基于自适应惩罚因子和用户评分行为的协同过滤算法

An Improved Collaborative Filtering Recommendation Algorithm Based on Adaptive Penalty Factors and User Rating Behavior

摘要:

针对传统协同过滤算法头部效应和推荐精度低2个问题，提出了一种基于自适应惩罚因子和用户评分行为的协同过滤算法，惩罚因子通过对热门项目进行惩罚修正了相似度的计算，自适应参数通过遍历的方式找寻不同数据集下的最优惩罚力度，有效缓解了传统算法的头部效应问题；用户评分行为通过考虑用户的评分时间差以及评分的分布特性差异，细化相似度的计算，提高了算法精度。采用4个公开数据集验证改进算法的效果，对于表现较好的MovieLens 1M数据集，保持推荐数目不变时，本文算法F₁分数相比传统算法平均提高约13.9%，有效提高了算法的推荐质量。同时，采用的倒排表构建项目—用户交互矩阵，有效提高了算法运算速度，在MovieLens 1M数据集下，较传统算法运行时间减少约72.1%。

Abstract:

To address the two key issues of head effect and low recommendation accuracy in traditional collaborative filtering algorithms， this paper proposes a collaborative filtering algorithm based on an adaptive penalty factor and user rating behaviors. The penalty factor adjusts similarity calculations by penalizing popular items， while the adaptive parameter identifies the optimal penalty intensity for different datasets through iterative searching， effectively mitigating the long-tail effect of traditional algorithms. User rating behavior further refines similarity calculations by considering time differences between ratings and variations in rating distribution characteristics， thereby improving algorithm accuracy. The proposed algorithm was validated on four public datasets. For the well-performing MovieLens 1M dataset， the algorithm achieved an average F₁ score improvement of approximately 13.9% over traditional algorithms while maintaining the same recommendation count， significantly enhancing recommendation quality. Additionally， the inverted index was employed to construct the item-user interaction matrix， improving computational efficiency. On the MovieLens 1M dataset， the runtime of the proposed algorithm was reduced by approximately 72.1% compared to traditional methods.