How we use AutoML, Multi-task learning and Multi-tower models for Pinterest Ads

Pinterest Engineering
13 min readintermediate
--
View Original

Overview

This article discusses how Pinterest utilizes AutoML, Multi-task Learning, and Multi-tower models to enhance its Ads marketplace. It highlights the importance of predicting user engagement and describes the technical implementations that lead to improved ad performance.

What You'll Learn

1

How to implement AutoML for feature engineering in ad prediction models

2

Why multi-task learning improves model efficiency and performance

3

How to apply multi-tower architecture to separate data sources in machine learning

4

When to use calibration techniques to enhance model predictions

Prerequisites & Requirements

  • Understanding of deep learning concepts and model calibration
  • Experience with machine learning frameworks and model training(optional)

Key Questions Answered

What is AutoML and how is it used in Pinterest Ads?
AutoML at Pinterest is a self-contained deep learning framework that automates feature injection, transformation, model training, and serving. It simplifies feature engineering by allowing models to learn directly from raw signals, resulting in improved click-through rate (CTR) predictions.
How does multi-task learning benefit ad engagement predictions?
Multi-task learning allows Pinterest to estimate various user engagement rates simultaneously, such as good clicks and scroll ups. This approach simplifies the system, reduces infrastructure costs, and enables knowledge transfer across different engagement objectives.
What are the key components of the multi-tower model architecture?
The multi-tower model architecture consists of a shared bottom layer that captures common knowledge from different data sources, while separate towers for each source isolate their unique characteristics. This structure helps improve performance by reducing interference between training examples from different sources.
What calibration techniques are used to improve prediction accuracy?
Pinterest employs a lightweight Platt Scaling model for calibration, which is trained with contextual, creative, and user profile signals. This approach has reduced calibration errors by up to 80%, enhancing the reliability of engagement predictions.

Key Statistics & Figures

Reduction in calibration error
80%
This reduction was achieved through the implementation of a lightweight Platt Scaling model for calibration.

Technologies & Tools

Machine Learning Framework
Automl
Used for automating feature engineering and model training in ad prediction.
Machine Learning Model
Deep Neural Networks (dnn)
Utilized for predicting user engagement rates in ads.

Key Actionable Insights

1
Implement AutoML to streamline feature engineering processes in your ad prediction models.
Using AutoML can significantly reduce the manual effort involved in feature engineering, allowing data scientists to focus on model performance rather than tedious data preparation tasks.
2
Adopt a multi-task learning approach to improve the efficiency of your machine learning models.
By training models on multiple related tasks simultaneously, you can leverage shared knowledge, which can lead to better performance and reduced computational costs.
3
Utilize a multi-tower architecture to handle diverse data sources effectively.
This structure allows you to maintain the unique characteristics of different data sources while still benefiting from shared learning, which can enhance overall model accuracy.
4
Incorporate calibration techniques to ensure your model predictions are reliable.
Calibration can significantly improve the confidence of your model's probability predictions, which is crucial for applications like ad ranking where accuracy impacts user experience.

Common Pitfalls

1
Relying too heavily on traditional feature engineering can lead to suboptimal model performance.
Many data scientists may underestimate the value of automated feature engineering, which can streamline processes and improve model accuracy.
2
Ignoring the importance of calibration can result in poorly calibrated predictions.
Without proper calibration, models may produce misleading confidence levels, which can negatively impact decision-making in applications like ad ranking.

Related Concepts

Machine Learning
Deep Learning
Feature Engineering
Model Calibration