Artificial Intelligence & Data Science · 1.2

Data Scientist

Build predictive models and extract deep insights using statistics, machine learning, and experimentation.

9–12 months7 phases33 skills

Foundations

Mathematicslinear algebra (vectors, matrices, dot products, eigenvalues/eigenvectors, SVD — the foundation of PCA and recommendation systems), calculus (derivatives, partial derivatives, chain rule, gradient — the math behind backpropagation), probability and statistics (Bayes' theorem with worked medical diagnosis example, distributions in depth, MLE and MAP estimation)

Python proficiency

Python proficiencyOOP (classes, inheritance, abstract base classes), decorators for logging and timing, generators for memory-efficient data processing, context managers, type hints, virtual environments with poetry

Data manipulation

Data manipulationadvanced Pandas (MultiIndex, method chaining with .pipe(), .eval() and .query() for performance, chunked reading for large files, memory optimization with .astype('category')), NumPy (vectorized operations, broadcasting, random generation for simulations)

SQL

SQLadvanced queries including recursive CTEs, window functions, query optimization using EXPLAIN ANALYZE, creating indexes for frequently queried columns

Phase 02 · 8h

Exploratory Data Analysis & Feature Engineering

EDA workflow

EDA workflowunivariate analysis (histograms with KDE, box plots, value counts), bivariate analysis (scatter plots with correlation annotations, contingency tables with chi-square), multivariate analysis (correlation matrix heatmaps, pair plots with target coloring, PCA/t-SNE 2D projections for cluster visualization), automated EDA using ydata-profiling (generates comprehensive HTML reports)

Feature engineering

Feature engineeringencoding categorical variables (one-hot with pd.get_dummies(drop_first=True), ordinal encoding with dictionaries, target encoding with sklearn's TargetEncoder using cross-validation to prevent leakage), numerical transformations (log1p for right-skewed data, StandardScaler for mean=0/std=1, MinMaxScaler for [0,1] range), binning with pd.cut() and pd.qcut(), interaction features (price_per_sqft = price / area), date feature extraction (year, month, day_of_week, is_weekend, days_since_event), PolynomialFeatures for non-linear terms

Feature selection

Feature selectioncorrelation analysis (remove one of two features with |r| > 0.9), mutual information (sklearn.feature_selection.mutual_info_classif), Recursive Feature Elimination (RFE wrapping a model), Lasso regularization (L1 drives unimportant coefficients to zero), feature importance from tree-based models (RandomForest.feature_importances_)

Handling imbalanced data

Handling imbalanced dataSMOTE (creates synthetic minority samples using interpolation — from imblearn.over_sampling import SMOTE), RandomUnderSampler, class_weight='balanced' parameter in sklearn models, evaluation with precision-recall curve and F1-score instead of accuracy

Missing data strategies

Missing data strategiesunderstanding missingness patterns (MCAR, MAR, MNAR), simple imputation (mean/median/mode), KNN imputation (sklearn.impute.KNNImputer), MICE (sklearn.impute.IterativeImputer), when to drop rows (<5% missing and MCAR) vs columns (>40% missing and non-critical)

Phase 03 · 16h

Machine Learning

Supervised learning — regression

Supervised learning — regressionlinear regression (review), Ridge (L2 regularization — shrinks coefficients, prevents overfitting with correlated features), Lasso (L1 — drives unimportant coefficients to zero for feature selection), ElasticNet (combines L1+L2), polynomial regression with PolynomialFeatures pipeline; evaluation with MAE, RMSE, R², Adjusted R²

Supervised learning — classification

Supervised learning — classificationlogistic regression (sigmoid function, probability output, decision boundary tuning), decision trees (Gini impurity, information gain, visualize with plot_tree, prone to overfitting without max_depth/min_samples_leaf), Random Forest (bagging ensemble of trees — RandomForestClassifier(n_estimators=100, max_depth=10)), XGBoost (gradient boosting with L1/L2 regularization, native missing value handling, built-in cross-validation with xgb.cv(), feature importance), LightGBM (leaf-wise growth — faster than XGBoost for >100K rows, uses GOSS and EFB), CatBoost (native categorical feature handling, ordered target encoding, minimal hyperparameter tuning), SVM (hyperplane with maximum margin, RBF kernel for non-linear boundaries, use LinearSVC for >100K samples), KNN (majority vote of k nearest neighbors, requires feature scaling)

Unsupervised learning

Unsupervised learningK-Means (elbow method for choosing k, silhouette score for cluster quality), DBSCAN (density-based, no k required, handles outliers), hierarchical clustering (dendrograms with Ward linkage), PCA (retain components explaining 95% variance with PCA(n_components=0.95)), t-SNE and UMAP for 2D visualization of high-dimensional data

Model evaluation

Model evaluationtrain_test_split with stratify, 5-fold cross-validation with cross_val_score, bias-variance tradeoff and learning curves, confusion matrix, precision/recall/F1, ROC-AUC (for balanced datasets) vs precision-recall AUC (for imbalanced), calibration curves and CalibratedClassifierCV

