L2

Question 1

What are missing values in datasets?

Answer 1

Missing values are NaN, null, ‘’, or placeholder values in datasets.

Missing values can affect data analysis and model performance.

Question 2

What is the purpose of missing values imputation?

Answer 2

To fill in missing data so that models can handle complete datasets.

Models cannot directly process missing data.

Question 3

What method replaces missing values with the mean or median?

Answer 3

Mean/Median imputation.
- Good for numerical data; fast

It is good for numerical data and is fast.

Question 4

When is KNN imputation used?

Answer 4

When the dataset is small to medium-sized.

• Find ‘k’ nearest neighbors & average their values

KNN finds ‘k’ nearest neighbors and averages their values.

Question 5

What does model-based imputation involve?

Answer 5

Training a regression/classification model to predict missing values.

• When relationships exist in data

This method is suitable when relationships exist in the data.

Question 6

What is iterative imputation?

Answer 6

A method that repeatedly models each feature with missing values using other features.

• More accurate but computationally heavy

It is more accurate but computationally heavy.

Question 7

What is the purpose of feature selection?

Answer 7

Reduces overfitting, speeds up training, improves interpretability, and reduces storage.

Effective feature selection is crucial for model performance.

Question 8

What does univariate statistics do in feature selection?

Answer 8

Tests each feature individually for relevance to the target.

Tools include f_classif, f_regression, chi2, and mutual_info_classif.

Question 9

What is mutual information used for in feature selection?

Answer 9

Part of univariate statistics - To assess relevance without assuming linearity, making it suitable for non-linear relationships.

It helps identify important features in complex datasets.

Question 10

What is model-based feature selection?

Answer 10

Using trained models to determine feature importance.

• Get best fit for a particular model
• Example models: Lasso regression (L1 penalty) / Tree-based models (feature importance from splits)
• Can be single-pass or iterative
  (single fit)
• Build a model, select features most impt to model
• Lasso, other linear models, tree-based models
• Multivariate – linear models assume linear relation

Example models include Lasso regression and tree-based models.

Question 11

What is forward selection in iterative model-based methods?

Answer 11

Starting with no features and adding the best one at each step.

It helps build a model gradually based on feature importance.

Question 12

What does backward elimination involve?

Answer 12

Starting with all features and removing the least useful one at each step.

This method aims to simplify the model by eliminating unnecessary features.

Question 13

What is Recursive Feature Elimination (RFE)?

Answer 13

An automated method for backwards removal using model feedback.

It is considered an expensive feature selection method.

Question 14

What is categorical data?

Answer 14

Data that falls into categories, such as color, country, or product type.

measurement levels: categorical – ordinal – interval – ratio

Categorical data is crucial for various types of analysis.

Question 15

What is One-Hot Encoding?

Answer 15

Adds one binary column per category to avoid ordinal assumptions.

• Avoids ordinal assumptions
  Works well for low-cardinality data (few unique values)

**Avoid using raw labels like 1, 2, 3—they introduce false order.

It is effective for low-cardinality data.

Question 16

What is Count-Based Encoding?

Answer 16

Replaces category with mean target value for that category.

• For regression: mean target value
• For classification: class probabilities (“people in this state have likelihood p for class 1”)
• Useful for high-cardinality features (e.g., ZIP codes, countries)

Useful for high-cardinality features like ZIP codes.

Question 17

What are digital images made of?

Answer 17

Grids (matrices) of pixels.

• Pixel = smallest unit; stores brightness or color

Each pixel is the smallest unit that stores brightness or color.

Question 18

What is the storage requirement for binary images?

Answer 18

1 bit per pixel.

• Each pixel is 0 / 1 (black/white)

Each pixel is either 0 (black) or 1 (white).

Question 19

What is the storage requirement for grayscale images?

Answer 19

1 byte (8 bits) per pixel.

• Pixel value: 0–255 (brightness scale)

Pixel values range from 0 to 255, representing brightness.

