Skip to content

AlgoMathITMO/Dynamic-classifier

Repository files navigation

Dynamic-classifier

Dynamic classifier for estimating the predictability of client's transactional behaviour
See Alexandra Bezbochina, Elizaveta Stavinova, Anton Kovantsev and Petr Chunaev: Dynamic classification of bank clients by the predictability of their transactional behavior; # 166 at ICCS 2022 Main Track.

Content

  • data_prepare.py - data preprocessing, which includes distillation of chosen categories, filling the missed weeks for each customer and calculation of indices for network training and evaluation; returns files D*table.csv and D*indtab.csv (* = 1 or 2 matching D1 or D2 dataset)
  • dynclass_show.ipynb - notebook which performs the dynamic classificator work on each training step;
  • dinclass_auto - does the same without pictures, but with collecting of errors values, returns file D*increm.csv with the metrics of accuracy and AUC ROC for each step;
  • dinckass_noincr - shows how the system works with the incremental learning turned off and also collects of errors values, returns file D*base.csv with the metrics of accuracy and AUC ROC for each step;
  • compare_base - makes plots of collected errors for incremental and not incremental models;
  • dynclass_collect - collects the predictability data by five categories to the predictability profile, returns file D*total.csv;
  • profiles.ipynb - makes pictures out of collected predictability profiles.

    Download the Raiffeisen dataset: https://drive.google.com/file/d/1H0guFNUdggdhrnFHUiWHTgb6y-QxA4tN/view?usp=sharing

  • About

    Dynamic classifier for estimating the predictability of client's transactional behaviour

    Topics

    Resources

    Stars

    Watchers

    Forks

    Releases

    No releases published

    Packages

    No packages published