摘要
近年来,可解释推荐因其不仅能提供个性化推荐,还能解释推荐原因而备受关注。基于Transformer的文本生成技术尽管可以生成高度自然的推荐解释,但其高模型复杂性和计算成本也带来了挑战。为应对这些问题,本文提出了一种基于Transformer架构的混合可解释推荐模型。该模型结合了合成坐标和Kolmogorov-Arnold网络(KAN),通过将用户和物品映射至低维嵌入空间,利用空间关系引导解释生成,实现了高效的评分预测和解释生成。实验结果表明,该模型在评分预测和解释生成任务中均达到了最先进的水平,且在BLEU-4和USR等关键指标上取得了显著提升。研究还揭示了通过优化嵌入向量和模块化设计的结合,能够有效提升可解释推荐系统的性能,同时降低模型复杂性、计算难度及资源消耗。
Explainable recommendations have gained attention for offering both personalized suggestions and clear explanations.While Transformer-based text generation produces natural explanations,it introduces challenges of model complexity and computational cost.This paper proposed a hybrid model combining synthetic coordinates and Kolmogorov-Arnold Networks(KAN)with Transformer architecture to address these issues.By mapping users and items into a low-dimensional space and leveraging spatial relationships,the model achieved efficient rating prediction and explanation generation.Experimental results showed state-of-the-art performance,with significant improve-ments in BLEU-4 and USR metrics.The study demonstrates that optimizing embeddings and using a modular design enhance system performance while reducing complexity and computational diffi-culty and resource consumption.
作者
艾均
徐戈辉
苏湛
苏杭
Jun Ai;Gehui Xu;Zhan Su;Hang Su(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai)
出处
《建模与仿真》
2025年第5期13-30,共18页
Modeling and Simulation
基金
国家自然科学基金项目(61803264)。
