Thread by @FlorisgoesF

“Data don’t lie”. But it typically requires a process of defining #research questions, hypotheses, methodology, interpreting and #dataviz that can introduce subjectivity and #bias. Scientific rigor and objectivity are key in #DataScience. Some #Tips for #DataScientists 🧵

Don’t dive straight into a dataset, domain knowledge is critical. Good #Science requires a theoretical understanding of a topic while #ignorance introduces bias. Sound domain knowledge enables you to ask the right questions and give relevant answers with #DataScience

Investigate the alternate hypothesis. Business questions asked to #DataScientists are often directive, as there already is a hypothesis. Don’t confirm this hypothesis without properly investigating the alternate option.

This is linked to confirmation bias. There is often so much information in a dataset, but we’re prone to focus on the parts confirming our initial hypothesis, overlooking other results. Don’t fall for this trap and conduct thorough #Analytics and #Science.

Try to identify the question behind the question. Most people tend to ask you relatively simple questions that are hard to generalize because they don’t understand the answers #DataScience could provide. They are specific but only exemplary of an overarching question.

Don’t make this a thought experiment though. Wandering of on your own induces bias. Instead try to close the communication gap with the business and discuss questions in depth, make #DataScience understandable and try to really understand the business problems.

With every decision you make, think about how your decision affects your results. Data you in- or exclude, preprocessing steps, the way features are constructed. You make dozens of decisions before training a single model, all potentially impacting results & explainability.

Watch out for statistical the significance trap. In #BigData, significance is easy to find. Look beyond the p-values though, as it’s rather about effect sizes and clinical/practical relevance. Know your statistical tests and the related effect size tests.

Make it stupid simple when you present data. Complex terminology and fancy #Datavisualization are great, and people will often trust you at first because it makes you look smart. This won’t last long however, as they will lose interest within minutes if they don’t understand you

Also, it’s easy to hide bias and ill-designed methodology behind complexity, as it does not invite easy criticism. Making it simple allows people to debate your approach. It makes you vulnerable, but people will respect and reward you for it, and it improves your #research

In the end, it’s about 3 things: awareness, thoroughness and simplicity. Awareness of the fact that every decision taken in a process can introduce subjective bias, being thorough in domain understanding & #research, and presenting results simple to prevent misinterpretation.