University of Melbourne · S2 2026 · FACULTY OF INFORMATION TECHNOLOGY

COMP20008 Elements of Data Processing

- one subject, every graph, every model, every mark
12 Chapters105-page Bible
Our own words - no uploaded lecturer files
Built to mirror S2 2026 · updated this semester
The Complete Exam Bible · S2 2026

COMP20008 Overview

Elements of Data Processing
— one subject, every step of the data pipeline, every exam mark
  • University of Melbourne
  • S2 2026
  • 12 chapters
  • Information Technology

COMP20008 Elements of Data Processing is the University of Melbourne's second-year data-science core — an end-to-end tour of the data pipeline: getting data in (types, formats, quality, crawling and scraping), cleaning and enhancing it (imputation, scaling, text pre-processing, Bag-of-Words and TF-IDF), interpreting it (clustering, PCA, correlation and mutual information, classification and regression, evaluation and feature selection), and using it responsibly (visualisation, ethics, intellectual property, k-anonymity and differential privacy, and modern LLMs, prompting and RAG).

It is assessed 50% by continuous project work and 50% by a closed-book 2-hour written exam, with a dual hurdle — you must score at least 20/50 on each half.

  • Assessed by Continuous / project assessment (data-processing & analysis projects across the semester, ~60 hrs; submissions ~week 5 and ~week 11) 50% + 1 more — Continuous / project assessment (data-processing & analysis projects across the semester, ~60 hrs; submissions ~week 5 and ~week 11), End-of-semester examination are a hurdle
  • Key terms The data pipeline, Bag of Words (BoW), TF-IDF, Mutual information & NMI
  • How to prepare Treat COMP20008 as a breadth subject where method-selection is the skill: the formula sheet is given…
Why read this guide

What you get in the full COMP20008 Exam Bible

  1. 12 chapters, 105 pages — every examinable topic, week by week
  2. 18 original practice questions, each with a model answer in the End-of-semester examination format
  3. 12 worked examples, solved the way COMP20008 marks
  4. 30 computed figures, redrawn from the course’s own numbers
  5. Exam-morning plan for the End-of-semester examination (50%) — what to revise last, where marks are lost

Free on this page: assessment table, chapter map, one worked example, key terms, FAQ. The rest is in the full exam bible.

COMP20008 · University of Melbourne
An independent, AskSia-authored study guide. AskSia is not affiliated with, endorsed by, or sponsored by University of Melbourne; the course code and name are used for identification only.
Assessment

How COMP20008 is assessed

ComponentWeightFormat
Continuous / project assessment (data-processing & analysis projects across the semester, ~60 hrs; submissions ~week 5 and ~week 11) · hurdle50%Project work applied to real datasets; in the current offering delivered as Assignment 1 (individual: coding & analysis + interactive oral via Zoom + critical-thinking submission) and Assignment 2 (group: report + slides + in-person group oral), plus weekly ungraded formative quizzes. A1/A2 sub-weights subject to confirmation
End-of-semester examination · hurdle50%Closed-book written exam, 2 hours, during the exam period; a formula sheet is printed on page 1; ~11–12 questions of 2–4 marks each — short-answer + small by-hand calculation + critical-evaluation across the whole syllabus
Other students are asking · COMP20008

What other students are asking about COMP20008

Where can I find lecture notes or study guides for COMP20008 at UniMelb?
The COMP20008 study guide in AskSia Library covers Elements of Data Processing in 12 chapters, with the assessment table, a chapter-by-chapter topic map, key terms and a worked example, free to read on this page.
Is there a final exam in COMP20008, and how much is it worth?
Yes. In the offering this AskSia Library guide is built from, End-of-semester examination carries 50% of the COMP20008 mark; the full split is Continuous / project assessment (data-processing & analysis projects across the semester, ~60 hrs; submissions ~week 5 and ~week 11) 50%, End-of-semester examination 50%.
What does COMP20008 cover on data cleaning, imputation & scaling?
In AskSia Library's COMP20008 guide, the Data Cleaning, Imputation & Scaling chapter explains week 2. The six data-quality dimensions; missing/inconsistent/duplicate/outlier problems; imputation (drop, mean/median/mode, model-based); min-max & z-score scaling; sampling.
What does COMP20008 cover on web crawling & scraping?
In AskSia Library's COMP20008 guide, the Web Crawling & Scraping chapter explains week 3. Crawling (link-following frontier, coverage, robots.txt) vs scraping (targeted field extraction); the seed-URL → fetch → parse → BeautifulSoup → clean-matrix pipeline; the visited-set flaw.
What does COMP20008 cover on text pre-processing & lemmatisation?
In AskSia Library's COMP20008 guide, the Text Pre-processing & Lemmatisation chapter explains week 3. Sentence splitting with exceptions; tokenisation & normalisation; lemmatisation vs stemming; stop-word removal; n-grams; why text has hidden structure.
Weights, dates and every chapter are set out with their sources further down this page.
Contents · every chapter, one map

