Data Science
The Complete Data Science Roadmap 2026: Master 72 Essential Topics Across 12 Core Patterns
Master data science with this comprehensive 2026 roadmap: 12 core patterns from data manipulation to deep learning, plus 287 practice problems.
Why Most Data Scientists Fail FAANG Interviews
You've built dashboards. You've run A/B tests. You've even deployed ML models to production.
But when the interviewer asks you to "explain the bias-variance tradeoff with a real example" or "design an experimentation framework for a company with 100M users," you stumble.
Sound familiar?
Here's the brutal truth: Data science interviews aren't about knowing pandas and scikit-learn. They're about understanding statistical fundamentals, knowing when to apply which algorithms, and reasoning about business impact.
Companies like Meta, Google, Netflix, and Uber don't want SQL query writers. They want data scientists who deeply understand statistical inference, can design robust experiments, optimize ML systems at scale, and translate insights into business value.
What if you had a complete roadmap of every data science concept that actually matters — organized, structured, and ready to master?
That's exactly what this guide delivers.
What You'll Learn
This comprehensive roadmap breaks down 72 essential data science topics organized into 12 fundamental patterns that form the backbone of modern data science at companies like Meta, Google, Netflix, Uber, and Airbnb.
We've analyzed thousands of data science interviews and mapped out exactly what you need to know — from foundational statistics to production ML systems.
By the end, you'll not only understand these concepts but also know how to practice them interactively with AI coaching on Thita.ai — where you can solve 287 real data science problems, get explanations, and simulate actual DS interviews.
What Is a Data Science Roadmap?
Think of this roadmap as your complete knowledge graph for data science mastery.
Instead of randomly jumping between YouTube tutorials on linear regression, A/B testing, and neural networks, you follow a structured path with clear prerequisites that builds knowledge progressively — from data manipulation fundamentals to production deep learning systems.
Example Learning Flow:
- Start with Data Manipulation & Preprocessing → Master Statistics & Probability → Learn Exploratory Data Analysis → Build Supervised Learning Models → Optimize with Model Selection & Validation → Scale to Production Systems
This systematic approach is exactly how senior data scientists at top tech companies prepare. They don't cram random tutorials — they build comprehensive mental models through structured learning.
What Makes This Roadmap Different:
- Maps directly to Thita.ai's 287 curated problems across 12 patterns
- Follows pedagogical prerequisites (each pattern builds on previous ones)
- Interview-focused with 25+ real FAANG interview questions
- Includes 12 real-world case studies from Netflix, Uber, Meta, and more
- Provides actionable 12-week learning plan
The 12 Core Patterns: Your Structured Learning Path
Each pattern represents a critical pillar of modern data science. We've organized them in recommended learning order with clear prerequisites and difficulty progression.
Total Content:
- 12 Core Patterns covering the complete DS landscape
- 72 Subpatterns (6 per pattern) for deep expertise
- 287 Practice Problems distributed across all topics
- Estimated Time: 16-25 weeks from beginner to expert
Below, we break down all 12 patterns with their essential topics, following the exact structure used by thousands of data scientists on Thita.ai.
1. Data Manipulation & Preprocessing (23 Problems)
Foundation: The critical first step in every data science project
Prerequisites: None — Start here! Difficulty: Beginner Learning Time: 1-2 weeks
Essential Subpatterns:
- Data Cleaning & Missing Values: Handling NaN, imputation strategies (mean, median, KNN, MICE), identifying data quality issues
- Outlier Detection & Treatment: Z-score, IQR method, isolation forests, winsorization, capping strategies
- Data Transformation & Scaling: StandardScaler, MinMaxScaler, RobustScaler, log transforms, Box-Cox, power transforms
- Data Aggregation & Grouping: GroupBy operations, pivot tables, multi-level indexing, window functions
- Feature Engineering: Creating interaction features, polynomial features, binning, encoding categorical variables
- Time Series Data Handling: Date parsing, resampling, rolling windows, lag features, time-based aggregations
Why This Matters: 80% of a data scientist's time is spent on data preparation. Master this first or everything else falls apart.
Real Interview Question (Meta): "You have a dataset with 30% missing values across multiple columns. Walk me through your approach to handle missing data and explain the trade-offs of each method."
💡 Pro Tip: Know when to use different imputation methods. Mean/median for MCAR (missing completely at random), KNN for MAR (missing at random), and domain-specific approaches for MNAR (missing not at random).
⚠️ Common Mistake: Applying StandardScaler before train-test split, causing data leakage. Always split first, then fit on training data only.
