基于平滑插值和自适应相似矩阵的推荐算法  

Recommendation algorithm based on smooth interpolation and adaptive similarity matrix

作　　者：高美珠 于万钧 陈颖 Gao Meizhu;Yu Wanjun;Chen Ying(School of Computer Science&Information Engineering,Shanghai Institute of Technology,Shanghai 201418,China)

机构地区：[1]上海应用技术大学计算机科学与信息工程学院,上海201418

出　　处：《计算机应用研究》2025年第4期1108-1114,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(61976140)。

摘　　要：针对协同过滤推荐过度依赖共同评分项目导致交互数据不足,及不同时间段共享同一相似矩阵无法准确度量用户相似度等问题,提出一种基于平滑插值和自适应相似矩阵的推荐算法。首先,在线性插值技术基础上,结合均值和标准差设定动态区间,并通过sigmoid函数平滑调整原始评分,消除用户评分习惯差异。其次,使用时序变换函数量化用户偏好遵循的不同动态模式和遗忘规律,增强用户偏好表示。最后,利用标签语义、标签质量、时序变换函数和相对评分差异信息熵,构建标签感知机制和全局评分机制,并利用生成的相似矩阵重构用户自适应相似矩阵。仿真实验结果表明,相较于其他基线算法,该算法推荐性能最优,召回率提升5.27和4.73百分点,归一化折损累计增益(NDCG)提升6.67和5.90百分点,验证了算法的有效性。Aimed at the problems where the collaborative filtering recommendation overly relied on common item ratings and lacked dense interaction data,and where sharing the same similarity matrix across different time periods could not accurately measure user similarity,this paper proposed a recommendation algorithm based on smooth interpolation and adaptive similarity matrix.Firstly,based on linear interpolation technique,dynamic intervals set the mean and standard deviation,and the sigmoid function smoothly adjusted the original ratings to eliminate differences in user rating habits.Next,the temporal transformation function quantified the different dynamic patterns and forgetting behaviors that user preferences followed,enhancing the representation of user preferences.Finally,the label semantics,label quality,temporal transformation function,and relative rating diffe-rence entropy constructed the label perception mechanism and global rating mechanism,and the generated similarity matrix was used to reconstruct the user-adaptive similarity matrix.The simulation results show that,compared to other baseline algorithms,the proposed algorithm achieves the best recommendation performance.The recall rate improved by 5.27 and 4.73 percentage points,and the normalized discounted cumulative gain(NDCG)improved by 6.67 and 5.90 percentage points,which verifies the effectiveness of the algorithm.

关 键 词：协同过滤 平滑插值 时序变换函数 标签语义 相对评分差异信息熵 标签感知机制 全局评分机制 自适应相似矩阵 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]