The aim of ReviVal project is to provide an accurate and scalable recovery technique for blocks of missing values in real world hydrological time series. This project is divided into two main parts.
The first part of this project deals with the accuracy issue. In fact, we propose to use an iterated version of matrix decomposition techniques (SVD, CD) to recover blocks of missing values in time series. Using the correlation value between time series, the proposed technique should learn, from the history of the time series that contains the missing blocks together with the histoty of the other time series the type, the shape and the amplitude of the missing bocks.
The second part of the project deals with the scalability issue. In fact, the state-of-the-art solution to perform the Centroid Decomposition (CD) computes a square correlation matrix with quadratic space complexity which makes it not scalable for time series with long history. We propose to reduce the space complexity from quadratic to linear. In order to achieve this goal, the correlation matrix should not be contructed and a vector should be used instead.
For each part of ReviVal project, an analytical and empirical study should be presented.