LightGBM using Enhanced and De-biased Item Representation for Better Session-based Fashion Recommender Systems

In this paper, we present our 5th place solution for the ACM RecSys 2022 challenge (http://www.recsyschallenge.com/2022/).The competition, organized by Dressipi, aims to predict the fashion item purchasing actions on a public dataset of 1.1 million online retail sessions. In the traditional sequence recommendation model, we mainly utilize the action sequence information to model the representations of items and users. However, the fashion categories and features of items change much more frequently in our task, and the bias caused by the popularity will greatly affect the representations learning for the users and items. In this work, our team, termed THLUO, devise a model, which injects the spatiotemporal features of each item in sessions to alleviate the bias problem and capture the latest fashion trend information hiding in the session. In more detail, we proposed a two stages model, which includes retrieval and re-ranking. In the retrieval stage, we adapt the positions and timestamp features into the item-CF model to eliminate the bias caused by the popularity. In the re-ranking stage, we not only adapt traditional feature engineering but also used the enhanced features created by neural net works and fusion them as inputs of LightGBM for final prediction. After careful experiments, our model’s result archive an outstanding score of 0.2062 in mean reciprocal rank metrics in the test dataset, finally ranked fifth in the competition.