Retraining vs Fine-tuning or Transfer Learning?
The article discusses the pros and cons of retraining machine learning models from scratch versus fine-tuning or transfer learning when new data comes in daily for an e-commerce clickstream project.
Why it matters
Determining the optimal model training strategy is crucial for maintaining accurate and up-to-date predictions in fast-moving e-commerce environments.
Key Points
- 1The project involves using XGBoost to find user intent, price sensitivity, and segment users
- 2The author is considering retraining the models from scratch using a mix of recent and older data
- 3Alternatively, the author is considering fine-tuning the models on the new daily data
- 4The author is looking for guidance on the best approach and resources to learn more
Details
The article discusses a project that involves using machine learning techniques like XGBoost and Linucp/Thompson sampling to analyze e-commerce clickstream data. The key question is whether it's better to retrain the models from scratch daily or to fine-tune the existing models on the new data. The author proposes a retraining strategy that uses 100% of the last 30 days of data, 50% of the 30-90 day data, and 10% of the 90-180 day data to avoid data accumulation and capture the latest trends. The author is looking for guidance on the best approach and resources to learn more about retraining versus fine-tuning or transfer learning in this type of scenario.
No comments yet
Be the first to comment