Hyperparameter tuning

Hyperparameter tuningGridSearchCV (exhaustive but expensive), RandomizedSearchCV (samples from distributions), Bayesian optimization with Optuna (define objective function, suggest_int/suggest_float with log=True for learning rates, study.optimize with n_trials=50)

Scikit-learn pipelines

Scikit-learn pipelinesPipeline([('scaler', StandardScaler()), ('model', RandomForestClassifier())]) — ensures preprocessing is applied consistently and prevents data leakage during cross-validation; custom transformers with BaseEstimator and TransformerMixin

Phase 04 · 23h

Advanced ML & Deep Learning

Neural networks

Neural networksperceptrons, activation functions (sigmoid, tanh, ReLU, Leaky ReLU), forward/backpropagation, loss functions (MSE for regression, binary cross-entropy for binary classification, categorical cross-entropy for multi-class), optimizers (SGD, Adam with default lr=0.001), learning rate schedulers (ReduceLROnPlateau)

PyTorch deep-dive

PyTorch deep-divetensors and autograd, building models by subclassing nn.Module, training loops (zero_grad → forward → loss → backward → step), DataLoader with custom Datasets, GPU training (model.to(device)), Dropout and BatchNorm for regularization, saving/loading models with state_dict

CNNs

CNNsconvolution layers, pooling, architecture progression (LeNet → AlexNet → VGG → ResNet with skip connections → EfficientNet with compound scaling), transfer learning (load pretrained ResNet-50, replace final layer, freeze early layers then fine-tune)

RNNs/LSTMs/GRUs

RNNs/LSTMs/GRUssequence modeling for text and time-series, LSTM gates (forget, input, output), bidirectional RNNs, vanishing gradient problem and how LSTMs solve it

NLP foundations

NLP foundationstokenization (word-level, BPE, WordPiece), word embeddings (Word2Vec, GloVe, FastText), TF-IDF for text classification, transformer architecture (self-attention, multi-head attention, positional encoding)

BERT and GPT

BERT and GPTfine-tuning BERT for classification/NER/QA using Hugging Face Transformers (BertForSequenceClassification, Trainer API), GPT architecture for text generation, Hugging Face ecosystem (Tokenizers, Datasets, Hub)

Recommendation systems

Recommendation systemscollaborative filtering (user-based and item-based with cosine similarity), matrix factorization (SVD using Surprise library), content-based filtering (TF-IDF on item descriptions), hybrid approaches

Time-series forecasting

Time-series forecastingARIMA (auto_arima from pmdarima for automatic parameter selection), Prophet (handling yearly/weekly seasonality and holidays), LSTM-based forecasting (sliding window approach in PyTorch), evaluation with MAE, RMSE, MAPE

Phase 05 · 8h

Experimentation & Causal Inference

Advanced experiment design

Advanced experiment designmulti-armed bandits (Thompson Sampling, Epsilon-Greedy — dynamically allocate traffic to better variants), sequential testing (monitoring results without inflating false positive rates), stratified randomization, guardrail metrics that must not degrade

Causal inference

Causal inferenceDifference-in-Differences (parallel trends assumption, regression with interaction term where β₃ is the causal effect), Propensity Score Matching (logistic regression to estimate treatment probability, match treated/control units with similar scores), Instrumental Variables (using a variable that affects outcome only through treatment), Regression Discontinuity (treatment assigned at a threshold — compare outcomes just above/below), Uplift modeling (T-learner and S-learner approaches for individual-level treatment effects)

Causal inference libraries

Causal inference librariesDoWhy (define causal model → identify → estimate → refute) and EconML (heterogeneous treatment effects)

Phase 06 · 7h

MLOps & Production

Model deployment with FastAPI

Model deployment with FastAPIbuild a REST API (POST /predict endpoint), validate input with Pydantic models, containerize with Docker (FROM python:3.12-slim), deploy on AWS ECS Fargate or Google Cloud Run

Experiment tracking with MLflow

Experiment tracking with MLflowlog parameters, metrics, and model artifacts with mlflow.start_run(), compare runs in the MLflow UI, register best models in the Model Registry, transition between Staging and Production stages

Monitoring

Monitoringdetect data drift and model drift using Evidently AI (DataDriftPreset and TargetDriftPreset comparison reports), set up alerting when distributions shift beyond thresholds

Feature stores

Feature storesFeast for managing features — define FeatureViews, store offline features in BigQuery for training, serve online features via Redis for real-time inference, ensure training-serving consistency

Phase 07 · 16h

Capstone

Capstone Project 1

Capstone Project 1End-to-end ML pipeline — churn prediction with EDA, 15+ engineered features, 5+ model comparison, Optuna tuning, MLflow tracking, FastAPI deployment in Docker, Streamlit dashboard for stakeholders

Capstone Project 2

Capstone Project 2Fine-tune BERT for a text classification task (fake news detection or intent classification) using Hugging Face Trainer API, evaluate with precision-recall curves, deploy on Hugging Face Spaces

Capstone Project 3

Capstone Project 3Causal analysis report — apply DiD and propensity score matching to an observational dataset, compare causal estimates to naive correlation, write a 5-page business audience report

Ready to start Data Scientist?

Start this roadmap to track progress and generate courses for each skill. Login required only to begin.