- 分享主题:TS - 论文标题:SecureML: A System for Scalable Privacy-Preserving Machine Learning - 论文链接:https://www.cs.utexas.edu/~rofuyu/papers/tr-mf-nips.pdf - 分享人:唐共勇 |
---|
1. Summary
【必写】,推荐使用 grammarly 检查语法问题,尽量参考论文 introduction 的写作方式。需要写出
- 这篇文章解决了什么问题?
- 作者使用了什么方法(不用太细节)来解决了这个问题?
- 你觉得你需要继续去研究哪些概念才会加深你对这篇文章的理解?
This paper analyzes two problems existing in the current time series analysis: a. the representation of large time series, that is, the time series with large N and T; b. The time series with missing values due to force majeure can be handled flexibly. Through the above two problems to measure the existing models, the existing DLM and AR can not deal with the above problems well. Due to the need to model high-dimensional time-series data, the TF matrix decomposition model is associated with it. It is good at modeling high-dimensional time series and can also model and predict time series when there are missing values. Therefore, this paper will build a time series prediction model based on TF. TF can only obtain the structural characteristics of time series, but can not capture the correlation between different time series, so it is necessary to introduce another structure to obtain the correlation of time series, and regularization can capture the correlation between time series. Existing studies have combined graph structure with TF, but graph networks can only show a positive correlation, and can only express the correlation between time series through a simple correlation matrix, which makes the captured correlation characteristics inaccurate and difficult to capture the long-term dependence of time series. Therefore, this paper will introduce Temporary Regulated to solve the problems encountered in the above graph network capture correlation, to build the TRMF model.
2. 你对于论文的思考
需要写出你自己对于论文的思考,例如优缺点,你的takeaways
优点:
1.可以解决高维数据以及数据缺失问题
2.将图正则化与矩阵分解结合起来,可以捕捉到时间序列之间的关系
缺点:
1.只能捕获线性的依赖
2.MF的近似分解是存在误差的
3. 其他
【可选】
MF分解