Effective and secure federated online learning to rank

TL;DR

The paper tackles privacy-preserving online learning to rank by integrating state-of-the-art OLTR with federated learning. It introduces FP-DGD, a federated adaptation of Pairwise Differentiable Gradient Descent (PDGD) using Federated Averaging, and augments it with $\epsilon$-differential privacy via clipping and Laplace noise to secure gradients. Empirical results on MQ2007 and MSLR-WEB10K show FP-DGD consistently outperforms the previous federated OLTR method FOLtR-ES and remains robust across privacy budgets, with modest trade-offs in highly privacy-conscious settings. The work demonstrates a scalable, privacy-aware approach to federated OLTR, providing practical guidance on parameter settings and releasing the accompanying code for reproducibility.

Abstract

Online Learning to Rank (OLTR) optimises ranking models using implicit user feedback, such as clicks. Unlike traditional Learning to Rank (LTR) methods that rely on a static set of training data with relevance judgements to learn a ranking model, OLTR methods update the model continually as new data arrives. Thus, it addresses several drawbacks such as the high cost of human annotations, potential misalignment between user preferences and human judgments, and the rapid changes in user query intents. However, OLTR methods typically require the collection of searchable data, user queries, and clicks, which poses privacy concerns for users. Federated Online Learning to Rank (FOLTR) integrates OLTR within a Federated Learning (FL) framework to enhance privacy by not sharing raw data. While promising, FOLTR methods currently lag behind traditional centralised OLTR due to challenges in ranking effectiveness, robustness with respect to data distribution across clients, susceptibility to attacks, and the ability to unlearn client interactions and data. This thesis presents a comprehensive study on Federated Online Learning to Rank, addressing its effectiveness, robustness, security, and unlearning capabilities, thereby expanding the landscape of FOLTR.

Figures (5)

• Figure 1: Offline performance (nDCG@10) across datasets, under different click models, averaged across all dataset splits and experimental runs. Shaded areas indicate the standard deviation.
• Figure 2: Offline performance in terms of MaxRR across datasets, under three different click models, averaged across all dataset splits and experimental runs. Shaded areas indicate the standard deviation.
• Figure 3: Offline performance on MSLR-WEB10K under three different click models, averaged across all dataset splits and experimental runs, for different number of clients $|C|$. We study FPDGD with differential privacy (w DP) and without (w/o DP).
• Figure 4: Investigation of the influence of batch size $|B|$ on offline performance for MSLR-WEB10K under three different click models, averaged across all dataset splits and experimental runs. Shaded areas indicate the standard deviation.
• Figure 5: Investigation of the influence of batch size $|B|$ on offline performance for MSLR-WEB10K under three different click models, averaged across all dataset splits and experimental runs, under fixed budget.

Theorems & Definitions (2)

• definition 1
• definition 2