Choose the analysis that matches your data
These tools cover different questions about a dataset, so start with the input and outcome you have in mind. Use Dataset Profiler to inspect inferred types, row and column counts, missingness, uniqueness, statistics, date ranges, top values, and data-quality warnings. Use SQL Query Tool for CSV and JSON when you need to filter, join, group, order, aggregate, or limit data in CSV, TSV, JSON, or JSONL files.
For a relationship between numeric columns, use Correlation Matrix Generator. It compares columns with Pearson, Spearman, or Kendall methods. Correlation describes association; it does not establish causation.
For records that may represent the same entity, use Fuzzy Duplicate Finder. It compares selected text columns with configurable similarity methods and groups likely matches for review. For unusual numeric values without a time-order requirement, use Outlier Detector. For observations that must be interpreted in timestamp order, use Time-Series Anomaly Detector.
Start with profiling or a targeted query
If you do not yet understand the shape or quality of the data, profile the dataset first. Its summary includes inferred column types, completeness, uniqueness, duplicate counts, numeric and text statistics, date ranges, top values, warnings, and per-column details. The profiler can export JSON, CSV, or HTML reports, but it reports findings rather than repairing the source.
When you already know the rows or combinations you need, query selected rows or combinations. The tool loads files as named tables and runs a documented, read-only SQL subset. It does not connect to external databases or modify the source files. Its table preview shows up to 500 rows, while an export can contain the full returned result.
Match the statistical question to the right detector
Choose the method according to the structure of the question, not just the word “anomaly.” The correlation matrix compares numeric columns and can show missing-value handling and significance values. It is useful for examining relationships among variables, not for labeling individual rows as wrong.
Outlier Detector scores unordered numeric rows with IQR, z-score, or modified z-score methods. You can optionally group the calculations by another column and export flagged rows. Time-Series Anomaly Detector instead orders observations by a timestamp and builds rolling or seasonal baselines from prior observations. It reports point-level results while handling duplicate timestamps, gaps, and timezone display; it does not forecast, fill gaps, interpolate, repair, or modify the source data.
A flag is a result for review, not proof that a value is incorrect. The method, grouping, missing values, timestamps, and chosen thresholds all affect how results should be interpreted.
Review possible duplicates before deciding what to keep
Use Fuzzy Duplicate Finder when exact equality is too strict and selected text fields can indicate that records are similar. It groups likely matches and lets you review each group: choose a record to keep, select field values from other records, or mark the group as not a duplicate. Nothing is deleted or merged automatically; the resolved dataset and group report are exported after review.
For a quick first check, check duplicate-row counts with the profiler. That count is useful for initial inspection, but it is not the similarity-based review workflow and does not decide which records represent the same entity.
A practical analysis workflow
- Inspect structure and data quality if the columns, completeness, or value distributions are not yet clear.
- Use the SQL tool to select or combine the relevant records when a targeted subset is easier to analyze.
- Choose the analysis that matches the data shape: correlations for numeric relationships, fuzzy matching for possible duplicate records, unordered outlier scoring for numeric rows, or time-series detection for timestamped observations.
- Review the displayed results, warnings, method, and thresholds before treating a flag or relationship as meaningful.
- Export the available report or result when you need to continue the review. For example, you can export a query result, review flagged outliers, or review time-series anomaly results.