Section: New Results

Non-negative Tensor factorization for spatio-temporal data analysis

Participant : Yufei Han.

This is a joint work with Fabien Moutarde from Mines ParisTech.

We investigate the use of non-negative tensor factorization for spatio-temporal data clustering and prediction. In general case, a spatio-temporal signal is represented as a set of multiple-variant temporal sequences. In the domain of intelligent traffic, the temporal records of traffic flow states (free-flowing/congestion) over a specific time duration with respect to hundreds of links in a transportation network can be considered as a simple but direct example of spatio-temporal signal. Both temporal causality between neighboring time sampling steps and spatial layout of the multiple-variant observation captured at each time sampling step are the focus of the spatio-temporal data analysis. Non-negative tensor factorization enables us to project the high dimensional spatio-temporal data into low-dimensional subspace and clustering/prediction can be then achieved on the derived subspace projection easily.

This year's highlights are

• A conference paper describing application of non-negative tensor factorization in traffic flow state prediction and clustering has been published and presented at ITS World Congress [30] ;

• The application of non-negative matrix factorization in clustering network-level traffic flow state in large-scale transportation network has been accepted for publication in a journal [11] .