What COMP20008 covers

The whole subject → one exam-ready map. Each topic links to its free chapter guide.

01

The Data Pipeline & Data Types

Weeks 1–2. The five-stage data pipeline; numerical vs non-numerical vs could-be-either; continuous/discrete; structured/semi-structured/unstructured; levels of measurement; JSON, XML and HTML.

The data pipeline — Input → Cleanup & Enhancement → Interpretation → Output → Use — is the spine of the whole subject, and every later week re-anchors to one of its five stages.

  • The five-stage data pipeline: Input → Cleanup & Enhancement → Interpretation → Output → Use
  • Numerical vs non-numerical vs could-be-either — and why "looks like digits" ≠ numerical
  • Continuous vs discrete numerical data
  • Levels of measurement: nominal, ordinal, interval, ratio
  • Structured vs semi-structured vs unstructured data
  • Semi-structured formats: JSON, XML and HTML
In the full guide: 6 pages · 2 figures
02

Data Cleaning, Imputation & Scaling

Week 2. The six data-quality dimensions; missing/inconsistent/duplicate/outlier problems; imputation (drop, mean/median/mode, model-based); min-max & z-score scaling; sampling.

Real data is dirty, so before any model runs you must audit quality and fix it.

  • The six data-quality dimensions: accuracy, completeness, consistency, timeliness, believability, interpretability
  • Common problems: missing values, inconsistent representation, duplicates, outliers, wrong/implausible values
  • Resolving inconsistencies: standardise naming, formats and equivalency; de-duplicate
  • Imputation: drop rows/cols; mean/median/mode; forward/backward fill; model-based (regression / k-NN)
  • The imputation trade-off: bias vs information loss
  • Min-max normalisation: x' = (x − min) / (max − min) → [0, 1]
In the full guide: 5 pages · 2 figures
03

Web Crawling & Scraping

Week 3. Crawling (link-following frontier, coverage, robots.txt) vs scraping (targeted field extraction); the seed-URL → fetch → parse → BeautifulSoup → clean-matrix pipeline; the visited-set flaw.

When the data you need lives on the web, you either crawl (systematically discover and fetch pages by following links to build coverage) or scrape (extract specific structured fields from a page's content).

  • Crawling: link-following frontier, breadth-first traversal, goal = coverage/indexing
  • Scraping: targeted extraction of structured fields, goal = data
  • robots.txt: the site owner declares which paths crawlers may/may not visit
  • The scraping pipeline: seed URL → fetch raw HTML → parse (html.parser → DOM)
  • Extraction: BeautifulSoup find_all() to select target nodes → get_text() to strip tags
  • Cleaning: tokenise & standardise the extracted text → clean matrix for analysis
In the full guide: 4 pages · 1 figure
04

Text Pre-processing & Lemmatisation

Week 3. Sentence splitting with exceptions; tokenisation & normalisation; lemmatisation vs stemming; stop-word removal; n-grams; why text has hidden structure.

Raw text has hidden structure, and turning it into clean tokens is the step before any Bag-of-Words or TF-IDF representation.

  • Why text has hidden structure and needs pre-processing before modelling
  • Stage 1 — sentence splitting at . ? ! with abbreviation and number exceptions
  • Stage 2 — tokenisation & normalisation: lowercase, strip punctuation
  • Stage 3 — lemmatisation: map words to a dictionary base form ("cats" → "cat", "chased" → "chase")
  • Lemmatisation vs stemming: dictionary base form vs crude suffix stripping
  • Stage 4 — stop-word removal and n-grams
In the full guide: 6 pages · 1 figure
05

Bag-of-Words, TF-IDF & Visualisation

Week 4. BoW document/word vectors (order discarded); windowed co-occurrence vectors; Euclidean distance; TF vs TF-IDF for distinguishing topics; chart-type choice and the binning effect.