Practice on Thita.ai: Start with Data Manipulation →
2. Statistics & Probability Fundamentals (24 Problems)
Foundation: The mathematical backbone of all data science
Prerequisites: Data Manipulation & Preprocessing Difficulty: Beginner Learning Time: 2 weeks
Essential Subpatterns:
- Descriptive Statistics: Mean, median, mode, variance, standard deviation, quartiles, skewness, kurtosis
- Probability Distributions: Normal, binomial, Poisson, exponential, uniform, understanding PDF/PMF/CDF
- Confidence Intervals: Calculation methods, interpretation, margin of error, sample size determination
- Hypothesis Testing: Null/alternative hypotheses, p-values, significance levels, Type I/II errors, power analysis
- A/B Testing: Experimental design, statistical significance, sample size calculation, multiple testing corrections
- Bayesian Statistics: Prior/posterior distributions, Bayes' theorem, conjugate priors, Bayesian A/B testing
Why This Is Critical: Statistics is the language of data science. You can't build reliable models without understanding statistical inference.
💡 Pro Tip: Understand the Central Limit Theorem deeply — it's the foundation for why many statistical methods work in practice.
⚠️ Common Mistake: Confusing p-value (probability of data given null hypothesis) with probability that null hypothesis is true.
Real Interview Question (Google): "Explain the difference between Type I and Type II errors using a real business example. How would you choose the significance level for an A/B test?"
Practice on Thita.ai: Master Statistics Fundamentals →
3. Exploratory Data Analysis (EDA) (24 Problems)
Foundation: Understanding your data before modeling
Prerequisites: Data Manipulation & Preprocessing + Statistics & Probability Fundamentals Difficulty: Beginner Learning Time: 1 week
Essential Subpatterns:
- Univariate Analysis: Analyzing single variables, frequency distributions, summary statistics, identifying patterns
- Distribution Analysis: Assessing normality, Q-Q plots, distribution fitting, transformation selection
- Data Visualization Techniques: Histograms, box plots, violin plots, scatter plots, heatmaps, pair plots
- Bivariate Analysis: Relationship between two variables, contingency tables, chi-square tests
- Correlation Analysis: Pearson, Spearman, Kendall correlations, correlation matrices, multicollinearity detection
- Pattern Recognition in Data: Identifying trends, seasonality, clusters, anomalies, data storytelling
Why This Matters: Good EDA reveals insights that inform modeling decisions and often uncovers business-critical findings before any modeling.
💡 Pro Tip: Always visualize your data before modeling. Anscombe's Quartet proves that identical summary statistics can hide completely different distributions.
Real Interview Question (Netflix): "You're given a dataset of user viewing behavior. Walk me through your EDA process to identify factors driving user engagement."
Common Pattern: EDA questions test your structured thinking and ability to extract insights, not just plotting skills.
Practice on Thita.ai: Master EDA Techniques →
4. Supervised Learning - Regression (25 Problems)
Foundation: Predicting continuous outcomes
Prerequisites: Statistics & Probability + EDA Difficulty: Intermediate Learning Time: 1-2 weeks
Essential Subpatterns:
- Linear Regression: OLS assumptions, coefficient interpretation, R², adjusted R², residual analysis
- Regression Evaluation Metrics: MSE, RMSE, MAE, MAPE, R², explained variance, choosing metrics for business context
- Polynomial & Ridge Regression: Polynomial features, regularization (L2), hyperparameter tuning, preventing overfitting
- Tree-based Regression: Decision trees, regression trees, tree pruning, interpretability vs accuracy
- Ensemble Methods for Regression: Random Forest, Gradient Boosting (XGBoost, LightGBM, CatBoost), bagging vs boosting
- Logistic Regression: Binary classification via regression, sigmoid function, log-odds, decision boundaries
Why This Is Critical: Regression is the foundation of predictive modeling. Most business problems involve predicting continuous values (revenue, churn probability, time to event).
💡 Pro Tip: Ridge regression (L2) shrinks coefficients smoothly, while Lasso (L1) can zero them out for feature selection. Know when to use each.
⚠️ Common Mistake: Not checking OLS assumptions (linearity, homoscedasticity, normality of residuals, independence). Violated assumptions invalidate your model.
Real Interview Question (Uber): "Design a model to predict ride demand 30 minutes ahead for each city zone. Which regression algorithm would you use and why?"
Practice on Thita.ai: Master Regression Techniques →
5. Supervised Learning - Classification (25 Problems)
Foundation: Predicting categorical outcomes
Prerequisites: Supervised Learning - Regression Difficulty: Intermediate Learning Time: 1-2 weeks
Essential Subpatterns:
- Binary Classification: Two-class problems, decision boundaries, probability calibration, threshold selection
- Classification Evaluation Metrics: Accuracy, precision, recall, F1-score, ROC-AUC, PR curves, confusion matrix
- Multi-class Classification: One-vs-rest, one-vs-one, softmax, handling class imbalance across multiple classes
- Tree-based Classification: Decision trees, Gini impurity, information gain, CART algorithm, interpretability
- SVM & Kernel Methods: Support vector machines, kernel trick, RBF/polynomial kernels, margin maximization
- Ensemble Methods for Classification: Random Forest, Gradient Boosting, AdaBoost, voting classifiers, stacking
Why This Matters: Classification powers everything from spam detection to fraud prevention to medical diagnosis. It's tested in 90% of DS interviews.
