Home / Exam patterns / MLP

Machine Learning Practice: exam pattern and behaviour

Computed from every MLP paper on the practice site (96 papers, 2785 questions), plus the pattern notes from your earlier study sessions.

What a paper looks like

Some uploads on the source site merge two parallel sets, which is why a few papers show double marks. MCQ = one right option, MSQ = one or more right options (strict marking), SA = type the number.

Which weeks the papers lean on

Share of papers whose concept summary mentions the week's topic (keyword match, so treat it as a guide, not a count of questions).

How much repeats

Read this carefully. Within one exam date the parallel sets are drawn from the same pool, so each different question appears on several papers (the 'copies' column). Across different dates, questions with identical wording or identical figure files are rare. Matching here is exact, so a re-typed or re-cropped version counts as different; the real overlap in question type is much higher than this table can see. The takeaway: old papers teach you the types and stories that come back with new numbers, they are not a list of answers to memorise.

Questions that came back on a later exam date (exact match)

What an MLP paper is like

MLP is the course where you are tested on scikit-learn and pandas details, mostly by reading code. The papers are longer than MLF and MLT (an End Term has a median of 37 questions and 100 marks, Quizzes about 22 questions and 50 marks), and the questions are short. Real stems from the papers:

• "What will be the shape of the feature matrix (X.shape) and the label vector (y.shape)?"
• "Choose the option to find the number of null values in the Contact column."
• "Which option computes the median of the Balance column for each category in Marital?"
• "Which code block correctly takes the learning rate as 'invscaling' / 'optimal'?"
• "What is the default loss value in SGDClassifier, and which classifier does it give?"
• "Which scikit-learn model supports incremental learning through partial_fit?"
• "How many different models will be trained during this grid search?"
• "Calculate the recall (up to 3 decimals) from the confusion matrix."
• "What is the main advantage of RandomizedSearchCV over GridSearchCV?"
• "Which method of classification needs more than $n$ classifiers?"
• "Assertion (A): in a scikit-learn Pipeline the last step must be a transformer... Reason (R): ..."
• "You're designing an MLPClassifier... which element introduces non-linearity?"

Topic weight, from the concept summaries

• Quiz 1 (26 papers): data preprocessing 18, cross-validation 16, ColumnTransformer 15, imputation 13, MSE 13, train-test split 13, SGDRegressor 11, one-hot encoding 10, Pipeline 9, GridSearchCV 9.
• Quiz 2 (21 papers; only 3 have a concept summary, so this list comes from reading the question stems): metrics from confusion matrices, imbalance (98% negatives), PR curve, partial_fit, one-vs-rest/one-vs-one, ensembles (Bagging base estimators, advantages), l1 LinearSVC gives sparse coef_, RandomizedSearchCV, Ridge alpha.
• End Term (49 papers): everything above, plus ensemble hyper-parameter grid counts, the "largest regularisation" ranking, KNN's computational bottleneck, HashingVectorizer vs CountVectorizer, pandas group-by filters, and neural network non-linearity.

How to prepare

The three counting skills that pay off every time: (1) number of fits $=$ combinations $\times$ folds ($+1$ refit); (2) number of classifiers: OvR $=K$, OvO $=K(K-1)/2$; (3) number of parameters in a network: $\sum(mn+n)$.