How to use and train AI models to predict tomorrow's trends ?
L'Oréal, the world's number one cosmetics company, has partnered with isahit to address this problem.
L'Oréal in a few words:
The world's leading beauty group thanks to major innovations
A presence in 150 countries
86,000 employees worldwide
505 patents filed in 2018 alone
Isahit in a nutshell:
First ethical data labeling platform for AI and Data processing in Europe
2000 projects with more than 1700 HITers
HITers established in more than 37 countries
300+ satisfied customers
Trendspotter: L'Oréal's project to detect new trends
In order to understand and detect new trends in the world of beauty, L'Oréal has developed a predictive tool, the trendspotter. This tool will retrieve data samples from social networks from which L'Oréal teams will be able to identify and analyze trends and report them to the product and marketing teams.
The 3 mains objectives of Trendspotter:
- Detect new trends before competitors do
- Predict the evolution of the trend in the future
- Illustrate the trend to gain speed
How does it work?
Step 1: data collection on social networks using APIs
Step 2 : a classifier that sorts and classifies the data corresponding to the beauty universe
The challenges encountered
The tool recovers a very large amount of data which presents some difficulties:
The uncertainty of the data:
The APIs will retrieve a very large amount of data that is not always related to the beauty world and therefore not relevant to L'Oréal. However, we need data that is sufficiently clean to train a classifier that will be able to detect which posts will talk about beauty in a context that interests L'Oréal. This is where isahit can help L'Oréal.
Managing different languages:
L'Oréal is present in 150 countries, so it is necessary to be able to analyze the data in a multitude of languages. It is therefore necessary to be able to find collaborators who speak the targeted languages.
The performance of a machine learning model is always inferior to the quality of the data; it is therefore essential to have quality data collected and a very meticulous labeling to guarantee the performance of the model.
The solution: knowing the data, thanks to isahit
In order to meet these challenges, isahit appeared to be the most appropriate solution because it offers:
- The possibility to work in several languages: isahit works with a very large community of HITers in 37 countries
- A very demanding quality review:
Isahit implements a very strict project monitoring, throughout its realization, trains its contributors and implements tests upstream to ensure that a constant quality is maintained.
- A great agility and a great capacity to adapt resources according to fluctuating needs
Specific processes put in place upstream of the project
In order to ensure that the project was carried out correctly, the L'Oréal and isahit teams carried out specific tests upstream.
On the L'Oréal side, the data science teams defined an action plan to obtain an efficient annotation process and to reduce subjectivity:
1. Extracting and processing a sample of data in-house in a known language
2. Hand labeling of these posts by several data scientists
3. Pooling of results and adjustment of disagreements
4. Elaboration of a presentation with precise rules and instructions
5. Sharing the instructions with isahit
On the isahit side, the dedicated team ensured that the following processes were put in place to guarantee the best possible quality of treatment:
1. Internal transmission of instructions and selection of contributors according to their skills and experience
2. Training of contributors (HITers) and management of their skills development
3. Implementation of a test on a sample of posts
4. Adjustment sessions of the instructions
5. Launching the project on a large volume and in several languages
And of course, throughout the project:
- Monitoring of financial costs from the isahit platform
- Weekly updates with the project manager via e-meeting
- Quarterly meetings with all the parties involved (sales, operations, etc.)
This strategy allowed for continuous monitoring of the following KPIs: project progress, quality rate, costs per project, time per task, updates, volumes, performance, finances.
Feedback from L'Oréal: Why choose isahit?
¤ This collaboration allowed a 50% reduction in the classification errors of L'Oréal's models
¤ isahit's ability to handle multiple languages (English, French, Spanish, Mandarin, Japanese, etc.) allowed them to cover all data
¤ Finally, the real consistency in the quality of the labeling, the reactivity and the flexibility of the support team largely contributed to the success of the project
A double victory: tech & social
Two years after the beginning of our collaboration, we are mutually very happy to be able to report the following results
- 1 million tasks completed
- 5 languages covered
- A positive and lasting impact on nearly 150 of our HITers!
...and the prize for the most beautiful alliance between a startup and a large group organized by Raise Sherpa & Bain & Company Paris ✨