Balancing Imbalanced Multi-Label Data | Dataford Interview Questions

Given an imbalanced multi-label dataset of image tags, write a data pipeline to reweight or sample instances for balanced mini-batch training.

Use PostgreSQL and return a deterministic mini-batch of six tagged images. Compute each image's sampling weight from the frequency of its associated tags, excluding images without a valid tag.

Output

One row per selected image, with image_id, image_name, sampling_weight, and batch_rank
Include six images with the highest weights, breaking ties by image_id ascending
Order by batch_rank ascending

Schema

images

Column	Type	Description
image_idPK	INT	Unique image identifier
image_name	VARCHAR(100)	Image file name

image_tags

Column	Type	Description
image_tag_idPK	INT	Unique image-tag relationship identifier
image_id	INT	Referenced image identifier
tag_id	INT	Image tag identifier; null indicates an invalid or missing tag

Your question is Balancing Imbalanced Multi-Label Data. Start with the requirements and the two tables in the Question tab.

Run and submit as often as you like. When you're ready, talk me through your approach or go straight to the code.

Problem