A stacking ensemble model for prediction of multi-type tweet engagements
S Goda, N Agata, Y Matsumura - Proceedings of the Recommender …, 2020 - dl.acm.org
S Goda, N Agata, Y Matsumura
Proceedings of the Recommender Systems Challenge 2020, 2020•dl.acm.orgThe RecSys Challenge 2020 is a competition with a task of predicting four types of user
engagements on Twitter: Like, Reply, Retweet and Retweet with comment. In this paper, we
describe Team Wantedly's approach to this challenge, which won the third place. We found
that the targets are highly correlated and it is important to use every engagement to predict
the other engagements. Therefore, we choose to stack LightGBM models to use this co-
occurrences effectively in the large dataset. Our final scores are as follows: 1.5266 (Retweet …
engagements on Twitter: Like, Reply, Retweet and Retweet with comment. In this paper, we
describe Team Wantedly's approach to this challenge, which won the third place. We found
that the targets are highly correlated and it is important to use every engagement to predict
the other engagements. Therefore, we choose to stack LightGBM models to use this co-
occurrences effectively in the large dataset. Our final scores are as follows: 1.5266 (Retweet …
The RecSys Challenge 2020 is a competition with a task of predicting four types of user engagements on Twitter: Like, Reply, Retweet and Retweet with comment. In this paper, we describe Team Wantedly’s approach to this challenge, which won the third place. We found that the targets are highly correlated and it is important to use every engagement to predict the other engagements. Therefore, we choose to stack LightGBM models to use this co-occurrences effectively in the large dataset. Our final scores are as follows: 1.5266 (Retweet PR-AUC), 30.06 (Retweet RCE), 0.1918 (Reply PR-AUC), 20.44 (Reply RCE), 0.7716 (Like PR-AUC), 24.76 (Like RCE), 0.0724 (Retweet with comment PR-AUC), 14.86 (Reply RCE). Our code is available at https://github.com/wantedly/recsys2020-challenge.