50 Data Science Interview Questions (With Answers)

Data science interviews test statistics, Python, machine learning intuition, business thinking, and communication skills. This guide covers the 50 most asked questions — with concise answers, code examples, and real interview scenarios.

Quick reference

Statistics & Probability

1. What is the difference between the mean, median, and mode? When do you use each?

Real example: Household income is right-skewed (a few billionaires inflate the mean). Median income better represents the typical household. House prices use median for the same reason.

2. Explain the Central Limit Theorem and why it matters.

CLT: The distribution of sample means approaches a normal distribution as sample size increases — regardless of the population distribution — given samples are independent and n ≥ 30 (rough rule of thumb).

Why it matters for data science:

• Justifies using normal-based hypothesis tests (t-tests, z-tests) on non-normal data
• Underpins A/B testing frameworks
• Means confidence intervals and p-values work even when the underlying metric (revenue, session length) is skewed
• Explains why bootstrap sampling works

3. What is a p-value? What does p < 0.05 actually mean?

A p-value is the probability of observing results at least as extreme as your data, assuming the null hypothesis is true.

• p = 0.03 means: if there were truly no effect, there's a 3% chance of seeing data this extreme by random chance
• p < 0.05 does NOT mean "95% probability the effect is real"
• p < 0.05 does NOT mean the effect is large or practically significant

Common misuses:

• Treating p-value as proof of truth (it's just evidence)
• Running until significant (p-hacking)
• Ignoring effect size alongside p-value
• Not correcting for multiple comparisons

4. What is the difference between Type I and Type II errors?

Power = 1 - β = probability of detecting a real effect. Increasing sample size reduces both errors.

5. Explain the difference between frequentist and Bayesian statistics.

Bayesian example: Spam filter. Prior = base rate of spam (60%). Likelihood = P(contains "free gift" | spam). Posterior = updated spam probability given the email's content.

6. What probability distributions should every data scientist know?

Interview tip: Know when data is "approximately normal" vs. when to check for skewness (use log-normal) or count data (Poisson).

7. What is correlation vs. causation? Give an example.

Correlation means two variables move together. Causation means one causes the other.

Classic example: Ice cream sales correlate with drowning deaths — both increase in summer. Removing ice cream wouldn't reduce drowning. The confounder is season (more swimming + more ice cream).

How to move from correlation to causation:

• Randomised controlled experiments (A/B tests)
• Instrumental variables
• Difference-in-differences
• Regression discontinuity
• Propensity score matching (observational data)

8. What is the law of large numbers?

As sample size increases, the sample mean converges to the population mean. This is why A/B tests need large samples, and why casino advantages are reliable at scale but not in small play sessions.

Contrast with CLT: LLN says the average stabilises; CLT describes the distribution of that average.

Python & pandas

9. How do you handle missing values in pandas?

import pandas as pd
import numpy as np

df = pd.DataFrame({
    'age':    [25, np.nan, 30, np.nan, 22],
    'salary': [50000, 60000, np.nan, 80000, 55000],
    'city':   ['NYC', 'LA', None, 'NYC', 'Chicago']
})

# Detect
df.isnull().sum()
df.isnull().mean()  # proportion missing per column

# Drop
df.dropna()                     # rows with any NaN
df.dropna(subset=['age'])       # only if age missing
df.dropna(thresh=3)             # keep rows with at least 3 non-NaN

# Fill — numeric
df['age'].fillna(df['age'].mean())       # mean imputation
df['age'].fillna(df['age'].median())     # median (better for skewed)
df['salary'].fillna(method='ffill')      # forward fill (time series)

# Fill — categorical
df['city'].fillna(df['city'].mode()[0])  # most frequent
df['city'].fillna('Unknown')

# Advanced: sklearn imputation
from sklearn.impute import SimpleImputer, KNNImputer
imp = KNNImputer(n_neighbors=5)
df[['age', 'salary']] = imp.fit_transform(df[['age', 'salary']])

