Overview
The article discusses how Pinterest utilized machine learning to enhance user engagement globally, particularly through improvements in the home feed experience. Key achievements include a 250% increase in localized Pins for international users and a 10% rise in daily Pin saves due to the implementation of gradient boosted decision trees (GBDT).
What You'll Learn
1
How to implement gradient boosted decision trees for content ranking
2
Why language and country matching improves user engagement
3
How to adaptively train machine learning models for better performance
Prerequisites & Requirements
- Understanding of machine learning concepts and algorithms
- Familiarity with Apache Hive and xgboost(optional)
Key Questions Answered
How did Pinterest increase the number of localized Pins for international users?
Pinterest increased the number of localized Pins for countries outside the U.S. by 250% over the past year by leveraging machine-learned models to better understand international content and user preferences. This localization effort significantly improved user engagement and retention.
What improvements were made to the home feed engagement?
The implementation of gradient boosted decision trees (GBDT) led to a more than 10% increase in the number of people saving Pins from their home feed daily. This was particularly pronounced for users outside the U.S., where relevant Pins saved increased by up to 18%, depending on the country.
What is adaptive training in GBDT models?
Adaptive training in GBDT models involves specifying different instance weighting schemes for different trees during training. This approach allows for tailored learning based on the importance of specific instances, leading to improved relevance predictions and user engagement.
Key Statistics & Figures
Increase in localized Pins
250%
Achieved for countries outside the U.S. over the past year
Increase in daily Pin saves
more than 10%
Resulting from the implementation of GBDT in the home feed
Increase in relevant Pins saved by international users
up to 18%
Depending on the country
Technologies & Tools
Some links below are affiliate links. We may earn a commission if you make a purchase.
Backend
Apache Hive
Used to generate training data for GBDT models
Machine Learning
Xgboost
Utilized to train GBDT models offline
Backend
Spark
Used for distributively training GBDT models on larger datasets
Key Actionable Insights
1Implement gradient boosted decision trees to enhance content relevance in user feeds.Using GBDT can significantly improve user engagement metrics, as evidenced by Pinterest's success in increasing daily Pin saves by over 10%. This technique is particularly effective when dealing with complex feature spaces.
2Incorporate language and country match features to cater to international users.By aligning content with users' language preferences, platforms can increase the number of saved Pins by 10 to 20%. This approach is crucial for global companies aiming to enhance user experience across diverse markets.
Common Pitfalls
1
Failing to localize content for international users can lead to decreased engagement.
Without considering language and cultural preferences, platforms risk alienating users, resulting in lower retention and interaction rates.
Related Concepts
Machine Learning Algorithms
Gradient Boosted Decision Trees
Content Localization Strategies