Start with the work, not a tool

Data science is a set of connected practices for turning data into defensible conclusions. Studying it means learning both the technical steps and the judgement behind them: how data was collected, what a metric measures, whether a model is appropriate, and what its results do—and do not—support.

Start by gathering the material you are expected to learn: course notes, slides, readings, worked examples, lab instructions and your own project work. Arrange it by topic rather than file type. A useful first map might cover data preparation, exploratory analysis, statistical reasoning, machine learning, evaluation and communication. Your syllabus or course outline should determine the final scope; do not assume that every data science course gives these areas equal weight.

Build a revision board from that material and keep each section narrow enough to review. A board on this topic in MySummaries ends up looking like this:

Data Science Core data science workflowStudy
Study map — foundations to evaluationCore data science workflow5 sections · 3 columns
Data splitting and leakage3 due
  • Split before fitting preprocessing steps; fit transformations on training data only.
  • Use validation data for model selection; keep test data for final evaluation.
  • A feature derived from the target or future information can leak the answer.
Regression metrics

MAE is the mean absolute error in the target's units. RMSE is the square root of mean squared error and penalises large errors more heavily.

Scaling and preprocessing
MethodTypical use
StandardisationCentre at mean 0 and scale to standard deviation 1
Min–max scalingMap values to a chosen range, often [0, 1]
One-hot encodingRepresent nominal categories without implying an order
Interpretation and communication2 due
  • Association alone does not establish causation.
  • Report the evaluation data, metric and relevant limitations.
  • Explain the result in the context of the decision it is meant to inform.
Classification metrics
  • Precision = TP / (TP + FP); recall = TP / (TP + FN).
  • F1 = 2 × precision × recall / (precision + recall).
  • Accuracy can mislead when classes are imbalanced.

The board is a map, not a substitute for working through the methods. Link each statement to an example from your course. For instance, if your notes say to scale features, add which estimator or distance calculation makes scale relevant. If the notes describe a result, record what data produced it and what conclusion the result supports.

Learn the workflow in order

Data science topics make more sense when studied as a process. For each project or worked example, trace the steps in order: define the question; inspect how the data was produced; clean and transform it; explore patterns; choose an analysis or model; evaluate it on suitable data; then communicate the result and its limits.

At each step, ask what decision is being made and what could go wrong. A missing value is not automatically an error to delete: the reason it is missing may matter. A high score is not automatically evidence of a useful model: check the split, baseline, metric and consequences of mistakes. A correlation is not proof that changing one variable will change another.

Use one dataset or case study from your course to practise the whole chain. Write down the prediction target or research question, the unit of observation, the available features and any time ordering. Then explain your cleaning choices and why your evaluation design fits the problem. For a time-dependent prediction, a random split can place later observations in training and earlier ones in testing; a time-aware split is often more appropriate. For grouped observations, consider whether records from the same person or site should stay in one partition. Follow the methods taught in your course and justify your choice.

Make a short “must know” checklist

Do not try to memorise every line in a notebook. Extract the facts and distinctions you need to retrieve without prompts: key formulas, assumptions, what a method does, and common failure modes. Keep the checklist small enough to revisit. These are the core points from the board:

Key pointsCore data science workflow
Prevent leakage: split data first, then fit imputation, scaling and feature-selection steps on training data; apply those fitted transformations to validation and test data.
For binary classification, precision = TP/(TP+FP), recall = TP/(TP+FN), and F1 = 2PR/(P+R); select metrics to reflect the error costs and class balance.
For regression, MAE reports average absolute error in target units; RMSE also uses target units and gives greater weight to large errors.
Standardisation uses training-set mean and standard deviation: z = (x − μ)/σ. Min–max scaling uses training-set extrema; neither method guarantees normality.
A predictive association is not, by itself, a causal effect; state the data, design and assumptions behind any causal claim.

Turn each checklist item into a question and answer it from memory before checking your notes. If you cannot state a formula, define its terms and work through a small numerical example. For instance, with 8 true positives, 2 false positives and 4 false negatives, precision is 8/10 = 0.8 and recall is 8/12 ≈ 0.67. That calculation also makes the denominators harder to confuse.

Practise retrieval, not rereading

Read a short section, close the source, and explain the idea aloud or in a few sentences. Then check the source for missing conditions and correct the explanation. This works particularly well for contrasts: precision versus recall, validation versus test data, prediction versus causal inference, and standardisation versus min–max scaling.

Flashcards are useful when each card tests one answer. They are less useful when they ask for an entire workflow or contain a long passage. A set from this board could look like this:

Cards — Core data science workflow7 due

When should you fit a scaler in a train–test workflow?

Fit it on the training data only, then use that fitted scaler to transform validation and test data.