10. How do you detect and handle outliers?

import numpy as np
import pandas as pd

data = pd.Series([10, 12, 11, 13, 100, 14, 9, 12])

# IQR method (robust)
Q1 = data.quantile(0.25)
Q3 = data.quantile(0.75)
IQR = Q3 - Q1
lower = Q1 - 1.5 * IQR
upper = Q3 + 1.5 * IQR
outliers = data[(data < lower) | (data > upper)]

# Z-score method (assumes normal distribution)
from scipy import stats
z_scores = np.abs(stats.zscore(data))
outliers_z = data[z_scores > 3]

# Modified Z-score (more robust)
median = np.median(data)
mad = np.median(np.abs(data - median))
modified_z = 0.6745 * (data - median) / mad

What to do with outliers:

• Investigate first — is it a data error or real?
• Cap/winsorize at 1st/99th percentile
• Use robust models (tree-based, quantile regression)
• Log transform to compress the scale
• Remove only if clearly erroneous

11. Explain groupby + agg in pandas.

df = pd.DataFrame({
    'region': ['US', 'EU', 'US', 'EU', 'US'],
    'product': ['A', 'A', 'B', 'B', 'A'],
    'revenue': [100, 200, 150, 300, 120],
    'units':   [10,   20,  15,  30,  12]
})

# Single aggregation
df.groupby('region')['revenue'].sum()

# Multiple aggregations
df.groupby('region').agg(
    total_revenue=('revenue', 'sum'),
    avg_revenue=('revenue', 'mean'),
    total_units=('units', 'sum'),
    num_products=('product', 'nunique')
)

# Multiple groupby keys
df.groupby(['region', 'product'])['revenue'].sum().unstack()

# Custom aggregation
df.groupby('region')['revenue'].agg(
    lambda x: x.quantile(0.75) - x.quantile(0.25)  # IQR
)

# Transform (keeps original index, returns same-length Series)
df['revenue_rank'] = df.groupby('region')['revenue'].rank(ascending=False)
df['pct_of_region'] = df['revenue'] / df.groupby('region')['revenue'].transform('sum')

12. How do you merge/join DataFrames in pandas?

customers = pd.DataFrame({'id': [1,2,3], 'name': ['Alice','Bob','Charlie']})
orders = pd.DataFrame({'order_id': [101,102,103], 'customer_id': [1,1,2], 'amount': [50,30,80]})

# INNER JOIN
pd.merge(customers, orders, left_on='id', right_on='customer_id', how='inner')

# LEFT JOIN — keep all customers, NaN if no orders
pd.merge(customers, orders, left_on='id', right_on='customer_id', how='left')

# Multiple keys
pd.merge(df1, df2, on=['key1', 'key2'], how='inner')

# Concat (stack vertically)
pd.concat([df1, df2], ignore_index=True)
pd.concat([df1, df2], axis=1)  # side by side

EDA & Visualisation

13. What is your approach to exploratory data analysis (EDA)?

A systematic EDA workflow:

import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

# 1. Shape & types
df.shape
df.dtypes
df.head()

# 2. Missing values
df.isnull().sum()
df.isnull().mean().sort_values(ascending=False)

# 3. Descriptive stats
df.describe()                          # numeric
df.describe(include='object')          # categorical

# 4. Target variable distribution
sns.histplot(df['target'])
df['target'].value_counts(normalize=True)  # class balance

# 5. Numeric distributions
df.hist(figsize=(12, 8), bins=30)

# 6. Correlations
sns.heatmap(df.corr(numeric_only=True), annot=True, fmt='.2f', cmap='coolwarm')

# 7. Categorical distributions
for col in df.select_dtypes('object').columns:
    print(df[col].value_counts())

# 8. Bivariate — numeric vs target
sns.boxplot(x='category', y='target', data=df)
sns.scatterplot(x='feature1', y='target', data=df)

# 9. Check for duplicates
df.duplicated().sum()

14. How do you choose the right chart type?

Rule: Use scatter for relationships, bars for comparisons, lines for trends. Avoid pie charts beyond 4-5 segments.

