The Demystifying Data Science conference is an annual free online conference hosted by Metis, a company that provides data science training for companies, individuals and institutions. As someone who has only very recently found their love for data science, I registered as soon as I found out about the event on Twitter, and seen as not everyone might have been able to follow the live event (often due to pesky jobs and time zones), I decided to collect the notes I took from the 8 amazing speakers that dropped tons of knowledge on all of us on Day 1. If you find this interesting, make sure to follow the speakers on Twitter, as well as myself at @alepoptosis.
“Don’t try to boil the ocean.”
Tarry spoke about the different levels of data science that lead to the industrial applications of AI as 5 plateaus, which need to be climbed progressively. These plateaus are:
1. Fundamentals of data science: statistics, probability, mathematics, computer programming. These are to be regularly revisited and refreshed, especially those that are relevant to the applications one is directly invested in at the time.
2. Data visualisation tools: these are technologies like Tableau, the matplotlib and seaborn packages for Python, ggplot for R and so on. He suggested to learn these tools when relevant for the project at hand, and to focus one’s energy on one of these at a time, to really understand it, instead of only having a vague knowledge of all of them.
3. Machine learning: supervised and unsupervised machine learning tasks and their relevant algorithms. Again, Tarry stressed the importance of not spreading oneself too thin, and that the most important thing is to make sure one is able to understand the data of the project at hand. Someone who reaches this plateau should already consider themselves a data scientist.
4. Deep learning
5. Applied AI
The point I think Tarry tried to get across - as well as many other speakers - is that nobody is ever going to be an expert in everything, therefore the most important thing is to have a good grasp of the fundamentals and be ready to dive into new tools and technologies when needed.
Kate listed a very interesting series of skills that would make an excellent data scientist, or data science team. I found these particularly interesting, as people (including me) are often at a loss when trying to think of skills to showcase in their CV, especially so-called “soft skills”. I transcribed the entire list here for future reference.
Business/Domain knowledge:
Technical skills:
Soft skills:
Apart from this list of skills, Kate also mentioned a very interesting initiative called Makeover Monday, which provides the community with a new dataset every week, inviting people to produce beautiful and informative ways to visualise that data. I certainly will do my best to participate whenever I can now!
Kunal presented the idea of structured thinking, i.e. bringing a framework to an unstructured problem. Data scientists are usually required to do so with complex problems, in an hopefully unbiased way, and sometimes with little experience - which is why Kunal suggests using structured thinking. He indicates that, especially, it should be used every time one does something new, something big, or something which has a very high cost of failure. He introduced an acronym he uses to turn a business problem into a data problem: TOSCAR. He indicates it as a technique to create a problem statement, which is something that should be defined before touching any data.
TOSCAR:
Kunal also gave two book suggestions:
“Show AND tell!”
Debunking statements such as “Show, don’t tell”, and “A picture is worth a thousand words”, Alberto talked about how data visualisation can be ambiguous, even if many think it is inherently intuitive to human beings.
He sees the problem as a lack of shared mental models, i.e. the way the creator of a visual interpreted the data behind it and expects it to be understood does not often match the way the viewer does.
This is not only something that applies to people that are new to data visualisation. Alberto talked about how, when seeing a visualisation, everyone creates a mental descriptions of it in their head. He finds that it is very easy to create a wrong one, at all levels of expertise, and recommended to always double check before drawing any conclusion.
For these reason, Alberto suggested testing our own visualisations with non-experts, trying not to bias them in the process, to make sure they can be easily accessed by anyone. When this is not the case, he believes it is important to make sure to explain the logic behind a graphic with as much text or words as one needs, without fear of sounding redundant.
Finally, he encouraged creators always try their best to balance oversimplification and overcomplication in their visualisations.
Apart from his upcoming book, How Charts Lie, Alberto also recommended that listeners check out the work of Hans Rosling, especially when it comes to his way of presenting visualisations, as well as his book Factfulness. Personally, I just started watching his talks, and it just made me even more eager to work on my storytelling skills.
Gabriela presented useful asset packs for Deep Learning:
Gabriela’s talk highlighted how the idea of creating and deploying a model is sometimes seen as easier than what it actually is. She outlined the necessary steps that precede the actual application of a deep learning model:
Gabriela also brought attention to the AI inclusivity project, which aims at increasing the representation and participation of gender minority groups to AI through education and mentorship. The project is still starting up, but I followed them at @AIinclusive to stay in the loop once they get going.
“You’ll never be an expert in all of data science!”
Emily’s talk was the one I was looking forward to the most, seen as I am so new to this world and would love to form a network of data scientists I can interact with and look up to. She asked, first of all, why is it important to network?”
Here are some of her tips for starting a network, and for doing so effectively:
Emily and Jacqueline, the last speaker, have written a book together called Build Your Career in Data Science.
Safiya presented the problem of bias in search algorithms, bringing important real-life cases in which Google searches have perpetuated racist terms and beliefs. Search engines should not be thought of as neutral filters, as they can reflect the bias of users and search results (like in the case of a fake news article showing up at the time of the results of the 2016 American election), or even be directly tampered with.
Safiya suggested what could be done to improve the situation:
Safiya wrote a book called Algorithms of Oppression which delves deeper into the topic of search engine bias.
Jacqueline’s talk was dedicated to debunking the myth that most of what a data scientist does has to do with building machine learning models. In reality, that only makes up about a 33% of the total work, while the remaining 67% is spent
Most importantly, Jacqueline specified how it is a lot more important to keep a project going than to get stuck on a complex model that ends up taking most of the analyst’s time, and eventually ends up killing the project itself. It’s better to settle on a system that is quicker to set up and maintain, in order to have time to focus on the more important parts of the project.
She also stressed the importance of communicating regularly (every 3-5 days) with the project stakeholders in order to keep them in the loop, especially in case the project is taking longer than expected.
I hope these notes helped you figure out what you missed out on, or refresh what you already heard. Overall, I loved attending the conference! The speakers were friendly and approachable, I deeply appreciated the great diversity among speakers, and I always felt welcome and involved. I can’t wait for the next opportunity to discover more about this community and the inspiring people in it.