💡 Pro Tip: For imbalanced datasets, use precision-recall curves instead of ROC-AUC. ROC-AUC can be misleadingly optimistic when classes are imbalanced.
Real Interview Question (Meta): "You're building a model to detect fake accounts. Your dataset has 99% real accounts, 1% fake. How do you approach this? What metrics do you use?"
Common Pattern: Interviewers test your understanding of class imbalance, metric selection, and business trade-offs (false positives vs false negatives).
Practice on Thita.ai: Master Classification →
6. Model Selection & Validation (24 Problems)
Foundation: Building reliable, generalizable models
Prerequisites: Supervised Learning - Regression + Classification Difficulty: Intermediate Learning Time: 1 week
Essential Subpatterns:
- Train-Validation-Test Split: Proper data splitting, stratification, time-based splits, preventing data leakage
- Overfitting & Underfitting: Recognizing symptoms, learning curves, model complexity trade-offs
- Bias-Variance Tradeoff: Decomposition, irreducible error, finding the sweet spot, relation to model complexity
- Cross-validation Techniques: K-fold CV, stratified K-fold, time series CV, leave-one-out, nested CV
- Model Comparison: Statistical tests for model comparison, paired t-tests, McNemar's test, practical vs statistical significance
- Hyperparameter Tuning: Grid search, random search, Bayesian optimization, automated hyperparameter tuning
Why This Is Important: A model that works on training data but fails in production is useless. Validation techniques ensure your model generalizes.
💡 Pro Tip: Use stratified K-fold for classification to ensure each fold has similar class distributions. Use time-based splits for time series data.
⚠️ Common Mistake: Tuning hyperparameters on the test set. The test set should be touched only once at the very end for final evaluation.
Real Interview Question (Amazon): "Explain the bias-variance tradeoff. How would you diagnose whether your model is suffering from high bias or high variance?"
Practice on Thita.ai: Master Model Validation →
7. Feature Selection & Dimensionality Reduction (22 Problems)
Foundation: Working with high-dimensional data
Prerequisites: Model Selection & Validation Difficulty: Intermediate Learning Time: 1 week
Essential Subpatterns:
- Feature Importance Analysis: Tree-based importance, permutation importance, SHAP values, interpreting feature contributions
- Filter Methods: Correlation-based selection, mutual information, chi-square test, variance threshold
- Wrapper Methods: Forward selection, backward elimination, recursive feature elimination (RFE), computational trade-offs
- Embedded Methods: Lasso (L1 regularization), ElasticNet, tree-based feature selection during training
- PCA & SVD: Principal Component Analysis, explained variance, singular value decomposition, eigenvalues/eigenvectors
- t-SNE & UMAP: Non-linear dimensionality reduction, visualization of high-dimensional data, manifold learning
Why This Matters: High-dimensional data leads to curse of dimensionality, overfitting, and computational inefficiency. Feature selection improves model performance and interpretability.
💡 Pro Tip: PCA is linear and preserves variance. t-SNE/UMAP are non-linear and preserve local structure. Use PCA for modeling, t-SNE/UMAP for visualization.
Real Interview Question (LinkedIn): "You have a dataset with 10,000 features and 1,000 samples. How would you approach feature selection? What methods would you use and why?"
Practice on Thita.ai: Explore Feature Engineering →
8. Unsupervised Learning (25 Problems)
Foundation: Finding patterns without labels
Prerequisites: Feature Selection & Dimensionality Reduction Difficulty: Intermediate Learning Time: 1-2 weeks
Essential Subpatterns:
- K-Means Clustering: Algorithm mechanics, elbow method, silhouette score, choosing K, initialization strategies
- Hierarchical Clustering: Agglomerative vs divisive, linkage methods (single, complete, average, Ward), dendrograms
- DBSCAN & Density-based Clustering: Density-based clustering, epsilon and min_samples, handling arbitrary shapes
- Principal Component Analysis (PCA): Variance maximization, dimensionality reduction, interpretation of components
- Anomaly Detection: Isolation Forest, Local Outlier Factor (LOF), statistical methods, one-class SVM
- Association Rules & Market Basket Analysis: Apriori algorithm, support, confidence, lift, frequent itemsets
Why This Is Critical: Not all data has labels. Unsupervised learning powers recommendation systems, customer segmentation, and anomaly detection.
💡 Pro Tip: K-means assumes spherical clusters of similar sizes. Use DBSCAN when you have arbitrary cluster shapes or noise points.
⚠️ Common Mistake: Using K-means on high-dimensional data without dimensionality reduction first. Distances become meaningless in high dimensions.
Real Interview Question (Airbnb): "Design a system to segment users based on their behavior. How would you choose the number of clusters? How would you validate the quality of your clustering?"