Once text is tokenised, Bag-of-Words turns it into count vectors (order discarded) that you can compare with Euclidean distance, and TF-IDF reweights those counts so distinctive words — not ubiquitous ones — drive similarity.

  • Bag-of-Words: a document as a vector of word counts over a fixed vocabulary; order discarded
  • Windowed co-occurrence vectors: a word represented by counts of context words within a window
  • Euclidean distance between two BoW vectors: d(x, y) = √( Σᵢ (xᵢ − yᵢ)² )
  • BoW positives (simple, interpretable, fixed-length, language-agnostic) vs negatives (loses order, sparse, common words dominate)
  • TF vs TF-IDF: down-weight words common across all documents, up-weight distinctive ones
  • Why pure TF clusters topics poorly and IDF fixes it
In the full guide: 8 pages · 2 figures
06

Clustering: k-means, VAT & the Elbow

Week 5. k-means (assign/recompute, SSE/inertia); the elbow method (advantage/limitation); outlier sensitivity and mitigation (k-medoids, robust/density methods); the VAT reordered distance matrix.

Clustering finds groups in unlabelled data.

  • k-means: assign each point to the nearest centroid → recompute centroids → repeat to convergence
  • The objective: minimise within-cluster sum of squared errors (SSE / inertia)
  • The elbow method: plot SSE vs k and pick the k at the bend
  • Elbow advantage (simple, visual) vs limitation (the bend can be ambiguous/subjective)
  • Alternatives to the elbow: silhouette, Calinski–Harabasz
  • Outlier sensitivity: a far outlier inflates SSE and pulls a centroid toward itself
In the full guide: 7 pages · 3 figures
07

Hierarchical Clustering & Dendrograms

Weeks 5–6. Agglomerative clustering; single vs complete vs average linkage; building dendrograms (y-axis = merge distance); cutting the tree for k clusters; NMI for comparing clusterings.

Unlike k-means, agglomerative hierarchical clustering needs no pre-set k: it repeatedly merges the two closest clusters, building a nested tree — a dendrogram whose y-axis is the merge distance — that you can "cut" at any level to get any number of clusters.

  • Agglomerative clustering: start with each point as its own cluster, merge the two closest, repeat
  • The dendrogram: a nested tree whose y-axis is the distance at which clusters merge
  • Cutting the tree: a horizontal cut at a given height yields a chosen number of clusters
  • Single linkage: distance between clusters = minimum pairwise distance (chaining)
  • Complete linkage: distance between clusters = maximum pairwise distance (compact clusters)
  • Average linkage: distance between clusters = mean pairwise distance
In the full guide: 5 pages · 1 figure
08

Correlation, Entropy & Mutual Information

Weeks 5–6. Pearson r (linear only); entropy & conditional entropy (log₂); mutual information & NMI; PCA (principal axes, variance explained, the 95%-on-PC1 caution); high-NMI-but-zero-Pearson (y = x²).

This chapter is about measuring how two variables relate — and why the choice of measure matters.

  • Pearson r: measures linear association only, range −1 ≤ r ≤ 1; blind to curved/non-monotone relations
  • Entropy H(X) = − Σ pᵢ log₂ pᵢ: bits of uncertainty; max when uniform, 0 when deterministic
  • Conditional entropy H(Y | X): uncertainty in Y once X is known
  • Mutual information MI(X, Y) = H(Y) − H(Y | X): reduction in uncertainty; captures any dependence; 0 iff independent
  • NMI = MI / min(H(X), H(Y)): rescales MI to 0 ≤ NMI ≤ 1
  • Pearson vs MI/NMI: y = x² has near-zero Pearson but high NMI
In the full guide: 7 pages · 3 figures
09

Supervised Learning: k-NN & Decision Trees

Week 7. Classification vs regression; train/test; k-NN (distance, k, scaling, cold-start/Jaccard pitfalls); decision trees (information gain/Gini, depth); high-cardinality split overfitting.

