about BUSS6002
01 · Exam map Course data + final instructions
| Final component | Marks | Fast implication |
|---|---|---|
| MCQ | 12 | definitions, traps, classify |
| Short answer | 27 | show formula + working |
| Python code | 6 | runnable syntax from memory |
Material permitted: One A4 sheet of handwritten notes DOUBLE SIDED. "Handwritten" means that the notes must be written directly by hand on the physical sheet of paper that you bring into the exam. (BUSS6002 final exam instructions and format (Semester 1 2026; confirm for Semester 2 on Canvas))
Priority: Copy formulas and decision triggers first; then notation traps; then the recall tables. Do not spend scarce handwriting space on introductions.
02 · Process + easy marks CH01–03
| Trigger | Answer |
|---|---|
| What / why / next / action? | descriptive / diagnostic / predictive / prescriptive |
| Big data | Volume · Variety · Velocity · Veracity; generation mechanism, not bytes |
| Decision context | simple / complicated / complex / chaotic: name cause–effect signature |
| Missingness | MCAR: none; MAR: observed driver; MNAR: missing value itself |
| CRISP-DM phase | Mark-winning job |
|---|---|
| Business understanding | problem, success, stakeholders, cost/ethics |
| Data understanding | collect, describe, explore, verify quality |
| Data preparation | clean, integrate, derive features |
| Modeling | choose, fit and tune method |
| Evaluation | check technical + business success |
| Deployment | deliver, monitor, maintain |
cor(x,y)=cov(x,y)/[sd(x)sd(y)] ∈ [−1,1]
cor = 0 does not imply independence. A circled box-plot point is extreme, not automatically erroneous. Investigate before deleting.
03 · Vectors + matrices CH04 + CH06
‖x‖₂ = √(xᵀx); ‖x‖₂² = xᵀx
d(x,y)=‖x−y‖₂; x ⟂ y ⇔ xᵀy=0
| Operation | Exam check |
|---|---|
| ax+by | equal lengths; scale each, then add |
| linear dependence | find coefficients not all zero giving Σaᵢxᵢ=0 |
| AB | (m×n)(n×p)→m×p; row·column |
| transpose | (Aᵀ)ᵢⱼ=Aⱼᵢ; generally AB≠BA |
σ̂(x)=(1/√n)‖x‖₂ (zero-mean)
cov̂(x,y)=⟨(1/n)x,y⟩ (zero-mean)
| Matrix object | Rule / role |
|---|---|
| I | AI=IA=A |
| A⁻¹ | AA⁻¹=I when inverse exists |
| det(A) | zero ⇒ singular; no unique inverse |
| data matrix X | rows observations; columns variables |
np.dot(x,y) # one scalar; vectorised
Dot product returns one number, not element-wise products. Norm/distance need the final square root; squared norm does not.
04 · K-means CH05
μᵢ = (1/|Cᵢ|)Σₓ∈Cᵢx
| Step | Do |
|---|---|
| 0 | choose k; initialise k centres |
| 1 assign | send each x to nearest μ by squared distance |
| 2 update | replace each μ by mean of new members |
| 3 repeat | until centres stop moving |
WCSS is non-increasing and bounded below by 0, hence convergence—but only to a local optimum. Run several starts; keep lowest WCSS. K-means does not choose k; use elbow/business actionability.
| Segmentation basis | Examples |
|---|---|
| demographic | age, income |
| geographic | location |
| psychographic | values, lifestyle |
| behavioural | usage, loyalty |
05 · OLS pipeline CH06
β̂ = arg minβ ‖y−Xβ‖₂² = (XᵀX)⁻¹Xᵀy
β̂₁=Sₓᵧ/Sₓₓ; β̂₀=ȳ−β̂₁x̄
R²=1−RSS/TSS=(TSS−RSS)/TSS
| Assumption / reading | Where it bites |
|---|---|
| E(ε|x)=0 | unbiased conditional mean; curve signals failure |
| var(ε|x)=σ² | constant variance; funnel signals failure |
| ε uncorrelated with x | estimation validity |
| Normal ε | needed for t/F inference, not for fitting β̂ |
approx. 95% CI: β̂₁ ± 2·SE(β̂₁)
| Question gives | Fast route |
|---|---|
| X and y matrices | normal equation (XᵀX)⁻¹Xᵀy |
| x̄,ȳ,Sₓₓ,Sₓᵧ | slope then intercept shortcut |
| fitted values | e=y−ŷ → RSS → R² |
| residual plot | shape → assumption → remedy |
Existence of (XᵀX)⁻¹ requires full column rank: redundant predictors make the normal equation singular. Centre the interpretation on units and the held-fixed predictors.
Interpret β̂ⱼ as the average change in y associated with +1 in xⱼ, holding other predictors fixed—never automatically causal.
