Our paper "Comparison of Transformer-Based Sequential Product Recommendation Models for the Coveo Data Challenge" is being presented at the SIGIR e'Com 21 Workshop today!15.07.2021
The paper "Comparison of Transformer-Based Sequential Product Recommendation Models for the Coveo Data Challenge" is our contribution to the Coveo Data Challenge organized by the e'Com Workshop at the SIGIR 21.
Providing good recommendations to keep users engaged and pre- dict their behavior are crucial components in today’s e-commerce businesses. To model user interests, various data sources can be utilized like user interactions including search behavior, product descriptions, but also product interactions like add-to-cart events or purchases. The SIGIR eCom - Coveo Data Challenge provides a new dataset containing such data points and calls for systems predicting the next product interaction as one of their challenge tasks. In this paper we report our approaches and results for the recommendation task of the challenge. We unify the various data sources from the Coveo Dataset to use it with sequential recommendation models and experiment with two datasets: One that includes all interac- tions and one that only consists of product interactions. For both datasets we train the transformer-based next-item recommender models SASRec and BERT4Rec. To integrate the available categori- cal metadata, we adapt KeBERT4Rec, which allows the addition of keyword descriptions, and experiment with two variants.