Feature Engineering

15. What is the difference between normalisation and standardisation?

from sklearn.preprocessing import StandardScaler, MinMaxScaler, RobustScaler

scaler = StandardScaler()
X_scaled = scaler.fit_transform(X_train)  # fit on train only!
X_test_scaled = scaler.transform(X_test)   # transform test

Critical: Always fit scalers on training data only. Fitting on test data causes data leakage.

16. How do you encode categorical variables?

import pandas as pd
from sklearn.preprocessing import LabelEncoder, OneHotEncoder

# One-hot
pd.get_dummies(df['city'], drop_first=True)

# Target encoding (manual with smoothing)
def target_encode(series, target, smoothing=10):
    global_mean = target.mean()
    stats = pd.DataFrame({'target': target, 'cat': series})
    agg = stats.groupby('cat')['target'].agg(['mean', 'count'])
    smooth = (agg['count'] * agg['mean'] + smoothing * global_mean) / (agg['count'] + smoothing)
    return series.map(smooth)

17. What is feature selection and what methods exist?

from sklearn.feature_selection import SelectKBest, f_classif, RFE
from sklearn.ensemble import RandomForestClassifier

# Filter: top k features by ANOVA F-test
selector = SelectKBest(f_classif, k=10)
X_selected = selector.fit_transform(X_train, y_train)

# Embedded: Lasso
from sklearn.linear_model import Lasso
lasso = Lasso(alpha=0.01).fit(X_train, y_train)
selected = [f for f, coef in zip(features, lasso.coef_) if coef != 0]

# Tree importance
rf = RandomForestClassifier().fit(X_train, y_train)
importances = pd.Series(rf.feature_importances_, index=features).sort_values(ascending=False)

Model Evaluation

18. When do you use accuracy vs. precision vs. recall vs. F1?

Rules of thumb:

• Cancer diagnosis: maximize recall (missing cancer is catastrophic)
• Spam filter: maximize precision (false positives annoy users)
• Credit scoring: AUC (need ranking, not one threshold)
• Class imbalance 1:100+: use PR AUC, not ROC AUC

19. What is cross-validation and why do you need it?

from sklearn.model_selection import cross_val_score, StratifiedKFold
from sklearn.ensemble import GradientBoostingClassifier

model = GradientBoostingClassifier()

# 5-fold CV
scores = cross_val_score(model, X, y, cv=5, scoring='roc_auc')
print(f"AUC: {scores.mean():.3f} ± {scores.std():.3f}")

# Stratified (preserves class proportion) — use for classification
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
scores = cross_val_score(model, X, y, cv=cv, scoring='f1')

# Time series: use TimeSeriesSplit (no future leakage)
from sklearn.model_selection import TimeSeriesSplit
tscv = TimeSeriesSplit(n_splits=5)

Why CV matters:

• Single train/test split has high variance — different random splits give different results
• CV averages over multiple splits for stable estimates
• Enables hyperparameter tuning without touching the test set
• Detects overfitting

20. How do you handle class imbalance?

from sklearn.utils.class_weight import compute_class_weight
from imblearn.over_sampling import SMOTE

# Class weights (fastest)
weights = compute_class_weight('balanced', classes=[0,1], y=y_train)
model = RandomForestClassifier(class_weight={0: weights[0], 1: weights[1]})

# SMOTE
sm = SMOTE(random_state=42)
X_resampled, y_resampled = sm.fit_resample(X_train, y_train)
# ⚠️ Only apply SMOTE to training data!

ML Fundamentals

21. Explain the bias-variance tradeoff.

Bias = error from wrong assumptions (underfitting). A linear model on non-linear data has high bias.

Variance = sensitivity to training data fluctuations (overfitting). A deep decision tree memorises noise.

Total error = Bias² + Variance + Irreducible noise

Remedies for high variance: regularisation, more data, simpler model, cross-validation. Remedies for high bias: more features, more complex model, better feature engineering.

