1/80
Looks like no tags are added yet.
What is regression in machine learning primarily used for?
Predicting a continuous numerical value
In regression modeling, the target variable is usually:
Continuous
The main goal of linear regression in ML is to:
Learn the weights of a linear equation relating inputs to outputs
In a simple linear regression model with one predictor, the line is parameterized by:
b0 and b1
What is another common name for prediction errors in regression?
Residuals
To find the best-fitting regression line, we usually try to minimize:
Residual Sum of Squares (RSS)
Which sequence best describes how least squares regression finds the best fitting line?
Visualize the goal, define the error, square and sum errors, find the minimum, solve for coefficients
Why is solving partial derivatives equal to zero often not feasible in many ML problems?
Because of high dimensionality, non-convexity, cost-function complexity, and stochastic training environments
Instead of solving analytically for stationary points in many ML problems, practitioners often use:
Gradient descent or parameter search
A derivative tells you:
The slope of a function at a point
A gradient is best described as:
The direction of fastest ascent represented as a vector of partial derivatives
In gradient descent, the parameters are updated in which direction?
In the direction opposite the gradient
What controls the magnitude of movement in gradient descent?
Learning rate
In the hill-climbing metaphor, the negative gradient tells you:
How to go downhill the fastest
When does gradient descent usually stop?
Once the pre-set convergence criterion is met
If the linearity assumption is violated in linear regression, the model may:
Produce biased predictions
A recommended pre-check for linearity is:
Scatterplots
In ML, independence of errors is generally less critical unless the data has:
A temporal or spatial structure
Normality of errors is less critical in ML unless the goal is to:
Make probabilistic predictions
If residuals deviate significantly from normality, one possible remedy is to:
Apply transformations such as log or square root to the target variable
Equal variance of errors mainly affects:
Coefficient estimation and interpretability
Which methods can help stabilize unequal variance of errors?
Weighted least squares (WLS) and transformations of the target
Multicollinearity in linear regression primarily:
Increases sensitivity to data changes
What are common checks or remedies for multicollinearity?
Check correlations, check VIF (variance inflation factor), use Ridge or Lasso
Which regression metric is intuitive, robust to outliers, and easy to explain?
MAE
Which metric punishes extreme errors more strongly and reflects risk?
RMSE
Which metric is scale-free and gives interpretable percentages?
MAPE
R² mainly indicates:
Explanatory power
How should you choose a regression performance metric?
Based on business objectives, data distribution, and cost of errors
What are good responses to poor regression performance?
Improve data quality, try different algorithms, adjust features, use better metrics and validation
A classic symptom of overfitting in linear regression is:
Low bias and high variance
Very large estimated coefficients in an overfit model often mean:
The model is capturing noise rather than genuine relationships
Ridge regression is also known as:
L2 regularization
What does Ridge regression add to the least squares objective?
A penalty based on squared coefficients
Why can Ridge regression improve generalization?
It adds some bias during training to reduce overfitting
Lasso regression differs from Ridge because it:
Shrinks some coefficients exactly to zero
A major advantage of Lasso is that it can:
Perform feature selection
Which are Lasso solvers or related approaches?
LARS, coordinate descent, PCD, ADMM
A practical issue with Lasso is that shrinking coefficients to zero may:
Introduce extra bias
A debiasing strategy for Lasso is to:
Run Lasso for feature selection, then Least Squares
When features are highly correlated, Lasso tends to:
Select among them arbitrarily
Elastic Net combines:
Ridge and Lasso regularization
Why are regularization methods sensitive to feature scaling?
All features should contribute equally to the penalty
What is regularization in linear regression?
Adding a penalty to constrain coefficients and reduce overfitting
Which learning style describes KNN most accurately?
Lazy or instance-based learning
In eager learning, the model typically:
Finds a global approximation before prediction
Instance-based (lazy) learning usually has:
Less effort in training and more in prediction
Key ingredients of distance-based models include:
Distance metrics, exemplars, and decision rules
The similarity assumption states:
Nearby points in feature space are more similar
Distance is an inverse proxy for similarity because:
Smaller distance means more similarity
Which distance metric is best for continuous data?
Euclidean
Manhattan distance is especially useful for:
Grid-like data
Chebyshev distance focuses on:
Maximum difference along any dimension
Cosine similarity is useful for:
Sparse, high-dimensional data like text
A centroid is:
The arithmetic mean position of all points
A medoid differs because it:
Must be a member of the dataset
In classification with nearest exemplars, a new point is classified based on:
The nearest exemplar or center
In KNN, after calculating distances, the next steps are:
Find neighbors and vote on labels
Why does 1-NN have low bias but high variance?
It perfectly separates training data
As k increases in KNN:
Bias increases and variance decreases
What is the advantage of KD-tree in KNN?
It prunes the search space
The KNIME workflow includes:
Preprocess text, vectorize, compute distance, KNN, optimize k, evaluate
The objective of text preprocessing is to:
Extract key terms and remove noise
Which are noise in text preprocessing?
Punctuation, stopwords, diacritical marks, casing
Tokenization is:
Breaking text into tokens
Stemming vs lemmatization:
Stemming may create non-words; lemmatization uses context
TF-IDF gives high score when:
Term is frequent in document but rare overall
TF measures:
Local frequency
IDF measures:
Global rarity
Why are N-grams useful?
They capture context and word order
N-gram importance measures:
Count and relative frequency
Why is cosine similarity used?
Efficient for sparse vectors
Distance from cosine similarity:
1 - cosine similarity
If vectors are very similar, angle is:
Close to 0 degrees
Which KNN metric statement is correct?
Manhattan is for grid-like data
The Elbow Method is used to choose:
Number of clusters k
What is a challenge of KNN in high dimensions?
Curse of dimensionality
How to mitigate high-dimensional issues?
PCA or feature selection
Why is feature scaling important?
Prevents large-scale features dominating distance
What does regression modeling describe?
Regression modeling describes the relationship between one or multiple features and a continuous target variable.
Which shapes correspond to common distance metrics?
Euclidean: circle, Manhattan: diamond, Chebyshev: square, Cosine: wedge.