#47 Paris Women in Machine Learning and Data Science: timeseries anomaly detection, sequential and reinforcement learning for demand side management, and how AI can fight cybersexism
Last week, we had the privilege of holding our 47th Meetup at Criteo. Criteo holds a special place in our hearts as they hosted our very first meetup back in 2017, making it emotional to return to where it all began!
The evening kicked off with Mariia Vladimirova, a research scientist at Criteo, who provided an introduction to the company and its activities. Following Mariia’s introduction, the usual presentation by Women in Machine Learning and Data Science took place, setting the stage for an engaging evening!
Natalie, the evening’s master of ceremonies, shared some exciting community updates. She announced a new partnership with EU Data Jobs, a job portal that also provides data science salary insights. They recently featured our meetups in one of their blog posts, helping to increase our visibility.
Additionally, Natalie highlighted an upcoming event organised by our friends at AI House, focusing on Generative AI. This meetup is scheduled to take place in Paris on April 30th.
Our opening speaker, Djohra Iberraken, a Data Scientist at GRDF, initiated the conference with a detailed presentation on anomaly detection and data imputation in time series.
As Europe’s largest natural gas distributor, GRDF is at the forefront of integrating advanced data science and artificial intelligence to enhance its operations. The focus on maintaining excellent data quality underpins their successful deployment of innovative business solutions, ensuring reliability and efficiency in natural gas delivery across numerous territories.
GRDF capitalises on diverse data sources, including those from Gazar meters and network sensors, collected on a daily and hourly basis. This substantial data volume supports critical business functions such as forecasting, predictive maintenance, and network sizing. The success of these applications hinges on a robust data quality framework.
During her talk, Djohra highlighted two primary challenges: anomaly detection and data imputation. Confronted with issues like the absence of continuous billing data and a lack of labels, she and her team have adopted both traditional and bidirectional approaches to address these problems. She deeps dive into the methodology used in her slides available below. These strategies have proven highly effective, enhancing both the predictability and stability of the distribution networks.
As they continue to refine these techniques, Djohra also emphasises the importance of balancing performance with resource management, a key consideration in industrial applications.
Tank you for your insightful presentation Djohra!
If you want to know more, check out her slides below:
Our second speaker, Margaux Brégère, gave us a second talk on energy management and more specifically demande side management. As electricity is difficult to store, it is crucial to strictly maintain the balance between production and consumption. The integration of intermittent renewable energies into the production mix has made the management of the balance more complex.
Demand side management traditionally relies on forecasting demand to adapt production accordingly. However, Margaux introduced a method that goes beyond mere forecasting. By harnessing the power of smart meters, which provide instant access to data and communication, her research proposes managing demand by sending incentive signals to control electricity usage actively.
Central to her methodology is the application of reinforcement learning, particularly the use of stochastic multi-armed bandits. This approach allows for an optimal balance between exploring consumer behaviour and exploiting known strategies to maximise efficiency, characterised by an exploration-exploitation trade-off.
Her studies utilise the Upper Confidence Bound (UCB) algorithm to manage and adjust incentive signals based on empirical data, aiming to minimise regret — the difference between the achieved and potential maximum rewards. This model is adept at adjusting price levels based on observed electricity demands and predicted outcomes, ensuring a precise management of energy consumption.
Margaux’s strategies show promising potential for enhancing efficiency in energy distribution systems while accommodating the growing influence of renewable energy sources.
Thank you Margaux for sharing your work with us!
If you want to know more, check out her slides below:
Our last speaker, Chloé Daudier, CEO and founder at Equilys, creative scientist and advocate for feminism, explores the role of artificial intelligence in combating cybersexism. With a personal touch, Chloé outlined the pervasive issue of online harassment and the potential for AI to mitigate it.
Chloé’s journey into this field was sparked by her encounter with a feminist who had been targeted with a barrage of hateful messages following a YouTube video release. Moved by this experience, Chloé leveraged her expertise in AI to address the issue.
Her presentation began by highlighting the vast amount of data available for analysis, illustrating the immense scale of online harassment with an example of 40,000 YouTube comments. Chloé explained to the audience the challenges she faced in sentiment analysis, including difficulties in determining the target of replies and understanding community-specific codes. Her trials with tools like ChatGPT did not yield significant results.
All of these led her to explore deeper analytical approaches like WordClouds and Topic Modeling, as well as delving into author behaviors, which led to the development of a Streamlit interface showcasing the top five comments from principal authors in the discussions. Chloe dispels the myth that harassers hide online: the vast majority are openly aggressive. She presents startling statistics: 84% of the victimes are women and girls, underscoring that this is not confined to feminists alone.
To conclude, Chloé not only highlights the current issues of cyberbullying but also presents interesting research orientations for further development. She calls for action, suggesting a more profound engagement of AI in monitoring and moderating online spaces to create a safer environment for all users.
Thank you, Chloé, for this very important talk!
If you want to know more, checkout her slides below:
Our next event is likely to take place in June. Stay tuned for updates!
If you do not want to miss our events, you can:
🔗 follow us on Twitter, Meetup, and LinkedIn
📑 check our Google spreadsheet if you want to speak 📣, host 💙, help 🌠
📍join our Slack channel for information, discussions, and opportunities
📩 send an email to the Paris WiMLDS team to paris@wimlds.org
🎬 subscribe to our WiMLDS Paris medium page and Youtube channel
📸 follow the global WiMLDS on Instagram, LinkedIn, and Facebook
🔥 share your company or lab’s job offers for free on the global WiMLDS’ website.