22. How does gradient boosting work?

Gradient boosting builds an ensemble of weak learners (typically shallow decision trees) sequentially, where each tree corrects errors of the previous.

F₀(x) = initial prediction (mean)
For m = 1 to M:
    rₘ = negative gradient of loss w.r.t. current prediction  ← residuals
    hₘ = tree fitted to residuals
    Fₘ(x) = Fₘ₋₁(x) + η × hₘ(x)    ← η = learning rate

Key hyperparameters:

• n_estimators — number of trees (more → lower bias, higher variance)
• learning_rate — shrinkage per tree (lower → need more trees, more regularised)
• max_depth — tree complexity (3-5 typical)
• subsample — stochastic gradient boosting (0.8 common)
• min_samples_leaf — regularisation

XGBoost, LightGBM, CatBoost all implement gradient boosting with hardware optimisations.

23. What is regularisation and when do you use L1 vs. L2?

Both L1 and L2 add a penalty to the loss function to discourage large weights.

from sklearn.linear_model import Ridge, Lasso, ElasticNet

ridge = Ridge(alpha=1.0)       # L2
lasso = Lasso(alpha=0.01)      # L1
enet  = ElasticNet(alpha=0.01, l1_ratio=0.5)

A/B Testing

24. Walk me through designing an A/B test.

Step-by-step:

1. Define hypothesis — "Changing CTA button from green to orange will increase click-through rate."
2. Choose metric — primary KPI (CTR), guardrail metrics (session length, bounce rate)
3. Calculate sample size:

   from statsmodels.stats.power import TTestIndPower
   # 80% power, 5% significance, detect 2% relative lift on 10% baseline CTR
   baseline = 0.10
   mde = 0.002  # minimum detectable effect (absolute)
   effect_size = mde / (baseline * (1 - baseline)) ** 0.5
   analysis = TTestIndPower()
   n = analysis.solve_power(effect_size=effect_size, alpha=0.05, power=0.80)
   

4. Randomise — hash user ID modulo 100, assign 0-49 → control, 50-99 → treatment
5. Run test — wait for predetermined sample size (don't peek and stop early)
6. Analyse — two-proportion z-test for CTR, t-test for continuous metrics
7. Ship or reject — check p-value + practical significance + guardrail metrics

25. What are common A/B testing pitfalls?

26. How do you decide if an A/B test result is practically significant?

Statistical significance (p < 0.05) alone is not enough — with large samples even tiny effects are significant.

• Effect size: Did CTR go from 10.0% to 10.1%? Is that worth deploying?
• Confidence interval: "The lift is between 0.5% and 2.1%" — entire CI above minimum acceptable?
• Business impact: Effect size × daily users × revenue per conversion = $X/month
• Cost of shipping: Engineering effort, complexity, maintenance
• Guardrail metrics: Did any secondary metric move negatively?

SQL for Data Science

27. How do you calculate a 7-day rolling average in SQL?

-- Daily active users with 7-day rolling average
SELECT
    date,
    dau,
    AVG(dau) OVER (
        ORDER BY date
        ROWS BETWEEN 6 PRECEDING AND CURRENT ROW
    ) AS rolling_7d_avg
FROM daily_stats
ORDER BY date;

28. How do you find user retention in SQL?

-- Day-1 retention: users who returned the day after registration
WITH cohort AS (
    SELECT
        user_id,
        DATE(MIN(event_time)) AS signup_date
    FROM events
    WHERE event_type = 'signup'
    GROUP BY user_id
),
activity AS (
    SELECT DISTINCT
        user_id,
        DATE(event_time) AS active_date
    FROM events
)
SELECT
    c.signup_date,
    COUNT(DISTINCT c.user_id)  AS cohort_size,
    COUNT(DISTINCT a.user_id)  AS retained,
    COUNT(DISTINCT a.user_id) * 1.0 / COUNT(DISTINCT c.user_id) AS d1_retention
FROM cohort c
LEFT JOIN activity a
    ON c.user_id = a.user_id
    AND a.active_date = c.signup_date + INTERVAL '1 day'
GROUP BY c.signup_date
ORDER BY c.signup_date;

