BUSS6002 Cheatsheet · Data Science in Business

University of Sydney · handwritten A4 planner for the final exam. Semester 2 formal examination period, from 16 November 2026 (the Exams Office publishes the exact date). Full BUSS6002 study guide →

A4 landscape · 7pt · 6 columns · ink + yellow · copy side 1 by hand 2 sides · BUSS6002 · University of Sydney
BUSS6002
Data Science in Business
University of Sydney · postgraduate coursework · Semester 2, 2026
BUSS6002 cheatsheet QR code – open the full BUSS6002 study guide
Full guide + ask Sia
about BUSS6002
SIDE 1/2   PRIORITY-ORDER PLAN FOR YOUR HANDWRITTEN A4 HANDWRITTEN A4 PLANNER Methods · formulas · decisions · assumptions

01 · Exam map Course data + final instructions

Mid-semester exam 25%Individual assignment 30%Final exam 45%
Final componentMarksFast implication
MCQ12definitions, traps, classify
Short answer27show formula + working
Python code6runnable 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

> Business goalData inspectData prepareModelEvaluateDeploy iterate ↻
CRISP-DM in order; iteration is allowed, but each phase has a distinct job.
TriggerAnswer
What / why / next / action?descriptive / diagnostic / predictive / prescriptive
Big dataVolume · Variety · Velocity · Veracity; generation mechanism, not bytes
Decision contextsimple / complicated / complex / chaotic: name cause–effect signature
MissingnessMCAR: none; MAR: observed driver; MNAR: missing value itself
CRISP-DM phaseMark-winning job
Business understandingproblem, success, stakeholders, cost/ethics
Data understandingcollect, describe, explore, verify quality
Data preparationclean, integrate, derive features
Modelingchoose, fit and tune method
Evaluationcheck technical + business success
Deploymentdeliver, monitor, maintain
EDA anchorsIQR = Q3 − Q1; fences = [Q1 − 1.5·IQR, Q3 + 1.5·IQR]
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

> vwprojᵥwxy xᵀy = ‖x‖‖y‖cosθ
Source redraw: dot product links algebra to angle; orthogonal means xᵀy=0.
Vector chain⟨x,y⟩ = Σxᵢyᵢ = xᵀy → scalar
‖x‖₂ = √(xᵀx); ‖x‖₂² = xᵀx
d(x,y)=‖x−y‖₂; x ⟂ y ⇔ xᵀy=0
OperationExam check
ax+byequal lengths; scale each, then add
linear dependencefind coefficients not all zero giving Σaᵢxᵢ=0
AB(m×n)(n×p)→m×p; row·column
transpose(Aᵀ)ᵢⱼ=Aⱼᵢ; generally AB≠BA
Statistics as inner productsμ̂ₓ=⟨(1/n)1,x⟩
σ̂(x)=(1/√n)‖x‖₂   (zero-mean)
cov̂(x,y)=⟨(1/n)x,y⟩   (zero-mean)
Matrix objectRule / role
IAI=IA=A
A⁻¹AA⁻¹=I when inverse exists
det(A)zero ⇒ singular; no unique inverse
data matrix Xrows observations; columns variables
Handwritten NumPy recallx=np.array([3,1,-2]); y=np.array([2,2,4])
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

Objective + updateWCSS = Σᵢ Σₓ∈Cᵢ ‖x−μᵢ‖₂²
μᵢ = (1/|Cᵢ|)Σₓ∈Cᵢx
StepDo
0choose k; initialise k centres
1 assignsend each x to nearest μ by squared distance
2 updatereplace each μ by mean of new members
3 repeatuntil 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 basisExamples
demographicage, income
geographiclocation
psychographicvalues, lifestyle
behaviouralusage, loyalty

