fewlab: fewest items to label for most efficient unbiased OLS¶

Problem: You have usage data (users × items) and want to understand how user traits relate to item preferences. But you can’t afford to label every item. This tool tells you which items to label first to get the most accurate analysis.

Overview¶

When You Need This¶

You have:

A usage matrix: rows are users, columns are items (websites, products, apps)
User features you want to analyze (demographics, behavior patterns)
Limited budget to label items (safe/unsafe, brand affiliation, category)

You want to run a regression to understand relationships between user features and item traits, but labeling is expensive. Random sampling wastes budget on items that don’t affect your analysis.

How It Works¶

The tool identifies items that most influence your regression coefficients. It prioritizes items that:

Are used by many people
Show different usage patterns across your user segments
Would most change your conclusions if mislabeled

Think of it as “statistical leverage”—some items matter more for understanding user-trait relationships.

Quick Example¶

from fewlab import items_to_label
import pandas as pd

# Your data: user features and item usage
user_features = pd.DataFrame(...)  # User characteristics
item_usage = pd.DataFrame(...)     # Usage counts per user-item

# Get top 100 items to label
priority_items = items_to_label(
    counts=item_usage,
    X=user_features,
    K=100
)

# Send priority_items to your labeling team
print(f"Label these items first: {priority_items}")

For a complete demonstration with performance analysis, see our interactive Jupyter notebook showing ~33% reduction in standard errors compared to random selection.

Installation¶

pip install fewlab

Requires: numpy, pandas

License¶

MIT