Recommendation systems with Mahout: introduction
•
4 recomendaciones•3,847 vistas
An introduction to recommendation systems: types of recommenders, types of data (input), user-user collaborative filtering, item-item collaborative filtering, running Mahout on a single node and on multiple nodes (on Hadoop)
Recomendados
Recomendados
Más contenido relacionado
Más de Adam Warski
Más de Adam Warski (9)
Último
Último (20)
Recommendation systems with Mahout: introduction
- 1. Adam Warski Recommendation systems 5/11/2013 Warszawa-JUG Introduction using Mahout @adamwarski
- 2. Agenda ❖ What is a recommender?! ❖ Types of recommenders, input data! ❖ Recommendations with Mahout: single node! ❖ Recommendations with Mahout: multiple nodes 5/11/2013 Warszawa-JUG @adamwarski
- 3. Who Am I? ❖ Day: coding @ SoftwareMill! ❖ Afternoon: playgrounds, Duplos, etc.! ❖ Evenings: blogging, open-source! ❖ Original author of Hibernate Envers! ❖ ElasticMQ, Veripacks, MacWire! ❖ http://www.warski.org 5/11/2013 Warszawa-JUG @adamwarski
- 4. Me + recommenders ❖ Yap.TV! ❖ Recommends shows to users basing on their favorites! ❖ ❖ ❖ from Yap! from Facebook! Some business rules 5/11/2013 Warszawa-JUG @adamwarski
- 6. Information … ❖ Retrieval! ❖ ❖ given a query, select items! Filtering! ❖ given a user, filter out irrelevant items 5/11/2013 Warszawa-JUG @adamwarski
- 7. What can be recommended? ❖ Items to users! ❖ Items to items! ❖ Users to users! ❖ Users to items 5/11/2013 Warszawa-JUG @adamwarski
- 8. Input data ❖ (user, item, rating) tuples! ❖ ❖ rating: e.g. 1-5 stars! (user, item) tuples! ❖ unary 5/11/2013 Warszawa-JUG @adamwarski
- 9. Input data ❖ Implicit! ❖ ❖ reads! ❖ ❖ clicks! watching a video! Explicits! ❖ explicit rating! ❖ favorites 5/11/2013 Warszawa-JUG @adamwarski
- 10. Prediction vs recommendation ❖ Recommendations! ❖ ❖ ❖ suggestions! top-N! Predictions! ❖ ratings 5/11/2013 Warszawa-JUG @adamwarski
- 11. Collaborative Filtering ❖ Something more than search keywords! ❖ ❖ ❖ tastes! quality! Collaborative: data from many users 5/11/2013 Warszawa-JUG @adamwarski
- 12. Content-based recommenders ❖ Define features and feature values! ❖ Describe each item as a vector of features 5/11/2013 Warszawa-JUG @adamwarski
- 13. Content-based recommenders ❖ User vector! ❖ ❖ ❖ e.g. counting user tags! sum of item vectors! Prediction: cosine between two vectors! ❖ profile! ❖ item 5/11/2013 Warszawa-JUG @adamwarski
- 14. User-User CF ❖ Measure similarity between users! ❖ Prediction: weighted combination of existing ratings 5/11/2013 Warszawa-JUG @adamwarski
- 15. User-User CF ❖ Domain must be scoped so that there’s agreement! ❖ Individual preferences should be stable 5/11/2013 Warszawa-JUG @adamwarski
- 16. User similarity ❖ Many different metrics! ! ❖ Pearson correlation: 5/11/2013 Warszawa-JUG @adamwarski
- 17. User neighbourhood ❖ Threshold on similarity! ❖ Top-N neighbours! ❖ ❖ 25-100: good starting point! Variations:! ❖ trust networks, e.g. friends in a social network 5/11/2013 Warszawa-JUG @adamwarski
- 18. Predictions in UU-CF 5/11/2013 Warszawa-JUG @adamwarski
- 19. Item-Item CF ❖ Similar to UU-CF, only with items! ❖ Item similarity! ❖ Pearson correlation on item ratings! ❖ co-occurences 5/11/2013 Warszawa-JUG @adamwarski
- 20. Evaluating recommenders ❖ How to tell if a recommender is good?! ❖ ❖ ❖ compare implementations! are the recommendations ok?! Business metrics! ❖ what leads to increased sales 5/11/2013 Warszawa-JUG @adamwarski
- 21. Leave one out ❖ Remove one preference! ❖ Rebuild the model! ❖ See if it is recommended! ❖ Or, see what is the predicted rating 5/11/2013 Warszawa-JUG @adamwarski
- 22. Some metrics ❖ Mean absolute error:! ❖ Mean squared error 5/11/2013 Warszawa-JUG @adamwarski
- 23. Precision/recall ❖ Precision: % of recommended items that are relevant! ❖ Recall: % of relevant items that are recommended! ❖ Precision@N, recall@N Recall Ideal recs Precision Recs 5/11/2013 Warszawa-JUG All items @adamwarski
- 24. Problems ❖ Novelty: hard to measure! ❖ ❖ leave one out: punishes recommenders which recommend new items! Test on humans! ❖ A/B testing 5/11/2013 Warszawa-JUG @adamwarski
- 25. Diversity/serendipity ❖ Increase diversity:! ❖ ❖ ❖ top-n! as items come in, remove the ones that are too similar to prior items! Increase serendipity:! ❖ downgrade popular items 5/11/2013 Warszawa-JUG @adamwarski
- 26. Netflix challenge ❖ $1m for the best recommender! ❖ Used mean error! ❖ Winning algorithm won on predicting low ratings 5/11/2013 Warszawa-JUG @adamwarski
- 30. Mahout single-node ❖ User-User CF! ❖ Item-Item CF! ❖ Various metrics! ❖ ❖ Pearson! ❖ ❖ Euclidean! Log-likelihood! Support for evaluation 5/11/2013 Warszawa-JUG @adamwarski
- 32. Mahout multi-node ❖ Based on Hadoop! ❖ Pre-computed recommendations! ❖ Item-based recommenders! ❖ ❖ Co-occurrence matrix! Matrix decomposition! ❖ Alternating Least Squares 5/11/2013 Warszawa-JUG @adamwarski
- 33. Running Mahout w/ Hadoop hadoop jar mahout-core-0.8-job.jar ! ! org.apache.mahout.cf.taste.hadoop.item.RecommenderJob ! ! --booleanData ! ! --similarityClassname SIMILARITY_LOGLIKELIHOOD ! ! --output output ! ! --input input/data.dat 5/11/2013 Warszawa-JUG @adamwarski