Matching Of Duplicate Company Names

Installation

# apt install required packages:
sudo apt update
sudo apt install -y zip htop pipenv
  
# clone repo:
git clone https://github.com/SashaMogilevskii/duplicate_names.git

# go to the folder:
cd duplicate_names

# create virtualenv:
pipenv install

# activate virtualenv:
pipenv shell

# run main.py:
cd src
pipenv run main.py

Run

cd src
pipenv run main.py

Architecture & example

Experiments setup

• Hardware

  • CPU count: 1
  • GPU count: 3
  • GPU type: Tesla T4 / Tesla A16 / GTX 1070TI (with CatBoost)

• Training data (in folder data):

  • Ex.1 (database.csv) with unique values name_1
  • Ex.2 (database.csv) with unique values name_1_upd

Preprocessing

• name_1_upd - unique values after processing.
• Preprocessing (services.py):
  • convert string to english
  • replace symbols
  • drop popular stop_words

Our results:

Model	Data	f1	Recall	Presicion
fuzzywuzzy	Ex.1	0.21	----	----
fuzzywuzzy	Ex.1 (unique)	0.42	----	----
tf_idf(400) + LR	Ex.1 (unique)	0.54	0.43	0.92
tf_idf(400) + LR	Ex.2 (unique)	0.63	0.48	0.92
tf_idf(400) + CatBoost	Ex.2 (unique)	0.78	0.52	0.94
tf_idf(550) + CatBoost	Ex.2 (unique)	0.78	0.54	0.94
fuzzywuzzy(PARTIAL_RATIO)	Ex.2 (unique)	0.56	0.47	----
RoBERTa (4 epochs)	Data con_engl	0.96	0.96	0.96

Optuna

Experience with the selection of hyperparameters

• Hyperparameters
  • max_length
  • batch_size
  • epochs

Tree

.gitattributes
.gitignore
.pre-commit-config.yaml
LICENSE
Pipfile
Pipfile.lock
README.md
pylama.ini
pyproject.toml
mypy.ini
data
   |-- data.csv
   |-- data_v2_10_oct.csv
   |-- data_v2_19_oct.csv
   |-- database.csv
models
   |-- __init__.py
   |-- catboost.pkl
   |-- tfidf1.pkl
notebooks
   |-- Baseline.ipynb
   |-- Baseline_model.ipynb
   |-- Baseline_model_10_20_CatBoost_unique.ipynb
   |-- Baseline_model_19_10(TF-IDF)).ipynb
   |-- Baseline_model_19_10.ipynb
   |-- Bert_model.ipynb
   |-- EDA  & Transform data(19_10).ipynb
   |-- EDA  & Transform data.ipynb
   |-- Model_10_22_CatBoost_unique.ipynb
   |-- Model_RoBERTa.ipynb
   |-- optuna_BERT.ipynb
src
   |-- configs
   |   |-- config.py
   |-- image
   |   |-- image_optuna.png
   |   |-- image_hyperparameters.png
   |   |-- image_1.png
   |-- main.py
   |-- models
   |   |-- __init__.py
   |   |-- base_model.py
   |   |-- catboost_model.py
   |   |-- fuzzywuzzy_model.py
   |-- preprocessing.py
   |-- pydantic_models.py
   |-- services
   |   |-- __init__.py
   |   |-- detection.py

Minimum required equipment

CPU: 1 vCPU, 4 Gb RAM

i5-8600K
~75 requests per sec

Scaling

Orchestra of workers using, for example, kubernetes (AWS) number of workers = ceil(number of company names / 3 * 10 ** 5)

Metrics

base_metric: F1-score. additional_metric: recall

justification of the metrics: Sample imbalance led us to choose recall as the main metric.

Contributors

1. Шакиров Ренат
2. Набатчиков Илья
3. Могилевский Саша

For contributing please use linterts and hooks the following commands:

black .
pre-commit run --all-files