06 · Residual decision CH06 Fig 3
| Plot | Diagnosis | Move |
|---|---|---|
| flat/even | adequate | no structure left |
| funnel | heteroscedastic | transform y, e.g. log y |
| curve/U | missed nonlinearity | add x² / basis term |
Residual eᵢ is the vertical observed-minus-fitted gap. RSS squares and sums those gaps. The least-squares line passes through (x̄,ȳ). A high R² describes in-sample fit only; it neither proves assumptions nor causality.
07 · Features + ethics CH07
| Symptom / case | Use / principle |
|---|---|
| positive right-skewed y; funnel | model log(y): symmetry + variance stabilisation |
| curvature | polynomial x²,x³,… |
| effect of x₁ depends on x₂ | interaction x₁x₂ |
| consent to collect? | Ownership |
| told how used? | Transparency |
| PII exposed/re-identifiable? | Privacy |
| purpose? | Intention |
| group harm? | Outcomes / disparate impact |
E(y|x)=exp(xᵀβ)·E[exp(ε)]
ŷcorr=exp(xᵀβ̂)·(1/n)Σexp(eᵢ)
By Jensen, E[exp(ε)]>1 for non-degenerate mean-zero error. Thus naive exp(xᵀβ̂) is downward biased; the smearing factor scales it up.
08 · Logistic regression CH08 Fig 1
p̂=1/[1+exp(−η)]; ŷ=1 if p̂>τ (default τ=0.5)
+1 in xⱼ multiplies odds by exp(βⱼ)
| GLM part | Linear | Logistic |
|---|---|---|
| response | continuous | Bernoulli |
| link g | identity | logit |
| activation g⁻¹ | identity | sigmoid |
| fit | OLS; closed form | MLE; iterative |
A probabilistic classifier retains p̂; thresholding discards confidence to create a label. BUSS6002 fixes τ=0.5 rather than asking you to optimise it.
Fit by Bernoulli maximum likelihood, not OLS. Positive η ⇒ p̂>0.5. Lower τ raises both recall and FPR.
09 · Classifier metrics CH08
| Actual + | Actual − | |
|---|---|---|
| Pred + | TP | FP · Type I |
| Pred − | FN · Type II | TN |
Recall/TPR=TP/(TP+FN)
Specificity/TNR=TN/(TN+FP)
FPR=FP/(FP+TN)=1−TNR
Precision=TP/(TP+FP)
F1=2TP/(2TP+FP+FN)
| Business question | Metric |
|---|---|
| Of real positives, caught? | recall |
| Of flags, trustworthy? | precision |
| Both matter, one score? | F1 |
| Rare positives? | do not rely on accuracy |
| ranking across thresholds? | ROC/AUC |
Propensity predicts natural action; campaign-response predicts action given exposure; uplift estimates the causal incremental effect using treatment vs control.
AUC near 1 means better ranking; 0.5 is chance. ROC runs from (0,0) to (1,1) and is monotone non-decreasing as the threshold relaxes.
10 · Select complexity CH09 Fig 1
complexity ↑ ⇒ bias ↓, variance ↑ ⇒ EPE U-shaped
polynomial: φᵢ(x)=xⁱ
radial: φᵢ(x)=exp[−(x−cᵢ)²/(2s²)]
| Data split | Only job |
|---|---|
| train | fit each candidate |
| validation | select lowest MSE |
| test | one final unbiased report |
11 · MLE + optimisation CH10 Fig 3
ℓ(θ)=logL(θ)=Σᵢlogp(yᵢ|θ)
θ̂=arg maxθ ℓ(θ)
| MLE line | Mark |
|---|---|
| write product likelihood | independence |
| take logs | product→sum |
| differentiate | score |
| set =0; solve | candidate θ̂ |
| second derivative <0 | maximum check |
Why log? Avoid underflow; simplify differentiation; log is increasing so argmax unchanged. Drop additive terms without θ.
| α | Failure |
|---|---|
| too large | overshoot/diverge |
| too small | very slow |
OLS has closed form. Logistic score is nonlinear, so minimise cross-entropy f=−ℓ iteratively; ∇f(β)=Xᵀ(p−y).
argmin is the parameter location; min is the achieved objective value. Independence is what permits L to factor into a product and ℓ into a sum.
12 · Scale by shape CH11
full gradient: O(n)/step; one-row SGD: O(1)/step in n
Big-O is an asymptotic upper bound on worst-case growth, not an exact operation count. SGD is noisy but updates much more often for tall data.
| Shape | Failure | Correct move |
|---|---|---|
| p>n | overfit; XᵀX singular | feature selection, projection, less flexibility |
| n≫p | each full pass expensive | blocks or one-row SGD |