Introduction
In this blog post, I present DICE (Deep Interest Network for Conversion Estimation), a novel approach to personalized conversion rate (CVR) prediction in recommendation system. DICE leverages users' historical interactions to provide more accurate CVR estimates, which are crucial for optimizing ad placements and bids.
Model Architecture
DICE builds upon recent advancements in attention mechanisms and sequential modeling. The core components of the architecture are:
- Input Embedding Layer:
- Processes user's past 90 days of purchases and 14 days of clicks
- Utilizes three types of embeddings: a) ASIN ID embedding b) Brand embedding c) Title-based text embedding (using M5 behavioral foundation model)
- Multi-head Attention Mechanism
- Cross-feature Interaction Layer: Captures complex relationships between different feature types.
- Prediction Layer: Outputs the final CVR estimate.
Key Innovations
- Personalization: Unlike previous models, DICE learns user-specific preferences from historical data.
- Temporal Awareness: Incorporates timestamp information to weigh recent interactions more heavily.
- Multi-dimensional Product Representation: Uses multiple embedding types to capture various aspects of products.
- Attention Mechanism: Allows the model to focus on the most relevant parts of a user's history for each prediction.
Conclusion and Future Work
DICE represents a significant step forward in personalized CVR prediction for e-commerce. By leveraging deep learning techniques and rich user history, it provides more accurate estimates that can improve ad relevance and efficiency.
Future work will focus on:
- Testing real-time purchase signal integration
- Exploring new attention architectures for diverse behavior types
- Exploring gating mechanism and experts architecture to dynamically adapt model parameters
- Further optimizing bid adjustment mechanisms
We believe these advancements will continue to push the boundaries of what's possible in personalized recommendation systems.