Plain-language explanation.
Data science in practice involves translating business questions into data problems, solving them with statistics and machine learning, and communicating the results to decision-makers. It requires not just technical skills but also communication, problem framing, and understanding of how organisations work. Most data science work is data cleaning, not machine learning.
Core concepts and standard treatment.
Data science project lifecycle: problem framing (translating vague business questions into measurable hypotheses), data discovery and acquisition (what data exists? what additional collection is needed? what are the legal and ethical constraints?), EDA and data cleaning, feature engineering, modelling and evaluation, interpretation and communication, deployment and monitoring, and stakeholder feedback loops. The 80/20 rule: approximately 80% of data science time is data wrangling, cleaning, and EDA — not modelling. Version control: Git/GitHub for code, DVC for data and models, MLflow for experiment tracking.
Deeper theory, debates and edge cases.
Python data science stack: NumPy (array operations), pandas (tabular data manipulation), matplotlib/seaborn/plotly (visualisation), scikit-learn (ML algorithms, pipelines, cross-validation, preprocessing), XGBoost/LightGBM/CatBoost (gradient boosting — typically best performance on tabular data), PyTorch/TensorFlow (deep learning), and FastAPI (model serving). R data science stack: tidyverse (dplyr, ggplot2, tidyr, purrr, readr, stringr, lubridate), Shiny (interactive dashboards), caret/tidymodels (ML pipelines), and R Markdown (reproducible reports). SQL remains the most universally required data science skill — most production data lives in SQL databases.
How it is applied in practice.
Data science team structures: embedded (data scientists in product teams — closer to business context, less peer support); centralised (centre of excellence — consistent standards, research, but slower to serve business partners); federated (both centre and embedded — complex coordination). Stakeholder communication: translating model outputs into business language (not "the F1 score is 0.87" but "the model catches 87% of fraud cases with 13% false positives, saving an estimated $3M annually"). Data science ethics in practice: bias monitoring in production models, documentation of model limitations, and escalation processes for high-stakes algorithmic decisions. The data scientist career path: analyst (SQL, EDA, reporting) to data scientist (ML modelling) to ML engineer (production systems) to research scientist (algorithm development) — or into management, product, or consulting tracks.