Data Science: From a Question to Defensible Evidence
Data science combines a domain question, data analysis and communication to support a decision.
- Author
- absequ
- Published
- Updated
- Category
- AI and data
- data science
- fundamentals
- practical guide
Data science combines a domain question, data analysis and communication to support a decision. A useful result may be a corrected measurement or a clear explanation, without any machine learning. Begin by identifying who will act on the result and what evidence would change their decision.
On this page
Key concepts
Define the unit of observation, the population of interest and how measurements were collected. Exploratory analysis looks for structure, anomalies and questionable assumptions before committing to a model. Keep raw records separate from cleaning decisions so the work can be explained and repeated.
Practical example
Suppose a team wants to understand slow support responses. Use synthetic tickets with creation time, first-response time and channel. Decide whether to measure elapsed hours or business hours, how to handle tickets still unanswered and whether automated replies count. A chart cannot resolve an undefined metric.
For response times of 1, 2, 2, 3 and 42 hours, the mean is 10 hours while the median is 2. Neither number alone describes the long wait. Report the distribution and inspect the 42-hour case before deciding it is a recording error. Compare like periods and channels before claiming the process improved.
Common mistakes
Do not replace missing values with zero without a reason: unknown duration and immediate response mean different things. Duplicates can inflate counts, and a change in the mix of ticket types can change an average without any operational improvement. Observational association does not by itself establish that an intervention caused an outcome.
Practice
Write a short analysis note containing the question, inclusion rules, cleaning decisions, summary, uncertainty and recommended next measurement. Use aggregated or synthetic examples when sharing; avoid exposing ticket text or personal identifiers in charts and exports.