Wide and Deep Learning in Crowdfunding Success Prediction: A Comparative Study of Deep Neural Networks and Linear Models in Predicting the Success of Reward-based Crowdfunding Campaigns

David George

Student thesis: Master thesis


The thesis studies to which extent deep learning models can predict whether crowdfunding campaigns will be successfully funded by drawing upon a dataset of 246,891 projects and 40 structured and text features from the crowdfunding platform Kickstarter. Though neural networks have been used in past research on crowdfunding success prediction, these were often limited to simple networks or arbitrary design choices. Hence, this thesis offers the first evaluation of state-of-the-art deep learning methods, such as sophisticated regularization, multi-branch networks, and activations/optimizers beyond ReLU and Adam. Further, the thesis studies the utility of deep learning in crowdfunding success prediction by comparing it to its shallow counterpart: a logistic regression with bag-of-words encoding. Such comparison of wide and deep learning has, so far, not been addressed in research, but is crucial to explore if deep learning can improve the performance of crowdfunding predictors and to pinpoint the models’ individual strengths and weaknesses. The findings suggest that deep learning is not a silver bullet, and also linear bag-of-words models can, given the right preprocessing, achieve a staggering accuracy of 82.5%. Nevertheless, deep neural networks possess certain properties, such as the ability to model non-linear relationships, an extensive hyperparameter space, and the ability to construct joint, multi-modal feature maps, that help them to outperform their shallow counterpart. Further, it is possible to combine the advantages of both models by jointly training them in a wide-and-deep-learning approach. The combined model achieves an accuracy of 83.1% and outperforms prior deep learning approaches by a large margin. Lastly, the thesis sheds light on the pitfall of data leakage in crowdfunding success prediction. It finds that prior work has utilized certain attributes that carry a non-trivial amount of information about the target variable and can bias the prediction model. By incorporating those leaking features, it is even possible to achieve an accuracy of up to 100%, given a sufficiently expressive model.

EducationsMSc in Business Administration and Information Systems, (Graduate Programme) Final Thesis
Publication date2021
Number of pages127
SupervisorsDaniel Hardt