What is Data Science?

"Reduce churn" is not a question. "Which subscriber segment cancels within 60 days, and why?" is. Most failed DS projects fail here, before a single line of code is written.

Pull from a warehouse, an API, a CSV export, or a live event stream. Know where every column actually came from before you trust it.

Handle missing values, fix types, remove duplicates, then look at distributions and correlations before assuming anything about the data.

Sometimes this is a GROUP BY and a chart. Sometimes it is a trained classification model. The method should match the question, not the other way around.

Would this hold up on next month's data? Is the sample size big enough? Statistics and a proper train/test split live here.

A model nobody understands or trusts changes nothing. The final output is a decision, a dashboard, or a deployed prediction — not a notebook.

Recommendation ranking

The "Because you watched…" row is a model predicting what you are most likely to finish, trained on billions of watch events like the ones in the watch_history table below.

Discover Weekly

A recommendation model comparing your listening patterns against millions of other users with statistically similar taste profiles.

Delivery time prediction

A regression model predicts exact delivery windows from historical order, traffic, and restaurant prep-time data — shown to you before you order.

Dynamic pricing

Surge pricing is a live model balancing rider demand against available driver supply in a specific zone, updated every few minutes.

Demand forecasting

Statistical forecasting models decide how much inventory to stock at each warehouse before a sale event, based on historical demand curves.

Churn prediction

A classification model flags subscribers likely to cancel next month — the exact analysis you will build yourself in Module 47 using StreamPulse.

25 subscribers across 8 countries. Basic, Standard, Premium plans. Signup dates from 2021–2025.

20 titles — Movies and Series across Drama, Sci-Fi, Comedy, Documentary and more. IMDb-style ratings and runtimes.

25 billing records. Active and cancelled subscriptions, monthly price by plan, and a cancel_reason for churned users.

100 watch events — which user watched which title, on what device, for how many minutes, and whether they finished it.

40 star ratings (1–5) with optional review text — some reviews are missing, on purpose, for you to clean later.

Data science is the practice of turning raw, often messy data into decisions through a combination of programming, statistics, and — when the question calls for it — predictive modeling. Data analytics typically focuses on answering "what happened?" using already-clean, structured data, usually through SQL queries, dashboards, and descriptive statistics.

The practical distinction is in scope. A data analyst usually works with data that is already collected, cleaned, and modeled in a warehouse — their job is answering business questions against it quickly. A data scientist frequently works further upstream: cleaning genuinely messy raw data, applying statistical rigor to validate whether a pattern is real, and building predictive models when the business question is "what will happen next" rather than "what happened."

In practice these roles blend heavily, especially at smaller companies. The clearest signal of a data science task versus an analytics task: if the deliverable is a trained model or a statistically validated experiment result, it is data science. If it is a report or dashboard on existing clean data, it is analytics.

Six stages. First, frame the question precisely — "reduce churn" is not answerable, but "which subscriber segment cancels within 60 days, and why" is. Second, collect the data from wherever it actually lives — a warehouse, an API, a stream of events. Third, clean and explore it: handle missing values, fix data types, remove duplicates, then look at distributions and correlations before assuming anything.

Fourth, analyze or model — sometimes this is a groupby and a chart, sometimes a trained classification model; the method should match the question's complexity, not the analyst's preference. Fifth, validate: would this hold up on next month's data, is the sample size large enough, does a statistical test support the conclusion. Sixth, communicate and ship — a model or insight nobody understands or acts on has changed nothing.

The most common project failure is skipping stage one — jumping straight to modeling before the actual business question is nailed down — or skipping stage five, presenting a pattern as fact without checking whether it is statistically meaningful.

Because a model — or any statistical conclusion — is only as trustworthy as the data it is built on. Real-world data arrives with missing values (a delivery_date that never got filled in because an order is still in transit), duplicate records (the same event logged twice by a retrying system), inconsistent formats (dates as strings in three different layouts), and outliers that may be real edge cases or simple data entry errors.

If these issues are not addressed first, every downstream step inherits the error. A churn model trained on data where 15% of cancellation dates are wrong will confidently learn the wrong pattern — and the model's high accuracy on that broken data gives false confidence rather than a warning sign.

Industry estimates commonly put data cleaning and preparation at 60–80% of a data scientist's actual time. This is not an inefficiency to be optimized away — it is the actual foundation the rest of the discipline depends on, which is why this course dedicates an entire section (Section 5) and the wrangling section that follows it (Section 6) to exactly this.

