Applied Data Science Capstone Project
Applied Data Science Capstone Project
The capstone domain tests end-to-end execution: framing a question, acquiring data, wrangling, exploratory analysis, modeling or clustering, evaluation, and presenting a reproducible result. The IBM capstone commonly uses location data and the Foursquare API to compare neighborhoods, so expect questions on API use, feature engineering, clustering, and communication.
Core rules
- Start with a stakeholder question and explicit success criteria.
- Document data provenance and respect API terms, rate limits, and licensing.
- Keep the project reproducible: ordered notebooks or scripts, pinned dependencies, fixed random seeds, and relative paths.
- Let exploratory analysis shape features; do not hunt for p-values after seeing results.
- Establish a baseline before a complex model.
- Deliver a GitHub repository, a clear notebook or report, and a presentation with limitations.
Worked example
Cluster neighborhoods by venue mix. Reasoning: geocode neighborhoods, query nearby venues, one-hot encode venue categories, and run k-means. Choose k using silhouette score plus interpretability. Profile each cluster by its most common venue types, then map clusters with folium. Evaluation is partly quantitative and partly business-facing: a cluster that is statistically stable but meaningless to stakeholders is not a useful deliverable.
Traps
- Leakage between training and evaluation data.
- Unreproducible scraping or API calls without saved raw data.
- Overclaiming causal impact from an observational analysis.
- Committing restricted data or credentials.
- Burying limitations instead of stating them.
- Submitting a notebook whose cells were run out of order.