Supervised learning predicts a label (classification) or a value (regression) from labelled examples.

  • Classification vs regression: predict a class/label vs a continuous value
  • Train/test split: learn from labelled data, evaluate on held-out data
  • k-NN: classify by majority vote of the k nearest neighbours (distance metric, scale features)
  • k-NN pitfalls: choice of k, irrelevant features, lazy learner, cold-start in recommenders
  • Jaccard-based user k-NN: new-user cold-start (tiny |A∩B|), niche-user irrelevance, drift
  • Decision trees: split on the feature with the best information gain (entropy reduction) or Gini
In the full guide: 8 pages · 2 figures
10

Regression, Evaluation & Feature Selection

Week 8. Linear regression (least squares, slope, R²); experimental design; k-fold cross-validation; confusion matrix to accuracy/precision/recall/F1; the precision-recall trade-off; feature selection; data leakage.

This chapter covers how to fit, evaluate and trust a predictive model.

  • Linear regression: ŷ = b₀ + b₁x by least squares; interpret slope; residuals; R²
  • Experimental design: train / validation / test split, hold-out
  • k-fold cross-validation: rotate the held-out fold to estimate performance more reliably
  • Confusion matrix: TP, FP, FN, TN
  • Metrics from the formula sheet: accuracy, precision, recall, F1
  • Precision vs recall trade-off: choose by the cost of false positives vs false negatives
In the full guide: 7 pages · 2 figures
11

Ethics, IP & Privacy

Week 9. Intellectual property (patent/copyright/trademark, licensing, GDPR); privacy levels; k-anonymity & quasi-identifiers; l-diversity vs the homogeneity attack; global differential privacy; bias & fairness.

The "data use" stage of the pipeline asks: just because you can process data, should you?

  • Intellectual property: copyright (expression, automatic), patent (inventions, registered), trademark (brand identifiers)
  • Control mechanisms: licensing, copyright, patents, trademarks, plus GDPR for personal data
  • Privacy levels: self, local (alter released record), global (alter answers to queries)
  • k-anonymity: generalise/suppress quasi-identifiers so each record matches ≥ k−1 others
  • Quasi-identifiers (age, postcode, employment) vs the sensitive attribute
  • l-diversity: each group has ≥ l distinct sensitive values — defends the homogeneity attack
In the full guide: 7 pages · 2 figures
12

LLMs, Prompting, RAG & Verifiable Data

Weeks 10–12. What makes an LLM generative (embeddings, hallucination); prompts vs fine-tuning; the AUTOMAT prompt schema; the RAG store and content-partitioning; blockchain tamper-evidence and digital-signature verification.

The subject closes on modern generative AI and on how to trust data.

  • What makes an LLM generative: web-scale text → embeddings (vectors) → neural network recombines to produce novel output
  • Hallucination: predicting plausible-sounding tokens with no grounding in fact → confident but wrong
  • Prompts vs fine-tuning: steer a frozen model in natural language (cheap) vs continued training (costly)
  • The AUTOMAT prompt schema: Act-as, User, Targeted-action, Output, Mode, Atypical cases, Topic
  • RAG: retrieve relevant external chunks → add to the prompt → generate a grounded answer
  • The content-partitioning problem: chunking and permissioning so the right (and only permitted) content is retrieved
In the full guide: 10 pages · 3 figures

The continuous half is currently delivered as an individual assignment and a group assignment, each with an oral assessment, plus weekly ungraded quizzes (A1/A2 sub-weights subject to confirmation).

The exam supplies a formula sheet on page 1, so marks come from picking the right method, doing one clean by-hand calculation, and critically evaluating the trade-offs — not memorising formulas.

COMP20008 cheatsheet for exam revision

AskSia publishes a free COMP20008 cheatsheet for exam revision, a one-page revision sheet that condenses the topics in this guide into the priorities, methods and drills to practise before the final.

The permitted-materials rule for this offering is not confirmed in the course materials we hold, so check the exam instructions on the LMS before relying on it in the exam room. Open the COMP20008 cheatsheet.

