Lightweight and Scalable Model for Tweet Engagements Predictions in a Resource-constrained Environment

Abstract

In this paper we provide an overview of the approach we used as team Trial&Error for the ACM RecSys Challenge 2021. The competition, organized by Twitter, addresses the problem of predicting different categories of user engagements (Like, Reply, Retweet and Retweet with Comment), given a dataset of previous interactions on the Twitter platform. Our proposed method relies on efficiently leveraging the massive amount of data, crafting a wide variety of features and designing a lightweight solution. This results in a significant reduction of computational resources requirements, both during the training and inference phase. The final model, an optimized LightGBM, allowed our team to reach the 4th position in the final leaderboard and to rank 1st among the academic teams. CCS CONCEPTS• Information systems → Recommender systems; • Computing methodologies → Classification and regression trees; Neural networks.

Multi-stage Ensemble Model for Cross-market Recommendation

“…We adopted 3 popular and successful implementations of GBDT: LightGBM 7 , XGBoost 8 and CatBoost 9 . Thanks to their flexibility and robustness, they can easily adapt to different types of features, obtaining challenge-winning results [2,5].…”

“…In the last stage of our model we perform a stacking ensemble using the same techniques described for the second stage in Section 4. 2.…”