05 · OLS pipeline CH06

Model + estimatory = Xβ + ε; ŷ = Xβ̂; e = y−ŷ
β̂ = arg minβ ‖y−Xβ‖₂² = (XᵀX)⁻¹Xᵀy
Simple regression shortcutSₓₓ=Σ(xᵢ−x̄)²; Sₓᵧ=Σ(xᵢ−x̄)(yᵢ−ȳ)
β̂₁=Sₓᵧ/Sₓₓ; β̂₀=ȳ−β̂₁x̄
FitRSS=Σeᵢ²; TSS=Σ(yᵢ−ȳ)²
R²=1−RSS/TSS=(TSS−RSS)/TSS
Assumption / readingWhere it bites
E(ε|x)=0unbiased conditional mean; curve signals failure
var(ε|x)=σ²constant variance; funnel signals failure
ε uncorrelated with xestimation validity
Normal εneeded for t/F inference, not for fitting β̂
Inference when requestedSE(β̂₁)²=σ²/Sₓₓ; t=β̂₁/SE(β̂₁)
approx. 95% CI: β̂₁ ± 2·SE(β̂₁)
Question givesFast route
X and y matricesnormal equation (XᵀX)⁻¹Xᵀy
x̄,ȳ,Sₓₓ,Sₓᵧslope then intercept shortcut
fitted valuese=y−ŷ → RSS → R²
residual plotshape → 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

GOODFUNNELCURVE even bandvar failure→ log ymean failure→ add x²
Source redraw: name the visible shape, the broken assumption and one remedy.
PlotDiagnosisMove
flat/evenadequateno structure left
funnelheteroscedastictransform y, e.g. log y
curve/Umissed nonlinearityadd 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 / caseUse / principle
positive right-skewed y; funnelmodel log(y): symmetry + variance stabilisation
curvaturepolynomial x²,x³,…
effect of x₁ depends on x₂interaction x₁x₂
consent to collect?Ownership
told how used?Transparency
PII exposed/re-identifiable?Privacy
purpose?Intention
group harm?Outcomes / disparate impact
BEFORE: funnelAFTER: even band transform log y →
Source redraw: a log-response transform can compress the tail and stabilise variance.
Back-transform correctlylog(y)=xᵀβ+ε
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

10.50ηη=0, p=0.5 p = 1/(1+exp(−η))
Source redraw: η=xᵀβ is log-odds; the sigmoid alone maps it into (0,1).
Link → activation → classlog[p/(1−p)] = η = xᵀβ
p̂=1/[1+exp(−η)]; ŷ=1 if p̂>τ (default τ=0.5)
+1 in xⱼ multiplies odds by exp(βⱼ)
GLM partLinearLogistic
responsecontinuousBernoulli
link gidentitylogit
activation g⁻¹identitysigmoid
fitOLS; closed formMLE; 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 +TPFP · Type I
Pred −FN · Type IITN
Read margins, not memoryAccuracy=(TP+TN)/N
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 questionMetric
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.

TPRFPRchanceAUCtop-left is better
Source redraw: ROC plots TPR against FPR as τ moves; AUC summarises ranking.

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

R² risesadj. R² may fallpredictors kfit stop
Source redraw: in-sample fit rewards extra terms; honest selection needs held-out data.
Generalisation targetEPE=E[(y−ŷ)²]=σ² + Bias² + Variance
complexity ↑ ⇒ bias ↓, variance ↑ ⇒ EPE U-shaped
Flexible but linear in βy=φ(x)ᵀβ+ε; β̂=(ΦᵀΦ)⁻¹Φᵀy
polynomial: φᵢ(x)=xⁱ
radial: φᵢ(x)=exp[−(x−cᵢ)²/(2s²)]
Data splitOnly job
trainfit each candidate
validationselect lowest MSE
testone final unbiased report

11 · MLE + optimisation CH10 Fig 3

LikelihoodL(θ)=p(data|θ)=∏ᵢp(yᵢ|θ)
ℓ(θ)=logL(θ)=Σᵢlogp(yᵢ|θ)
θ̂=arg maxθ ℓ(θ)
MLE lineMark
write product likelihoodindependence
take logsproduct→sum
differentiatescore
set =0; solvecandidate θ̂
second derivative <0maximum check

Why log? Avoid underflow; simplify differentiation; log is increasing so argmax unchanged. Drop additive terms without θ.

θ₀minstep against ∇f θₖ₊₁=θₖ−α∇f
Source redraw: gradient points uphill; subtract it to minimise f=−ℓ.
αFailure
too largeovershoot/diverge
too smallvery slow

OLS has closed form. Logistic score is nonlinear, so minimise cross-entropy f=−ℓ iteratively; ∇f(β)=Xᵀ(p−y).

Bernoulli log-lossf(β)=−Σᵢ[yᵢlogpᵢ+(1−yᵢ)log(1−pᵢ)]

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

> Which dimension is big? WIDE: p > nreduce pXᵀX singularTALL: n ≫ pblock rows / SGDfull pass costly
Shape decides the remedy: large p and large n are different problems.
Complexity + SGDO(1)<O(log n)<O(n)<O(n log n)<O(n²)<O(2ⁿ)
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.