Practice on Thita.ai: Master Unsupervised Learning →
9. Business Analytics & Metrics (24 Problems)
Foundation: Translating data into business value
Prerequisites: EDA + Supervised Learning - Regression + Classification Difficulty: Intermediate Learning Time: 1 week
Essential Subpatterns:
- Business KPI Analysis: Defining KPIs, north star metrics, metric trees, OKRs, tracking business health
- Customer Segmentation: RFM analysis, cohort analysis, CLV (Customer Lifetime Value), behavioral segmentation
- Marketing Analytics: Attribution modeling, funnel analysis, CAC (Customer Acquisition Cost), ROAS, incrementality testing
- Churn Prediction: Survival analysis, time-to-event modeling, retention curves, proactive churn prevention
- Price Optimization: Demand elasticity, price sensitivity analysis, dynamic pricing, A/B testing pricing strategies
- Recommendation Systems: Collaborative filtering, content-based filtering, matrix factorization, hybrid approaches
Hottest Topic in 2026: Data scientists are increasingly expected to understand business impact, not just build models. This is what separates senior DS from junior.
💡 Pro Tip: Always frame your analysis in terms of business impact. "The model has 95% accuracy" is less compelling than "This model will save $2M annually by reducing false positives."
Real Interview Question (Spotify): "Daily active users dropped 10% last week. Walk me through how you'd investigate this. What metrics would you look at? What hypotheses would you test?"
Common Pattern: Product/business sense questions test your ability to think like a stakeholder, not just a modeler.
Practice on Thita.ai: Master Business Analytics →
10. Time Series Analysis (24 Problems)
Foundation: Modeling temporal data and forecasting
Prerequisites: Supervised Learning - Regression + Model Selection & Validation Difficulty: Advanced Learning Time: 1-2 weeks
Essential Subpatterns:
- Time Series Components: Trend, seasonality, cyclical patterns, noise decomposition (additive vs multiplicative)
- Trend & Seasonality Analysis: Moving averages, seasonal decomposition (STL), differencing, detrending
- Forecasting Techniques: Exponential smoothing, Holt-Winters, simple moving average, weighted moving average
- ARIMA Models: AutoRegressive Integrated Moving Average, ACF/PACF, order selection, stationarity testing
- Time Series Cross-validation: Rolling window, expanding window, time-based splits, avoiding data leakage
- Advanced Time Series Models: SARIMA, Prophet, LSTM for time series, ensemble forecasting
Why This Matters: Time series powers demand forecasting, anomaly detection, financial modeling, and capacity planning. It's critical for companies like Uber, DoorDash, and Netflix.
💡 Pro Tip: Always check for stationarity using ADF test before applying ARIMA. Non-stationary series need differencing.
⚠️ Common Mistake: Using standard cross-validation for time series data. This causes data leakage. Always use time-based splits.
Real Interview Question (DoorDash): "Design a system to forecast food delivery demand for the next week across all restaurants in a city. What features would you use? How would you handle seasonality?"
Practice on Thita.ai: Master Time Series Analysis →
11. Natural Language Processing (24 Problems)
Foundation: Understanding and modeling text data
Prerequisites: Unsupervised Learning + Feature Selection & Dimensionality Reduction Difficulty: Advanced Learning Time: 2 weeks
Essential Subpatterns:
- Text Preprocessing: Tokenization, stemming, lemmatization, stop word removal, text normalization
- Bag of Words & TF-IDF: Count vectorization, term frequency-inverse document frequency, n-grams, feature extraction
- Word Embeddings: Word2Vec (CBOW, Skip-gram), GloVe, FastText, contextual embeddings vs static
- Sentiment Analysis: Polarity detection, aspect-based sentiment, handling negation, domain-specific sentiment
- Named Entity Recognition: Entity extraction, NER models, information extraction, knowledge graphs
- Topic Modeling: Latent Dirichlet Allocation (LDA), Non-negative Matrix Factorization (NMF), discovering themes
Growing Field: NLP powers search, chatbots, content moderation, and recommendation systems. Essential for modern DS roles.
💡 Pro Tip: TF-IDF captures word importance better than raw counts by downweighting common words. Use it over Bag of Words for most applications.
Real Interview Question (Twitter): "Design a system to detect toxic comments in real-time. How would you handle slang, misspellings, and context-dependent toxicity?"
