A Hybrid Approach of Statistics and Embeddings for Multilingual and Multi-Locale Recommendation
Keywords: recommender systems, sequential recommendation
TL;DR: solution of team gpt_bot in KDD Cup 2023, 3rd place in Task2 and 4th place in Task1
Abstract: To encourage the development of multilingual recommendation systems, Amazon pushlished a multilingual and multi-locale shoppingsession dataset, and KDD Cup 2023 challenge on Multilingual Session Recommendation Challenge was hosted based on this dataset.In this paper, we present our solution for this competition. Following a widely-used setting in recommender system, our solutionconsists of two stages: recalling and ranking. In the first stage, weutilize various recalling strategies to retrieve a set of candidateproducts, including covisit matrix based collective filtering, graphembedding based I2I searching, text transformer based I2I searchingand BPR based U2I searching. In the second stage, we develop amodel to predict the probability of each user engaging with thecandidate products. This model is an ensemble of two Catboostmodels, which include various statistical features and embeddingsimilarity features. Finally, we achieved 4th place in Task1 and 3rd place in Task2
Submission Number: 5
Loading