"What unites people? Armies? Gold? Flags?", says Tyrion Lannister in the finale of GoT, addressing the most powerful lords & ladies of Westeros as they sit down to choose the next Queen or King. Nodding the head sideways, he continues – "Stories… There’s nothing in the world more powerful than a good story. Nothing can stop it. No enemy can defeat it. And who has a better story than Bran the Broken.".
Today, as AutoML gets prominence, coding Data Science models is getting easier than ever due to high level APIs like Keras, Tensorflow Estimator, Fastai etc. as well as plethora of AutoML tools like DataRobot, DataIQ, H2O, AWS Sagemaker, GCP AutoML etc.
By 2020, more than 40% of data science tasks will be automated, resulting in increased productivity and broader usage by citizen data scientists. – Gartner
In this scenario, one skill that has become all the more important in a great Data Scientist’s repertoire is the Art of Story-telling.
As you build the most sophisticated models, leveraging the most complex algorithms and merging hundreds of complex data sources, one factor is a key determinant of whether your model gets to see the light of the day or dies a quick death leaving you cursing the uninitiated : your ability to convince the people who matter the most – the business heads.
Only 1 in 10 data science projects actually make it to production – VentureBeat
We all love to hear stories. But story-telling is no easy task. If you’ve tried to narrate a story impromptu to your kid, you know what I am talking about. It requires introducing a protagonist and an antagonist, having inciting action, building a conflict and then bringing resolution.
Regions that are not traditionally thought to be part of a "language network" in the brain become consistently activated when people listen to narratives. Example areas are the precuneus/posterior cingulate cortex, and mPFC. – Susana Martinez-Conde et al in the Journal of Neuroscience
In order to build a good story for our model, let’s draw a leaf or two from the Golden Circle popularized by Simon Sinek – an author & thinker I admire tremendously – in his seminal book ‘Start with Why’. Sinek’s Golden Circle comprises of Why, How & What; starting with the ‘Why’, going into the ‘How’ and finally addressing the ‘What’ when trying to inspire people. We’ll just tweak it a little bit in the context of analytical story-telling.
As you prepare to convince the business heads to adopt your model for their decision making, look to build a story comprising of these 3 elements : Why, How, What.
1. WHY: Why do you think this model is important?
Which business problems does it address? Which business KPIs does it impact and how? As the business adopts the model, what outcomes could they expect? You should clearly be able to address and depict a world without the model (i.e. the current state) and a world with the model (i.e. the future state), and help them visualize what both of these worlds will look like (you don’t need to get into the quantitative details yet).
_Through 2022, only 20% of analytic insights will deliver business outcomes – Gartner analyst Andrew White_
2. HOW: How did you go about building the model?
Most often, we emphasize too much on the actual model output – accuracy, ROC, MAPE, MSE, R-squared and the list continues. Yes those are important, but more important is how did you arrive at these. Business wants to understand the process you used – How have you defined the target variable, i.e. what exactly are you trying to predict? How did you source the data? How did you transform it? Which assumptions did you use? How did you get to the end point? How have you taken care of the biases during model training? Have you incorporated complexities in the model at the cost of interpretability? And finally, if they do decide to go with it, how are you going to implement it? How will the change management be addressed?
_Through 2022, over 75% of organizations will use DNNs for use cases that could be addressed using classical ML techniques. – Gartner analyst Andrew White_
3. WHAT: What is your model output?
Finally, once your audience have understood why you think your model is the next best thing after Einstein’s Theory of General Relativity (pun intended) and how have you gone about building it, you’ve built enough tension for everyone to relish the aha moment. That’s where you make the revelation – What are the accuracy measures of your model? What are the explanatory variables that play a key role in predicting the desired outcome? How does your model address extreme values? What are the breaking points of your model? How sensitive is your model to fluctuations in key variables? Are the model-determined relationships consistent with the intuitive relationships that we already know of? What actions can the business take leveraging model predictions? What is the benefit of a right decision based on model output? And conversely, what’s the cost of a wrong decision (hoping it is very small)?
But it doesn’t end here. The proverbial nail in the coffin is where you are able to make a statement indicating how much $ value (or € or £ or ¥ or ₹) can the business generate by implementing the model, either in form of cost savings or revenue increase. How you measure that is a different topic, for some other time. You have to ease the anxieties related to potential model degradation in future, and how will you address it. You also have to assure them of auditability & traceability of model predictions at any point of time.
So, next time your kids ask you to tell a story, don’t shy away. They’re just helping you practice a skill that may just help you cross the line between a ‘model in the lab’ and a ‘model in production’. Happy story-telling!!!
Originally published at https://www.linkedin.com.