Recommender using Solr for recommendation queries and Mahout to generate the similarity matrix model. This implementation has several unique features:
- Creates models for single actions recommendations and two-action cross-recommendations.
- Both models are available as pre-calculated values (all recommendations for all users) and for online query using potentially recently generated action data not yet incorporated in the models.
- Ingests text files in CSV or TSV form including arbitrary data along with actions, user, and item ids. Turns these 'log' files into a form compatible with Mahout.
- Uses Mahout to pre-calculate the necessary item similarity matrix and all recommendations for all users.
- Turns the Mahout item similarity matrix into text docs in Solr format--in Solr terms this is one doc (item id) per row and the doc contents in a column (similar items stored as item ids).
- Encodes user history in CSV files with one row per user (user id) and items for a given action in columns (item ids).
- Once indexed by Solr (outside the scope of this project) using a user's action history as a query on the item similarity matrix will yield recommendations as an ordered list of results.
Read full article from pferrel/solr-recommender
No comments:
Post a Comment