29. Write a query to find the second-highest salary per department.

-- Using DENSE_RANK
WITH ranked AS (
    SELECT
        employee_id,
        department,
        salary,
        DENSE_RANK() OVER (PARTITION BY department ORDER BY salary DESC) AS rk
    FROM employees
)
SELECT department, salary AS second_highest_salary
FROM ranked
WHERE rk = 2;

Business Acumen

30. Daily active users dropped 20% yesterday. How do you investigate?

A systematic diagnosis framework:

1. Is the drop real?
   - Check data pipeline: ETL delay? Missing events? Logging bug?
   - Check other metrics: if all metrics dropped proportionally → likely data issue

2. Scope the problem
   - By platform: iOS / Android / Web
   - By geography: specific country or global?
   - By user segment: new vs. returning users?
   - By traffic source: organic / paid / email?

3. What changed?
   - Any deployments or feature flags yesterday?
   - Marketing campaign started or ended?
   - Competitor launched something?
   - External event (holiday, news event)?
   - Push notification failure?

4. Form hypotheses ranked by likelihood
   - Most likely: data pipeline issue
   - Next: feature deployment caused friction
   - Then: external factor (holiday in key market)

5. Validate each hypothesis with data

6. Communicate findings + recommended action

31. How would you measure the success of a new feature?

Framework: Define metrics before launching.

Goal metric (primary): Directly measures feature value

• New search feature → queries per session, search success rate

Engagement metric: Feature adoption

• % of active users who used the feature

Quality metric: Is the experience good?

• Feature-specific satisfaction, error rate, P95 latency

Guardrail metrics: Ensure you haven't broken anything

• Core engagement (DAU, session length), revenue, retention

Evaluation:

• Short-term: A/B test for 2+ weeks
• Long-term: cohort analysis (do users who adopted retain better?)

Practical ML Scenarios

32. How do you prevent data leakage?

Data leakage occurs when training data contains information that wouldn't be available at prediction time — causing falsely optimistic CV scores that fail in production.

Common sources:

• Target in features: Including columns derived from the label
• Future data in past prediction: Using next day's features to predict today
• Global statistics before splitting: Fitting a scaler on the full dataset
• Duplicate records across splits: Same user in train and test

# ❌ WRONG: leakage
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)  # all data!
X_train, X_test = train_test_split(X_scaled)

# ✅ CORRECT: no leakage
X_train, X_test, y_train, y_test = train_test_split(X, y)
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)   # fit on train only
X_test  = scaler.transform(X_test)         # transform test

33. How do you approach a churn prediction problem?

1. Define churn: no activity for 30 days? Cancelled subscription? Context-dependent.

2. Frame as classification: binary (churned/active) or survival analysis

3. Build features:
   - Recency (days since last action)
   - Frequency (actions per week)
   - Monetary (spend per month)
   - Engagement trajectory (declining?)
   - Support tickets, plan type, usage of key features

4. Handle class imbalance (often 5-20% churn)
   - Use class_weight='balanced'
   - Evaluate with F1, PR AUC

5. Model choice:
   - Logistic regression (interpretable baseline)
   - Gradient boosting (usually best performance)
   - Survival analysis if time-to-churn matters

6. Calibrate probabilities (CalibratedClassifierCV)

7. Business action: at what threshold do we intervene?
   Optimise threshold for precision/recall tradeoff given intervention cost

8. Monitor: retrain monthly, watch for drift in feature distributions

34. How would you build a recommendation system?

Cold start problem:

• New user → use demographics + onboarding preferences
• New item → use content features

Evaluation: Offline (precision@K, recall@K, NDCG) + Online A/B test (CTR, conversion).

Communication & Scenarios

35. How do you explain a complex model to a non-technical stakeholder?

Use the 3-level explanation:

1. What does it do? — "Given a customer's history, the model predicts how likely they are to leave us in the next 30 days."

