xLSTM architecture feasibility for recommendations

The primary idea of Recommeder system is to present users with the most relevant items such as Music, Movies, Software, Services by learning large scale interaction datasets. In this thesis, we will investigate the feasibility of Extended Long Short Term Memory (xLSTM) architecture for sequential recommendation tasks. We benchmark xLSTM against strong existing baselines, including Transformer based models (such as BERT4Rec, SAS4Rec) and recurrent alternatives, on five public datasets: MovieLens 100K, MovieLens 1M, MovieLens 10M, MUSIK4all, and Amazon Software. We also evaluate and compare models using top k ranking metrics (Recall@10, MRR@10, NDCG@10) with training time and further conduct an in depth analysis on the model architectural embedding weights and performance results.Our results show that, xLSTM attains competitive accuracy relative to Transformer baselines, particularly on medium and large scale sequential datasets and while offering favorable memory scaling and stable training. On smaller sequences, performance gaps narrow or favor Transformers models, highlighting the trade offs between capacity and data sequence settings. Overall, our results indicate that xLSTM is a strong sequence model at scale competitive with Transformer baselines on larger sequential like MovieLens ML-1M/10M, memory efficient and training stable, yet bit less effective on the smallest sequences like MovieLens ML 100K. Finally, these results and findings will point to a practical path for deploying xLSTM in modern large scale sequential recommender systems.