Worked example · free

Bag-of-Words word vectors + Euclidean distance (closed-book, formula-sheet style)

Q [4 marks]. In the text "The fox saw the hen and chased it. The hen turned and chased the fox.", build Bag-of-Words context vectors for "fox" and "hen" using a window of 2 words either side (excluding the target itself), then compute the Euclidean distance between them. You may leave the answer as a square root.
Bag-of-Words turns text into vectors.Bag-of-Words turns text into vectors. Fix a shared vocabulary, then count how often each word appears in each document. Word order is discarded — only the counts survive. Each ROW is a document vector; documents become points you can measure distance between.Two documentsD1: “the cat sat onthe mat”D2: “the dog sat onthe log”order is discarded → just countscountword→thecatsatonmatdoglogD12111100D22011011D1 = (2, 1, 1, 1, 1, 0, 0)D2 = (2, 0, 1, 1, 0, 1, 1)each row is a point → measure d(D1, D2)
Bag-of-Words turns text into vectors. Fix a shared vocabulary, then count how often each word appears in each document. Word order is discarded — only the counts survive. Each ROW is a document vector; documents become points you can measure distance between.
  • 1 markCollect the context. "fox" occurs twice — context {the, saw} and {chased, the}; "hen" occurs twice — context {chased, it, the, turned} and {and, chased}. So fox: the=2, saw=1, chased=1; hen: chased=2, it=1, the=1, turned=1, and=1.
  • 1 markAlign on the shared vocabulary {the, saw, chased, it, turned, and} and pad missing words with 0: fox = (the 2, saw 1, chased 1, it 0, turned 0, and 0); hen = (the 1, saw 0, chased 2, it 1, turned 1, and 1).
  • 1 markApply the Euclidean-distance formula d(x, y) = √( Σᵢ (xᵢ − yᵢ)² ): d = √( (2−1)² + (1−0)² + (1−2)² + (0−1)² + (0−1)² + (0−1)² ).
  • 1 markEvaluate the sum of squares: each squared difference is 1, so d = √(1 + 1 + 1 + 1 + 1 + 1) = √6.
d(fox, hen) = √6 — you may leave the answer as a square root; the exam states you do not have to reduce a root to a decimal.
Sia tip — Build the union vocabulary first and pad missing words with 0 before subtracting — the most common mark-loser is forgetting a word that appears for only one of the two targets, which throws every later squared difference off.
Glossary

Key terms

The data pipeline
The spine of the whole subject: Data Input → Data Cleanup & Enhancement → Data Interpretation (models) → Data Output (visualisation) → Data Use (ethics). Every week re-anchors to one of these five stages, so naming the stage a question lives in is half the answer.
Bag of Words (BoW)
Represents a document (or a word's context) as a vector of counts over a fixed vocabulary; word order is discarded. Simple, interpretable, fixed-length and language-agnostic, but loses order/semantics and is sparse and high-dimensional, with common words dominating.
TF-IDF
Term frequency weighted by inverse document frequency, tf·idf with idf(t) = log(N / df(t)). It down-weights words common across all documents and up-weights distinctive ones, so topic-specific words drive similarity — which is why it clusters documents by topic far better than raw TF.
Mutual information & NMI
Mutual information MI(X,Y) = H(Y) − H(Y|X) measures the reduction in uncertainty about Y from knowing X; it captures any dependence (including non-linear) and is 0 only if X and Y are independent. NMI rescales MI to 0–1 by dividing by min(H(X), H(Y)).
k-anonymity vs l-diversity
k-anonymity generalises/suppresses quasi-identifiers so every record is indistinguishable from at least k−1 others — it hides who. l-diversity additionally requires at least l distinct sensitive values within each group — it hides what, defending against the homogeneity attack that defeats k-anonymity alone.
FAQ

COMP20008 FAQ

Is COMP20008 hard?

It is broad rather than deep. The challenge is the unusually wide surface — data types and quality, cleaning, web scraping, text and Bag-of-Words, clustering and PCA, correlation and mutual information, classification and regression, evaluation, privacy, LLMs/RAG and blockchain — almost all examined as short-answer plus a tiny by-hand calculation plus "critically evaluate", not coding.

With a formula sheet provided, the difficulty is breadth and method-selection, not memorising maths. Students who keep up week-to-week and practise the recurring question types find the exam very predictable.

Is the COMP20008 exam a hurdle?

