Examhaus — Achieve with confidence.

IB Computer Science A4.2: Data preprocessing (HL only)

New IB Computer Science syllabus · first exams May 2027 · Theme A · HL only

A model is only as good as its data: "garbage in, garbage out". A4.2 covers how data is cleaned and prepared before training, which features to keep, and why having too many features causes problems.

StatementTopicCommand term
A4.2.1Significance of data cleaningDescribe
A4.2.2Role of feature selectionDescribe
A4.2.3Importance of dimensionality reductionDescribe

Key points

A4.2.1 Data cleaning

Poor data quality directly lowers model performance and introduces bias. The techniques named in the guide:

  • Handle outliers: cap them, remove them, or keep them if they're genuine rare events.
  • Remove or merge duplicates.
  • Identify incorrect data and filter out irrelevant data.
  • Fix badly formatted data: dates, units, inconsistent text.
  • Deal with missing data in one of three ways:
    • delete the row or column: simple, but loses information
    • impute a value, such as the mean, median or mode: keeps the rows, but can distort the data
    • predict the missing value from other features: more accurate, but more complex
  • Normalization: rescales values to a fixed range, usually 0–1 (min–max).
  • Standardization: rescales to a mean of 0 and a standard deviation of 1 (z-score). It's less sensitive to outliers.

A4.2.2 Feature selection

Keep the most informative features and drop the rest. The three strategy types in the guide:

  • Filter methods: score each feature with a statistic before training.
  • Wrapper methods: test combinations of features by training models.
  • Embedded methods: selection happens inside the training algorithm itself.

A4.2.3 Dimensionality reduction

The curse of dimensionality: as the number of features grows, these problems appear:

  • overfitting
  • higher computational cost
  • sparse data
  • less meaningful distance measures
  • harder visualisation
  • a need for more samples
  • higher memory use

Reducing the number of variables while keeping the relevant information counters these problems.

Note: specific statistical techniques such as PCA and LDA are beyond the scope of the IB course. You need the concept, not the method.

Exam-style question

Describe the difference between normalization and standardization, including when each is appropriate. [5]

Mark scheme

Award [1] for each point, up to [5]:

  • Normalization rescales values to a fixed range, usually 0–1, using min–max scaling;
  • It's appropriate for distance-based algorithms (such as K-NN and k-means) and for neural network inputs;
  • Standardization rescales a column to a mean of 0 and a standard deviation of 1 (a z-score);
  • It's appropriate when data is roughly normally distributed and for algorithms that expect zero-centred input;
  • Standardization is less sensitive to outliers than min–max normalization;

Common mistakes that cost marks

  • "Just delete every row with a missing value." This can throw away most of the data, and biases what's left.
  • Confusing normalization with standardization. One rescales to a fixed range; the other uses the mean and standard deviation.
  • Confusing feature selection with dimensionality reduction. Selection keeps some of the original features; reduction creates fewer new ones.

Frequently asked questions

Is A4.2 on the SL syllabus?
No. The whole of A4.2 is HL only.

Do I need to know PCA?
No. The guide states PCA and LDA are beyond the scope of the course.

Go further