Company Interview Questions · Analytics & AI Firms

ZS Associates: Data Science Associate

❓ 7 questions🏢 ZS Associates🆕 Recent reports

ZS Associates: Data Science Associate 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 for ZS Associates Advance Data Science Associate roles (Bengaluru, February and May 2026; the May candidate received an offer), a Mumbai report (May 2025, selected) and a campus challenge report. The questions are paraphrased, and the model answers are written by EasyLearnSpace.

Role	Advance Data Science Associate / Data Science Associate
Rounds	HackerRank online assessment (ML, stats, SQL, Python) → HR screen → ML classification case study (build a solution) → case debrief (resume + ML concepts) → final round (puzzles, ML, GenAI, AI frameworks)
Questions reported	Linear vs logistic regression, why ReLU, precision vs recall, evaluating a RAG app, MLflow, outliers, medium SQL, guesstimates
Outcomes	Offer (May 2026, rated difficult); selected (Mumbai 2025); 1 not selected

01Rounds at a glance

Case study: you get a classification dataset, build a model, then defend it in the debrief.
Final round: puzzles and guesstimates (consulting DNA) plus GenAI concepts and AI frameworks.
ZS is a pharma and healthcare analytics leader, so domain examples help.

02Questions and model answers

What is the difference between linear and logistic regression?

	Linear regression	Logistic regression
Target	Continuous	Binary (or multi-class with softmax)
Output	ŷ = wᵀx + b, unbounded	p = σ(wᵀx + b) in (0, 1)
Loss	MSE (least squares)	Log-loss (cross-entropy)
Interpretation	Coefficient = change in y	Coefficient = change in log-odds (exp → odds ratio)
Assumptions	Linearity, homoscedastic errors	Linear in the log-odds; independent observations

Why is ReLU preferred in neural networks?

No saturation for positive inputs: the gradient is 1, which avoids the vanishing gradients of sigmoid and tanh, so deep networks train faster.
Cheap to compute (max(0, x)), and produces sparse activations.
Downside: "dying ReLU" (units stuck at 0). Fixes are Leaky ReLU, PReLU, ELU and GELU/SiLU (used in Transformers), plus a good initialisation (He) and learning rate.

Precision vs recall: when do you prioritise each? (Give a pharma example.)

Precision: when false positives are costly. For example, flagging physicians for an expensive field-sales visit means we want flagged doctors to really be high-potential.
Recall: when misses are costly. For example, detecting adverse-event mentions in medical text, where missing one is a compliance risk.
Balance with F1 or a cost-weighted threshold, and show the precision-recall curve to stakeholders so they can choose the operating point.

How do you evaluate a RAG application?

Retrieval: recall@k, MRR, context precision (are the retrieved chunks relevant?).
Generation: faithfulness (claims supported by the context), answer relevance, correctness against ground truth.
Tools: RAGAS, DeepEval, TruLens, LangSmith evaluations; a golden set of 100+ questions written with domain experts.
Online: user feedback, citation click-through, escalation rate, latency and cost.
Regression testing: re-run the suite on every prompt, model or index change.

What is MLflow and how do you use it?

An open-source platform for the ML lifecycle:
Tracking: log parameters, metrics and artefacts per run, and compare experiments.
Projects: reproducible runs.
Models: a standard packaging format, served via REST.
Model Registry: versions plus stages (staging/production), approvals and lineage.
Usage: mlflow.autolog() in training, register the best model, and deploy from the registry in CI/CD. Newer versions also trace and evaluate LLM apps.

How do you treat outliers in a modelling dataset?

Investigate first: an error (fix or remove) or a real extreme (keep it; it may be important).
Detect: IQR rule, z-score (on roughly normal data), visual checks, Isolation Forest for multivariate outliers.
Treat: cap/winsorise, log transform, robust scalers or losses (Huber), tree-based models.
Document the impact on metrics, and never "clean away" the signal you're trying to predict.

Guesstimate: how many diabetes patients visit endocrinologists in Bengaluru each month?

Answer (show your structure):

Bengaluru population about 1.4 crore → adults about 70% ≈ 98 lakh.
Diabetes prevalence in urban adults about 12% → about 11.8 lakh diabetics; about 60% diagnosed ≈ 7 lakh.
Share seeing an endocrinologist (rather than a GP) about 20% ≈ 1.4 lakh patients; visits every 3 months → about 47,000 visits a month.
State your assumptions and sanity-check them: about 300 endocrinologists × about 160 visits a month ≈ 48,000. That matches.

03How to prepare for ZS

Practise a 2–3 hour classification case: clean, model, evaluate, then present your choices clearly.
Revise ML basics and SQL; prepare puzzles and guesstimates.
Know RAG evaluation and MLflow. GenAI is now part of the final round.

Sources

Created by Ajeetkumar