Company Interview Questions · Banking & Fintech

Mastercard & Visa: Data Scientist

❓ 6 questions🏢 Mastercard & Visa🆕 Recent reports

Mastercard & Visa: Data Scientist Interview Experiences (2025–26). Rounds, real reported questions and model answers, compiled from public candidate reports (sources at the end). Practice mode hides every answer: say your answer first, press Show answer, then rate yourself. Use 🔁 Review again to build your revision list and 🎲 Random question for a quick drill.

★ Must-know 145📝 Assignments
About this interview

Compiled from public candidate reports: Mastercard Data Scientist (New Delhi, offer; Gurgaon, April 2025; a 4-round process with a business case and programming test, offer, May 2025; Dublin, 2026) and Visa Data Science (Washington DC super day, September 2025; CodeSignal tests, 2025–26). The questions are paraphrased, and the model answers are written by EasyLearnSpace.

Mastercard Visa
Role Data Scientist / Data Scientist I Data Scientist
Rounds HR screen → 2 technical (business case + programming test) → culture and values. Campus: a single interview Online coding (CodeSignal: data prep + ML model) → phone screen → super day (4 interviews)
Reported questions How precision and recall change with the threshold; derive loss functions; gradient descent; why log loss; explain a customer-segmentation project Which unsupervised methods you have used; fraud, risk and personalisation cases; find the middle of an array
Outcomes 2 offers, 2 not selected Not selected in these reports (little feedback)

01Rounds at a glance

02Questions and model answers

Q1
How do precision and recall change as you move the classification threshold?
  • Raise the threshold: fewer positives are predicted, so precision usually rises and recall falls.
  • Lower it: more positives, so recall rises and precision falls.
  • The PR curve traces this trade-off. Choose the threshold by business cost (fraud loss vs declined good transactions) or by a target (e.g. recall ≥ 80% at the highest precision).
Q2
Why do we use log loss for classification instead of MSE?
  • Log loss = the negative log-likelihood of a Bernoulli model. It is the principled maximum-likelihood objective for probabilities.
  • With a sigmoid output it is convex in the weights (MSE + sigmoid is not), and its gradients don't vanish when the model is confidently wrong. Gradient (p − y)·x gives strong corrections.
  • It heavily penalises confident mistakes and encourages calibrated probabilities.
Q3
Derive the gradient-descent update for logistic regression.
  • Loss: L = −(1/n) Σ [y log p + (1−y) log(1−p)], with p = σ(wᵀx).
  • Using dσ/dz = σ(1−σ): ∂L/∂w = (1/n) Σ (p − y) x.
  • Update: w ← w − η · (1/n) Xᵀ(p − y). Add λw for L2.
  • Shortcomings to mention: sensitivity to the learning rate, feature scaling, and slow convergence on ill-conditioned data. Second-order methods (Newton/IRLS, L-BFGS) converge faster for small d.
Q4
Explain a customer-segmentation project end to end.
  • Goal: segment cardholders for targeted offers.
  • Features: RFM, category mix (travel, dining, groceries), channel (online/POS), cross-border share, tenure. Log-transform skewed spend, then standardise.
  • Method: k-means (choose k with the elbow, silhouette and business interpretability) or a GMM; PCA for visualisation.
  • Personas: "frequent travellers", "online shoppers", "essentials-only", "dormant".
  • Validation: stability across months; uplift from a targeted campaign A/B test (e.g. +12% activation in the travel segment).
Q5
Which unsupervised methods have you used, and for what?
  • Clustering: k-means (segments), DBSCAN (density-based, finds outliers), hierarchical (taxonomies).
  • Dimensionality reduction: PCA (compression, denoising), t-SNE/UMAP (visualisation).
  • Anomaly detection: Isolation Forest, autoencoders (reconstruction error for fraud).
  • Topic modelling / embeddings: LDA, or embeddings + clustering on complaint text.
  • Association rules: market-basket analysis.
  • For each, say why it fits and how you validated it without labels.
Q6
Design a fraud-detection model for card-not-present (online) transactions.
  • Signals: device fingerprint, IP/geo mismatch, email age, velocity per card/device/merchant, basket and amount anomalies, merchant risk, 3-D Secure result.
  • Model: gradient boosting plus graph features (shared devices/cards), with a real-time feature store; a latency budget under 50 ms.
  • Labels: chargebacks (delayed) and confirmed fraud; handle label delay and class imbalance.
  • Decision: approve / step-up (3DS/OTP) / decline, with thresholds by merchant risk.
  • Monitoring: fraud basis points, false-decline rate, drift; weekly retraining; explainability reason codes.

03How to prepare

Sources

Created by Ajeetkumar