Yes. The handbook sets a dual hurdle: you must score at least 20/50 on the continuous project assessment and at least 20/50 on the end-of-semester exam, on top of passing overall. The headline split is 50% continuous / 50% exam, so the exam is both the largest single stake and a barrier you cannot afford to fail.

Do I need to memorise the formulas?

No. A formula sheet is printed on page 1 of the exam — Euclidean distance, Pearson r, entropy and conditional entropy, mutual information and NMI, accuracy/precision/recall/F1, Jaccard and cosine. Marks come from choosing the right method, substituting the numbers correctly with one clean by-hand calculation, and explaining the trade-off.

The exam even notes you don't have to reduce a square root to a decimal — answers are expressions, not memorised numbers.

Is COMP20008 a coding exam?

Not in the written exam. Coding (pandas, BeautifulSoup, scikit-learn) is assessed through the continuous project half and the workshops.

The closed-book exam is short-answer, small by-hand computation and critical evaluation — for example classifying data types, building a BoW vector, reading a dendrogram, or naming the two steps of differential privacy — so practise reasoning and calculation by hand, not writing code under exam conditions.

What does the continuous assessment involve?

In the current instance the 50% continuous half is two assignments on real datasets. Assignment 1 is individual — a coding-and-analysis submission, an interactive oral assessment over Zoom, and a critical-thinking submission. Assignment 2 is a group project — a report, slides and an in-person group oral (content marked at the group level, communication individually).

There are also weekly ungraded quizzes for formative practice. The exact sub-weights inside the 50% are subject to confirmation against your unit outline.

Chapter questions

Questions students ask in each COMP20008 chapter

Why does the same value sometimes count as numerical and sometimes not?
Because the deciding test is whether arithmetic on the value is meaningful, not whether it is written with digits. "3010" as a postcode is a label — averaging postcodes is nonsense — so it is non-numerical; "3010" as a count of items supports addition and averaging, so it is numerical. Always ask what the number represents before you classify it. Chapter 01 · The Data Pipeline & Data Types
When should I drop rows versus impute missing values?
Drop when missingness is rare and looks random, so removing a few rows barely affects the dataset; impute when too many rows would be lost or when the values are systematically missing in a way you can model. The trade-off is the heart of the answer: dropping loses information and can bias the sample, while imputing fabricates plausible values and can dampen real variance. State the trade-off, not just the method. Chapter 02 · Data Cleaning, Imputation & Scaling
What is the core difference between crawling and scraping?
Crawling is about breadth — discovering and fetching many pages by following links, with coverage as the goal — while scraping is about depth on one page — extracting specific fields from its content, with data as the goal. Crawling is a breadth-first traversal of the link graph; scraping is targeted parsing of a single page's DOM. In practice you often crawl to find pages, then scrape each one. Chapter 03 · Web Crawling & Scraping
What is the difference between lemmatisation and stemming?
Both reduce inflected words to a common form, but lemmatisation uses dictionary and grammatical knowledge to return a valid base word ("chased" → "chase", "better" → "good"), while stemming just strips suffixes by rule and may return a non-word ("studies" → "studi"). Lemmatisation is more accurate and is preferred; stemming is faster and cruder. The exam frequently wants this exact contrast. Chapter 04 · Text Pre-processing & Lemmatisation
What information does Bag-of-Words throw away, and does it matter?
It discards word order, so "dog bites man" and "man bites dog" give the same vector, and it loses most semantics. It is also sparse and high-dimensional, and common words can dominate. Whether that matters depends on the task: for coarse topic similarity BoW (especially with TF-IDF) works well; for anything needing order or meaning (sentiment with negation, syntax) you need richer representations or n-grams. Chapter 05 · Bag-of-Words, TF-IDF & Visualisation
Why can't I just pick the k that minimises SSE?
Because SSE always decreases as k increases — with k equal to the number of points, every point is its own cluster and SSE is 0. Minimising SSE would always pick the largest k, which is meaningless. The elbow method instead looks for the k where SSE stops falling sharply, balancing fit against the number of clusters; silhouette and Calinski–Harabasz are alternatives that score cluster quality directly. Chapter 06 · Clustering: k-means, VAT & the Elbow
How is hierarchical clustering different from k-means?
k-means needs you to fix k up front, randomly initialises centroids, and partitions the data into exactly k flat clusters by minimising SSE. Agglomerative hierarchical clustering needs no k in advance: it merges the closest clusters step by step into a full tree (dendrogram), and you choose the number of clusters afterwards by cutting the tree at a height. Hierarchical also gives a nested structure showing how groups relate, which k-means does not. Chapter 07 · Hierarchical Clustering & Dendrograms
How can two variables have zero correlation but be strongly related?
Pearson r only measures linear association, so a relationship that is strong but curved or symmetric can give r ≈ 0. The classic example is y = x² over a symmetric range of x: y is fully determined by x, yet as x rises y falls then rises, cancelling the linear trend so r ≈ 0. Mutual information captures this dependence and is high, which is why the exam contrasts the two measures. Chapter 08 · Correlation, Entropy & Mutual Information
Why does feature scaling matter so much for k-NN?
k-NN decides neighbours by distance, so a feature measured on a large scale (income in dollars) will dominate the distance and swamp a small-scale feature (age in years), purely because of its units. Without scaling, the "nearest" neighbours are chosen mostly by the big-range feature, distorting the classification. Scaling each feature (min-max or z-score) puts them on a comparable footing so distance reflects genuine similarity. Chapter 09 · Supervised Learning: k-NN & Decision Trees
Why can a model with 97% accuracy still be bad?
Because accuracy is misleading on imbalanced data. If 95% of transactions are legitimate, a model that predicts "not fraud" for everything scores 95% accuracy while catching zero fraud. The metric you trust depends on the task: for fraud or disease screening, recall (catching the rare positives) matters far more, and you should report precision, recall and F1 rather than accuracy alone. Chapter 10 · Regression, Evaluation & Feature Selection
What is the difference between copyright, a patent and a trademark?
Copyright protects the expression of creative works — text, art, music, code — and arises automatically when the work is created. A patent protects a novel invention or process, must be applied for and granted, and lasts a limited term. A trademark protects identifiers of a brand — names, logos, slogans. They cover different things, so the same product can involve all three (patented mechanism, copyrighted manual, trademarked name), plus licensing and GDPR for personal data. Chapter 11 · Ethics, IP & Privacy
What makes a large language model "generative"?
It is trained on web-scale text that is broken into fragments and converted into embeddings — vectors of numbers — and a neural network learns to recombine them to produce new text. "Generative" means it creates novel output token by token rather than looking up and returning a stored answer. The same idea extends to images and other modalities. The 2024 Q9a wants this chain: web text → embeddings → neural recombination → novel output. Chapter 12 · LLMs, Prompting, RAG & Verifiable Data
Study strategy

