Success Prediction on Crowdfunding with Multimodal Deep Learning

Chaoran Cheng, Fei Tan, Xiurui Hou, Zhi WeiIJCAI 2019 · DOI

TL;DR

This paper tackles the challenge of predicting crowdfunding project success using a multimodal deep learning approach that integrates text, images, and metadata from project profiles. By leveraging these diverse data sources, the authors demonstrate that image features significantly improve prediction accuracy, especially for projects with limited textual descriptions, offering a more comprehensive and timely prediction tool for creators and platforms.

What Problem Does This Paper Address?

Crowdfunding platforms like Kickstarter have become vital for individuals and businesses to raise funds. Predicting whether a project will succeed (i.e., reach its funding goal) is crucial for both creators and platforms. While many existing approaches focus solely on textual descriptions or rely on post-launch dynamic data (like comments or funding progress), this paper addresses the limitations of these methods. Specifically, it highlights that:

  1. Text-centric bias: Most prior work overlooks the significant role of visual information (images) in attracting backers and conveying project appeal.
  2. Late prediction: Existing methods often require data that only becomes available after a project has been live for some time, preventing timely prediction and proactive adjustments by creators or strategic decisions by platforms.
  3. Limited multimodal integration: There's a need for a principled way to combine heterogeneous information from text, images, and metadata to capture complex interactions that drive success.

The paper aims to develop a model that can predict project success using only pre-launch profile information, thereby enabling early predictions and leveraging the full spectrum of available data, especially visual content.

Methods

The core of this paper is the development of a Multimodal Deep Learning (MDL) framework. This framework consists of three distinct branches, each designed to process a different modality:

  • Textual Branch: This branch uses word embeddings (GloVe) to represent textual descriptions, capturing semantic relationships between words.
  • Visual Branch: This branch leverages pre-trained Convolutional Neural Networks (CNNs), specifically VGG16, to extract features from project images. These features are then aggregated to represent the visual content of a project.
  • Meta Branch: This branch processes metadata such as campaign category (using one-hot encoding) and funding goal (using binning to create discrete ranges).

At the end of each branch, the extracted features are concatenated to form a unified multimodal representation. This combined feature vector is then fed into a fully connected layer for final prediction. The model is trained using a cross-entropy loss function. The authors also explore different aggregation strategies for visual features and compare their approach against baseline methods like Support Vector Machines (SVM) using Bag-of-Words (BoW) for text and Bag-of-Visual-Words (BoVW) for images.

Datasets

The study uses a dataset scraped from Kickstarter, the largest crowdfunding platform in the U.S. The data was collected over two weeks and includes textual profiles, images, category, funding goal, and funding dates for campaigns launched before 2015.

Key aspects of the dataset and its preparation:

  • Filtering: Only successful and failed projects were kept. Textual profiles with fewer than 12 words and images smaller than 200 pixels were removed. Projects in skewed categories were also excluded.
  • Data Splitting: To evaluate performance on shifting distributions, the data was split by year: 2015-2016 for training, 2017 for validation, and 2018 for testing.
  • Feature Dimension Truncation: To manage computational cost and address image fusion sparsity, text descriptions were truncated to the 0.9 quantile of word count (725 words), and image numbers were limited to the 0.9 quantile (18 images).

No new benchmark datasets were introduced; the focus was on analyzing an existing platform's data.

Key Insights and Ideas

  • Images Matter, A Lot: The most striking insight is the significant contribution of image features to success prediction, especially for projects with minimal text. This validates the intuition that visuals are a powerful tool for attracting backers.
  • Multimodality is Key: Combining text, images, and metadata yields superior prediction performance compared to using any single modality alone. The interaction between these different data types is crucial.
  • Deep Learning Outperforms Traditional Methods: The deep learning-based feature extraction (especially from pre-trained CNNs for images) significantly outperforms traditional methods like BoVW, highlighting the power of transfer learning in computer vision for this domain.
  • Pre-posting Prediction is Feasible: The framework's reliance on pre-launch data makes it valuable for early prediction, benefiting both creators (for profile refinement) and platforms (for project selection).
  • Metadata Adds Distinctive Signals: While text and images are intuitive, metadata (like category and funding goal) also provides valuable, distinctive signals for prediction.

Key Results

  • The proposed MDL-TIM (Text, Image, Metadata) model achieved the best performance across various metrics (Recall, Precision, F1-Score, AUC).
  • Compared to the baseline SVM-BoW (text only), the MDL-TIM model showed a significant improvement in AUC (0.8326 vs. 0.7356).
  • An ablation study revealed that removing text data (IMG/Meta) resulted in the worst performance, indicating text's general importance. However, the MDL-Image/Meta model surprisingly outperformed MDL-Text/Meta in scenarios with more images (image level 4), underscoring the growing importance of visuals.
  • The study found that higher image levels generally led to better prediction, while the benefit of more text was not as consistently observed, further emphasizing the impact of visual content.

Limitations and Future Work

  • Image Fusion Sparsity: The authors acknowledge that dealing with the sparsity of image fusion remains a challenge, even with their approach.
  • Data Scarcity for Certain Modalities: While the dataset was substantial, the paper notes that the number of words and images in some profiles were limited, potentially affecting the robustness of predictions for those cases.
  • Computational Cost: The use of deep learning models, especially for image processing, can be computationally intensive. The paper mentions truncating feature dimensions to manage this.
  • Future Work: The authors suggest that further exploration of different fusion techniques and more sophisticated multimodal representations could lead to even better performance. Investigating the impact of other modalities like videos or creator social network information could also be fruitful.

Summary updated

BibTeX
@inproceedings{cheng2019success,
  author = {Chaoran Cheng and Fei Tan and Xiurui Hou and Zhi Wei},
  title = {{Success Prediction on Crowdfunding with Multimodal Deep Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/299},
}
Figure