one-class collaborative filtering

User 337 | 10/2/2015, 1:55:53 AM

I have a one-class collaborative filtering task, where no ratings are given, only binary information on what was bought or viewed. I've used GraphChi previously. GraphLab Create is a step forward in many ways--good work! However, I'm having difficulty training meaningful FactorizationRecommender models on one-class data; all recommendations are very close to 1. I haven't seen any options specifically for one-class tasks in the documentation. Am I missing something?

Comments

User 19 | 10/2/2015, 2:02:25 AM

Hi jdjensen,

We do offer one-class collaborative filtering methods. (We typically call this setting as "implicit data", rather than "explicit data" where you have ratings.) Try checking out the rankingfactorizationrecommender.create or [itemsimilarity.create](https://dato.com/products/create/docs/generated/graphlab.recommender.itemsimilarityrecommender.create.html#graphlab.recommender.itemsimilarity_recommender.create); these handle your setting by default.

Please let us know if you have questions, Chris


User 337 | 10/2/2015, 8:14:17 PM

Got it. Thanks!


User 940 | 10/6/2015, 5:58:06 PM

Good to hear!

Cheers! -Piotr