RECSYS CHALLENGE 2016
LEARNING TO RANK IN RECSYS 2016
▸ Learning to Rank Framework
▸ Listwise, Pairwise, Pointwise
▸ Relevance Judgements
▸ Training & Evaluation dataset
▸ LambdaMART
▸ Gradient tree boosting & NDCG loss function
RECSYS CHALLENGE 2016
DATA EXPLORATION
▸ Understanding data.
▸ Probability of Item being interacted "Popularity".
▸ Probability of Item being interacted by his terms titles
▸ Probability of recurring interactions.
▸ Probability of a user interacting Items and geographical
distance
RECSYS CHALLENGE 2016
RETRIEVERS
▸ Job title/Job tag match
▸ Relevant items per interaction
▸ Interactions made by user
▸ Impressions shown to user
▸ Collaborative filtering, Item to Item/User to User
▸ Clustered users.
RECSYS CHALLENGE 2016
LEARNING PROCESS
▸ Speed up the training process.
▸ Reducing the number of features applied.
▸ Reducing the number of elements used to train.
▸ Internal evaluation.
▸ NDCG@30 metric.
RECSYS CHALLENGE 2016
RESULTS & CONCLUSIONS
Run Internal Official Improved
Baseline 22849 30721
Baseline BM25 (b=0.2 / k1=1.2) 23591 36857 20 %
Interactions made by the user 162846 229949 523 %
Training process 176312 241422 5 %
Impressions with decay 400217 434433 79,94 %
Item popularity 435325 475940 10 %
Clustering 442049 488192 3 %
Collaborative filtering 457391 507022 4 %
Best effort @ 2000 480396 533232 5 %
Best effort @ 9000 482083 535899 1 %