.png){kind=icon}
Return
Collaborative knowledge and personalized preference alignment for sequential recommendation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116234.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Sequential recommender systems model users' historical interactions as temporal sequences to predict their potential items of interest. While existing LLM-based sequential recommendation methods effectively enhance the performance of these systems by leveraging rich implicit world knowledge, they still suffer from two critical limitations: the collaborative knowledge gap and preference modeling bias. To overcome the challenges and fully unlock the potential of LLMs for recommendation, this work proposes CollaPPA, a novel framework that effectively integrates rich collaborative knowledge while capturing nuanced user preferences. Different from prior works, CollaPPA jointly trains on sequential recommendation alongside three recommendation-related tasks: Top-N recommendation, explanation generation, and attribute generation. This integrated multi-task learning strategy enables the model to strengthen the core capabilities of sequential recommendation while establishing cross-modal connections between behavioral interactions and non-behavioral knowledge. Furthermore, we propose a personalized preference alignment method that adopts multi-negative contrastive learning, optimizing one preferred item against multiple dispreferred candidates to closely mirror real-world ranking scenarios. Extensive experiments on three public datasets demonstrate the superiority of our CollaPPA, achieving performance gains of 4.6% to 20.6% over state-of-the-art baselines. Our source code is released at https://github.com/815382636/CollaPPA.
Keywords:
Sequential recommendation
Large Language Model
Personalized preference alignment
Journal
K
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available