While my background is more data management than data science, I spent (way too) many years with a major data science software vendor. From my perspective,
- People often don't understand that ML models are never truly done and require regular evaluation, tweaking, and evaluation to see if it is still the champion model. How to discuss? Discuss the concept of an analytics model lifecycle and ensure they devote resources to maintenance in addition to development.
- Lack of accounting for ModelOps and not planning how to implement the model in the real world. How to discuss? Advise them to extend development time and engage the IT/Ops team early on to feed data efficiently to the model and ensure model results are distributed/used appropriately.
- Not accounting for model bias. People often view models as sterile, unattached code, not realizing they can inherit the biases of their creator. How to discuss? Bring up the topic early and often. Half the battle with bias is knowing that it could be there and attempting to seek it out.
