Abstract:

To address the issues of rating fluctuation and limited linear prediction ability in traditional collaborative filtering algorithms， a dual-channel recommendation algorithm based on Fast Fourier Transform （FFT） denoising and reverse-path collaborative enhancement named FRP-DCR （Frequency-domain denoising and Reverse-Path enhanced Dual-Channel Recommendation） was proposed， which integrates FFT-based denoising and reverse-path collaborative enhancement. First， user rating sequences were denoised in the frequency domain using FFT， so as to extract long-term preference trends， thereby reducing the impact of short-term emotional fluctuations on recommendation. Furthermore， a time decay mechanism was introduced to increase the algorithm's ability to represent temporal dynamics. Second， a reverse-path collaborative enhancement structure was designed during the construction of collaborative feature vectors， to simulate indirect user-item-user collaborative relationships， thereby addressing the inability of traditional methods to capture weak connections between users due to their reliance on direct user interactions. Finally， a regularized Broad Learning System （BLS） was applied to model the main rating structure， and a gradient boosting framework based efficient decision tree algorithm was adopted to model the nonlinear residuals， thereby improving prediction accuracy. The proposed algorithm， FRP-DCR， was compared with five baseline methods， including QP-DNN-TF-IDF （Quadratic Polynomial-Deep Neural Networks-Term Frequency-Inverse Document Frequency）， Light Graph Convolutional Network （LightGCN）， and Neural Collaborative Filtering （NCF）， to verify its performance. Experimental results on the MovieLens dataset show that FRP-DCR reduces the Mean Absolute Error （MAE） by 0.029 5 to 0.667 5， and reduces the Root Mean Square Error （RMSE） by 0.036 6 to 0.798 3. The algorithm achieves good performance in both metrics.

Key words: Fast Fourier Transform (FFT), reverse-path collaborative enhancement, regularized Broad Learning System (BLS), decision tree, nonlinear residual

摘要：

针对传统的协同过滤推荐算法中，用户评分存在波动性及线性预测能力受限的问题，提出一种基于快速傅里叶变换（FFT）去噪与反向路径协同增强的双通道推荐算法FRP-DCR （Frequency-domain denoising and Reverse-Path enhanced Dual-Channel Recommendation）。首先，通过FFT对用户评分序列进行频域去噪，以提取长期偏好趋势，从而减少短期情绪波动对推荐的影响，并引入时间衰减机制，以提高算法对评分动态变化的表征能力；其次，在构建协同特征向量时，设计反向路径协同增强结构，以模拟用户-物品-用户之间的间接协同关系，从而解决传统协同过滤算法仅依赖用户的直接交互行为而导致的难以捕捉用户间的弱连接关系的问题；最后，采用正则化宽度学习系统（BLS）与基于梯度提升框架的高效决策树算法，分别建模主线评分结构与非线性残差部分，以提高预测精度。通过与QP-DNN-TF-IDF （Quadratic Polynomial-Deep Neural Networks- Term Frequency-Inverse Document Frequency）、轻量级图卷积网络（LightGCN）和神经协同过滤（NCF）算法等5种基线方法进行比较，对本文算法FRP-DCR的性能进行验证。实验结果表明，在MovieLens数据集上，所提算法FRP-DCR在平均绝对误差（MAE）上降低了0.029 5~0.667 5，在均方根误差（RMSE）上降低了0.036 6~0.798 3，即在2个评价指标上均表现出良好性能。

关键词: 快速傅里叶变换, 反向路径协同增强, 正则化宽度学习系统, 决策树, 非线性残差