10 Feature Engineering Techniques (Pune DS, 2026)
The short version
The list
- 1
One-hot encoding for categorical features
Convert categorical columns to multiple binary columns: `pd.get_dummies(df, columns=['city'])` creates city_Pune, city_Mumbai, city_Delhi. Use drop_first=True to avoid multicollinearity with linear models. Common gotcha: high-cardinality columns (10,000+ unique values) blow up dimensionality.
Why it matters: Asked at ~85% of Pune data scientist rounds. Foundation technique; expected to know cold.
Best for: Low-cardinality categoricals; tree-based + linear models.
- 2
Label encoding for ordinal categoricals
Map ordered categories to integers preserving order: low=0, medium=1, high=2. Use sklearn's OrdinalEncoder with explicit category order. Don't use LabelEncoder for input features (it doesn't handle unseen values cleanly + assumes meaningless integer ordering).
Why it matters: Asked at ~60% of Pune rounds. Common follow-up: when to use ordinal vs one-hot encoding for the same column.
Best for: Ordered categories (low/med/high, S/M/L/XL, never/rarely/often).
- 3
Target encoding for high-cardinality categoricals
Replace categorical values with the target's mean for each category: cities replaced by mean churn rate per city. Critical: compute on training set only + apply to test set to avoid leakage. Use smoothing for low-sample categories (Bayesian target encoding) to prevent overfitting.
Why it matters: Asked at ~45% of Pune product company + analytics consultancy rounds. Senior-fresher differentiator over one-hot.
Best for: High-cardinality categoricals; tree-based models; classification + regression.
- 4
Scaling: StandardScaler vs MinMaxScaler
StandardScaler: z-score normalisation (mean=0, std=1) — preserves outliers' relative position. MinMaxScaler: [0, 1] range — sensitive to outliers. Use Standard for most cases (linear models, neural networks); MinMax for image data + cases needing bounded range. Tree-based models don't need scaling.
Why it matters: Asked at ~70% of Pune rounds. Walking through when each is appropriate signals understanding.
Best for: Linear models, neural networks, distance-based algorithms (KNN, K-Means).
- 5
Feature interactions + polynomial features
Create combinations: price * quantity = revenue. PolynomialFeatures generates all degree-N combinations automatically. Useful when domain knowledge suggests interactions matter (income × education for credit risk). Caution: explodes feature count + risk overfitting.
Why it matters: Asked at ~35% of Pune rounds. Domain-knowledge application signal.
Best for: Linear models capturing non-linear relationships; small feature sets.
- 6
Binning continuous variables
Convert age into buckets (0-18, 18-30, 30-50, 50+) via pd.cut() or pd.qcut() (equal-frequency bins). Useful for: non-linear relationships in linear models, monotonic constraints in tree models, business-rule interpretability. Trade-off: information loss.
Why it matters: Asked at ~30% of Pune rounds, especially BFSI + risk-modelling contexts where binning supports regulatory interpretability.
Best for: Non-linear relationships; regulatory contexts; interpretability requirements.
- 7
Date/time feature extraction
From a timestamp extract: year, month, day, dayofweek, hour, is_weekend, is_holiday, days_since_event, time_since_signup. Captures seasonality + recency patterns. Use sin/cos transforms for cyclical features (hour-of-day, day-of-year) so model knows 23h is close to 0h.
Why it matters: Asked at ~50% of Pune rounds. Universal applicability + creative-application signal.
Best for: Any time-series data; user behaviour patterns; seasonal effects.
- 8
Handling missing values: imputation strategies
Simple: mean/median/mode imputation per column. Better: KNN imputation (sklearn.impute.KNNImputer) — uses similar rows' values. Better still: tree-based imputation via IterativeImputer or model-based (predict the missing value from other features). Always create a 'was missing' flag column — missingness itself often carries information.
Why it matters: Asked at ~65% of Pune rounds. The 'was missing' flag is the senior-fresher discriminator.
Best for: Real-world messy datasets (~always required).
- 9
Log transformation for skewed distributions
`np.log1p(x)` for right-skewed columns (income, prices, counts). Compresses large values + spreads small values; helps linear models that assume normality. log1p over log because it handles zeros gracefully. Reverse with expm1() for predictions.
Why it matters: Asked at ~40% of Pune rounds. Walking through 'before histogram → after histogram' signals real EDA experience.
Best for: Right-skewed continuous features; income / prices / counts.
- 10
Feature selection: filter, wrapper, embedded
Filter methods: correlation/chi-squared/mutual information (fast, model-agnostic). Wrapper methods: recursive feature elimination with a model (slow, accurate). Embedded methods: tree-based feature importance, L1 regularisation coefficients (good middle ground). Use filter for initial screening + embedded for final selection on tree-based models.
Why it matters: Asked at ~45% of Pune product company + analytics consultancy rounds. Demonstrates pragmatic ML workflow understanding.
Best for: High-dimensional datasets; interpretability requirements; model latency optimisation.
How we built this list
Techniques ranked by Pune data scientist + ML engineer interview-frequency from Archer Infotech's placement-cell debriefs over 2024-2026 cycles + production-use prevalence at Pune analytics consultancies (ZS Associates, Tiger Analytics, Mu Sigma) + product company ML teams (Druva, Helpshift, BrowserStack ML, Persistent product). Focuses on tabular data feature engineering (most common Pune ML work); deep learning feature engineering (embeddings, augmentation) covered separately in ML-engineering tracks.
FAQs
Common questions about feature engineering techniques.
How much feature engineering do I need for Pune Data Analyst vs Data Scientist roles?
Data Analyst: foundation 3 (one-hot encoding, scaling, date/time extraction) at conceptual depth. Data Scientist: all 10 at working depth + ability to walk through trade-offs. ML Engineer: + deeper production patterns (feature pipelines, feature stores, versioning). The bar rises sharply with role tier.
Should I learn manual feature engineering or use automated tools (AutoML, featuretools)?
Manual first — automated tools are excellent productivity boosters but only after you understand what they're doing. Pune interview rounds probe your understanding of why a transformation matters, not just whether you applied it. Learn manual techniques on 5-10 real datasets, then use featuretools / Featurewiz / Open-FE for production.
What's the most-failed feature engineering question at Pune data scientist interviews?
Data leakage from target encoding + statistics computed on the full dataset. Candidates compute mean/median/target-encoding using the entire dataset including test rows. Correct pattern: fit transformer on training set only, then transform train + test separately. This signals understanding of train/test boundary integrity.
Should I use feature engineering with deep learning models too?
Less for unstructured data (images, audio, text — deep learning learns features automatically). Still useful for tabular data with deep learning — most kaggle wins on tabular use thoughtful feature engineering + a deep model. For LLM applications: prompt engineering and structured output design replace traditional feature engineering at the application layer.