Noções básicas de aprendizado de máquina

A useful model must perform the intended task on new inputs. Memorizing a dataset or producing an impressive demonstration is insufficient evidence of that ability.

Principais conclusões

• Define the task before the architecture.
• Compare against a simple baseline.
• Evaluate failures and downstream consequences.

Begin with a concrete prediction or decision-support task. Predicting a number is regression; assigning a category is classification. Grouping unlabeled examples is clustering. Generating new text or images has different objectives and evaluation methods. Avoid choosing a fashionable architecture before defining the output. A practical workflow has data collection, preparation, model fitting, evaluation, deployment, and monitoring. Errors can arise in any stage. A model trained on well-formed records can fail when a production service changes units or swaps two input columns. Establish a baseline before fitting a complex model. For forecasting, the previous value may be a useful baseline; for classification, the most common class provides a minimum comparison. A baseline exposes whether the extra complexity contributes useful information. Use training examples to fit parameters and separate examples to assess performance. Keep the final test set out of repeated tuning. Choose metrics that reflect the consequences of mistakes, and inspect actual failed cases. A system that performs well on average may still be unusable for rare but essential cases.

1. Construct a toy dataset with 80 ordinary messages and 20 urgent messages. Always predicting ordinary gives 80% accuracy.
2. A model scoring 82% might add little value if it still misses most urgent messages.
3. Count urgent messages correctly identified and ordinary messages incorrectly escalated. Decide which tradeoff meets the actual workflow.

These illustrative counts show how a baseline and task-specific metrics make evaluation more informative.