Python for Machine Learning
The 2026 Skills Guide

Python's dominance in machine learning is not accidental. It emerged from a combination of factors: an expressive, readable syntax that lets researchers iterate quickly; a rich scientific computing ecosystem built on NumPy and SciPy; and early adoption by the deep learning community via Theano, which shaped every framework that followed.

Today, every major ML framework — PyTorch, TensorFlow, JAX — exposes a Python API as its primary interface. The HuggingFace ecosystem, LangChain, LlamaIndex, and every LLM toolkit in wide use are Python-first. Knowing Python well is not optional for UK AI engineers — it is the minimum.

What separates candidates at interview is not whether they know Python, but how well they know it. Writing vectorised NumPy operations instead of Python loops. Understanding memory layout (contiguous vs non-contiguous arrays). Writing typed, tested, reviewable ML code rather than notebook-quality scripts. These distinctions matter.

NumPy's ndarray is the universal data container for numerical computation in Python. Every ML framework stores tensors in a format that is directly interoperable with NumPy arrays. Understanding how NumPy works is therefore foundational — not just for data manipulation, but for understanding how PyTorch tensors, Pandas DataFrames, and scikit-learn inputs are stored and processed.

Key NumPy skills for ML engineers:

• Vectorisation — operating on entire arrays at once using NumPy's ufuncs (universal functions), rather than Python loops. A loop over a million-element array is typically 100× slower than the equivalent NumPy operation.
• Broadcasting — NumPy's mechanism for applying operations across arrays of different shapes by implicitly expanding dimensions. Understanding broadcasting eliminates the need for many explicit reshaping operations.
• Array manipulation — reshape(), transpose(), stack(), concatenate(), advanced indexing with boolean masks and integer arrays.
• Linear algebra — np.linalg provides matrix multiplication (@ operator), eigendecomposition, SVD, and norm calculations. Essential for understanding the mathematics behind ML algorithms.
• Random number generation — np.random.Generator (the modern API, not np.random.seed()), reproducible splits, shuffling arrays in sync.

Pandas is used throughout the ML pipeline: exploratory data analysis, feature engineering, data cleaning, train/validation splitting, and results analysis. The skill ceiling is high — it is easy to write working but slow Pandas code that becomes a bottleneck at scale.

Pandas patterns ML engineers must know:

• Vectorised operations over .apply() — .apply() with a Python function iterates row-by-row. For string operations use .str accessor methods; for numeric transformations use arithmetic operators and built-in aggregation methods. Reserve .apply() for genuinely complex transformations with no vectorised alternative.
• GroupBy for feature engineering — .groupby().transform() for group-level aggregates broadcast back to the original DataFrame shape. .groupby().agg() for multiple simultaneous aggregations. Essential for time-series feature engineering and building aggregated training features.
• Efficient merges — understanding the join types (inner, left, right, outer) and their performance characteristics. Using categoricals and sorted merge keys for performance on large DataFrames.
• Memory management — downcasting numeric columns (pd.to_numeric(..., downcast='integer')), using categorical dtype for high-cardinality string columns. A DataFrame of floats that should be int16 can use 4× more memory than necessary.
• Avoiding data leakage — never computing train statistics before the train/test split. Fitting preprocessors on train data only and applying to test via scikit-learn Pipeline.

scikit-learn's consistent estimator API is one of the best-designed interfaces in any ML framework. Understanding it deeply — not just calling .fit() and .predict() — is important both for practical ML work and for code interviews.

• Pipeline — sklearn.pipeline.Pipeline chains transformers and an estimator into a single object that can be cross-validated and hyperparameter-tuned without data leakage. Using ColumnTransformer inside a Pipeline to apply different preprocessing to numeric, categorical, and text features is standard practice.
• Cross-validation — cross_val_score() and cross_validate() for k-fold evaluation. Understanding stratified k-fold for imbalanced classification and TimeSeriesSplit for temporal data.
• Hyperparameter search — GridSearchCV for exhaustive search, RandomizedSearchCV for larger spaces. Understand the refit parameter and how to extract the best estimator.
• Custom transformers — inheriting from BaseEstimator and TransformerMixin to write reusable, Pipeline-compatible transformers. Essential for encapsulating feature engineering logic.
• Model selection metrics — precision, recall, F1, ROC AUC, average precision (PR AUC) for classification; MAE, RMSE, R² for regression. Knowing when accuracy is a misleading metric (class imbalance).