Practice on Thita.ai: Explore NLP Techniques →
12. Deep Learning Fundamentals (25 Problems)
Foundation: Neural networks for complex patterns
Prerequisites: Supervised Learning - Classification + Unsupervised Learning + Feature Selection Difficulty: Expert Learning Time: 2-3 weeks
Essential Subpatterns:
- Neural Network Basics: Perceptrons, activation functions, forward propagation, loss functions, network architecture
- Backpropagation & Optimization: Gradient descent, chain rule, optimizers (SGD, Adam, RMSprop), learning rate scheduling
- Deep Learning for Tabular Data: Deep neural networks for structured data, embeddings for categorical features, entity embeddings
- Convolutional Neural Networks (CNNs): Convolution layers, pooling, filters/kernels, image classification, transfer learning
- Recurrent Neural Networks (RNNs): LSTM, GRU, sequence modeling, vanishing gradients, bidirectional RNNs
- Transfer Learning: Pretrained models, fine-tuning, feature extraction, domain adaptation
Critical for 2026: Deep learning is becoming table stakes for senior DS roles, especially for computer vision, NLP, and complex tabular data.
💡 Pro Tip: Don't use deep learning for small tabular datasets (<10K rows). Gradient boosting (XGBoost, LightGBM) will outperform and train faster.
⚠️ Common Mistake: Not normalizing inputs before training neural networks. This leads to slow convergence and poor performance.
Real Interview Question (Google): "When would you choose a deep neural network over gradient boosting for a tabular dataset? What are the trade-offs?"
Practice on Thita.ai: Master Deep Learning →
Real-World Case Studies: Learning from Production Systems
Studying how top companies solve data science problems is invaluable. Here are 12 production systems to learn from:
Netflix Recommendation System
Architecture: Collaborative filtering + content-based filtering, matrix factorization, deep learning for personalization Key Decisions: Hybrid recommender combining multiple signals, handling cold start with content features, A/B testing every change Scale: 200M+ subscribers, billions of ratings, 80% of viewing from recommendations Trade-offs: Accuracy vs diversity (filter bubble), real-time vs batch recommendations, explainability vs performance
Uber Demand Forecasting
Architecture: Time series forecasting per city zone, ARIMA + LSTM ensemble, external features (weather, events) Key Decisions: Hierarchical forecasting (city → zone), handling seasonality and trends, real-time adjustments Scale: Forecasts across 10,000+ cities globally, 15-minute granularity Trade-offs: Forecast horizon vs accuracy, global model vs local models, computational cost vs precision
Meta Newsfeed Ranking
Architecture: Multi-stage ranking (candidate generation → ranking → diversification), gradient boosted trees, deep learning Key Decisions: Predict engagement (likes, comments, shares), balancing relevance and diversity, personalization at scale Scale: 3B+ users, trillions of ranking decisions daily Trade-offs: Engagement vs user wellbeing, filter bubbles vs discovery, model complexity vs latency
Amazon Product Search & Recommendations
Architecture: Query understanding + retrieval + ranking, behavioral signals, collaborative filtering Key Decisions: Semantic search with embeddings, personalized ranking, frequent pattern mining for "bought together" Scale: 350M+ products, billions of searches monthly Trade-offs: Relevance vs revenue, personalization vs privacy, speed vs accuracy
Spotify Music Recommendations
Architecture: Collaborative filtering (user-user, item-item), NLP on metadata, audio feature extraction, deep learning Key Decisions: Discover Weekly (exploration), Daily Mix (exploitation), handling new artists/songs (cold start) Scale: 500M+ users, 100M+ songs, weekly personalized playlists Trade-offs: Popular vs niche recommendations, artist diversity vs user preference, explicit vs implicit feedback
Airbnb Dynamic Pricing
Architecture: Demand prediction + price elasticity modeling, gradient boosting, personalized price tips Key Decisions: Feature engineering (location, seasonality, amenities), handling sparse data, balancing host and guest interests Scale: 7M+ listings globally, dynamic pricing suggestions updated daily Trade-offs: Booking rate vs revenue, host autonomy vs algorithmic optimization, short-term vs long-term revenue
Google Search Quality
Architecture: Query understanding, document ranking (PageRank + content signals + behavioral signals), BERT for semantic understanding Key Decisions: Relevance vs diversity, handling ambiguous queries, personalization vs consistency Scale: 8.5B searches per day, indexing trillions of pages Trade-offs: Click-through rate vs user satisfaction, short-term engagement vs long-term trust
LinkedIn People You May Know (PYMK)
Architecture: Graph algorithms (2nd/3rd degree connections), collaborative filtering, feature engineering from profile data Key Decisions: Balancing network size and quality, spam prevention, diversification across industries/roles Scale: 900M+ members, billions of connection suggestions Trade-offs: Connection quantity vs quality, privacy vs recommendation accuracy, computational cost vs freshness
DoorDash Delivery Time Prediction
Architecture: Ensemble of gradient boosting + neural networks, real-time features (traffic, restaurant busyness) Key Decisions: Multi-task learning (prep time + delivery time), handling outliers, continuous model retraining Scale: Millions of deliveries daily across thousands of cities Trade-offs: Prediction accuracy vs latency, conservative estimates vs aggressive promises, global vs restaurant-specific models
Stripe Fraud Detection
Architecture: Real-time classification with gradient boosting, anomaly detection, network analysis for fraud rings Key Decisions: Precision vs recall trade-off (blocking legitimate transactions vs fraud losses), explainability for disputes Scale: Processing billions in payments, sub-100ms latency requirement Trade-offs: False positives (customer friction) vs false negatives (fraud losses), model complexity vs interpretability
Pinterest Visual Search
Architecture: CNN for image embeddings, approximate nearest neighbor search (FAISS), multi-modal understanding Key Decisions: Object detection + embedding generation, handling style vs content similarity, cross-modal search (text → image) Scale: 400M+ users, billions of pins, visual search across all images Trade-offs: Search speed vs accuracy, style similarity vs content similarity, computational cost of embedding generation
Twitter Timeline Ranking
Architecture: Two-stage ranking (candidate selection → heavy ranking), engagement prediction, recency vs relevance Key Decisions: Predicting positive engagement (replies, retweets) not just clicks, balancing chronological and algorithmic Scale: 400M+ users, millions of tweets per minute Trade-offs: Engagement vs misinformation, echo chambers vs discovery, real-time vs batch processing
How to Learn Data Science Effectively with Thita.ai
Knowing the roadmap is just the start. Deliberate practice with real-time feedback is what converts knowledge into interview success and production capability.
