KAIST GSDS Interview Presentation: Multimodal Purchase Prediction
📝 Project Overview
This project was presented as part of the interview for KAIST GSDS admission. Drawing from my background in Business and Computer Science, I proposed a research agenda to model consumer purchase behavior using multimodal data—images, texts, and behavioral logs—while integrating cultural factors such as consumer nationality.
The goal was to explore how cultural characteristics influence model performance and whether locally fine-tuned models outperform global ones in e-commerce applications.
🔬 Methodology & Research Direction
- Data Types:
- Text: Reviews, product titles, search terms
- Image: Product photos, user-uploaded images
- Behavioral: Click logs, page dwell time, purchase conversion
- Metadata: Price, brand, discount rate, consumer nationality
- Models:
- Multimodal Transformers (e.g., MMBT, CLIP)
- Attention-based interpretability and contribution analysis
- Research Questions:
- Can multimodal models outperform single-modal ones in predicting purchase behavior?
- What features or modalities contribute most to consumer decisions?
- Do culturally localized models perform better than global ones?
- How does model performance change when consumer nationality is explicitly modeled?
📈 Key Thoughts
- Multimodal approaches can better capture ‘complex’ purchase behaviors, especially in culturally diverse populations.
- National and cultural traits significantly influence features like price sensitivity, product preference, and visual attention.
- Model interpretability (e.g., via attention weights) could provide actionable insights for both marketing and recommendation systems.
🎤 Personal Role
This was a solo project proposal developed specifically for the KAIST GSDS interview process. I designed the topic, conducted background literature review, and created all presentation materials to demonstrate alignment with the school’s research direction.