Detailed Data Science interview questions covering statistics, SQL, Python, data cleaning, visualization, experiments, metrics, machine learning basics, and business analysis.
Data Science is the practice of using data, statistics, programming, domain knowledge, and machine learning to discover insights and support decisions. A data scientist does not only build models; they define the problem, collect and clean data, analyze patterns, communicate findings, and help teams act on evidence.
Data Analytics focuses on answering business questions from data, such as what happened and why. Data Science is broader and includes analytics, experimentation, predictive modeling, machine learning, and decision support.
A typical project starts with problem framing, then data collection, data validation, exploratory data analysis, cleaning, feature engineering, modeling or statistical analysis, evaluation, communication, deployment if needed, and monitoring. In interviews, explain that the workflow is iterative: findings from EDA or error analysis often send you back to improve data quality, metric definitions, features, or assumptions.
Exploratory Data Analysis, or EDA, is the process of understanding a dataset before formal modeling or decision-making. It includes checking shape, missing values, distributions, outliers, correlations, target balance, time patterns, and segment-level behavior.
import pandas as pd
df = pd.read_csv("customers.csv")
print(df.shape)
print(df.info())
print(df.describe(numeric_only=True))
print(df.isna().mean().sort_values(ascending=False).head())Missing values can be handled by deletion, simple imputation, model-based imputation, adding missing indicators, or treating missing as its own category. The best method depends on why the value is missing. If income is missing because a user skipped the field, that missingness may be meaningful.
df["age"] = df["age"].fillna(df["age"].median())
df["city"] = df["city"].fillna("Unknown")
df["income_missing"] = df["income"].isna().astype(int)Outliers are observations that are unusually far from the rest of the data. They may be valid rare events, data entry errors, fraud signals, or measurement issues.
Data cleaning fixes or standardizes issues that make analysis unreliable. It includes handling missing values, duplicate rows, inconsistent categories, invalid dates, impossible values, mixed units, whitespace, incorrect types, and broken joins.
Data profiling is a systematic summary of data quality and structure. It checks row counts, column types, null percentages, uniqueness, ranges, duplicate records, invalid values, category frequencies, and distribution changes. Profiling is useful before EDA, before pipeline deployment, and after upstream data changes.
Sampling bias occurs when collected data does not represent the population you want to understand. For example, analyzing only active app users may ignore churned users, leading to overly positive conclusions.
Descriptive statistics summarize the main properties of data. Common measures include mean, median, mode, variance, standard deviation, percentiles, minimum, maximum, and frequency counts. They help quickly understand scale, spread, skewness, and unusual values.
summary = df.groupby("plan")["monthly_spend"].agg(
count="count",
mean="mean",
median="median",
p90=lambda x: x.quantile(0.90)
)
print(summary)Inferential statistics uses sample data to make conclusions about a larger population. It includes confidence intervals, hypothesis tests, regression inference, and estimation.
Hypothesis testing evaluates whether observed data provides enough evidence against a default assumption called the null hypothesis. For example, the null hypothesis may be that a new landing page has the same conversion rate as the old one.
A p-value is the probability of observing results at least as extreme as your sample result, assuming the null hypothesis is true. It is not the probability that the null hypothesis is true.
A confidence interval gives a range of plausible values for a population parameter. For example, a 95 percent confidence interval for conversion lift may be 1.2 percent to 3.8 percent.
Statistical power is the probability that a test detects a real effect when one exists. Low power increases the chance of missing meaningful changes. Power depends on sample size, effect size, variance, and significance level.
A/B testing compares two or more versions of a product, page, message, or model by randomly assigning users to groups and measuring outcomes. For example, a company may test whether a new checkout page increases purchases.
Correlation means two variables move together. Causation means one variable directly influences another.
Feature engineering creates or transforms variables so patterns become easier to analyze or model. Examples include extracting weekday from a timestamp, calculating customer tenure, creating average order value, counting failed login attempts, or aggregating transactions over the last 30 days.
Data leakage happens when analysis or model training uses information that would not be available at decision time. For example, predicting churn using a cancellation_reason field leaks future information because the reason is known only after churn.
Safe joins require checking key uniqueness, row counts before and after the join, duplicate keys, missing matches, and whether the join type matches the business question. A bad join can multiply rows and inflate metrics. For example, joining customers to orders without aggregating orders first can duplicate customer-level fields.
SELECT
c.customer_id,
c.segment,
COUNT(o.order_id) AS order_count,
SUM(o.amount) AS total_spend
FROM customers c
LEFT JOIN orders o
ON c.customer_id = o.customer_id
GROUP BY c.customer_id, c.segment;Important SQL concepts include filtering, aggregation, joins, subqueries, common table expressions, window functions, date functions, ranking, conditional aggregation, and handling nulls. Data scientists use SQL to extract clean datasets, build metrics, analyze funnels, create cohorts, and validate dashboards.
Window functions calculate values across related rows without collapsing rows like GROUP BY does. They are useful for ranking, running totals, moving averages, session analysis, retention analysis, and deduplication. Common functions include ROW_NUMBER, RANK, DENSE_RANK, LAG, LEAD, SUM OVER, and AVG OVER.
SELECT
user_id,
order_date,
amount,
SUM(amount) OVER (
PARTITION BY user_id
ORDER BY order_date
) AS running_spend
FROM orders;Pandas is used for loading, cleaning, transforming, grouping, joining, and summarizing tabular data. Common operations include filtering rows, selecting columns, handling missing values, grouping by categories, merging datasets, pivoting, and calculating new features.
active_users = (
df[df["status"] == "active"]
.groupby("plan", as_index=False)
.agg(users=("user_id", "nunique"), avg_spend=("spend", "mean"))
.sort_values("avg_spend", ascending=False)
)
print(active_users)Duplicate detection depends on the definition of uniqueness. Sometimes duplicates are exact row matches.
duplicates = df[df.duplicated(subset=["email"], keep=False)]
latest = df.sort_values("updated_at").drop_duplicates("email", keep="last")A metric is a quantified measure used to track performance or behavior. Metric definition matters because different definitions can lead to different decisions. For example, active users may mean users who logged in, users who performed a core action, or users who spent at least five minutes.
