【摘要】 因子分解机(Factorization Machine,FM)是利用矩阵分解思路针对推荐系统中数据稀疏情况设计的机器学习算法.传统的FM模型参数是通过梯度下降方法进行优化求解,但针对数据集训练样本较少的情况,梯度下降方法不能保证参数收敛到全局最优.粒子群算法(Particle Swarm Optimization,PSO)是一种快速启发式算法,具有全局搜索的特性.为提高FM模型的表现能力,首先基于PSO算法确定全局最优位置,然后利用梯度下降优化FM参数,本文提出了PSO-FM算法.在数据集Diabetes进行实验对比,结果表明,改进后的基于粒子群的因子分解机算法PSO-FM在模型训练速度和预测准确度上都优于传统的因子分解机FM算法.更多还原

【Abstract】 Factorization machine(FM) is a new machine learning algorithm based on the matrix factorization for sparse data in recommender system.The solution of parameters in the traditional FM model is based on the optimization method of gradient.However, in the case of small amount of samples,the optimization method of gradient can not guarantee convergence to global optimum.Particle swarm optimization(PSO) is fast heuristic algorithm.It has global search ability.In order to improve the performance of FM,we proposed the PSO-FM algorithm,which determining the global optimal location based on PSO firstly,then optimizing FM parameters by gradient descent.We compared PSO-FM with FM on diabetes dataset,and the result shows that PSO-FM can improve training speed and accuracy.更多还原