基于深度学习的推荐算法逐渐成为推荐系统领域的主流研究方向。然而,大多数现有工作仅基于单一的用户与物品交互数据,并且缺乏可解释性。本文对用户评论进行充分挖掘,并且额外引入物品信息来缓解冷启动问题并提高推荐算法的准确性。该算法利用BERT预训练模型来处理文本数据,并将用户与物品的评论特征与矩阵分解得到的潜在特征相融合,最后在评分预测任务中使用Kolmogorov-Arnold网络进行特征学习。通过本文算法与其他基线算法在亚马逊评论数据集上进行对比,该算法与基线算法相比显著提高了评分预测的精度以及准确率和召回率。本研究通过深入挖掘用户评论文本和物品描述信息,证明其在提升推荐算法准确性方面的显著效果,为推荐系统的研究提供了新的思路。Recommendation algorithms based on deep learning have emerged as a prominent research in the field of recommender systems. However, most existing approaches rely solely on user-item interaction data and lack interpretability. This article thoroughly explores user reviews and incorporates additional item information to alleviate the cold-start problem and enhance the accuracy of recommendation algorithms. The proposed approach employs the BERT pre-trained model to process textual data and integrates review-based features of users and items with latent features obtained through matrix factorization. Finally, the Kolmogorov-Arnold network is utilized for feature learning in the rating prediction task. Comparative experiments on Amazon review datasets demonstrate that the proposed algorithm significantly outperforms baseline methods in terms of rating prediction accuracy and recall. By deeply mining user review texts and item descriptions, this study validates their substantial impact on improving recommendation accuracy and offers new insights for recommender system research.
