特征漂移约束算法在推荐系统中的优化

Optimization of recommendation system by feature drift constraint algorithm

在时间矩阵分解方法的基础上,利用概念漂移检测捕获随时间动态变化的用户兴趣和项目偏好特征,可以有效提高个性化推荐算法的准确性。为此,该文提出特征漂移约束(feature drift constraint,FDC)算法,首先,根据输入样本的评级反馈构建评级矩阵的时间序列,采用矩阵分解方法将评级矩阵分解为用户特征矩阵和项目特征矩阵;其次,在输入新的评级样本后训练模型,采用随机梯度下降方法获得优化的学习参数,计算概念漂移的动态特征加权用于调整模型;最后,结合用户兴趣特征向量和项目偏好特征向量内积计算得到预测的项目评级,实现项目推荐。仿真结果表明,与MF、TSVD++、TMF和MCFTT算法相比,特征漂移约束算法在推荐准确性和概念漂移检测的有效性方面均有较好提升。

Based on the time matrix decomposition method,using concept drift detection to capture user interest and item preference features that change dynamically over time can effectively improve the accuracy of the personalized recommendation algorithm.To this end,a feature drift constraint algorithm(FDC)algorithm is proposed.Firstly,the time series of the rating matrix is constructed according to the rating feedback of the input samples,and the rating matrix is decomposed into the user feature matrix and the item feature matrix by the matrix decomposition method.Secondly,train the model after inputting new rating samples,use the stochastic gradient descent method to obtain the optimized learning parameters,calculate the dynamic feature weight of the concept drift to adjust the model.Finally,combine the user interest feature vector and the item preference feature vector inner product to calculate forecast project rating and realize project recommendation.The simulation results show that compared with the MF,TSVD++,TMF and MCFTT algorithms,the feature drift constraint algorithm has improved the accuracy of recommendation and the effectiveness of concept drift detection.