Classification of business activities by machine learning: the case of France

Webinar on automatic coding

26 March 2026

About me

• Who am I ?
  • Data scientist at Insee
  • INSEE statistician civil servant
  • working now at the Business Statistics Directorate
  • SIRENE business register

Hi, I’m Nathan Randriamanana. I’m a data scientist at INSEE.

Everyone now, what’s a data scientist ? A data scientist works at the interface between statistics, programming, and domain knowledge. In general, we analyze data, build models, and sometimes use machine learning or computational methods to provide insights that help organizations make informed decisions.

I’m a civil servant, which means I work as an official statistician for the government. I’m currently based in the Business Statistics Directorate, and part of my work involves managing the SIRENE business register, which tracks the full life cycle of companies in France from their creation, through modifications, to their closure, and sometimes even reactivation.

About my job

• I wear two hats:
  • 🎩 Application Administrator: Acting as the functional lead for the SIRENE register. I bridge the gap between business needs and IT by writing functional specifications for system maintenance and evolution.
  • 🧢 Data Scientist: Managing the end-to-end ML workflow for APE classification from model training and delivery to its integration into the production application.

Example from presentation at the Cloud Native Days France 2026 conference regarding cloud technologies

Linkedin

Table of contents

1️⃣ Context

2️⃣ State of play: How do we assign activity codes today ?

3️⃣ Why did we end up using Machine Learning ?

4️⃣ fastText: a simple and efficient classifier

5️⃣ Beyond fastText: introducing TorchTextClassifiers

6️⃣ Questions ?

1️⃣ Context

What is SIRENE ?

System for the Identification of the Register of Enterprises and Establishments

• Sirene is the French national company registry

• When a company registers:

  • sole authority for issuing SIREN and SIRET numbers
  • a main activity code is attributed (APE code)

• Milestones and challenges faced:

  • Refactoring of the Sirene information system Sirene 4
  • NACE revision ➡️ dual coding in NACE rev 2 and 2.1

• End of March 2024:

  • Officially switch from Sirene 3 to Sirene 4

• Processing 5 million registrations across 2025–2026 🚀

• Consequences: Ideal moment to innovate (but under the constraint!)

A one-stop shop

Another key actor: National Industrial Property Institute

• INPI is the operator of the one-stop shop (Single Window)
• This replaces the old paper-based CFE (Business formality centers)
• It is the primary data source for the Sirene registry
• Implementation of the PACTE Law (2019) to simplify the French economy

PACTE law

Since January 1, 2023, all business formalities must be filed online via a single portal.

The flow” of formalities

The administrative landscape

• Siren number: company directory identification system
• Principal activity code (APE)
• Classification on a daily basis
• Different administrations
• Different information systems
• Requirements: quick, responsive and flexible to updated instructions

2️⃣ State of play: How do we assign activity codes today ?

Assign an activy code: different processes

– Thanks to a literal description of the activity of the enterprise during an administrative procedure ● The most important one : when an enterprise is set up, a form is filled in and a large amount of information is sent to Insee. This includes a literal description of the activity by the firm ● New forms can be transmitted if the enterprise changes its activity – Claims : if the enterprises disagree with the activity code given by Insee – Surveys : structural statistical surveys (the so-called ESA, EAP) in which enterprises are asked about the breakdown of their activity

Assign an activy code: two outcomes

• Automatic coding
• Human review

Near-ubiquity of ML

Automatic activity classification

Human review

Model

• Text classification model which uses additional categorical variables
• Historically, a system expert named SICORE as automatic label coding system
• First ML usecase with fastText library
• Originally trained on legacy data annotated partly by the coding engine and partly manually

3️⃣ Why did we end up using Machine Learning ?

