DataTaunew | comments | leaders | submitlogin
Ask DT: How to play in an organisation where data is 3rd?
4 points by Nadav 2861 days ago | 2 comments
After caving in to all the hype someone in the company finally decided to hire a data scientist (YOU). Unfortunately once you join you realise that your opinion is only heard/considered after that of the product managers and designers, if at all.

Any advice how to handle/change that?



5 points by BenoitParis 2860 days ago | link

I'd do the following:

1. Go to the CFO. Understand how the revenue is made, what the margins are. Have it known you went to the CFO.

2. Pick a signal related to something the CFO cares about. Retention, ARPU, behaviors (cohorts, etc). Start measuring it (you don't improve what you don't measure).

3. Build a visualization tool over it. Have managers and designers guide you. You want their collaboration. Have them know this is about recognising their work in hard data form (pdf great: easily forwardeable to the CFO for the budget reviews).

4. Follow what they do, and have them see the fruit of their labor in the graphs. Hopefully they start experimenting and judge of an outcome by your graphs.

5. Follow their experiments, and provide guidance. When they start introducing rules over behavior, talk about having classification do it, and how it will properly capture the long tail. Maybe let them do their hard coded behaviors, but provide how an algorithm would do better.

Also, prefer standard tools. Pick a standard ETL, a standard BI tool, and standard pointy-clicky charting. A lot of managers can talk SQL and pivot tables. Let them own the data you prepare them. Then you can bring in fine analysis and machine learning heavy artillery.

For machine learning, develop a few simple abstractions that they can wrap their heads around. Don't go all decision tree forests, kernel tricks and ROC; but more: lets say there are empty cases in this worksheet, a human would fill them by hand like that, but we can have a statistical box do it to the perfection.

--------

Aside: There is a story about a LinkedIn Data Scientist who went to the parallel org structure (CFO I think) to build the degree of separation from your profile to the one you where browsing. He did it out of ad space, by injecting his own Javascript, with his own database. Maybe you can go that route too, but it is more risky than the `accept and collaborate` path I described earlier.

-----

2 points by Nadav 2860 days ago | link

Thank you very much for the detailed answer. I've done quite a bit of what you are suggesting, but your answer sparked some additional ideas.

I guess my main frustration comes from a feeling that 'this is not what I've signed up for' (dashboarding...). I was under the impression that the company is interested in a deeper analysis of their data and making decisions based on that. Hopefully through a process similar to the one you are describing above, they'll shift towards that.

-----




RSS | Announcements