基于循环时间卷积网络的序列流推荐算法

针对循环神经网络(Recurrent Neural Network,RNN)模型在序列流推荐中只能从宏观上捕捉序列的演变模式,忽略了物品(Item)间内部的微观联系,无法长程建模序列数据的变化规律的问题,提出了多维度序列建模算法循环时间卷积网络(Recurrent Temporal Convolutional Network,RTCN)。首先,将每个物品表示成定长向量,采用多层因果卷积和扩张卷积操作扩大感受野范围,建立序列元素间的长程依赖关系。利用残差连接网络提取不同层次的特征信息,解决反向传播中梯度衰减甚至消失的问题。综合设计时间卷积网络(Temporal Convolutional Network,TCN)提取序列流中前后物品间的局部特征,将物品信息映射到隐藏空间,得到细粒度的特征向量。为进一步建立元素间的宏观联系,将特征向量依次输入门限循环单元(Gated Recurrent Unit,GRU),迭代更新现有隐藏状态并预测下一时刻的输出。RTCN通过时间卷积网络,从输入序列流提取出长时间、多维度、细粒度的局部关联特征;经过门限循环网络,建模序列间的长距离依赖关系,捕捉序列元素的演变模式,并预测下一个出现的物品。利用网站、手机应用和音乐3个不同场景中的数据对模型进行了实验。实验结果显示,RTCN模型在召回率(Recall)和平均排序倒数(MRR)两个指标上比RNN模型高出6%~13%,比传统推荐算法高出9%~59%。通过对比不同的损失函数,模型在交叉熵损失函数下表现最优。此外,由于TCN中的卷积层具有多通道的结构,当数据维度丰富时,该模型对物品和用户的上下文信息具有很强的综合能力。

Session-based Recommendation Algorithm Based on Recurrent Temporal Convolutional Network

Since the Recurrent Neural Network(RNN)generally models transition patterns,ignores the inner connection of items and can’t model the long-term evolving patterns of sequential data in session-based recommendations.A Recurrent Temporal Convolutional Network(RTCN)was proposed.Firstly,each item in the sequence is embedded as a vector,the multi-layer casual convolutions and dilated convolutions are applied so that the receptive field is improved and the long-term connections are established.A residual network is stacked to extract features from different layers.Therefore,the gradient vanishing or even disappearing in back propagation can be solved.With above operations,a well-designed Temporal Convolutional Network(TCN)is established.It extracts local features from sequence items,maps item information into latent space and generates fine-grained feature vectors as results.To further explore the connections between items in macroscopic way,the feature vectors are feed into Gated Recurrent Unit(GRU).After multiple iterations and updates to hidden states,the model can make a prediction of the next item.RTCN can extract long-time,multi-dimension,fine-grained local features from inputs by adapting temporal convolutional network.It also models the long-distance connections between items,captures the transition patterns and infers the next items by using GRU networks.The experimental results demonstrate that the RTCN model outperforms 6%~13%than RNN-based model and 9%~59%than other traditional recommendation methods under the metrics of Recall and Mean Reciprocal Rank(MRR).By comparing different definitions of loss,RTCN performs best under the cross entropy loss function.Meanwhile,due to the TCN multi-channel structure,the proposed model has a high potential capacity to embedding context features of items and users when the dataset information is rich.