Adapter-Based Personalization for Modeling Short-Term Preference Drift in Multimodal Recommendation Systems

Authors:

Anubhav Pratap Singh¹ · Vinay Kumar Dwivedi¹ · Chandra Shekhar Gautam¹ · Akhilesh A. Waoo¹

^{¹Department of Computer Science and Engineering,}

^{AKS University, Satna, Madhya Pradesh, India}

Abstract - With the rapid growth of digital platforms such as e-commerce, video streaming, and social media, recommendation systems have become essential for delivering personalized content and reducing information overload. Recent multimodal recommendation systems improve recommendation quality by combining different data sources such as text, images, videos, and user interaction behaviour. However, most existing recommendation approaches primarily rely on long-term historical preferences and assume that user interests remain stable over time. In real-world scenarios, user preferences frequently change for short periods due to temporary contextual factors such as mood, trends, seasonal events, or immediate needs, creating the challenge of short-term preference drift. Existing session-based and adaptive recommendation methods partially address this issue but often overemphasize recent interactions, ignore stable long-term preferences, or require computationally expensive full model retraining. To address these limitations, this study proposes an adapter-based personalization framework for multimodal recommendation systems that efficiently captures short-term preference drift while preserving long-term user knowledge. The proposed framework integrates multimodal feature learning, long-term preference representation, and short-term session behavior analysis using lightweight adapter layers that enable rapid adaptation without retraining the complete backbone model. Experimental results demonstrate that the proposed model achieved superior recommendation performance with a Precision@10 of 0.73, Recall@10 of 0.69, and NDCG@10 of 0.71. In addition, adaptation time was reduced from 420 ms to 150 ms, while trainable parameters decreased from 120 million to 12 million. These results show that the proposed framework provides an efficient, scalable, and practical solution for dynamic personalized recommendation systems.

Key Words:  Adapter-Based Personalization, Computational Efficiency, Multimodal Recommendation Systems, Personalized Recommendation, Short-Term Preference Drift, User Behaviour Modelling