ShapeFailureCorrect move
p>noverfit; XᵀX singularfeature selection, projection, less flexibility
n≫peach full pass expensiveblocks or one-row SGD
https://www.asksia.ai/cheatsheet/usyd-buss6002 · Full BUSS6002 study guide →
Your exam allows one HANDWRITTEN A4 (both sides): copy side 1 by hand; do not bring this printout.
BUSS6002Forward this to your tutorial group
v2 · 2026-10-16 · Shared by ________
Semester 2 formal examination period, from 16 November 2026 (the Exams Office publishes the exact date).
BUSS6002
Worked drills · traps · fast recall
Data Science in Business · University of Sydney · use before the exam
BUSS6002 cheatsheet QR code – open the full BUSS6002 study guide
Full guide + ask Sia
about BUSS6002
SIDE 2/2   5 FULL WORKED DRILLS · METHOD CHOOSER · PYTHON RECALL · TOP TRAPS PRACTISE · CHECK · REPEAT All numerical values from the course guide

13 · Drill 1: box plot CH03 worked example

Given sorted A$ totals: 28, 35, 42, 46, 58, 60, 75, 86, 90, 150, 210.

  1. n=11 ⇒ median Q2 = 6th value = 60.
  2. Position rule (n+1)/4 ⇒ Q1=3rd=42; Q3=9th=90.
  3. IQR=90−42=48; 1.5IQR=72.
  4. Fences: 42−72=−30; 90+72=162.
  5. Only 210>162, so 210 is circled; 150 remains inside.
Answerfive-number summary=(28,42,60,90,210)
fences=(−30,162); circled: 210 only
28Q1 42Q2 60Q3 90150210
Source redraw with the worked values: the whisker ends at 150; 210 is only flagged.
TrapCorrection
circle = deleteno: investigate context first
whisker = maxonly most extreme value inside fence
mean resistantmedian/IQR resist extremes; mean/sd do not

14 · EDA fast recall CH03

MomentMeaningRobust partner
1stlocation / meanmedian
2ndscale / variance; report sdhalf-IQR
3rdsymmetry / skewGalton skewness
4thtails / kurtosis; Normal=3Moors kurtosis

Mean chases the long tail: right-skew ⇒ mean>median; left-skew ⇒ mean<median. Q-Q points near y=x support Normal residuals; tail departures signal skew/heavy tails.

Mini drillCorrect classification
500-listener price surveysmall data: bounded, deliberate, once-off
continuous play/skip/search/device logs from tens of millionsbig data: Volume + Velocity + Variety
quarterly revenue reportsmall data
9-person listening labsmall data

Answer structure: state the four-V test; eliminate bounded deliberate sources; select the continuous multi-source logs. Rigour and byte count do not decide the label.

Missing-data cueMechanism
independent of every variableMCAR
depends on visible age/groupMAR

15 · Drill 2: vectors CH04 worked example

a=[3,1,−2]ᵀ, b=[2,k,4]ᵀ.

  1. ⟨a,b⟩=(3)(2)+(1)k+(−2)(4)=6+k−8=k−2.
  2. Orthogonal ⇒ inner product zero: k−2=0 ⇒ k=2.
  3. Then b=[2,2,4]ᵀ and ‖b‖₂=√(2²+2²+4²)=√24.
  4. √24=2√6≈4.90 (3 s.f.).
Answer⟨a,b⟩=k−2; k=2; ‖b‖₂=2√6≈4.90
> baaᵀb=0⇒ 90°
Source redraw: “orthogonal” is an instruction—set the dot product to zero and solve.
Self-checkExpected
inner product typescalar
norm signnon-negative
‖b‖ vs ‖b‖²√24 vs 24

16 · Drill 3: K-means CH05 worked example

A=[1,1], B=[2,1], C=[6,5], D=[7,5]; k=2; μ₁=[1,1], μ₂=[2,1].

Pointd² to μ₁d² to μ₂Assign
A01C₁
B10C₂
C4132C₂
D5241C₂

C₁={A}; C₂={B,C,D}. μ₁=[1,1]; μ₂=[(2+6+7)/3,(1+5+5)/3]=[5,3.67]. New WCSS ≈24.7, lower than the start; next pass flips B into C₁.

Write: squared distance; mean of new members; objective non-increasing; local optimum.

