Efficient and explainable sequential recommendation with language model

Abstract

Motivated by the outstanding success of large language models (LLMs) in a broad spectrum of NLP tasks, applying them for explainable recommendation become a cutting-edge recently. However, due to the inherent inconsistency in the information and knowledge focused, most existing solutions treat item recommendation and explanation generation as two distinct processes, incurring extensive computational costs and memory footprint. Besides, these solutions often pay more attention to the item-side (i.e., item attributes and descriptions) for explanation generation while ignoring the user personalized preference. To close this gap, in this paper, we propose a personalized explainable sequential recommendation model, which aims to output the recommendation results as well as the corresponding personalized explanations via a single inference step. Moreover, to mitigate the substantial computational cost, we devise a rescaling adapter and a Fast Fourier Transform (FFT) adapter for parameter-efficient fine-tuning (PEFT). Theoretical underpinnings and experimental results demonstrate that compared with prevalent PEFT solutions, our adapter possesses three merits: (1) a larger receptive field across the entire sequence for long-term dependency modeling; (2) element product in orthogonal bases for noise attenuation and signal amplifying; (3) better alignment and uniformity properties for precise recommendation. Comprehensive experiments on three public datasets against nine sequential recommendation solutions and three explanation generation solutions illustrate our Pleaser outperforms the strong baselines significantly with only 5% parameter fine-tuning. Code available at https://github.com/WHUIR/PLEASER.