Why DataRobot Sucks
- Limited machine learning algorithms and limited ways to optimize them for use cases
- Limited flexibility of supporting data features
- Limited Sub-selection model results from the limited choice of models
- Ugly visualization for performance and comparisons
- The blender option for ensemble methods and choices is limited
- To do anything complex it is too limited
- Doesn't replace the value of handcrafted models and the value of feature engineering process
- Why automate feature engineering especially when the process is what allows one to build a generalizable model and with a better understanding of the business cases
- Limited ways to evaluate the choice of model
- Limited ways of import/export model for build/deployment
- Tight coupling to a third-party way of doing things
- Doesn't follow the formal data science method
- In fact, doesn't even replace 99% of the work
- GDPR and governance issues when processing data through the third-party models
- Benchmarks on models are not available for public peer review
- Cost far exceeds the benefit
- Many of the models provided have less than optimal outcomes - low confidence scores
- Not good for complicated analysis, best to keep use cases as simple as possible
- Productivity gain is only with very limited and simple cases
- Invariably, most business cases have noisy data and require custom models where it will be unworkable and useless to work against complexity and ambiguity
- One needs to learn another third-party tool and be willing to trust the model solution blindly
- Quick wins and successes is not the answer nor the solution if it does not solve the business case
- The reality is there is no free or easy lunch for most business cases and one has to put in the time and effort
- No need to cheat one's way through the process
- Unconvincing autopilot feature - there is no such thing as autopilot in machine learning in fact there is no real community standards or even patterns defined so how can one jump decades in progress
- Solution is useful to people that can't code, don't like to code, don't understand the data science method, and prefer simple drag and drop options
- Data input is treated in most cases as a table, not workable for noisy unstructured data
- Often will end up with overfitted models when the whole point of machine learning is to build generalizable models
- No flexible options for transfer learning on unseen data
- And, that isn't even the end of it...
