Traffic data imputation via tensor completion based on soft thresholding of Tucker core

José Henrique De Morais Goulart, Alain Kibangou

Abstract:

Technological limitations and practical difficulties cause inevitable losses of traffic data in the typical processing chain of an intelligent transportation system. This has motivated the development of imputation algorithms for mitigating the consequences of such losses. As
the involved datasets are usually multidimensional and bear strong spatio-temporal correlations, we propose for traffic data imputation a tensor completion algorithm which promotes parsimony of an estimated orthogonal Tucker model by iteratively softly thresholding its core. The motivation of this strategy is discussed on the basis of characteristics typically possessed by real-world datasets. An evaluation of the proposed method using speed data from the Grenoble south ring (France) shows that our algorithm outperforms other imputation methods, including tensor completion algorithms, and delivers good results even when the loss is severely systematic, being mostly concentrated in long time windows (of up to three hours) spread along the considered time horizon.