基于门控循环单元与主动学习的协同过滤推荐算法

针对传统协同过滤推荐算法存在无法反映用户短时兴趣的问题,提出一种基于门控循环单元(gated recurrent unit, GRU)神经网络与主动学习的协同过滤推荐算法。在采用GRU神经网络的基础上,将数据进行时序化处理,反映用户兴趣变化,并利用主动学习动态采样数据中的高质量的数据进行GRU神经网络的训练,使模型快速建立。在MovieLens1M数据集上的试验结果表明:加入主动学习的GRU模型的推荐算法比基于用户的协同过滤推荐算法(user-based collaborative filtering, UCF)、基于马尔科夫模型的协同过滤推荐算法(markov chain, MC)、基于隐语义模型的协同过滤推荐算法(latent factor model, LFM)算法有更高的短时预测率、召回率、项目覆盖率以及用户覆盖数,能够有效预测用户短时兴趣,提升精度,发掘长尾物品,且与原始GRU模型相比能够以更少的迭代次数达到相同效果。

GRU-based collaborative filtering recommendation algorithm with active learning

The traditional collaborative filtering recommendation algorithm failed to reflect short-term user interest. In order to reflect the short-term interests of users better, a collaborative filtering recommendation algorithm based on Gated Recurrent Unit (GRU) neural network with active learning was proposed. Based on the GRU neural network, the algorithm processed the data into time-series data to reflect the change of the user's interest and used active learning to sample high-quality data dynamically for accelerating the training of GRU neural network. The result on MovieLens1M dataset showed that the GRU model with active learning could obtain higher short-term prediction success rate, recall rate, item coverage, and user coverage compared with the user-based collaborative filtering method (UCF), the markovian chain model-based collaborative filtering method (MC) and the matrix factory-based collaborative filtering method (LFM), so it could effectively predict the short-term interest of users, improve the accuracy, discover the long-tail items. Meanwhile, it could achieve the same effect with fewer iterations compared with the original GRU model.