to improve predictive accuracy for minority classes.