How to study for the exam

Treat COMP20008 as a breadth subject where method-selection is the skill: the formula sheet is given, so your edge is picking the right tool fast and explaining the trade-off. (1) Anchor every topic to the five-stage data pipeline (input → cleanup → interpretation → output → use) so you always know which stage a question lives in.

(2) Drill the recurring exam patterns directly — they repeat almost verbatim year to year: classify data types, name two data-quality problems, crawling vs scraping, BoW vectors + Euclidean distance, TF vs TF-IDF, single/complete-linkage dendrograms, the elbow + outliers, Pearson-vs-NMI (the y = x² archetype), confusion-matrix metrics + which to optimise, data leakage, the k-anonymity/l-diversity homogeneity attack, the differential-privacy two-step, AUTOMAT, RAG, and blockchain verification.

(3) For every calculation write the formula in symbols, substitute the numbers, then state the answer with one line of interpretation — examiners reward the reasoning chain and the "so what".

(4) Practise the "critically evaluate" muscle: each method has a standard limitation (Pearson misses non-linearity, the elbow is subjective, k-means is outlier-sensitive, high PCA variance ≠ relevance, leakage gives MSE 0), and naming it earns the mark.

(5) Keep up with the workshops and assignments — the by-hand skills they build are exactly what the closed-book exam tests, and the dual 20/50 hurdle means neither half can be neglected.

A+Everything unlocked
Unlocks this Bible + all 185 of your subjects at University of Melbourne, plus 1,000+ more Bibles across the AskSia library.
Sia - your COMP20008 tutor, unlimited, worked the way the exam marks it
The full 105-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works