K-means claimVerdict
supervisedfalse: there is no target y
centroid must be observedfalse: it is a coordinate-wise mean
chooses kfalse: k is input
objective strictly fallsfalse: non-increasing; may hold
global optimum guaranteedfalse: initialisation matters

Squared distance and ordinary distance rank centres identically, but squaring removes the root and makes the mean the clean update minimiser.

16B · Matrix product drill CH06 worked example

GivenA=[[2,1],[3,0]], B=[[1,4],[5,2]]

Row·column gives AB=[[2·1+1·5, 2·4+1·2],[3·1+0·5, 3·4+0·2]]=[[7,10],[3,12]].

17 · Drill 4: OLS + R² CH06 worked example

x̄=10, ȳ=50, Sₓₓ=200, Sₓᵧ=900, TSS=5000.

  1. β̂₁=Sₓᵧ/Sₓₓ=900/200=4.5.
  2. β̂₀=ȳ−β̂₁x̄=50−4.5×10=5.
  3. Fitted line: ŷ=5+4.5x.
  4. SSR=β̂₁Sₓᵧ=4.5×900=4050.
  5. R²=SSR/TSS=4050/5000=0.81; RSS=950.
  6. +1 in x is associated with average +4.5 in y; x explains 81% of variation.
(x̄,ȳ)ŷ=β̂₀+β̂₁xe=y−ŷ
Source redraw: the OLS line passes through (x̄,ȳ) and minimises squared vertical residuals.
TrapRepair
causal wordingsay “associated with”
R²=81 units81% of variation explained
Normality needed to fitonly needed for inference

18 · Regression diagnosis script CH06–07

SeeStateThen
random even bandmean + variance plausiblemodel adequate
U/curveE(ε|x)=0 failspolynomial/basis term
fan/funnelvar(ε|x)=σ² failslog response
right-skewed ytail + variance issuelog y; correct back-transform

Answer in three clauses: shape → assumption → remedy. “The model is bad” earns little without the named failure.

Second OLS speed drillx=(1,2,3,4), y=(2,2,4,5)
x̄=2.5; ȳ=3.25; Sₓₓ=5; Sₓᵧ=5.5
β̂₁=1.1; β̂₀=0.5; ŷ=0.5+1.1x
fitted=(1.6,2.7,3.8,4.9)

Residuals=(0.4,−0.7,0.2,0.1), so RSS=0.70. The fitted values and residual signs provide a quick arithmetic check.

Finish the fit checkTSS=Σ(yᵢ−ȳ)²=6.75
R²=1−0.70/6.75≈0.896
OLS quantityMeaning
eᵢ=yᵢ−ŷᵢvertical residual
RSS=Σeᵢ²loss minimised
R²fraction of y variation explained in sample
SE(β̂)sampling variability for inference
Transform clueFeature move
right-skewed positive ylog y, then smear on return

19 · Drill 5: confusion matrix CH08 worked example

y=[1,1,0,1,0,0,1,0,1,0]; ŷ=[1,0,0,1,1,0,1,0,1,1].

Actual 1Actual 0Total
Pred 1TP=4FP=26
Pred 0FN=1TN=34
Total5510
  1. Accuracy=(4+3)/10=0.70.
  2. Recall=4/(4+1)=0.80.
  3. Precision=4/(4+2)=0.667.
  4. F1=8/(8+2+1)=8/11=0.727.
Margin memoryrecall: actual-positive column TP/(TP+FN)
precision: predicted-positive row TP/(TP+FP)
FPR: actual-negative column FP/(FP+TN)

Always: draw the 2×2 before calculating. Accuracy can look excellent when the rare positive class is never found.

20 · Logistic worked prediction Course worked example

β₀=−2.0, β₁=2.0, β₂=1.0; x₁=1.5, x₂=1.

  1. η=−2+(2)(1.5)+(1)(1)=2.0 log-odds.
  2. p̂=1/[1+exp(−2)]; exp(−2)≈0.135.
  3. p̂=1/1.135≈0.88.
  4. 0.88>τ=0.5 ⇒ ŷ=1, churn risk.
> η=2.0log-oddssigmoid1/(1+e⁻η)p̂=0.88ŷ=1 never call η a probability
Source redraw: linear predictor first, sigmoid second, threshold last.
CoefficientInterpretation
βⱼ>0raises log-odds and p
exp(βⱼ)odds multiplier for +1 in xⱼ
τ lowerTPR↑ and FPR↑

20B · Ethics case drill CH07 worked example

A résumé screen down-ranks certain postcodes, cannot explain rejections, and later shows systematic ethnic-group disadvantage.

