[Paper Review] Federated Learning for Ranking Browser History Suggestions

Florian Hartmann, Sunah Suh+3 more|arXiv (Cornell University)|Nov 26, 2019

Privacy-Preserving Technologies in Data18 references38 citations

TL;DR

The paper demonstrates a privacy-preserving federated learning system to optimize Firefox URL bar history and bookmark suggestions, replacing handcrafted frecency weights and reducing typing effort by over half a character. It shows robustness for deployment and reports large-scale real-user training.

ABSTRACT

Federated Learning is a new subfield of machine learning that allows fitting models without collecting the training data itself. Instead of sharing data, users collaboratively train a model by only sending weight updates to a server. To improve the ranking of suggestions in the Firefox URL bar, we make use of Federated Learning to train a model on user interactions in a privacy-preserving way. This trained model replaces a handcrafted heuristic, and our results show that users now type over half a character less to find what they are looking for. To be able to deploy our system to real users without degrading their experience during training, we design the optimization process to be robust. To this end, we use a variant of Rprop for optimization, and implement additional safeguards. By using a numerical gradient approximation technique, our system is able to optimize anything in Firefox that is currently based on handcrafted heuristics. Our paper shows that Federated Learning can be used successfully to train models in privacy-respecting ways.

Motivation & Objective

• Demonstrate that Federated Learning can be used in a major software product for privacy-preserving model improvement.
• Develop a robust optimization process suitable for deployment with limited feedback signals from user interactions.
• Replace handcrafted heuristic weights in the frecency-based ranking with learnable parameters while ensuring safety and interpretability.
• Provide mechanisms for gradient estimation and secure aggregation suitable for real-world Firefox deployment.
• Make the approach reusable for optimizations beyond the URL bar by maintaining a general, computable gradient framework.

Proposed method

• Implement a federated learning loop where a subset of clients computes local updates from privately held interactions and sends updates to a server for aggregation.
• Use a pointwise SVM-style ranking loss adapted for the frecency-based ranking to train the model to favor the selected item with a margin Delta.
• Adopt Rprop as the gradient-descent optimizer to bound updates and adapt per-weight step sizes for stability and interpretability.
• Approximate gradients via finite-difference methods to accommodate non-differentiable or hard-to-change code paths within Firefox, enabling gradient estimates without rewriting core code.
• Utilize Shield for rapid prototyping and a controlled upgrade path, with simulations to validate ideas before real deployment.
• Apply safeguards: initialize from the existing frecency model, bound updates, enforce nonnegative weights, and maintain newer visits’ weights higher than older ones.

Experimental results

Research questions

• RQ1 Can Federated Learning effectively improve the ranking of URL bar suggestions in a major software product without collecting user data?
• RQ2 How can the optimization be made robust and stable enough for real-world deployment with limited, noisy, user-driven feedback?
• RQ3 Can a black-box, handcrafted heuristic-based ranking system be re-parameterized and improved using gradient-based learning in a privacy-preserving manner?
• RQ4 What gradient estimation methods and privacy-preserving aggregation strategies are practical for large-scale browser deployments?

Key findings

• The new model reduces the number of characters typed before selecting a suggestion by over half a character.
• The system was deployed to a large fraction of Firefox Beta users and trained with data from roughly 360,000 participants.
• The optimization ran for 137 iterations across a near-three-day online phase, with millions of gradient updates and billions of server interactions.
• The approach demonstrates that Federated Learning can be applied to a major product outside simulations, with practical safeguards and stability considerations.
• The design uses a pointwise adaptation of a ranking loss and an Rprop-based optimizer to ensure stable, bounded updates.
• An anonymized dataset and open-source code were prepared to accompany the study (dataset forthcoming).