Question 1

Your product team wants to test whether changing a CTA button from blue to orange increases conversion rate. After 2 weeks, the orange group achieves 3.2% conversion vs 2.9% for blue (n=10,000 per group). How do you analyze this and what decision do you make?

Accepted Answer

Analysis: (1) Two-proportion z-test or chi-squared test, (2) Calculate p-value (with delta = 0.3pp, n=10K per group, p-value ≈ 0.03), (3) Check pre-defined statistical power and minimum detectable effect, (4) Check confounders (traffic source, device type, day of week), (5) Practical significance: is 0.3pp meaningful for the business? Strong candidates immediately ask: was power analysis done upfront? What was the pre-defined alpha? Is this multiple testing (were other variants tested)? And critically: what does 0.3pp at that traffic volume mean for annual revenue impact?

Question 2

You have a dataset with 500 features to predict customer churn for a telecom company. Describe your end-to-end feature selection process, including how you prevent data leakage.

Accepted Answer

Three method groups: (1) Filter methods: correlation with target, mutual information, variance threshold - fast, model-agnostic, (2) Wrapper methods: RFE (Recursive Feature Elimination) - more accurate but expensive, (3) Embedded methods: Lasso (L1 regularization), tree-based feature importance from LightGBM/XGBoost. Combine and validate with cross-validation. Data leakage prevention: fit feature selection ONLY on training set, then transform validation/test. For churn, specifically warn about temporal leakage - no features computed from future data (e.g., how many calls the customer made next month).

Question 3

A Vietnamese logistics company wants to predict delivery delays based on 50 features (weather, driver history, route, cargo type). They have 200,000 records over 2 years. Would you choose deep learning or traditional ML and why?

Accepted Answer

Decision framework across 4 factors: (1) Data size: 200K records with 50 features - not large enough for deep learning to dominate, (2) Feature type: tabular data - tree-based models (LightGBM, XGBoost) typically win on tabular, (3) Interpretability: logistics needs explainable decisions, (4) Latency and resources: is real-time inference needed? Does the team have ML expertise? Correct answer: with 200K tabular records, gradient boosting (LightGBM/XGBoost/CatBoost) is the first choice. Deep learning is only worth considering if there is raw text (driver notes), images, or complex sequential patterns that cannot be captured by manual features.

Question 4

Your XGBoost loan rejection model achieves AUC = 0.91 but the legal team requires explainable rejection reasons for each loan (per State Bank of Vietnam regulations). How do you handle this?

Accepted Answer

Solution: (1) SHAP values compute each feature's contribution to each prediction, (2) For each rejection: display top 3-5 features with the most negative SHAP values (negative impact), (3) Map technical feature names to business language ('debt_to_income_ratio' → 'Debt-to-income ratio exceeds safety threshold'), (4) Use SHAP TreeExplainer (faster than KernelExplainer for tree models), (5) Store SHAP values for audit purposes. SHAP vs. LIME comparison: SHAP has better mathematical consistency (Shapley values have axiomatic foundations), TreeExplainer is O(TL^2) vs. brute-force O(TLD). Bonus: distinguish global vs. local interpretability - SHAP beeswarm plot for global patterns.

Question 5

You need to forecast weekly retail sales for 500 SKUs at a Vietnamese supermarket chain 8 weeks ahead. Data has strong seasonality (Tet, April 30, September 2) and some SKUs have missing values due to stockouts. How do you approach this?

Accepted Answer

Strategy: (1) Handle stockout-induced missing values: do NOT interpolate - mark as zero-demand (different from true missing), (2) Model selection: with 500 SKUs - global models beat local models (LightGBM with cross-SKU features, N-BEATS, TFT), (3) Feature engineering: lag features, rolling statistics, Vietnamese lunar calendar (when is Tet each year?), (4) Hierarchical reconciliation if forecasting at category and store level, (5) Evaluation: MAPE is inappropriate for zero values, use RMSSE. Vietnam-specific bonus: Tet lunar calendar shifts each year (3-4 weeks relative to Gregorian) - encode weeks before/after Tet by lunar calendar; short intense spikes like Mid-Autumn Festival mooncakes; seasonal goods differ between Northern and Southern Vietnam.

Question 6

You need to build a sentiment classification system for product reviews on Shopee Vietnam. What are the unique NLP challenges of Vietnamese, and would you fine-tune PhoBERT or use an LLM like Gemini?

Accepted Answer

Vietnamese NLP challenges: (1) Tone marks change meaning completely: 'ma' vs. 'mà' vs. 'má', (2) Word segmentation - Vietnamese lacks spaces between compound words ('học sinh' = 1 token), (3) Teencode and abbreviations common in reviews ('ko' = 'không', 'dc' = 'được'), (4) Frequent code-switching (Vietnamese-English mixing). PhoBERT vs. LLM: PhoBERT (fine-tuned) wins on: high volume (cost-effective), low latency, domain customization for Shopee. Gemini wins on: zero-shot/few-shot without labeled data, nuanced understanding, explainability. Bonus: mention VnCoreNLP for word segmentation, xlm-roberta-base as multilingual baseline.

Question 7

Your food delivery app wants to test whether surge pricing during peak hours affects 30-day retention. Design the experiment and explain the key issues you must watch out for.

Accepted Answer

Design: (1) Randomization unit: user-level (not order-level) to avoid spillover, (2) Stratify by user cohort (new vs. returning, city), (3) Primary metric: 30-day retention; guardrail metrics: revenue per user, order frequency, (4) Minimum runtime: at least 1-2 weeks to capture weekly patterns, but 4+ weeks to measure 30-day retention, (5) Pre-experiment power analysis: with expected 2% retention drop, what n is needed? Key issues: (1) Network effects - users discuss prices with each other (Hawthorne effect and contamination), (2) Novelty effect - initial behavior change due to new pricing awareness, (3) Seasonality - Vietnamese peak hours (11am-1pm, 5pm-7pm) vs. weekends differ, (4) Survivorship bias in the retention metric.

Question 8

Your forecasting model shows that a 20% marketing budget increase will lift revenue by 8-12% with a wide confidence interval. The CFO asks: 'Why is the confidence interval so wide? Are you sure?' How do you respond and present this?

Accepted Answer

Presentation: (1) Translate CI to business language: '90% probability that growth falls in 8-12%; even the worst case at +8% is still positive ROI', (2) Explain CI width causes: high historical variance, few comparable data points, uncertainty from market factors, (3) Propose uncertainty reduction: pilot test with smaller budget, (4) Scenario analysis: best/base/worst case. Strong candidates understand the CFO is not asking about statistics - they are asking 'can I trust the decision I'm about to make on this?' The answer should focus on the decision framework: 'Even in the worst-case scenario, is this investment still worth it?'