Profiling Metrics And Engine Output
After the source is accepted, the tool builds a profile from the CSV and optional descriptor context.
Dataset Overview
The overview answers the first operational questions:
- How many rows and columns are present?
- How many cells are missing?
- What percentage of rows are duplicates?
- How many columns are numeric, categorical, datetime, identifier, text, boolean, or unknown?
- How many columns carry quality flags?
- How many potential outlier cells were detected?
These values are not the final interpretation. They are triage signals that tell the team where to inspect next.
Column Type Inference
Every column receives an inferred type. Type inference allows the profile to choose appropriate statistics. A numeric field can receive min, max, mean, standard deviation, quartiles, median, p95, and outlier counts. A datetime field can receive min date, max date, date range, and unique count. Text fields emphasize length and uniqueness. Categorical fields emphasize top values and cardinality.
Range, mean, median, p95, outliers
Skew, units, impossible values
Check bounds and business meaning
Top values, unique count, rare levels
Mixed labels or unexpected categories
Normalize categories before analysis
Min date, max date, date range
Timezone, stale records, gaps
Confirm period and update cadence
Unique percent and duplication pattern
False keys or join collisions
Validate key rules before merging
Missing Values
Missingness is tracked at dataset and column level. The profile can also capture special missing counts when the engine recognizes values that often mean missing, such as blank-like or sentinel values.
Interpret missingness in context:
- small missingness may be acceptable for descriptive analysis
- high missingness can make a column unusable
- systematic missingness can reveal process gaps
- missingness concentrated in one group or time period can bias conclusions
Duplicates
Duplicate rows can mean repeated events, data export issues, or legitimate repeated records. Profiling reports the duplicate count and percentage, but the domain owner must decide whether duplicates are errors.
Outliers And Extremes
Outlier counts highlight records outside expected distribution ranges. Outliers are not automatically wrong. They may be rare but valid cases, data-entry errors, unit problems, or high-value signals.
Quality Flags
Quality flags are short warnings attached to columns. They help reviewers scan for likely issues such as high missingness, high cardinality, constant values, possible identifiers, unusual distributions, or sparse categories.
Use flags to prioritize review. Do not delete or transform data solely because a flag exists.