A Uniform Ensemble Model for Multilingual Recommendation Systems
Qinglun Wang, Yuan Ji · 2023
Incomplete or missing user profiles can disrupt the effectiveness of recommender systems. Therefore, it is vital to establish a uniform ensemble model, which tends to recognize diverse customer shopping intentions. This paper presents our recall models consisting of traditional recommendation algorithms (i.e. itemCF, userCF, swing, etc.), deep graph recommendation algorithms (LightGCN), and embedding recommendations (word2vec embedding, cross encoder). This model is verified via multilingual shopping session datasets, namely Amazon-M2. Within KDD CUP 2023 Tasks 1 and 2, we use the sample data processing and model framework by low-cost devices. On the final leaderboard, the Onewo team is in 19th in Task1 and 21st in Task2. Every strategy is applied to match the top 100 items. Furthermore, we adopt a combination of group and 3 LGB models to re-rank recalled items for every strategy. By strategies, the Mean Reciprocal Rank (MRR) scores for tasks 1 and 2 of phase 2 are 0.33347 and 0.35541 respectively.