Kickstarter Crowdfunding Campaign Dataset

Crowdfunding campaigns frequently fail to reach their funding goals, posing a significant challenge for project creators. To address this issue and empower future crowdfunding stakeholders, accurate prediction models are essential. This study evaluates the relative significance of diverse modalities (visual, audio, and text) in predicting campaign success. By utilizing Bidirectional Encoder Representations from Transformers (BERT) to capture rich contextual information within these modalities and employing eXtreme Gradient Boosting (XGBoost) for predictive modelling, we aim to identify crucial factors influencing campaign outcomes. The experiment results show that contextual meaning in each visual modality and audio modality reaches an F1 score of 84%, and text modality reaches an F1 score of 80%. The combination of the three modalities reached an F1 score of 81%. Among the three modalities, text was identified as the most representative feature. In particular, our findings also show that project owners should focus on emotional sequences in the project description to increase the success rate.