This is the code associated to the paper "Toward a framework for seasonal time series forecasting using clustering" published by C. Leverger et al. in IDEAL 2019 Conference, Manchester.

• Install Khiops on your machine,
• Install python requirements from requirements.txt,
• Replace the string C:/ProgramData/Anaconda3/python.exe by your own python executable path on runit.sh file,
• Run ./runit.sh in a shell,
• After execution, find and check results on res folder.

Note that without Khiops software installed on your machine it is pointless to try to run this piece of software.

Khiops licenses and download available online for research purposes (3 months trials), see https://khiops.diod.orange.com/.

Seeds provided directly on the code.

R scripts for AETSF provided in R folder.

• "om": Orange Money project hits per seconds,
• "om_cpus": Orange Money project CPUs per seconds,
• "aw": Australian walk,
• "melb_temp": temperature in Melbourne,
• "it": internet traffic,
• "rf_stjean": St Jean-related dataset (river flow and rainfall), datasets number 6 & 7 in the paper.
• "rf_niagara": river flow of Niagara,
• "hc": Hourly Consumption electricity, kaggle dataset.

In order to use with your own data, datasets must be recoded. Time series data must be transformed to csv files which have 4 columns, and the header must follow indicated norms and naming conventions. The first column, named "date_", is the date of the current season; this column must follow the format "dd/mm/YYYY". The second column "val_" is the values of the time series. Third column "n_day_" is the identifier of the season (1 being the first season, 2 the second, etc.). Finally, the column "time_" represents the index of the value inside the season (1 being the first value of the considered season, 2 the second, etc.) Moreover, Khiops, the software which is used for co-clustering, requires a particular data format: the data frames must be alphanumerically sorted following the column "n_day_".

We first started the development only analysing days, thus the name of "n_day_" column which is not generic.

Crafting time series clustering for performances comparisons using https://github.com/rtavenar/tslearn / see the original paper "Tavenard, Romain. "tslearn: A machine learning toolkit dedicated to time-series data." (2017)."