[1G5-OS-13c] OS-13

6:10 PM - 6:25 PM JST(9:10 AM - 9:25 AM UTC)

[1G5-OS-13c-03] Predicting Conversion from Inquiry FormsLLM-based Feature Extraction and External Information Retrieval

〇TAKEMI OHAMA¹ (1. Nyle Inc.)

Keywords:

Marketing,business,machine learning

This study proposes a method for predicting contract probability from B2B inquiry forms using Large Language
Models (LLMs). In B2B contexts, inquiry forms typically contain minimal information to prevent user drop-off,
making traditional machine learning approaches challenging due to limited features and small sample sizes. We
addressed these challenges by: (1) extracting motivation and capability scores from free-text using Claude 3 Haiku,
(2) autonomously collecting company attributes through AI agent web search using only company name and email
domain, and (3) applying cyclic encoding and cross-terms to prevent overfitting. Using 1,794 samples from an
SEO consulting service, our CatBoost model achieved Test AUC 0.8503 with Train-Test Gap of 2.94pt. The
company attributes collected by AI agent contributed 59.7% of feature importance, demonstrating the effectiveness
of external information augmentation for small B2B datasets.