As we’ve written about in previous APEX of Innovation posts, a primary driver behind data science project failures is poor planning and lack of clarity about the problem at hand. These and other mistakes can take weeks or months to properly address, curtailing the initiative’s success and undermining stakeholders’ overall confidence in data science.
To avoid this scenario, consider the following tips to improve your data science workflow:
1. Set the Right Objective
Machine learning algorithms are adept at identifying the optimal solution, but they can’t determine if your company is maximizing the right thing at the right time. Corporate priorities and values evolve as the company grows and changes, and it’s important to make sure that your objective function reflects these goals. Taking the time to periodically review this and make any necessary adjustments helps ensure that you are optimized for the right problem and avoid headaches down the road.
2. Get on the Same Page
Another consideration is taking the time to present insights in terms easily accessible to business leaders so that they spend less time translating and more time actualizing the data. Consider the differences between the following sentences: “We saw a 100 point increase in accuracy in the test set of 100,000 examples,” compared to “If we had these improvements in place, we would have saved $20,000 in the last quarter.” Whenever possible, replace technical jargon with phrases more closely aligned with bottom-line impact.
3. Allow Room for Discovery
The biggest breakthroughs in data science often come from the exploration of new approaches and opportunities. Giving data scientists time to experiment with new avenues and investigate different angles can deliver big gains in the form of new capabilities, better models, and faster time to results.
4. Talk to Your Consumer
It’s impossible to build an effective model if you don’t understand your end-user and the problems they’re trying to solve. As such, it’s critical that you talk to your consumer to get a grasp on their key pain points before you dedicate any time to designing your model.
5. Optimal Solutions Tend to be Suboptimal
Highly optimized solutions typically cost more to implement and maintain and tend to be less flexible. In addition, optimized solutions are only optimized if the objective function is entirely accurate and unlikely to change. As discussed above, this is not generally the case, given the speed at which business evolves, which is why building simpler solutions whenever possible is often the “optimal” choice.
Head over to VentureBeat for more information on the considerations outlined above.