The objective of the competition was to create the best recommeder system for a streaming service by providing 10 recommended items. The evaluation metric was MAP@10.
In this repo we report the various experiments made and the various evolutions of our Recommender System.
Link to the official website of the challenge
The end result was a hybrid Recommender System formed by
- RP3betaRecommender
- Two versions of SLIMElasticNetRecommender specialized for different portions of users.
See the submission recommender here.
We used both an implicit and an explicit matrix, normalizing the values using a dynamic Logistic function according to the item bias and user bias, see the implementation here of the funciton explicitURM.
- MAP@10 = 0.06021 on Kaggle's public leaderboard