2021 Fiscal Year Annual Research Report
A scalable privacy-preserving information retrieval system based on federated optimization, on-device intelligence and semantic matching
| Project/Area Number |
19H04215
|
|---|
| Research Institution | University of Tsukuba |
|---|
Principal Investigator |
于 海涛 筑波大学, 図書館情報メディア系, 准教授 (30751052)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
吉川 正俊 京都大学, 情報学研究科, 教授 (30182736)
康 シン 徳島大学, 大学院社会産業理工学研究部(理工学域), 助教 (80777350)
|
|---|
| Project Period (FY) |
2019-04-01 – 2024-03-31
|
| Keywords | Metric Optimization / Policy Gradient / Differential Privacy |
|---|
| Outline of Annual Research Achievements |
This year our first task is the on-device intent-detection model. We found that performing diversified ranking can be adopted as the key technique for solving queries with multiple intents. To this end, we proposed a new diversified ranking model based on direct metric optimization. The results show that our model achieves the state-of-the-art performance. Our second task is the framework design for federated optimization in the context of information retrieval. To this end, we focus on how to integrate learning-to-rank and federated learning. Our experimental results show that this direction is applicable.
|
| Current Status of Research Progress |
Current Status of Research Progress
3: Progress in research has been slightly delayed.
Reason
Due to the impact of COVID-19, the Government of Japan have declared a state of emergency several times. The research work has to be conducted either online or at home. When using the research room, the number of students is limited. As a result, the efficiency is impacted to some extent.
|
| Strategy for Future Research Activity |
The major research objective of this year is to develop the federated information retrieval model. The recent privacy preserving machine learning techniques, e.g., local differential privacy, will be employed in order to preserve the privacy when collecting model updates from individual clients. Secondly, a multi-round evaluation will be conducted to test and refine the federated optimization framework. As the third task, we will construct the prototype system by seamlessly integrating the models proposed in previous years.
|
| Remarks |
This open-source project, referred to as PTRanking (Learning-to-Rank in PyTorch) aims to provide scalable and extendable implementations of typical learning-to-rank methods based on PyTorch.
|