Here's how Thita.ai's AI-powered platform accelerates your data science mastery:
1. Follow the Structured Learning Path
Visit Thita.ai's Data Science Learning Path to see all 12 patterns with 72 subpatterns, organized exactly as outlined in this roadmap. Progress through topics systematically with clear prerequisites.
2. Solve 287 Real Data Science Problems
Access 287 curated DS problems covering:
- Data manipulation and preprocessing challenges
- Statistical inference and hypothesis testing
- Machine learning model building and optimization
- Time series forecasting scenarios
- Business analytics and metrics problems
- Production ML system design
3. Get AI Coaching in Real-Time
- Ask the AI coach to explain any concept from the roadmap
- Discuss trade-offs: "Should I use random forest or gradient boosting for this problem?"
- Get hints when stuck on statistical problems
- Understand multiple solution approaches
- Learn at your own pace with personalized guidance
4. Practice Data Science Interviews
- Simulate real DS interviews with an AI interviewer
- Get evaluated on technical depth, statistical reasoning, and business sense
- Receive detailed feedback on your modeling approaches
- Practice explaining complex concepts to non-technical stakeholders
- Build confidence before real interviews
Pro Tip: Use Interview Mode for realistic pressure practice, then switch to Learn Mode for deep conceptual exploration.
Start Practicing DS Interviews →
5. Track Your Progress Across 12 Patterns
- See your mastery across all 12 core patterns
- Identify knowledge gaps and focus study time
- Build momentum with consistent learning
- Monitor improvement over time
- Get personalized recommendations on what to study next
Your 12-Week Action Plan: From Zero to Data Science Mastery
Don't just read — take action with this structured 12-week plan that maps to the 12 patterns:
Week 1: Data Manipulation & Preprocessing
Focus: Pattern 1 - Data Manipulation & Preprocessing
Master the foundation of all data work.
Topics to Cover:
- Data cleaning and missing value handling
- Outlier detection and treatment
- Data transformation and scaling
- Pandas operations (groupby, merge, pivot)
Practice Tasks:
- Clean a messy real-world dataset
- Implement multiple imputation strategies
- Compare scaling methods for different algorithms
- Handle outliers in multiple ways and compare impact
Goal: Solid pandas skills and data preparation expertise
Week 2: Statistics & Probability Fundamentals
Focus: Pattern 2 - Statistics & Probability Fundamentals
Build the mathematical foundation for all inference.
Topics to Cover:
- Descriptive statistics and distributions
- Probability theory and common distributions
- Confidence intervals and hypothesis testing
- p-values and statistical significance
Practice Tasks:
- Calculate confidence intervals manually
- Conduct various hypothesis tests (t-test, chi-square)
- Interpret p-values correctly in business context
- Identify appropriate statistical tests for different scenarios
Goal: Deep understanding of statistical inference
Week 3: Exploratory Data Analysis
Focus: Pattern 3 - Exploratory Data Analysis (EDA)
Learn to extract insights before modeling.
Topics to Cover:
- Univariate and bivariate analysis
- Correlation analysis and visualization
- Distribution analysis and transformation
- Pattern recognition in data
Practice Tasks:
- Perform comprehensive EDA on multiple datasets
- Create compelling visualizations that tell stories
- Identify data quality issues through EDA
- Extract business insights from exploratory analysis
Goal: Systematic EDA process and insight extraction
Week 4-5: Supervised Learning (Regression & Classification)
Focus: Patterns 4-5 - Supervised Learning - Regression & Classification
Master predictive modeling fundamentals.
