recommender-system/

mindfull

In general, recommendation systems assume that each transaction made by a user reflects a single purchase intent (mono-intent). However, in reality, a transaction may contain several underlying intentions, such as routine shopping, consumption for specific purposes, or household needs, which cannot be effectively captured by a mono-intent model. Previous studies have proposed multi-intent approaches such as topic modeling, transformer-based, and clustering techniques. However, these methods assume that each intent has the same weight, especially when products are consumed simultaneously and do not have explicit labels that correlate between products, thereby reducing context understanding and personalization capabilities. Based on this gap, we propose a framework for implicit multi-intent detection in shopping transactions using the extensive knowledge of Large Language Models (LLMs) through the In-Context Learning (ICL) prompting technique. As for the preference assessment mechanism, we use the Position-Based Grouping (PBG) method to estimate user preferences based on the order of items added to the cart. The results of our experiments on the Instacart dataset show that our proposal is capable of producing a significant performance improvement compared to existing sequential recommendation systems, where our best model is able to increase Recall by up to 122% and MRR by up to 217%, indicating that it is more effective in capturing user preference trends for specific intentions in the purchase sequence.

For our method you guys can download the full source code in Here

And for the comparison with baseline, results, or analysis you can download it Here