A team plans to skip exploratory work and begin modelling immediately, arguing that the algorithm will find whatever matters. What is the practical risk?
Select an answer to reveal the explanation.
Short Explanation
Skip exploration and you learn about duplicated rows, a region that stopped reporting, and a column that quietly encodes the answer after the model is in production. Leakage looks like excellent validation scores right up until it does not.
Full Explanation
The correct choice is the one saying data quality problems, skew, and leakage go unnoticed, so the model learns artefacts and validates well while being unusable. Exploration is where you discover that a column is largely null after a schema change, that one segment is heavily over-represented, or that a feature encodes the target - leakage that produces impressive validation results and fails completely on live data. None of these announce themselves during training.
The option claiming the model cannot be trained at all until an exploratory query has registered the table in the catalog is incorrect because catalog registration is an administrative act, unrelated to whether anyone has explored the data.
The option saying training will consume more compute, but the resulting model will be equally valid, is incorrect because it treats the issue as cost. The risk is a model that is confidently wrong, which no amount of compute corrects.
The option saying there is no real risk if the training data volume is large enough is incorrect because volume amplifies systematic problems rather than diluting them. More rows carrying the same bias produce a more confident bad model.
For the exam, remember exploration as the step that establishes whether the data can support the model, before anyone fits one.