A switching-based deep learning framework for personalized and adaptive E-commerce recommendations

In recommendation systems, it is bit challenging to address diverse user profiles, particularly when users demonstrate varied interaction histories and preferences. These chalenges include data sparsity and the cold start problem which requires the innovative solutions to ensure precise and efficient recommendations. Cross-domain recommendations is yet another challenge posed by the diverse user profiles. When users have significantly varied the interests across different domains, recommending the relevant items that cater to their diverse choices becomes very complex. Personalization and diversity trade-offs are also very crucial, which require a trade-off between diverse recommendations and capturing the individuality of certain users. Another serious concern is the Real-time Adaptability because of user’s changing preferences and behaviours. Most of the existing systems can not dynamically adapt to these shifts and are also not capable of handling cross-domain recommendations which requires accurate modelling of the user’s multidimensional preferences to provide up-to-date suggestions. Our innovative system is designed to accommodate diverse user profiles based on their interaction history and engagement levels with the items. Considering the highly dynamic digital landscape, we’ve categorised users into three major groups, namely newbies, light users, and heavy users. This paper proposes an innovative approach within the recommendation systems domain by building switching-based hybrid recommendation systems to cater to the needs of different user profiles. By optimising multiple objectives, including abundant but less informative signals such as product views, the proposed system demonstrates improved performance across different user segments. Experimental results show a reduction in validation loss from 0.3414 to 0.1545 for absolute new users, improvements in HR@10 from 0.30 to 0.60 and NDCG@10 from 0.35 to 0.65 for light users, and strong predictive performance for implicit feedback with an accuracy of 0.91, precision of 0.89, recall of 0.88, and F1 score of 0.89. Experimental results indicate that, relative to the baseline methods, the proposed framework achieves a reduction of approximately 54.7% in validation loss for absolute new users, while HR@10 and NDCG@10 for light users increase by 100% and about 85.7%, respectively.