Training On-Device Ranking Models from Cross-User Interactions in a Privacy-Preserving Fashion Marc Najork Google LLC, 1600 Amphitheatre Pkwy, Mountain View, CA 94043, USA najork@google.com ABSTRACT as a ranker) fit nicely into such a framework; other aspects (e.g. en- Personal search is concerned with surfacing content relevant to an forcing k-anonymity thresholds on query and document n-grams) information need (as expressed by a query) from a user’s personal will require new research. The same holds true for other search im- information repository. Since personal corpora are typically much provements that involve learning, such as improving recall through smaller than public ones (particularly the web), recall is more of an synonym expansions trained from query reformulations or result issue. Moreover, since documents are not shared among users, cross- co-clicks [10]. user interaction signals (such as co-clicked results for identical or We hope that this abstract will inspire researcher in Information similar queries) cannot be leveraged in a straightforward manner. Retrieval to explore this exciting new frontier of privacy-safe on- When limited to a single user, interaction signals are typically device personal search. too sparse to be useful as labels or as features in learned ranking functions. REFERENCES [1] Martín Abadi, Úlfar Erlingsson, Ian J. Goodfellow, H. Brendan McMahan, Ilya Bendersky et al. [3] recently described a methodology for lever- Mironov, Nicolas Papernot, Kunal Talwar, and Li Zhang. 2017. On the Protection aging user interactions in the form of clicked search results in a of Private Information in Machine Learning Systems: Two Recent Approches. In way that allowed them to aggregate interactions across the entire 30th IEEE Computer Security Foundations Symposium (CSF). 1–6. [2] Qingyao Ai, Keping Bi, Cheng Luo, Jiafeng Guo, and W. Bruce Croft. 2018. Unbi- user base of a personal search service, by projecting both queries ased Learning to Rank with Unbiased Propensity Estimation. In 41st International and documents into a shared, dense feature space, and training a ACM SIGIR Conference on Research & Development in Information Retrieval (SIGIR). ranking function on these features using result clicks as relevance 385–394. [3] Michael Bendersky, Xuanhui Wang, Donald Metzler, and Marc Najork. 2017. judgments. Using clicks as relevance labels requires accounting for Learning from User Interactions in Personal Search via Attribute Parameteri- the inherent selection bias in click logs, which can be measured zation. In 10th ACM International Conference on Web Search and Data Mining (WSDM). 791–799. through short-lived result randomization experiments on a portion [4] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Bren- of users [7, 12] or learned jointly with the ranking function [2, 13]. dan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. 2016. In the past several years there has been a lot of interest in training Practical Secure Aggregation for Federated Learning on User-Held Data. CoRR abs/1611.04482 (2016). arXiv:1611.04482 machine-learned models in a federated fashion, suitable for on- [5] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam D. Smith. 2006. Cali- device training and inference [8]. To prevent leakage of personal brating Noise to Sensitivity in Private Data Analysis. In 3rd Theory of Cryptogra- information, one can leverage ideas from differential privacy, where phy Conference (TCC). 265–284. [6] Robin C. Geyer, Tassilo Klein, and Moin Nabi. 2017. Differentially Private noise is added to any training record proportional to the sensitivity Federated Learning: A Client Level Perspective. CoRR abs/1712.07557 (2017). of that record [5]. Several recent works have studied the topic of arXiv:1712.07557 [7] Thorsten Joachims, Adith Swaminathan, and Tobias Schnabel. 2017. Unbiased learning with differential privacy in a federated setting [1, 4, 6]. Learning-to-Rank with Biased Feedback. In 10th ACM International Conference In the same time period there has been tremendous interest in on Web Search and Data Mining (WSDM). 781–789. the IR community on privacy-preserving IR, manifested by three [8] Jakub Konecný, Brendan McMahan, and Daniel Ramage. 2015. Federated Opti- mization: Distributed Optimization Beyond the Datacenter. CoRR abs/1511.03575 workshops and two tutorials; see https://privacypreservingir.org (2015). arXiv:1511.03575 for a good overview. [9] Dmitry Lagun, Chih-Hung Hsieh, Dale Webster, and Vidhya Navalpakkam. 2014. Can we adapt the ideas from on-device learning using privacy- Towards Better Measurement of Attention and Satisfaction in Mobile Search. In 37th International ACM SIGIR Conference on Research and Development in preserving federated shared models to personal information re- Information Retrieval (SIGIR). 113–122. trieval? Fundamentally, ranked retrieval from personal corpora [10] Cheng Li, Mingyang Zhang, Michael Bendersky, Hongbo Deng, Donald Metzler, and Marc Najork. 2018. Embedding-based Synonyms for Personal Search. (2018). involves three types of data, all of which are privacy sensitive: Under submission. documents (e.g. files, photos, messages, music, videos etc); queries [11] Shuang Song, Kamalika Chaudhuri, and Anand D. Sarwate. 2013. Stochastic (including query reformulations and refinements over the course of gradient descent with differentially private updates. In IEEE Global Conference on Signal and Information Processing (GlobalSIP). 245–248. a search session), and implicit feedback such as click and attention [12] Xuanhui Wang, Michael Bendersky, Donald Metzler, and Marc Najork. 2016. signals [9]. Much of the existing work in privacy-safe federated Learning to Rank with Selection Bias in Personal Search. In 39th International learning has focused on marrying stochastic gradient descent-style ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). 115–124. optimizations with differential privacy (see e.g. [11]). Some portions [13] Xuanhui Wang, Nadav Golbandi, Michael Bendersky, Donald Metzler, and Marc of the framework for jointly estimating position bias and training Najork. 2018. Position Bias Estimation for Unbiased Learning to Rank in Personal Search. In 11th ACM International Conference on Web Search and Data Mining a ranking function [13] (e.g. using gradient boosted decision trees (WSDM). 610–618. DESIRES 2018, August 2018, Bertinoro, Italy © 2018 Copyright held by the owner/author(s).