Unstructured data includes images, text, sounds, and similar content that does not fit neatly into numerical or categorical fields. Product descriptions, support emails, incident notes, photos, and videos can all be useful project resources, but their quality needs to be checked before relying on them. section 5.2.1 section 6.1.4
Start with a data survey that maps risks rather than attempting a full analysis:
- Locate potentially useful fields and inspect a sample of records for meaningful content. section 5.2.4
- Count empty or NULL records and watch for placeholders such as “.” or “n/a”. section 5.2.4
- For images, render and scan a selection; repeated or unexpected content may warrant investigation. section 5.2.4
select incident_description from incidents limit 5; >"HVAC failure meant loss of cold air for floor 12 to 25", "HVAC intermittent for several hours, users complaining", "HVAC reported as making rattling sound", "Loss of cold air on 5th floor, seems isolated", "Users reporting excessive heat on 5th floor"
Sample incident descriptions show why a field may be useful yet difficult to turn into structured information: locations appear in prose, while incidents may also help explain unusual sensor readings. section 5.2.4
Follow the survey with deeper exploratory data analysis when needed. A median image is not a useful counterpart to a median customer record. For labeled images, inspect class counts, orientation, positioning, backgrounds, and missing labels; systematic human review or quickly constructed ML can help expose these patterns. section 6.1.4 section 5.2.4