Question 20

What is the storage requirement for color (RGB) images?

Answer 20

3 bytes per pixel.

• 3 values per pixel (Red, Green, Blue)

Each pixel contains three values for Red, Green, and Blue.

Question 21

What are the types of image tasks in machine learning?

Answer 21

1. Classification: Label whole image (e.g., Fracture / No fracture)
2. Detection: Locate object within image
3. Segmentation: Classify each pixel (e.g., tumor pixels vs normal)

An example is classifying an image as ‘Fracture’ or ‘No fracture’.

Question 22

What is the definition of precision in evaluation metrics?

Answer 22

Precision = TP / (TP + FP)

Precision measures the accuracy of positive predictions.

Question 23

What is the definition of recall in evaluation metrics?

Answer 23

Recall = TP / (TP + FN).

Recall measures the ability to find all relevant instances.

Question 24

What does F1 Score represent?

Answer 24

F1 score = harmonic mean of precision and recall.

F = 2 x (precision x recall) / (precision + recall)

It balances both metrics for better evaluation.

Question 25

What is accuracy in evaluation metrics?

Answer 25

Accuracy = (TP + TN) / Total ## Footnote Accuracy can be misleading with class imbalance.

Question 26

What is tokenization in text data processing?

Answer 26

Splitting text into words (tokens). - list of tokens turns into input for in additional processing → parsing / text mining - can swap out sensitive data (E.g. Typically payment card / bank accnt no. —with a randomized no. in same format) ## Footnote It is a crucial step for further text analysis.

Question 27

What is stemming?

Answer 27

Reducing a word to its root form by trimming. (e.g., "studying" → "study") ## Footnote Example: 'studying' becomes 'study'.

Question 28

What is lemmatization?

Answer 28

Using grammar and vocabulary to reduce a word to its proper base form. (e.g., "studies" → "study") ## Footnote Example: 'studies' becomes 'study'.

Question 29

What does Bag-of-Words (BoW) represent?

Answer 29

Each document is represented by a vector with a value for each word in the vocabulary. - Represents each sentence / document as a vector with a value for each word in vocab --> Binary: word present / absent in doc --> Count: how often word appears in doc --> Popular approach: Term Frequency x Inverse Document Frequency (TF-IDF) - Length = vocab size - Values = word presence (binary), count, weight - Each component → no. of times that word appears in sample (e.g. in a sentence / document) - Useful for sentiment analysis (e.g. recognizing words such as “great”, “terrible” …) ## Footnote It can be binary or count-based.

Question 30

What does TF-IDF stand for?

Answer 30

Term Frequency-Inverse Document Frequency. - Emphasizes rare but impt words ✔ Helps reduce impact of common words (e.g., “the”) ✔ Highlights informative terms ✔ Highly interpretable → each word is an independent feature BUT - All structure is lost → Crucial info may be lost (e.g. “I passed the IML exam, but I failed the Computational Linguistics exam”) - Misspellings (e.g. “machine” & “machnie” counted as diff word - Some expressions consists of diff (multiple) words (e.g. product review with word “worth” -> What if the review said “not worth” vs “definitely worth”) ## Footnote It emphasizes rare but important words.

Question 31

What is the formula for TF in TF-IDF?

Answer 31

TF(t, d) = count of term t in document d / total terms in d. ## Footnote It measures how frequently a term appears in a document.

Question 32

What is the formula for IDF in TF-IDF?

Answer 32

IDF(t) = log(total docs / # docs with term t). IDF of a rare word high / IDF of a frequent word low → having effect of highlighting words that are distinct **TF-IDF(t, d) = TF × IDF ## Footnote It highlights the importance of rare words.

Question 33

What is an n-gram in text processing?

Answer 33

Multi-word tokens that capture short phrases instead of single words. Examples: 1-gram: “not”, “worth” 2-gram: “not worth” (captures sentiment better) ## Footnote Examples include 1-grams and 2-grams.