It’s still early days in data science, and it’s still difficult to know how to get started. Indeed, a 2018 O’Reilly survey finds that half of organizations are still in the “exploring” phase. Given the hype around data science, many companies are approaching it as a solution looking for a problem. Without a clear use case, very little business value will be gained.
If you’re just starting out on the data science journey, here are some keys to starting successful data science projects.
Step back. Get perspective. What is your true north? How does your data science problem relate to this?
You should always strive to solve problems aligned with your business’s true north. This is easier said than done, especially given the transformational potential of data technologies. You will be tempted to do data science for the sake of data science. Focus instead on solving core business problems. Make sure that the problem you’re solving is aligned with your business’s true north.
Start with the end in mind. Identify ONE problem (preferably small) that needs to be solved.
Data science projects work best when you focus on a single problem, preferably something small and discrete. In this way, you can obtain, analyze and model the data in a way that doesn’t contaminate the results.
Starting with a small, single objective also provides your data team the design patterns and experience for more sophisticated next projects.
What does “done” look like? What are good results?
Establish a finish line for initial success. Define “done” clearly and succinctly. Your first data science project doesn’t need to be a big win. For example, you may decide to create a simple model to predict customer churn. Decide in advance the metrics you will use to determine if this model is performing well. Without a notion of “done”, it’s easy to wander aimlessly.
Get executive and organizational buy-in.
Lack of executive and organizational buy-in is a major reason that data projects fail. Get support from the top-down and bottom-up, and rally your team around the problem. Communicate regularly with stakeholders and give them skin in the game.
As executives and the organization gain more trust in the business’s ability to execute on data projects, a virtuous cycle of support and data value is created.
Assess your data team, capabilities and resources. Tell them the WHAT. Let the team figure out the HOW.
A good data team will support the functions of data science, data engineering, data analysis, and data architecture. If the team is small, team members may need to play multiple roles, but data engineers, data analysts, and data architects have as much value as the data scientist. It’s all about building a cohesive team with all the essential moving parts.
Once you’re confident that you have the team and resources to execute on your data goals, make sure the team is aligned on WHAT you’re trying to achieve. Let the data team figure out HOW to accomplish this goal. This is as much management 101 as anything else. Make sure that you consult stakeholders who will use the tools day to day before making technology decisions.
Invest in building a solid data foundation.
It’s tempting to jump into a data science project with nothing more than a Python notebook, some libraries (Sklearn/Tensorflow/Pytorch/etc), and boundless enthusiasm. For exploratory data analysis or toy data projects, this is fine. But in order for the business to get real and sustained value from its data initiative, you will need to establish a solid data foundation.
The data science hierarchy of needs is an insightful blueprint for realizing value from your data.
Please understand that it takes substantial investments of organizational effort, capital and time to build a data foundation. It requires an ongoing journey of continuous improvement. Given the lack of clarity or a timeline for data ROI, it’s tempting to cut corners or become impatient. But in a world where winning with data is a key competitive advantage, do you think you want to take this investment lightly?
Start simple. Use simple algorithms before moving to advanced approaches.
Avoid starting with sophisticated approaches. Just because deep learning is all the rage doesn’t mean it’s the first approach you should take. In a lot of cases, simple approaches like linear or logistic regression work great and are very simple to implement.
Plan for production from the beginning of the project.
Will your model be exposed through a web service? Should it trigger emails to customers? In and of themselves, models do not solve business problems. Decide early what it means to put your models into production and make sure that your team has the resources to deploy. Data science done in isolation on laptops does not add value to the business. A production solution should include automated data pipelines, model training and serving, as well as monitoring and alerting for the model, pipelines and production service.
Watch out for resume driven development.
New data tools pop up on a daily basis, and it’s easy for data teams to get shiny object syndrome for the latest and coolest data tool. This is a dangerous temptation. A dirty secret of engineers and data scientists is that they may choose tools in order to improve their resumes, not because they are the best fit for the problem at hand. Resume-driven development is a real thing, and it can have unintended consequences. Make sure your data team has a process for vetting and implementing data tools that align with the objectives and problems being solved. Making an upfront commitment for best-fit will avoid the consequences of a suboptimal data stack down the road.
Avoid technology for the sake of technology. Use the simplest solution that will deliver value to the business. See what you can glean from regression before applying more sophisticated ML. If you need to forecast website traffic, start with Facebook Prophet rather than building a sophisticated bespoke time series model. Utilize a fully managed cloud ML platform like AWS Sagemaker in lieu of standing up a TensorFlow cluster on Kubernetes. Pursue custom, cutting edge solutions only after you’ve identified clear business value beyond a simple solution. Take into account long term operational and maintenance costs, not just the upfront resources required for a proof of concept.
Your journey with data science will create value for your business by taking a smart, methodical approach. Build a solid data foundation. Experiment, implement, learn, and iterate. Best of all. Have fun. You’re a pioneer in the early days of data science.
If you are working on data science projects and want some advice on taking it into production, don’t hesitate to reach out.