Search, Recommendations and Ranking with Machine Learning

Search, Recommendations and Ranking with Machine Learning

This is a description of a project type we work on: helping people find the right thing in a large collection, using matching and ranking that improve with use.

The Challenge

Search returned results that were technically matches but practically useless: the right word in the wrong context, popular items crowding out relevant ones, and no way for the business to see which queries were failing its users.

Our Approach

Discovery examined real queries and what people clicked afterwards, because that behaviour is the only honest relevance signal available. The content itself was audited for the fields that should influence ranking: titles, categories, attributes, freshness and availability.

Solution

A search service combining structured filters with text matching, ranked by signals the business can reason about: relevance, availability, popularity and recency. Where the collection and the behaviour justify it, a learned ranking model is trained on the recorded interactions, with the simple rules kept as the baseline it must beat.

Technology

An indexing approach chosen for the size and shape of the collection, a service layer exposing search and suggestions to both web and mobile clients, and logging that records queries, results and outcomes for evaluation.

Implementation

Released behind a comparison: both rankings served, outcomes measured, and the new ranking promoted only when it demonstrably improved on the old one.

Results

Query success, abandonment and click-through are measured from the logging, and published only where verified and approved.

Search that misses? Tell us about your project and we will start with your queries.

Project details in this article are deliberately general. Client names, verified results and references are published only with the client's approval.