基于双层网络挖掘用户偏好的序列化推荐算法

Session-based Recommendation Algorithm Based on Two-tier Network Mining User Preferences

下一项推荐在电商领域是具有挑战性的任务之一。针对现有的序列化推荐模型中把上下文交互物品视为对下一项待推荐物品的重要程度为一样,以及用户长短期偏好在推荐过程中所占比重不清楚的问题,提出了双层网络挖掘用户偏好的序列化推荐算法。在第一层注意力网络结构中,采用带有注意力机制的神经记忆网络结合用户历史会话嵌入计算当前序列中物品的注意力权重。在第二层注意力网络中,将带权重的上下文信息结合门控循环单元的当前隐藏层输出组成的神经网络记忆单元,再用注意力机制计算出当前记忆单元所占的权重,即短期偏好的权重。然后结合用户历史购物信息和当前上下文的嵌入通过注意力机制计算出当前上下文的长期偏好所占权重。通过在电商数据集JDATA上的测试,所提方法在推荐准确率有较大的提升。

The next recommendation is one of the challenging tasks in the field of e-commerce.In the existing serialization recommendation model,the importance of contextual interactive items to the next item to be recommended is the same,and the proportion of users'long-term and short-term preferences in the recommendation process is the same.An improved algorithm of users'long-term and short-term preferences based on attention network is proposed.The proposed algorithm uses a two-layer attention network.In the first layer of attention network structure,the neural memory network with attention mechanism combined with user historical session embedding is used to calculate the attention weight of items in the current sequence.In the second layer of attention network,the weighted context information is combined with the current hidden layer output of the gated loop unit to form a neural network memory unit,and then the attention mechanism is used to calculate the weight of the current memory unit,that is,the weight of short-term preference.Then,combined with the user's historical shopping information and the embedding of the current context,the weight of the long-term preference of the current context is calculated through the attention mechanism.Through the test on the open e-commerce data set JDATA,the recommendation accuracy of the proposed method has been greatly improved.