Dual stream adaptive sequential recommendation

Abstract Recommendation systems are AI-tools for assisting users to choose the items they may like. Based on what a user has clicked, rated, or purchased in the past, the system learns to predict what the user may want next. This paper proposes the Dual-Stream Adaptive Sequential Recommendation system (DSASR). DSASR employs two independent causal transformer encoders: one captures short-term preferences from recent interactions; the other captures stable long-term preferences from historical behavior. A Consistency Gate computes the preference drift to represent how much the recent preference has shifted from long-term history, and uses it to compute a personalized blend weight conditioned on user demographic attributes including age, gender, occupation, and a learned identity embedding. For training, combined Binary Cross-Entropy (BCE) and Bayesian Personalized Ranking (BPR) loss have been used, directly optimizing the ranking metrics used at evaluation. Experiments have been conducted on MovieLens 100K, MovieLens 1 M, and Amazon Beauty datasets. The model has been evaluated on metrics such as Hit Rate at N (HR@N), Normalized Discounted Cumulative Gain at N (NDCG@N), and Mean Reciprocal Rank at N (MRR@N). The results show that DSASR consistently outperforms existing methods, achieving HR@10 = 0.7253, NDCG@10 = 0.4650, MRR@10 = 0.3627 for MovieLens 100K, HR@10 = 0.8043, NDCG@10 = 0.5832, MRR@10 = 0.5130 for MovieLens 1 M, and HR@10 = 0.4353, NDCG@10 = 0.2686, MRR@10 = 0.2172 for Amazon Beauty. The improvement is due to the Consistency Gate computing a personalized blend weight for each user, which takes user demographic attributes as well as preference drift into account, making the recommendations more personalized than existing systems. The code is available in https://github.com/skarifahmed/DSASR .