AI0-001 AI Models and Data Engineering Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Target encoding

  Why this is correct

  Target encoding replaces each category with the mean of the target variable for that category. It is effective for high-cardinality features because it reduces dimensionality and captures the relationship between the category and the target. However, it can lead to overfitting if not properly regularized, especially with rare categories. Techniques like smoothing or adding noise can mitigate this. In this scenario, target encoding is a suitable method for high-cardinality categorical features.

• ✗

  Label encoding

  Why it's wrong here

  Label encoding assigns a unique integer to each category. While it is memory-efficient, it imposes an ordinal relationship that may not exist, which can mislead models that interpret numerical values as ordered. For high-cardinality features, label encoding can still result in many unique integers, but the arbitrary ordering can introduce bias. It is generally not recommended for nominal categorical variables unless the model can handle categorical features natively. Thus, label encoding is not the best choice here.

• ✓

  Frequency encoding

  Why this is correct

  Frequency encoding replaces each category with its frequency or count in the dataset. It is a simple and effective method for high-cardinality features because it reduces dimensionality and does not impose an artificial order. It can capture the importance of frequent categories, which may be informative. However, it does not consider the target variable, so it may not capture predictive relationships as well as target encoding. In this scenario, frequency encoding is a suitable choice for high-cardinality categorical features.

• ✗

  Binary encoding

  Why it's wrong here

  Binary encoding converts each category into binary code and then splits the digits into separate columns. It reduces dimensionality compared to one-hot encoding and is more memory-efficient. However, it can still create many columns for very high cardinality and may not capture the relationship with the target as effectively as target encoding. It also introduces a somewhat arbitrary binary representation. While it is a valid technique, it is not as appropriate as target or frequency encoding for high-cardinality features in this context.

• ✗

  One-hot encoding

  Why it's wrong here

  One-hot encoding creates a binary column for each category, which can lead to a very high-dimensional feature space when cardinality is high. This increases memory usage and computational cost, and can cause the curse of dimensionality, negatively impacting model performance. While it is simple and preserves all information, it is generally not recommended for high-cardinality features. Therefore, one-hot encoding is not the most appropriate choice in this scenario.