We consider a retail store classification problem supported by a real word dataset. It includes yearly sales from several thousand stores with dirty categorical features on product labels and product hierarchies. Despite the fact that classification is a well-known problem of machine learning, current baseline methods are inefficient due to the dirty nature of data. As a consequence, We propose a practical solution using an intermediary global approximate classification of products, based on local products hierarchies. Activities for a subset of stores are human-labeled to serve as ground truth for validation, and enable semi-supervision. These experiments show the effectiveness of our approach compared to baseline methods.
This work is carried out thanks to the support of the ANRT through the CIFRE (Industrial Agreements for Training through Research) program. The project is supported by Bimedia company and LIAS laboratory (ISAE-ENSMA).
The key directories for experiments are organised as:
- datasets: Contains the Bimedia sales dataset and the validation dataset (with labelled stores) provided for the experiments.
- thesaurusBT: Contains all the files related to the experiments.
- thesaurusBT/algortihms: Contains the Python implementations of models used in our experiments or adaptation of implemented models that cannot be used as they are.
- thesaurusBT/experiments/lib: Contains the Python scripts where most of Python functions are defined.
- thesaurusBT/experiments/preprocessing: Contains the Python scripts and data ressources required for the execution of pre-processing algorithms.
- thesaurusBT/experiments/processing: Contains the Python scripts and data ressources required for the execution of processing algorithms.
- thesaurusBT/experiments/results: Contains the Python scripts for the evaluation and the reporting of workflows (pre-processing + processing) results.
The contribution is developped with the Python programming language. The minimal software requirements for the installation of this package are:
- Python 3
- PIP
- Git
- All operating systems that support Python
At the root Git repository, execute the following command to install the thesaurusBT Python package:
pip install -e .For complete experimental workflow execution, run thesaurusBT Python package. The arguments define the steps to include into the workflow during its execution.
Available arguments are provided by the following command:
python -m thesaurusBT -hAn example of the execution of a complete workflow of sales processing for the labeling of store activities:
python -m thesaurusBT -pre pivot -pro br_svm -cb offIn this example, the pre-precessing technic specified by the argument -pre is an aggregation pivot, the processing technic used for the multi-labelling is a Binary Relevance model with SVM as base classifier specified by the argument -pro and CatBOOST feature selection is not used as specified by the argument -cb.
The results of the preprocessing and processing steps are stored in the outputs directory.
For the execution of all the combinations of experimental worflows, run thesaurusBT Python package with this command:
python -m thesaurusBT -allThe results of the preprocessing and processing steps are stored in the outputs directory.
In order calculate and report the scores archived by the experimentations, run thesaurusBT Python package with this command:
python -m thesaurusBT -resThis command display on the terminal the scores of each workflow exectuted, and create a file results.csv, saved in the outputs directory.
Details the license agreement of TME: LICENSE
- Maxime PERROT (core developer) (Bimedia and LIAS/ISAE-ENSMA)
- Mickael BARON (LIAS/ISAE-ENSMA)
- Brice CHARDIN (LIAS/ISAE-ENSMA)
- Stéphane JEAN (LIAS/University of Poitiers)