A scalable privacy-preserving information retrieval system based on federated optimization, on-device intelligence and semantic matching

今年我们的主要任务是开发基于大规模数据集的联邦信息检索模型。为此，我们首先将信息检索视为用户和搜索引擎系统之间的交互过程，而不是每个查询的独立排名。其次，我们探索了如何有效地整合在线学习排名和联合学习。具体地说，在线学习排名使我们能够应对上述互动过程。联合学习使我们能够通过在设备上学习排名模型来处理隐私问题。在几个基准排序数据集上的一系列实验结果表明，该方向是可行的。

This year our main task is to develop the federated information retrieval model based on a large-scale dataset. To this end, we firstly view information retrieval as an interactive process between users and the search engine system rather than an independent ranking per query. Secondly, we explore how to effectively integrate online learning-to-rank and federated learning. Specifically, online learning-to-rank enables us to cope with the aforementioned interactive process. Federated learning enables us to deal with the privacy issue by learning the ranking model in an on-device manner. Based on a series of experiments on several benchmark ranking datasets, our experimental results show that this direction is applicable.