Are you solving the right AI problem?
Solving the right problem is as important as solving the problem right
Every problem is not suitable for AI! Problem framing is important to ensure success in designing and deploying an AI product. In developing AI products, the following is my checklist to ensure we are solving the “right” problem.
Verified there is quantifiable business value in solving the problem
To deliver a well-tuned robust ML model deployed in production can range from 6–24 months depending on the complexity. Being crystal clear about the strategic value of the project is critical.
Verified that simpler alternatives (such as hand-crafted heuristics) are not sufficient to address the problem
Simpler alternatives are often overlooked. Also, these heuristic-based approaches should be used as a baseline for accuracy to compare the effectiveness of the AI solution
Ensure that the problem has been decomposed into the smallest possible units
Instead of thinking of the business problem as one single model, instead, the goal should be to think of the overall solution that should be decomposed as a sequence of models and methods.
Define how the AI output will be applied to accomplish the desired business outcome
Simply predicting customer churn is of little value unless it can be applied by the customer support agent to proactively reach out and manage the customers at risk of churning.
Clear measurable metric(s) to measure the success of the solution
The metric could be direct or an intermediate/proxy. For instance, a direct metric such as increasing customer happiness may not always be measurable. A proxy metric can be the number of times the referral link is shared.
Clear understanding of precision versus recall tradeoff of the problem
If you have a model predicting patients with a potential of heart attack, what are you optimizing: precision or recall. In this example, a high recall that minimizes False Negatives is more important than minimizing False Positive.