2. How accurate is it? — "Of the customers it flags as high-risk, 65% actually churn. We catch 80% of all churners."

3. What should we do with it? — "If we sort customers by this score and call the top 500 highest-risk ones, our retention team can prevent ~$200k in monthly revenue loss."

Avoid: accuracy percentages without context, technical jargon, black-box framing.

36. What is the difference between a data scientist, data analyst, and ML engineer?

In practice, roles overlap heavily at startups. Large companies have more specialisation.

Remaining Interview Topics

37. What is PCA and when do you use it?

Principal Component Analysis finds orthogonal axes (principal components) that capture maximum variance. Used for:

• Dimensionality reduction before modelling (speed, remove correlated features)
• Visualisation (2D/3D projection of high-dim data)
• Noise reduction

from sklearn.decomposition import PCA
import matplotlib.pyplot as plt

pca = PCA(n_components=2)
X_2d = pca.fit_transform(X_scaled)

print(pca.explained_variance_ratio_)  # variance per component
print(pca.explained_variance_ratio_.cumsum())  # cumulative
plt.scatter(X_2d[:, 0], X_2d[:, 1], c=y)

Limitations: Only captures linear relationships. For non-linear: t-SNE or UMAP (visualisation only). PCA requires scaling first.

38. What is the difference between K-means and hierarchical clustering?

from sklearn.cluster import KMeans
from sklearn.metrics import silhouette_score

# Elbow method for k selection
inertias = []
for k in range(2, 11):
    km = KMeans(n_clusters=k, random_state=42).fit(X_scaled)
    inertias.append(km.inertia_)

# Silhouette score (higher = better, range -1 to 1)
km = KMeans(n_clusters=3, random_state=42).fit(X_scaled)
print(silhouette_score(X_scaled, km.labels_))

39. Explain precision-recall tradeoff with a concrete example.

Cancer screening model:

• At threshold 0.3: Recall=0.95, Precision=0.20 → catch 95% of cancers, but 80% of positives are false alarms → acceptable (missing cancer is catastrophic)
• At threshold 0.7: Recall=0.60, Precision=0.75 → miss 40% of cancers but fewer false alarms → not acceptable

Spam filter:

• At threshold 0.3: Recall=0.98, Precision=0.60 → catch nearly all spam but 40% of flagged are real emails → bad
• At threshold 0.8: Recall=0.70, Precision=0.97 → miss some spam but rarely flag real emails → acceptable

The optimal threshold depends on the cost ratio of false positives to false negatives.

40. How do you monitor a deployed model?

from scipy.stats import ks_2samp

# Kolmogorov-Smirnov test for distribution drift
train_feature = X_train['age']
live_feature   = new_data['age']
stat, p_value  = ks_2samp(train_feature, live_feature)
if p_value < 0.05:
    print("Drift detected in 'age'")

41. What is the difference between bagging and boosting?

42. What is the difference between a confusion matrix and ROC curve?

Confusion matrix: shows TP/FP/TN/FN at one threshold — gives precision, recall, accuracy at that point.

ROC curve: plots TPR vs. FPR across all thresholds — threshold-independent performance measure. AUC = probability that model ranks a random positive higher than a random negative.

When to prefer each:

• Confusion matrix: when you've decided on a threshold and want to understand specific errors
• ROC/AUC: when comparing models across all thresholds, or doing model selection
• PR curve: when class imbalance is severe (positive class rare)

43. What is transfer learning and when do you use it?

Transfer learning uses a model pre-trained on a large dataset as a starting point, then fine-tunes on your (smaller) target dataset.

from transformers import AutoTokenizer, AutoModelForSequenceClassification

# Pre-trained BERT fine-tuned for sentiment
model = AutoModelForSequenceClassification.from_pretrained(
    "bert-base-uncased", num_labels=2
)
# Fine-tune on your domain data

When to use:

• Limited labelled data in your domain
• Target task similar to pre-training task
• Pre-training data domain overlaps with yours (NLP: general text → medical text)

