SFB 876 - News

Spatio-Temporal Random Fields (STRF)

STRF estimates prediction models for spatio-temporal data. The parameter storage is compressed by removing uninformative parameters in a systematic way. For finding the best parameters via maximum likelihood estimation, a separable optimization algorithm that can be performed independently in parallel in each graph node is implemented in STRF.

The software is released under 4-clause BSD License. This is the most recent implementation of the STRF model that is presented in the article Spatio-temporal random fields: compressible representation and distributed estimation .

Quick start

The following will train a model on smartphone utilization data:

  1. Download and extract the source version of STRF.
  2. Uncompress the source code, change to the corresponding directory and type make to build the learner STRF and the predictor STRFp.
  3. Run ./STRF --data data/APPS_CELLS_BAT_G_3600s.csv --vertexset data/APPS_BAT_G.V --primary day,month,year --timekey slot --vertexw 0 --spatialw 0 --maxiter 25 --omodel apps.model --alpha 0.25 --damp 0.75 to estimate a spatial graph, train a STRF and store the learned model to the file ./apps.model.
  4. The prediction accuracy of the model can be evaluated with ./STRFp --data data/APPS_CELLS_BAT_G_3600s.csv --vertexset data/APPS_BAT_G.V --primary day,month,year --timekey slot --model apps.model --alpha 0.25 --damp 0.75
A detailed description of all command line options will be added soon.

Binary Version

The binary version of STRF that can be downloaded below, requires no additional libraries. For the CUDA accelerated version, a CUDA capable GPU and a recent CUDA driver (tested with 331.49) are required. The files were compiled on Ubuntu 13.04 with g++-4.7.3-1ubuntu1 and CUDA compilation tools, release 5.5, V5.5.0, respectively. The CUDA version is compiled for devices with compute capabilities 1.3, 2.0, 3.0 and 3.5.


Binary versions for windows will be added soon.