Machine Learning Assessment: Sample Questions | Bryq
Beginner
Intermediate
Advanced
These are three sample questions from Bryq’s Machine Learning assessment, a technical test for hiring machine learning engineers, data scientists and AI engineers. The full assessment covers three levels: beginner, intermediate and advanced, from core algorithms and learning types to feature engineering, data preprocessing and model optimization. Each question is multiple choice, with one correct answer out of four options.
Beginner
Intermediate
Question
A dataset has a “color” column with the values red, green and blue. Before training a linear model, how should this column usually be prepared?
Select your answer
One-hot encode it into a separate 0/1 column for each color.
Replace the values with 1, 2 and 3.
Normalize it to a range between 0 and 1.
Drop it, since models can’t use text values.
WHY IS THIS THE CORRECT ANSWER
One-hot encoding turns each category into its own 0/1 column, so the model treats the colors as separate, unordered values. Mapping them to 1, 2 and 3 would wrongly suggest that blue is “greater” than red, and dropping the column throws away potentially useful information.
Intermediate
Intermediate
Question
A fraud detection model flags 100 transactions, of which 80 are actually fraudulent. The dataset contains 200 fraudulent transactions in total. What is the model’s recall?
Select your answer
40%
80%
50%
20%
WHY IS THIS THE CORRECT ANSWER
Recall measures how many of the actual positives the model found: 80 correctly flagged ÷ 200 total fraud cases = 40%. The 80% figure is precision, the share of flagged transactions that were really fraudulent.
Advanced
Intermediate
Question
Before splitting data into training and test sets, a team scales every feature using the mean and standard deviation of the full dataset. What is the main problem with this approach?
Select your answer
Test set statistics leak into training, making evaluation results overly optimistic.
Scaling reduces model accuracy on numerical features.
Scaling is only needed for tree-based models.
It causes the model to underfit the training data.
WHY IS THIS THE CORRECT ANSWER
Fitting the scaler on the full dataset lets information from the test set shape how the training data is transformed, so the test set is no longer truly unseen. Preprocessing steps should be fitted on the training data only, ideally inside a pipeline, and then applied to the validation and test data.