A North Star metric is the primary metric that represents long-term customer value and business success. For a streaming platform it might be weekly engaged viewers.
Leading indicators move before the final business outcome and can help teams act early. Lagging indicators measure results after the fact. For example, trial activation may be a leading indicator for paid conversion, while quarterly revenue is a lagging indicator. Good dashboards often include both.
Cohort analysis groups users by a shared starting point or behavior and tracks them over time. For example, users who signed up in January can be compared with users who signed up in February to understand retention.
Funnel analysis measures how users move through a sequence of steps, such as visit, signup, add to cart, checkout, and purchase. It helps identify where users drop off. A strong answer should mention step definitions, ordering rules, time windows, segmentation, and whether users can repeat or skip steps.
Churn analysis studies why customers stop using a product or service. It can include churn rate calculation, segmentation, survival analysis, feature usage patterns, customer feedback, and churn prediction models. The first challenge is defining churn correctly: for a subscription product it may be cancellation, while for a marketplace it may be inactivity for a certain number of days.
Segmentation divides users, customers, products, or events into meaningful groups. Segments can be rule-based, such as new versus returning users, or model-based, such as clusters from behavior data. Segmentation helps teams understand different behaviors and avoid average metrics hiding important differences.
Choose charts based on the message. Use line charts for trends over time, bar charts for category comparisons, histograms for distributions, scatter plots for relationships, box plots for spread and outliers, and heatmaps for matrix-style patterns.
A good dashboard has a clear audience, clear metric definitions, useful filters, readable charts, visible time periods, and a layout that supports scanning. It should show key metrics first, then diagnostic breakdowns. Avoid overcrowding, unclear colors, vanity metrics, and charts without actionable context.
Data storytelling is communicating insights through a clear narrative supported by evidence. It connects the business question, data, analysis, finding, impact, and recommendation.
Time series analysis studies data ordered by time, such as daily sales, hourly traffic, or monthly revenue. Important concepts include trend, seasonality, autocorrelation, stationarity, lag features, moving averages, and forecast error. Time-series validation should respect chronological order instead of randomly splitting rows.
Linear regression predicts a continuous outcome and helps estimate relationships between variables. It can be used for forecasting, baseline modeling, and explanation.
Logistic regression is used for classification, especially binary classification. It estimates the probability of an event such as churn, fraud, conversion, or loan default.
Regression predicts continuous values, such as revenue or delivery time. Classification predicts categories, such as churned or not churned. The target type determines the algorithm and metric. Regression may use MAE, RMSE, or R-squared. Classification may use accuracy, precision, recall, F1, ROC AUC, or PR AUC.
Classification evaluation depends on business cost. Accuracy works only when classes are balanced and error costs are similar.
from sklearn.metrics import classification_report, roc_auc_score
y_pred = model.predict(X_test)
y_prob = model.predict_proba(X_test)[:, 1]
print(classification_report(y_test, y_pred))
print("ROC AUC:", roc_auc_score(y_test, y_prob))Regression models are evaluated using metrics such as MAE, MSE, RMSE, R-squared, and MAPE. MAE is easy to explain because it is average absolute error. RMSE penalizes larger errors more heavily.
Train-test split separates data used for model training from data used for final evaluation. The test set should represent unseen data.
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(
X,
y,
test_size=0.2,
random_state=42,
stratify=y
)Cross-validation evaluates a model over multiple train-validation splits. In k-fold cross-validation, the data is divided into k folds, and each fold is used once for validation.
Overfitting happens when a model learns noise or accidental patterns from training data and performs poorly on new data. Symptoms include high training performance and low validation performance.
Underfitting happens when a model is too simple to capture real patterns. It performs poorly on both training and validation data.
Bias is error from overly simple assumptions. Variance is error from sensitivity to training data noise. High-bias models underfit, while high-variance models overfit.
Feature scaling puts numeric features on comparable scales. It is important for algorithms that depend on distance or gradients, such as KNN, SVM, logistic regression, neural networks, PCA, and k-means. Tree-based models usually do not require scaling. Common methods include standardization and min-max normalization.
A data pipeline moves and transforms data from sources to usable outputs such as tables, dashboards, reports, features, or models. A reliable pipeline includes validation, logging, error handling, freshness checks, lineage, and monitoring.
Reproducible analysis means another person can rerun the work and get the same result using the same data, code, environment, and assumptions. It requires versioned code, documented data sources, fixed random seeds where needed, dependency tracking, clear notebooks or scripts, and saved outputs.
Start with the business question and answer, then explain the evidence, impact, confidence, limitations, and recommended action. Avoid leading with technical details unless they are necessary.
A complete case study starts by clarifying the goal, defining metrics, checking data quality, performing EDA, segmenting results, testing assumptions, building a baseline or model if needed, evaluating impact, and recommending action. In interviews, explain tradeoffs and ask clarifying questions. The best answer is not just technically correct; it is useful for the business decision.
Use curated frontend, Java, Python, or cloud and DevOps packs with 200 questions, model answers, code examples, and a seven-day plan.
Combine four related topic areas into one focused role-preparation path.
Practise explaining code, design decisions, limitations, and failure modes.
Move from baseline review to a timed mock interview and final checklist.
Explore 500+ free tutorials across 20+ languages and frameworks.