Topics to Cover:
- Linear and logistic regression
- Tree-based models (decision trees, random forest)
- Gradient boosting (XGBoost, LightGBM)
- Evaluation metrics for regression and classification
Practice Tasks:
- Build regression models for continuous prediction
- Implement classification for binary and multi-class problems
- Compare different algorithms on same dataset
- Understand when to use which algorithm
Goal: Production-ready supervised learning skills
Week 6: Model Selection & Validation
Focus: Pattern 6 - Model Selection & Validation
Ensure your models generalize to unseen data.
Topics to Cover:
- Train-validation-test splits
- Cross-validation techniques
- Bias-variance tradeoff
- Hyperparameter tuning
Practice Tasks:
- Implement k-fold cross-validation
- Diagnose overfitting vs underfitting
- Tune hyperparameters with grid/random search
- Compare models statistically
Goal: Rigorous model validation expertise
Week 7: Feature Engineering & Selection
Focus: Pattern 7 - Feature Selection & Dimensionality Reduction
Optimize your feature space for better models.
Topics to Cover:
- Feature importance analysis
- Filter, wrapper, and embedded methods
- PCA for dimensionality reduction
- t-SNE and UMAP for visualization
Practice Tasks:
- Engineer features for real-world problems
- Apply PCA and interpret components
- Use SHAP values for feature importance
- Implement recursive feature elimination
Goal: Advanced feature engineering skills
Week 8: Unsupervised Learning
Focus: Pattern 8 - Unsupervised Learning
Find patterns without labels.
Topics to Cover:
- K-means and hierarchical clustering
- DBSCAN for arbitrary cluster shapes
- Anomaly detection methods
- Association rules mining
Practice Tasks:
- Segment customers with clustering
- Detect anomalies in transaction data
- Implement market basket analysis
- Evaluate clustering quality with multiple metrics
Goal: Unsupervised learning mastery
Week 9: Business Analytics & Metrics
Focus: Pattern 9 - Business Analytics & Metrics
Connect data science to business impact.
Topics to Cover:
- KPI analysis and metric design
- Customer segmentation and LTV
- Churn prediction
- Recommendation systems
Practice Tasks:
- Design metrics for a business problem
- Calculate customer lifetime value
- Build a churn prediction model
- Implement collaborative filtering
Goal: Business-focused data science skills
Week 10: Time Series Analysis
Focus: Pattern 10 - Time Series Analysis
Model temporal data and forecast the future.
Topics to Cover:
- Time series decomposition
- ARIMA modeling
- Prophet for forecasting
- Time series cross-validation
Practice Tasks:
- Forecast demand for a business
- Decompose time series into components
- Build ARIMA models from scratch
- Compare forecasting methods
Goal: Production-ready forecasting skills
Week 11: Natural Language Processing
Focus: Pattern 11 - Natural Language Processing
Process and understand text data.
Topics to Cover:
- Text preprocessing pipelines
- TF-IDF and word embeddings
- Sentiment analysis
- Topic modeling with LDA
Practice Tasks:
- Build sentiment classifier
- Implement text similarity search
- Extract topics from document collection
- Create text feature representations
Goal: NLP fundamentals for DS applications
Week 12: Deep Learning & Interview Prep
Focus: Pattern 12 - Deep Learning Fundamentals + Mock Interviews
Complete your journey and prepare for interviews.
Topics to Cover:
- Neural network fundamentals
- CNNs and RNNs
- Transfer learning
- When to use deep learning vs traditional ML
Practice Tasks:
- Build neural networks for tabular data
- Implement image classification with CNNs
- Study all 12 case studies deeply
- Complete 10+ mock interviews on Thita.ai
- Review 25+ common interview questions
Goal: Interview confidence and complete DS mastery
Pro Tips from Data Scientists at Top Companies
From a Meta Data Scientist:
"Product sense is as important as technical skills. When you present findings, always connect them to business impact. We don't hire people who just run models — we hire problem solvers who use data."
From a Google Research Scientist:
"Master statistics deeply. You can't build reliable ML systems without understanding confidence intervals, hypothesis testing, and experimental design. Surface-level knowledge gets exposed in interviews."
From a Netflix ML Engineer:
"A/B testing is critical. Every feature we ship is tested. Know how to design experiments, calculate sample sizes, handle multiple testing, and interpret results conservatively."
From an Uber Data Scientist:
"Time series is everywhere at Uber — demand forecasting, ETA prediction, pricing. Master ARIMA, understand seasonality, and know when simpler models (exponential smoothing) beat complex ones."
From an Airbnb Analytics Lead:
"Causal inference matters more than prediction in many business contexts. Understand the difference between correlation and causation. Learn propensity score matching, difference-in-differences, and instrumental variables."
Common Mistakes to Avoid
1. Jumping to Modeling Without EDA Always explore your data first. You'll often find data quality issues, unexpected patterns, or insights that inform your modeling approach.
2. Using Accuracy for Imbalanced Datasets For fraud detection (1% fraud), a model that predicts "no fraud" for everything gets 99% accuracy but is useless. Use precision, recall, F1-score, or AUC-ROC instead.
