Before training a model on two years of support tickets, an analyst counts rows per month, plots ticket length and checks how often each field is empty. The checks reveal a month with half the usual volume and a category that disappears after a system change, both of which would have shaped the model.
The NIST/SEMATECH e-Handbook of Statistical Methods describes EDA as an approach to data analysis that uses a variety of techniques, mostly graphical, to maximize insight into a data set, uncover its structure, detect outliers and anomalies and test underlying assumptions. The handbook names John Tukey’s 1977 book Exploratory Data Analysis as the seminal work.
EDA starts with the file, not its description. When CUBIG examined the UCI Online Retail dataset (November 2026), 24.93% of transaction lines had no customer ID, although the repository’s metadata answers “No” to “Has Missing Values?”. The analysis appears in missing data in sales records.
Common EDA checks:
- counts of rows, distinct values and empty cells for each field
- distributions and outliers, using histograms, box plots or summary statistics
- relationships between fields, such as correlations or cross-tabulations
- changes over time, such as volume by month or a category that appears or disappears
How is exploratory data analysis different from data profiling?
- Data profiling produces standard statistics for each column, such as types, value ranges and missing counts, often automatically.
- Exploratory data analysis is broader and question-driven: it uses those statistics and charts to understand what the data says about a specific problem.
- Profiling can be one step inside EDA, and a profile usually prompts the questions EDA then follows up.
Where does exploratory data analysis break in AI work?
EDA breaks when its findings stay in a notebook. An analyst may notice that a field changed meaning in March or that one region is missing most values, but if the note is not recorded with the data version, the next model trained on the same file can miss the same problem. EDA on the full dataset can also leak test information into decisions about which features to keep.
EDA is most useful when it is tied to a task. In CUBIG’s Syntitan, Ready for Raw Data runs a structured first pass, a diagnosis on six axes such as integrity and context, and records the result with the Version. Whether the data suits a particular model or decision still needs questions that only that task can answer.