Troubleshooting & Problem Resolution Flashcards

A model's performance varies widely depending on how data is split into train and test sets. What technique gives a more reliable estimate?

Answer: k-fold cross-validation

Cross-validation averages performance over multiple splits and reduces split-related variance.

Both training and test errors are high for a model. What does this indicate?

Answer: Underfitting or high bias

Poor performance on both sets means the model is too simple or features are inadequate.

Association rule mining returns millions of rules, most of them trivial. What is the best way to manage this?

Answer: Raise minimum support and confidence and filter by lift

Higher thresholds and lift filtering keep only strong, meaningful rules.

A neural network's loss becomes NaN during training. Which is the most likely cause to check?

Answer: A learning rate that is too high or unscaled inputs

Exploding gradients from a high learning rate or unscaled data commonly produce NaN loss.

A text-mining model performs poorly because the vocabulary is dominated by words like 'the' and 'and'. What preprocessing step helps?

Answer: Stop-word removal and TF-IDF weighting

Removing common words and weighting by TF-IDF emphasizes informative terms.

A clustering result places one huge cluster and many tiny ones. Which action is the most reasonable to try?

Answer: Revisit feature scaling, the value of k, or the algorithm choice

Scaling, k selection, and algorithm suitability all affect cluster balance and shape.

A dataset has 500 features but only 300 records, and the model is unstable. What is a sound response?

Answer: Apply feature selection or dimensionality reduction

High dimensionality relative to sample size causes instability, so reduce features.