liveBook · Manning

De-biasing technique

As explained in Building Reliable AI Systems

A method for reducing systematic distortions in data or model behavior so learned patterns better reflect the underlying signal.

Clicking a listing after a search creates an implicit query–listing relevance signal. But users click listings near the top more often, even when they are not more relevant. De-biasing techniques aim to keep this position bias from distorting the semantic relationships learned by embeddings. section 4.2.6

Position bias in click data

A top-ranked listing may receive a click because of its placement rather than its relevance to the query. Treating that click as an unqualified relevance signal can reinforce position bias in the embeddings. section 4.2.6

How Building Reliable AI Systems covers it

How deep it goes
Identifies top-position click bias and the goal of correcting it, but does not specify methods or implementation.
Examples
Uses clicks on search listings to illustrate how prominent placement can distort relevance signals.
What you can do after
Understand why click-trained embeddings need correction for position bias.
How it connects
Links implicit feedback and position bias to embedding models.

Also explained in

Compare how 2 books explain de-biasing technique →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Building Reliable AI Systems. The section links point to the parts of the book it draws on.