Implementation of SSMF: Shifting Seasonal Matrix Factorization,
Koki Kawabata, Siddharth Bhatia, Rui Liu, Mohit Wadhwa, Bryan Hooi.
NeurIPS, 2021.



  • Please RUN python3, which starts processing DISEASE dataset and produces the SSMF results into “out” directory.
  • Also, please RUN python3 --max_regimes 1 to compare its performance with SMF, which does not assume regime shifts.

Input for SSMF

SSMF expects 3-dimensional numpy.ndarray, whose last dimension corresponds to time points. If you have multi-column dataframes, utils.list2tensor helps convert your data to 3-dimensional array.

Commnad line options

  • --output_dir:
    Path of output directory (default: ‘out’)
  • --periodicity:
    Length of periodicity (default: 52, for yearly patterns of weekly data)
  • --n_components:
    Rank of matrix factorization (default: 10)
  • --max_regimes:
    Maximum number of regimes (default: 50, set at 1 for SMF)
  • --learning_rate:
    Size of gradient steps in each update (default: 0.2)
  • --max_iter:
    Number of iterations to estimate new regimes (default: 1)
  • --float_cost:
    Number of bits in MDL (default: 32)
  • --update_freq:
    Interval of regime estimation (default: 1)


You will need to install following libraries,



If you use this code for your research, please consider citing our paper.

    title={SSMF: Shifting Seasonal Matrix Factorization},
    author={Koki Kawabata, Siddharth Bhatia, Rui Liu, Mohit Wadhwa, Bryan Hooi},


View Github