Homepage Pop‑up Recommendation System for Car Purchase Intent: Background, Feature Engineering, Model and Strategy Optimization, and Results

1. Background Introduction

AutoHome is an all‑in‑one automotive platform offering viewing, selection, purchase and usage services. To more efficiently meet users’ car‑buying needs, the homepage pop‑up uses precise targeting to push vehicle series that match users’ purchase intent, improving conversion from traffic to leads and shortening the purchase path.

The core challenges are: (1) accurately identifying users in the car‑buying stage, (2) determining which vehicle series they are interested in, and (3) recognizing users who want to seek better purchase channels through the platform.

2. Project Experience

Feature Optimization – Features include user‑dimensional, vehicle‑dimensional, cross features, statistical and sequential features. User dimensions leverage a rich tag system covering demographics, device, lifecycle stage, activity level, behavior preferences, and inferred purchase intent. Cross features simulate typical purchase‑intent user journeys, and vehicle‑level pre‑training and clustering capture intrinsic vehicle relationships.

Feature fusion adds important features to the model’s upper layers, employs various cross‑feature mechanisms and attention to capture salient user‑vehicle interaction patterns.

Model Optimization – Phase 1 used a classic DeepFM architecture with FM for automatic cross‑feature learning and a deep network for high‑order features. Phase 2 incorporated FM, DCN, FiBiNet and other cross‑feature networks in parallel on the wide side, plus attention mechanisms. Phase 3 introduced an LHUC module with strong bias features for fine‑grained user segmentation and personalized parameters.

Additional customizations include auxiliary loss to learn user lead‑generation intent and online adjustments of loss weighting to prioritize high‑score negative samples, thereby improving lead prediction accuracy.

Strategy Optimization – In practice, the pop‑up strategy applies time‑decay to prioritize real‑time user preferences, filters out vehicle series mismatching user budgets, and adjusts pop‑up frequency based on user activity.

3. Project Effect

The precise targeting transformed random pop‑up delivery into intent‑based delivery, boosting conversion rate (CVR) by 355% and significantly enhancing commercial monetization efficiency.

4. Future Planning

Future work includes using uplift modeling to assess the impact of pop‑up formats on user retention, and further leveraging user‑profile data across scenarios to improve consistency and user experience.

Author: Yu Tingting, Data Platform – Data Application Team, responsible for user‑profile algorithm optimization at AutoHome.