All 7 cards
When should you fit a scaler in a train–test workflow?Fit it on the training data only, then use that fitted scaler to transform validation and test data.
What is precision for a binary classifier?TP / (TP + FP): the fraction of predicted positives that are true positives.
What is recall for a binary classifier?TP / (TP + FN): the fraction of actual positives that are identified.
Write the F1 formula in terms of precision P and recall R.F1 = 2PR / (P + R), the harmonic mean of precision and recall.
What does MAE measure?The mean absolute difference between predictions and observed values, in the target's units.
Why can accuracy be a poor summary for an imbalanced classification problem?A model can predict the majority class often and achieve high accuracy while missing many minority-class cases.
Does a correlation between two variables prove that one causes the other?No. Association alone does not establish causation; consider the study design, confounding and other assumptions.

That is a MySummaries deck, filled with data science material. Yours is written from your own notes. Start free

Study the deck as a drill, not a list to skim. Read the first question, answer it before revealing the back, then grade how well you recalled it using the available grade buttons. A correct answer that needed substantial prompting is not the same as fluent recall. Continue through the cards, and return to the difficult ones after a gap. Spaced repetition is most useful when your rating reflects what you actually retrieved, rather than how familiar the answer looked.

Work with data and explain your choices

Conceptual recall needs to be paired with practical work. Use a small dataset or a course exercise and complete one task at a time: inspect types and missingness; summarise distributions; make a plot; fit a simple baseline; evaluate it; and explain the result. Keep a brief lab record with the question, code or method, outcome, interpretation and one limitation.

When a result surprises you, check the path that produced it rather than changing several things at once. Inspect the data split, preprocessing, target definition, baseline and metric. For a classifier, compare a confusion matrix with the summary score. For regression, inspect residuals as well as MAE or RMSE. A metric compresses information; it does not explain every error or establish that a model is suitable for use.

Practise turning technical output into a short explanation for a reader who did not build the analysis. State what was measured, on which data, what the result means in context, and what it cannot establish. If a model is being considered for a decision, discuss the costs of false positives and false negatives and whether the data represent the setting where it would be used. Keep uncertainty and limitations visible rather than hiding them behind a single score.

Review weak areas with a repeatable cycle

Once a week, review the questions you missed, the code errors you had to fix and the explanations that were vague. Label each problem: missing fact, confused definition, unsuitable method, calculation error, leakage risk or unsupported interpretation. Then return to the relevant source and make a targeted correction. Re-test the point later without notes.

Use short mixed practice as well as topic-by-topic work. A mixed session could ask you to choose a metric, identify a leakage risk, interpret a plot and distinguish a predictive claim from a causal one. Switching between related ideas helps expose confusions that are easy to miss when you study each topic in isolation. Keep the questions aligned with your course material; the right method depends on the problem, data and assumptions.

For a focused session of about an hour, spend 10 minutes recalling the previous session, 20 minutes studying one concept from your notes, 20 minutes applying it to a worked example or dataset, and 10 minutes checking and recording errors. Adjust the timing to your workload. The important feature is the sequence: retrieve, learn, apply, check. Avoid counting time spent passively rereading as proof that you can use the method.

Build a listening recap from your notes

An audio recap can help you revisit a topic while walking or between study blocks, but use it to reinforce material you have already checked. Choose a narrow section, such as evaluation metrics, and make sure the explanation includes definitions, a worked distinction and a limitation. The lecture should follow the board's content, not replace the source material. A listening screen on this topic in MySummaries could read back a recap like this:

Lecture — Core data science workflow8 min
Choosing and interpreting evaluation metricsA short route from confusion-matrix counts to a defensible metric choice.
02:4608:12
Speed1×1.25×1.5×2×

Transcript · tap any word to jump there

Start with the question the model is meant to answer. In binary classification, true positives and false positives contribute to precision: true positives divided by all predicted positives. True positives and false negatives contribute to recall: true positives divided by all actual positives. The denominators answer different questions, so write them out before selecting a metric.

Suppose a classifier identifies 8 true positives, produces 2 false positives, and misses 4 actual positives. Precision is 8 divided by 10, or 0.8. Recall is 8 divided by 12, or about 0.67. The model's precision is higher than its recall in this example: among the cases it labelled positive, most were positive, but it failed to identify some actual positives.

Do not choose accuracy by habit. If one class is much more common, predicting that class can yield high accuracy while performing poorly on the minority class. Compare suitable metrics with a baseline and inspect the confusion matrix. The costs of different errors, the data and the intended use determine which summaries are informative. A score describes performance on an evaluation set; it does not prove that a model will work equally well elsewhere or that an association is causal.

After listening, pause and reproduce the precision and recall calculations without looking. Then explain why accuracy might mislead in an imbalanced dataset. If you cannot do both, return to the relevant board section and practise the cards or example again. The aim is not to accumulate study materials; it is to make sound choices and explain them using the evidence in front of you.

How MySummaries helps

Build a revision board from your own PDFs, slides or photographed handwritten notes, then use its sections to create flashcards with spaced repetition, written mock exams marked against your notes and audio lectures in an examiner voice. That gives you a way to move from your course material to active recall and targeted review. Open MySummaries.