PrincipleOne-sentence justification
Outcomespostcode acts as a proxy and produces disparate impact regardless of legitimate efficiency intent
Transparencyrejected candidates cannot understand or contest how their data drove the decision

21 · Validation MSE CH09 worked example

Validation y=(8,12,20,15); ŷ=(10,11,17,14).

Run the arithmetice=y−ŷ=(−2,1,3,1)
e²=(4,1,9,1); Σe²=15
MSE=(1/4)·15=3.75

MSE is the mean, not the sum. Compute it for every candidate; choose the lowest validation MSE. Test MSE is reported once after selection.

Bias²VarianceEPEcomplexity → best
Source-grounded redraw: bias falls, variance rises and EPE is minimised in the middle.
RegimeBiasVarianceSignal
under-fithighlowtrain + validation poor
goodbalancedbalancedvalidation minimum
over-fitlowhightrain low, validation rises

22 · Poisson MLE CH10 worked example

yᵢ iid Poisson(λ); sample (2,4,3,5,1), n=5, Σyᵢ=15.

Five mark linesL(λ)=∏ᵢ λʸⁱe⁻λ/yᵢ!
ℓ(λ)=(Σyᵢ)logλ−nλ−Σlog(yᵢ!)
dℓ/dλ=(Σyᵢ)/λ−n=0
λ̂=(Σyᵢ)/n=ȳ=15/5=3
d²ℓ/dλ²=−(Σyᵢ)/λ²<0 ⇒ maximum

Same skeleton: Bernoulli MLE = sample proportion; exponential rate MLE = 1/ȳ. OLS equals MLE under Gaussian errors.

ℓ maximumℓ′=0, ℓ″<0λ
Source redraw: solve the stationary point, then prove it is a maximum.
Why log L?Pay-off
many probabilities <1avoid underflow
productbecomes differentiable sum
log increasingsame argmax
Tall-data extension: n=2,000,000full batch reads all rows ⇒ O(n)/step
SGD uses one random row ⇒ O(1)/step in n
up to n cheap updates per full-pass budget

23 · Method chooser CH03–11 synthesis

> What is the task?No target → groupsTarget y → predictK-means + scalechoose k: elbowselect: val lossnumeric y → OLSbinary y → logit
Start from the response and task, not from a favourite algorithm.
Question wordingImmediate move
“orthogonal”xᵀy=0; solve
“nearest centroid”compare squared distances
“best line”minimise RSS / OLS
“probability of class 1”η then sigmoid
“rare positive”precision / recall / F1
“unseen performance”validation loss; reserve test
“most plausible parameter”likelihood → log → differentiate
“millions of rows”tall; SGD / row blocks
“p>n”wide; cut predictors

24 · Train / select / report CH09

1 · Train2 · Validate3 · Refit4 · Test
fit all candidateschoose lowest losschosen form on train+valreport once

The test set is not a second validation set. Repeated use leaks information and makes the reported performance optimistic.

25 · Python from memory CH04, 08, 09

NumPy essentialsimport numpy as np
x=np.array([3,1,-2])
np.dot(x,y)   # inner product
mse=np.mean((y-yhat)**2)
sklearn skeletonmodel.fit(X_train,y_train)
p=model.predict_proba(X_val)[:,1]
yhat=(p>0.5).astype(int)

Vectorise rather than loop. Check brackets, axis, array shapes and imports. Code marks reward runnable syntax, not prose.

26 · Final trap scan All chapters

  • AB≠BA; check inner dimensions.
  • K-means: squared distance; local, not guaranteed global.
  • Regression slope: association, not cause.
  • η=xᵀβ is log-odds, not probability.
  • Training error is not a selection criterion.
https://www.asksia.ai/cheatsheet/usyd-buss6002 · Full BUSS6002 study guide →
Your exam allows one HANDWRITTEN A4 (both sides): copy side 1 by hand; do not bring this printout.
BUSS6002Forward this to your tutorial group
v2 · 2026-10-16 · Shared by ________
Semester 2 formal examination period, from 16 November 2026 (the Exams Office publishes the exact date).

Want one for YOUR exact syllabus?

Sia is your free desktop study agent. Drop your University of Sydney BUSS6002 slides — Sia builds a sheet tailored to YOUR exam. Better than this library because it knows YOUR materials.

↓ Download Sia · Free