GemNet2 : A breakthrough in online matching
Today, I'm very thrilled to announce a major breakthrough in applying AI to human resource management by the HrFlow.ai (ex: Riminder.net) team.
#GemNet2 has been validated as a new state-of-the-art for the ‘online matching & next job prediction problem’ achieving :
• 96.84% weighted-accuracy (Vs. 83% in 2018) and 99.31% AUC on a test set of 81986 job offers
• an inference speed of 10 million predictions per second for profiles and jobs (Vs. 5000/23 seconds in 2015, cf. Eric Cohen report).
The current result has been 5 years in the making.
I started GemNet (General Enterprise Model Network) as a research project in 2015 under the supervision of Pr. Matthew Blaschko (KU Leuven), Pr. Pawan kumar (DeepMind), Pr. Gilles Faÿ (CentraleSupelec) and Pr. Alessandro Lazaric (Facebook).
It has required an extraordinary amount of sacrifices from the team, and I'm very proud of the latest achievements lead by the amazing Thomas ZHU, to push the boundaries of the stats-of-the-art.
The embeddings retrieved from the model for profiles and jobs shows also have excellent transfer learning capabilities as the training starts :
- > 82% AUC for profiles
- > 95% AUC for jobs
We hope it will have a big impact, especially during the current Pandemic as :
• millions of people are not able to find work
even sectors from IT to healthcare are struggling to fill open position.
• companies are rushing to re-deploy their workforce
These results wouldn't be possible without :
• The dedicated financial support of Bpifrance and Région Ile-de-France
• The training capacity and sponsorship of Amazon Web Services
• These results wouldn't be possible without our challengers:
Criteo (GemNet0 2016), EDF (GemNet1 2018), Safran & Talentsoft (GemNet2 2020).