Data Science Teams & Practice

It's a blog after a 1.5-year gap... 

Recently, I had the chance to attend a webinar/zoom panel discussion on Building Effective Data Science Teams organized by RStudio. It was an interesting conversation - I highly recommend spending the hour to watch if you are interested in this field.  I'll summarize some key takeaways from the discussions for me and some thoughts from my side. 

1. Communication, Communication, Communication!

The panel could not have stressed more the need for MORE communication between the data science team and the various stakeholders (business/functional team, internal customers, etc.) 

Data Science(DS) teams/Software Developers tend to like to work in a cave and only like to emerge with a shiny (pun intended!) new object - this is a high-risk game. More communication with the stakeholders allows for the project to be focused/aligned with everyone's expectations and more importantly, creates a sense of 'joint-ownership' of the project and the outputs. Getting buy-in and adoption/implementation of any analytics work is a big thing!

2. Share Negative Results

This could very well be a sub-topic of the communication topic. There is a tendency to not share negative results or lack of the breakthrough model/output with the stakeholders. However, the point made by the panelists is that this transparency and communication builds the credibility of the team's work within the larger organization. A finer point on this is also to present the right bounds of applicability of the models/output. 

3. Discuss/Debate the 'How' 

It is equally or more important to discuss/debate/argue the experimental design choices made rather than the outputs. If the right contributions are made into the 'How' part, the outputs tend to be better. It is putting a lot more thought and effort into the right stages of the analysis pipeline.  

This concept plays into the current book that I am reading - Framers. The authors here argue that this can be the human advantage in the future AI and machine-dominated world. Here is a related podcast of the authors discussing the book. 

4. Centralized vs. Dispersed/Embedded Data Science teams

I don't think there is a right answer here. It largely depends on the scale/size, and maturity of the data science function within an organization. I would make the argument that the end goal will be to achieve the scale within an organization to have the 'dispersed/embedded' setup i.e. each business/functional unit to have their own teams. The main reason for this is the following: For most organizations, Data Science/Analytics is a means to an end. The DS teams will become better at what they do when they increase their expertise in the functional area.