基于受限波兹曼机的推荐算法研究

摘要

本文针对推荐系统中传统推荐算法在处理较稀疏数据效果表现不佳的问题，将一种最新的机器学习方法带入到推荐算法中，利用多层波兹曼机组成的深度结构模型与传统最近邻推荐方法相结合，形成一种新的推荐模型算法。本文还利用有限步吉布斯采样的最小化散度差(Constrastive Divergence)方法解决了该模型的无监督训练问题，并且通过预训练和反馈微调使得模型的训练得以实现；最后利用深度结构抽取的抽象特征结合最近邻方法进行预测推荐。另外，本文利用传统基于相似度最近邻方法，矩阵分解方法和新模型算法在相关数据集上进行多组实验，实验结果表明该算法不仅在稀疏的数据上也表现出更好的效果，并且拥有着更快的收敛速度。%Aiming at solving the problems of poor performance in recommendation system when using traditional content-based or collaborative-iflter based methods,a new recommendation model was proposed to deal with recommendation task in this article. This new model was a deep structure composed of several layers of restricted boltzmann machine which were learned using a unsupervised learning method called Constrastive Divergence algorithm adopting limited steps of gibbs sampling,besides, other strategies such as pre-training and ifne-tune were used to make the model trained possible. At last, this article carried out several experiments among traditional matrix decomposition and the new model,the result turned out that the new model not only performed well in speed in iteration,and even performed better in sparse data compared to the traditional ones.