The methodology: a legacy

• Methodology is not new
• Automatic coding for difficult cases and manuel check for the easiest one
• Coding systems since 1981 ! 🧙🏻‍♂️

The shock

• data not clean anymore ➡️ Teams got overwhelmed
• automation drops to 30%

Display an exemple of activity description from business formality centers

  BAKERY

Display an exemple of activity description from the one-stop-shop

 I am looking to developpe in the cleaning and maintainance of busines offices: changeing bins, vacuuming, and mopping

4️⃣ fastText: a simple and efficient classifier

A package from FAIR lab

• Up until November 2022, we used an automatic label coding system, called Sicore
  • Based on a training file of encoding examples
  • Drawbacks : if the label did not match an encryption example, no code suggestion was returned. It was then coded manually by a human being
• Since November 2022, we have implemented a new model based on machine
• Very accurate even with literal descriptions that have never been coded before
• A 100 % result even with a low accuracy rate (always outputs an answer …)
• However we have decided to maintain a manual check if the accuracy is not good

Feature extraction

• Word embedding: a method of vectorisation.
• Pre-trained embeddings available in open-source.
• We learn our own word embeddings.
• Additionally, embeddings for:
  • word n-grams and character n-grams.

Linear classifer

• 2 classification methods:
  • Softmax: a single multiclass classifier.
  • One-vs-all: multiple binary classifiers.
• Optimisation: stochastic gradient descent algorithm.
• Loss function: cross-entropy.

fastText model

• fastText: very simple and fast (C++) “bag of n-grams” model.

Handling categorical variables

• Concatenation of the text description with the names and values of the auxiliary variables:

Text	NAT	TYP	EVT	SUR
Cours de musique	NaN	X	01P	NaN

🢃

“Cours de musique NAT_NaN TYP_X EVT_01P SUR_NaN”

• Imperfect method: 3-grams “AT_” or “T_0” used.

Preprocessing

• Preprocessing essential for natural language processing.
• Constraints: simple, light and easily reproducible in Java .

Transformation	Text description
Input	3 D: La Deratisation - La Desinsectisation - La Desinfection
Lower-case conversion	3 d: la deratisation - la desinsectisation - la desinfection
Punctuations removal	3 d la deratisation la desinsectisation la desinfection

Preprocessing

Transformation	Text description
Input	3 D: La Deratisation - La Desinsectisation - La Desinfection
…	…
Numbers removal	d la deratisation la desinsectisation la desinfection
One-letter word removal	la deratisation la desinsectisation la desinfection
Stopwords removal	deratisation desinsectisation desinfection

Preprocessing

Transformation	Text description
Input	3 D: La Deratisation - La Desinsectisation - La Desinfection
…	…
NaN removal	deratisation desinsectisation desinfection
Stemming	deratis desinsectis desinfect

Modeling

5️⃣ Beyond fastText: introducing TorchTextClassifiers

fastText: In Production, but Archived

• fastText: The go-to for text classification at Insee.
• Reliable and high-performing; deployed in 2021 for APE codification…
• …but the repository has been archived since March 19, 2024.

The Stakes

• Lack of library maintenance: eventual risks regarding stability and compatibility.
• Most importantly: it stalls modernization.
• Meanwhile, a highly dynamic Deep Learning/NLP ecosystem: PyTorch, Hugging Face, etc.

2025: Transitioning to PyTorch!

• PyTorch model closely mirrors the fastText architecture: a smooth transition.
• Discussed here:
  • PyTorch allows for customized architectures tailored to our needs (handling categorical variables).
  • Improved training monitoring.
  • Opportunities for modernization: explainability, calibration, higher-performance models…
• A dedicated package to decouple model definition from its implementation.

From torchFastText to torchTextClassifiers

• Evolution from a single package into a toolkit (a unified framework) for text classification with categorical variables.
  • Inspired by other PyTorch toolkits: Pythae for VAEs, TorchSeg for segmentation…
• Conceptualization of the various components within a text classification model.
• Seamless integration with the Hugging Face ecosystem.

Library scope and positioning

From a production standpoint

• If you want to know more, get here and star this repo 🌟

Architecture

Features

• Explainability 🔍 🤖 📊
• Multi-label support 🏷️
• Modular design 🧩⚙️

Get started

1. check out the example notebook 📚 🚀
2. import your own data in Load and preprocess data
3. Choose your section based on your goal
   • torchTextClassifiers: a wrapper to handle them all ☺️ ➡️ Fast implementation: just grab-and-go 🛍️
   • Let’s dive into the different components 🤓 (experts) ➡️ More control: modular customization 🔧 👌 😎

Some tips for data import

Just update the import section and match your column names.
Works out-of-the-box with Parquet or CSV for instance (Can ask our AI chat)

6️⃣ Questions ?