Recommender systems are notoriously difficult. Even if a company can find data science talent, building a good recsys means overcoming a number of big challenges, all at once.  Even in the simplest of cases, recommenders involve massive volumes of ever-changing data and the complexity of large-scale data processing, among a variety of other factors.

Only a select few have mastered the art of recommendation engines.  Fortunately for you, we’ve compiled some of the very best presentations we’ve seen to help alleviate the burden of starting completely from scratch.

[If you want a free, custom-built recommender, advised by one of the authors below (guess which!), sign up here: www.mortardata.com/recommender.]

Building Large-scale Real-world Recommender Systems by Xavier Amatriain

Xavier discusses the anatomy of Netflix’s personalization (“everything is personalization”), their data and models (including logistic/linear regression, elastic nets, matrix factorization, and Markov chains), consumer data science (using offline and online testing), and the company’s recommender stack.

Key Lessons Learned Building Recommender Systems for Large-scale Social Networks by Christian Posse

Christian shares some of the wisdom he acquired during his time working on recommender systems at LinkedIn, including the importance of ensuring that recommendations make strategic sense, provide a great user experience, address and appropriately weigh relevant (often competing) objectives, challenges from slicing/labeling data, and the value of measuring/testing.

Music Recommendation and Discovery RemasteredOscar Celma and Paul Lamere

Oscar and Paul cover the unique challenges of building recommender systems for music (including huge inventory, low cost per item, low consumption time, highly contextual usage), what makes a good music recommendation, and the pros/cons of various recommender strategies. Oscar and Paul also evaluate current commercial and research systems and also looking at innovative new (at that time) techniques.


Building a Real-time, Solr-powered Recommendation Engine by Trey Grainger

Trey (who authored Solr in Action) provides an overview of search and matching concepts using Solr, including content-based (attribute-based, hierarchical, textual similarity, and concept-based) and behavioral-based approaches for building recommender systems. He also addresses why CareerBuilder chose Solr instead of Mahout for their recsys.

Committing to Recommendations (plus video!) by Eric Colson (disclosure: Eric is a Mortar Data advisor)

Eric speaks about Stitch Fix’s efforts to create the perfect recommender system – one so good that it allows the company to ship it to them without the customer ever seeing it. Eric walks through the company’s approach to algorithms, the unique data it collects, and how it combines algorithms with human judgment and vice versa.

[In case you missed it, we’re building custom recommender systems for free.]

Free, Custom-Built
Recommender System