Data Preprocessing

Data cleaning is about fixing problems in data, not preserving what's already good. Incomplete data (A) refers to missing values or gaps — a core cleaning target. Noisy data (D) refers to errors, outliers, and random variance that distort analysis. Intentionally incorrect data (E) covers deliberate errors or falsifications that still need to be detected and corrected. Options B and C ("accurate" and "consistent") describe desirable outcomes of cleaning — they are goals, not problems being addressed.

The standard data quality dimensions include: Accuracy (A) — does the data correctly reflect reality? Completeness (C) — is all required data present? Timeliness (E) — is the data up-to-date and available when needed? These are widely accepted ISO/industry quality measures. Redundancy (B) is a storage/design concern (duplicate data), not a quality dimension itself. Obsolescence (D) describes data becoming outdated, which is the absence of timeliness — the concept is captured by E rather than being a separate metric.

The standard automated imputation techniques are: Global constant (A) — replace all missing values with a placeholder like "Unknown" or 0. Mean/median (B) — substitute the attribute's central tendency (mean for continuous, median for skewed data). Decision tree induction (E) — train a model on non-missing rows to predict missing values. Option C (deleting the entire dataset) is never a valid imputation strategy — it destroys information. Option D (attribute correlations) sounds plausible but is not a standard named technique in this taxonomy; the correlation insight is embedded within prediction models like regression or decision trees.

The three canonical noise-smoothing techniques are: Binning (A) — groups nearby values into bins and replaces them with the bin mean, median, or boundary. Regression (C) — fits a mathematical function to data and uses predictions to replace noisy values. Clustering (E) — identifies outliers as points that don't fit any cluster and flags them for removal or replacement. Duplication (B) creates more copies of data — it has nothing to do with noise reduction. Aggregation (D) is a data reduction technique (e.g., rolling averages), but is not classified as a noise-smoothing technique in this course's framework.

This is a tricky question! The course specifically identifies two root causes: Equipment malfunction (A) — a sensor or recording device failing mid-collection leaves gaps. Data entry mistakes (C) — a human operator skips a field or enters an invalid value that gets removed during validation. Options D and E might seem reasonable in the real world, but in the context of this course they are not listed as canonical causes. Deliberate omissions (D) relate more to data suppression/anonymisation, and transmission timeouts (E) are an infrastructure issue, not a data collection/entry cause in this taxonomy.

The key phrase here is "neighboring data points." Binning (A) explicitly groups sorted values into local neighborhoods (bins) and smooths each value using its bin neighbors — this is the textbook definition of local smoothing. Linear Regression (D) fits a global line using all data points and predicts smoothed values — it uses all neighboring context implicitly. The other options are either incorrectly described or use different mechanisms: K-means is flat (not hierarchical), and decision trees are for classification/prediction, not ensemble outlier detection as described.

A is correct — this is a precise description of binning: sort → partition → smooth locally. B is correct — linear regression imputation works well for low-variance targets but degrades when there is high unexplained variance (poor R²). E is correct — decision trees use information gain (for classification) or Gini impurity and can be applied to rank/select features. C is incorrect — outliers detected by clustering are removed or treated, not merged into nearby clusters; adding them in would defeat the purpose. D is incorrect — equal-width binning actually performs worse on skewed data because rare extreme values dominate entire bins; equal-depth (quantile) binning handles skew better by ensuring equal counts per bin.

The two types of binning are easily confused: Equal-width (B) divides the value range into k intervals of the same size — e.g., if values range from 0–100 and k=5, each bin covers 20 units regardless of how many data points fall in it. Equal-depth (equal-frequency) (described by option A) ensures each bin contains the same number of records. Options C, D, and E describe neither — adaptive/dynamic partitioning is a different concept entirely.

Feature selection has three method families: Filter (statistical tests, fast, model-agnostic), Wrapper (use an actual model to score subsets), and Embedded (selection built into model training). The key advantage of wrapper methods (B) is that they account for feature interactions with a specific model — they search for the subset that literally makes the model perform best. This comes at a cost: they are computationally expensive (ruling out A) and struggle with large datasets (ruling out D). Option E is the opposite of what wrappers do, and C is nonsensical.

RFE is a backward elimination approach: it starts with all features, trains the model, identifies the least important feature, removes it, and repeats until the target number of features is reached. So: B is correct (iterative removal of least important). C is correct (features get a ranking from most to least important as a by-product). E is correct (RFE is model-agnostic — it works with SVMs, random forests, logistic regression, etc., as long as the model can produce feature importance scores). A is wrong — that describes forward selection, not RFE. D is wrong — RFE uses a target number of features to stop, not a statistical threshold.