3. Data Leakage via Improper Splitting Fit scalers, imputers, and encoders only on training data, then transform validation/test data. Fitting on all data leaks information.
4. Ignoring the Bias-Variance Tradeoff Complex models (deep learning, large ensembles) have low bias but high variance. Simple models (linear regression) have high bias but low variance. Choose based on your data size and problem.
5. Not Checking Statistical Assumptions Linear regression assumes linearity, homoscedasticity, normality of residuals, and independence. Violated assumptions invalidate your model.
6. Confusing Correlation with Causation Ice cream sales correlate with drowning deaths (both peak in summer), but ice cream doesn't cause drowning. Use causal inference methods for causal claims.
7. Over-relying on Feature Importance from Trees Tree-based feature importance is biased toward high-cardinality features. Use permutation importance or SHAP values for more reliable results.
8. Not Considering Model Interpretability Black-box models (deep learning, large ensembles) may perform better but are harder to debug and explain. For regulated industries (finance, healthcare), interpretability matters.
25+ Commonly Asked Data Science Interview Questions
Here are real interview questions from Google, Meta, Amazon, Netflix, Uber, and other top tech companies:
Statistics & Probability
- Explain the Central Limit Theorem and why it matters for data science.
- What's the difference between Type I and Type II errors? Give a business example.
- How would you calculate the required sample size for an A/B test?
- Explain p-values. What does a p-value of 0.03 mean?
- When would you use a t-test vs a z-test vs a chi-square test?
Machine Learning Fundamentals
- Explain the bias-variance tradeoff with a real example.
- How does random forest reduce overfitting compared to a single decision tree?
- What's the difference between bagging and boosting?
- When would you use precision vs recall as your primary metric?
- Explain gradient descent. Why do we use mini-batch instead of full batch?
Model Evaluation & Validation
- How do you detect and prevent data leakage?
- Explain cross-validation. When would you use stratified K-fold?
- Your model has great training accuracy but poor test accuracy. What's wrong?
- How do you handle imbalanced datasets?
- What metrics would you use for a recommendation system?
Applied Problems
- Design an experiment to test whether a new feature increases user engagement. (Meta)
- Daily active users dropped 10% yesterday. How do you investigate? (Spotify)
- Build a model to predict which users will churn in the next 30 days. (Netflix)
- How would you detect credit card fraud in real-time? (Stripe)
- Design a demand forecasting system for food delivery. (DoorDash)
Business & Product Sense
- How would you measure the success of Facebook's newsfeed?
- Uber wants to predict surge pricing. What features would you use?
- You're launching a new product feature. How do you A/B test it?
- How would you build a recommendation system for an e-commerce site?
- Explain a data science project you've done in terms a CEO would understand.
Advanced Topics
- When would you use deep learning vs gradient boosting for tabular data?
- Explain how LSTM solves the vanishing gradient problem in RNNs.
- How does TF-IDF work and why is it better than Bag of Words?
- What are the assumptions of linear regression? How do you check them?
- Explain Principal Component Analysis. How do you choose the number of components?
Practice these on Thita.ai with AI interviewer feedback and detailed solutions!
Conclusion & Next Steps
Stop the scattered learning — start with structure.
These 72 data science topics organized into 12 core patterns (with 287 practice problems and 12 real-world case studies) are your complete roadmap to mastering data science in 2026.
The reality: Most aspiring data scientists waste months watching random YouTube tutorials and reading scattered blog posts without a clear path. The top 1% follow a structured roadmap with prerequisites and practice with real problems.
Which group will you join?
Start your journey today:
- Open Thita.ai's Data Science Learning Path → — Follow the structured curriculum with 287 problems across 12 patterns
- Practice with AI Coach → — Get personalized explanations and guidance on any DS topic
- Do Mock DS Interviews → — Simulate real interviews and get detailed feedback
- Follow the 12-Week Plan — Structured learning from foundations to mastery
Your next data science interview will feel like a roadmap you've already traveled.
Ready to Master Data Science?
Join thousands of engineers who've landed offers at Meta, Google, Netflix, Uber, and more using Thita.ai's structured, AI-powered approach.
Last Updated: January 8, 2026
Related Articles
- The Complete AI & LLM Roadmap 2026: 80+ Topics to Master Modern AI Engineering
- The 90 DSA Patterns That Cover 99% of Coding Interviews
- The Ultimate System Design Roadmap: 120 Topics to Master Every Interview
- AI Interview Practice: Free Mock Interview Simulator
- Free ATS Resume Checker: AI-Powered Resume Analyzer
External Resources
- Thita.ai Data Science Learning Path
- Thita.ai AI Mock Interviews
- Thita.ai AI Coach
- Kaggle Datasets
- Papers with Code - ML
- Google Research Blog
Questions? Join our Discord community and connect with thousands of data scientists preparing for FAANG interviews.