Examining the Success and Failure of Crowdfunding Campaigns using Explainable AI

Abstract

The paper investigates factors that contribute to success or failure of crowdfunding campaigns using binary classification and explainable AI techniques, specifically SHAP and Counterfactual Explanations. A dataset of completed Kickstarter campaigns was used to train two Extreme Gradient Boosting (XGBoost) classification models, one using textual data alone from the campaigns’ description and the second using categorical, numerical and textual features from the campaigners’ profiles. Findings indicate that sentence length in conjunction with text complexity are associated with campaign success. Certain categorical data such as the project’s type show a strong link to success, while textual terms (“stretch goals”) that convey both elements of ambitiousness and risk are also strongly correlated with success. We enhance implications through a third counterfactual explanations model that generates suggestions on how failed projects could have altered the outcome to a favourable one by improving the language and textual features of the proposed idea’s description.