Common in:

• NLP (BERT, GPT fine-tuning)
• Computer vision (ResNet/EfficientNet backbone for custom classification)

44. What are embeddings and how are they used?

An embedding is a dense low-dimensional vector representation of a high-dimensional entity (word, user, product).

from sentence_transformers import SentenceTransformer
import numpy as np

model = SentenceTransformer('all-MiniLM-L6-v2')
sentences = ["Paris is beautiful", "I love France", "Python is great"]
embeddings = model.encode(sentences)  # (3, 384) arrays

# Cosine similarity
from sklearn.metrics.pairwise import cosine_similarity
sim = cosine_similarity(embeddings)

Uses:

• Semantic search: find similar documents
• Recommendation: user/item embeddings for collaborative filtering
• Anomaly detection: flag items far from cluster centres
• Feature input for downstream models

45. How do you approach a time series forecasting problem?

# 1. Plot and understand
import matplotlib.pyplot as plt
ts.plot()
# Identify: trend? seasonality? irregular spikes?

# 2. Decompose
from statsmodels.tsa.seasonal import seasonal_decompose
decomp = seasonal_decompose(ts, model='additive', period=7)
decomp.plot()

# 3. Check stationarity
from statsmodels.tsa.stattools import adfuller
stat, p, *_ = adfuller(ts)
# If p > 0.05: non-stationary → difference the series

# 4. Model options:
# - Statistical: ARIMA, SARIMA, Exponential Smoothing (Prophet)
# - ML: Features (lag, rolling stats, calendar) → GBM
# - DL: LSTM, Temporal Fusion Transformer, N-BEATS

from prophet import Prophet
model = Prophet(yearly_seasonality=True, weekly_seasonality=True)
model.fit(df)  # df has 'ds' (date) and 'y' (value) columns
future = model.make_future_dataframe(periods=30)
forecast = model.predict(future)

Anti-patterns table

Data science vs. related roles

Common mistakes

FAQ

Q: What programming languages do data scientists use most in 2025?

Python dominates (80%+) with pandas, NumPy, sklearn, matplotlib, and LightGBM. SQL is the second must-have. R is still used in academia and pharma. Julia is growing for numerical computing. Most production ML is Python.

Q: Do I need a PhD to become a data scientist?

No. A BS/MS in statistics, CS, maths, or engineering is sufficient for most industry roles. A PhD helps for research-focused positions (Meta AI Research, DeepMind) or cutting-edge algorithm work. Industry values portfolio projects, Kaggle results, and practical ML skills over academic credentials.

Q: What is the difference between a data science notebook and production code?

Notebooks (Jupyter) are for exploration, EDA, and prototyping — linear execution, no tests, messy OK. Production code is modular, version-controlled, tested, logged, with error handling and scheduled retraining. Most ML projects start as notebooks and get refactored into production Python packages or ML pipelines (MLflow, Airflow, Kubeflow).

Q: Is deep learning replacing classical machine learning for data scientists?

For tabular data (most business problems): no. Tree-based methods (LightGBM, XGBoost, CatBoost) still outperform neural nets on structured data. Deep learning dominates for images, text, audio, and video. Data scientists working on business analytics use classical ML 90% of the time.

Q: What Kaggle competitions are most useful for data science interviews?

Tabular competitions (Porto Seguro, Titanic, Housing Prices) for feature engineering. NLP competitions (Jigsaw, feedback) for text. Time-series competitions (Store Sales, M5) for forecasting. Being in the top 10% on any moderately competitive competition is impressive — focus on learning over ranking.

Q: How do I transition from data analyst to data scientist?

Learn statistics (hypothesis testing, probability, regression) beyond basic reporting. Get comfortable with scikit-learn and building end-to-end ML pipelines. Complete a Kaggle competition. Build a project that predicts something (churn, price, demand) with proper train/test evaluation. Study A/B testing methodology deeply. The main gaps are usually statistics and coding depth, not domain knowledge.