A cohort retention analysis. Suppose StreamPulse wants to know how subscriber retention differs between users who signed up in January versus March. This requires no machine learning model at all: it requires grouping subscribers by signup month (a cohort), calculating what percentage of each cohort is still active at 30, 60, and 90 days, and visualizing the resulting retention curves.

The output — a clear chart showing that the March cohort retains 15% better than January, likely correlated with a pricing change that took effect in February — is a fully legitimate, high-value data science deliverable built entirely from pandas, groupby aggregation, and a well-chosen chart. This exact analysis is covered as a full case study in Module 51.

Machine learning is one tool in the data science toolbox, appropriate when the goal is genuinely predictive — "will this specific user churn" rather than "how did this group behave historically." Reaching for a model when a groupby and a chart would answer the question just as well is a common junior mistake, not a sign of rigor.

Python is the dominant language, with pandas for data manipulation, NumPy for numerical computation, matplotlib and seaborn for visualization, scipy for statistical tests, and scikit-learn for classical machine learning models — all covered progressively through this course. SQL remains essential for pulling data out of a warehouse before any of the Python tooling gets involved.

Beyond the core libraries, a typical workflow also touches Jupyter notebooks or an IDE for exploration, Git for version-controlling analysis code, and increasingly cloud notebook environments (Google Colab, Databricks, SageMaker) when the data or compute needs exceed a laptop. This course's live in-browser Python playground gives you the pandas, NumPy, matplotlib, and general Python experience directly — the same code you write here runs unchanged in a real Jupyter notebook or production script.

For deep learning specifically — image, text, and generative model work — PyTorch and TensorFlow dominate, but that is a distinct specialization covered in this platform's separate AI/ML track, not the classical, statistics-grounded data science covered here.

Cause: A typo in a variable name — 'user' was typed instead of 'users' (the DataFrame preloaded in this playground). Python is case- and spelling-exact: a variable must match exactly how it was created, with no fuzzy matching. This is the single most common error beginners hit in their first week of Python.

Fix: Check the exact name of the variable you are trying to use — in this playground, the five preloaded DataFrames are users, titles, subscriptions, watch_history, and ratings (all lowercase, exact spelling). Click 'Tables' in any playground to see the full list with their column names.

Cause: You tried to access a column that does not exist under that exact name — here 'First_Name' was typed instead of 'first_name'. Unlike SQL, pandas column names are case-sensitive and must match exactly, including underscores versus spaces.

Fix: Run users.columns to see the exact column names available, or users.head() to see them as a table header. Copy the exact spelling and case from there — do not guess or assume a column is named the way you would expect.

Cause: You tried to combine a text value and a number directly, for example 'Age: ' + 25. Python will not silently convert the number to text the way a spreadsheet formula might — it requires you to be explicit about the conversion.

Fix: Wrap the number in str() to convert it first: 'Age: ' + str(25) — or better, use an f-string which handles the conversion automatically: f'Age: {25}'. F-strings (prefixing a string with f) are the standard, more readable way to combine text and values in modern Python.

Cause: A typo in a method or attribute name — 'shpae' instead of 'shape'. Python raises an AttributeError (not a syntax error) because 'shpae' is a perfectly valid variable name as far as the parser is concerned — it simply does not exist as something you can call on a DataFrame.

Fix: Check the spelling carefully — the error message shows you exactly what was typed, so compare it letter by letter against what you meant. Common correct DataFrame attributes: .shape (no parentheses, it's a property), .columns, .dtypes, .head(), .describe(). A code editor with autocomplete catches most of these before you even run the code.

Cause: Python uses indentation (whitespace at the start of a line) to define code blocks — there is no equivalent of curly braces {} like in JavaScript or Java. This error means a line has more or less leading whitespace than Python expects at that point, often from mixing tabs and spaces, or accidentally indenting a line that should be at the top level.

Fix: Make sure every line inside the same block (like inside an if statement or a for loop) uses exactly the same indentation — the convention is 4 spaces per level, never tabs. Most code editors can be configured to insert spaces when you press Tab, which prevents this error entirely.

Using the preloaded StreamPulse DataFrames, write one line of pandas that returns the average imdb_rating for each genre in the titles table, so you can see which genre StreamPulse's catalog rates highest.