一种结合自编码器与强化学习的查询推荐方法

Query Suggestion Method Based on Autoencoder and Reinforcement Learning

查询推荐的目的是发掘搜索引擎用户的查询意图,并给出相关查询推荐。传统的查询推荐方法主要依靠人工提取查询的相关特征,如查询频率、查询时间、用户点击次数和停留时间等,并使用统计学习算法或排序算法给出查询推荐。近年来,深度学习方法在查询推荐问题上获得了广泛应用。现有的用于查询推荐的深度学习方法大多是基于循环神经网络,通过对查询日志中所有查询的语义特征进行建模以预测用户的下一查询。但是,现有的深度学习方法生成的查询推荐上下文感知能力较差,难以准确捕捉用户查询意图,且未充分考虑时间因素对查询推荐的影响,缺乏时效性和多样性。针对上述问题,文中提出了一种结合自编码器与强化学习的查询推荐模型(Latent Variable Hierarchical Recurrent Encoder-Decoder with Time Information of Query and Reinforcement Learning,VHREDT-RL)。VHREDT-RL引入了强化学习联合训练生成器和判别器,从而增强了生成查询推荐的上下文感知能力;利用融合查询时间信息的隐变量分层递归自编码器作为生成器,使得生成查询推荐有更好的时效性和多样性。AOL数据集上的实验结果表明,文中提出的VHREDT-RL模型获得了优于基准方法的精度、鲁棒性和稳定性。

The purpose of query suggestion is to explore the query intent of search engine users and provide relevant query suggestion.Traditional query suggestion methods mainly rely on manually extracting relevant features of queries,such as query frequency,query time,user clicks and dwell time,etc.,and use statistical learning algorithms or ranking algorithms to give query suggestion.In recent years,deep learning methods have been widely used in query suggestion problems.The existing deep learning methods for query recommendation are mostly based on recurrent neural networks,which predict the next query of the user by modeling the semantic features of all queries in the query log.However,the existing deep learning methods have poor context awareness of query suggestion,it is difficult to accurately capture user query intentions,and the influence of time factors on query suggestion is not fully considered,and it lacks timeliness and diversity.In response to the above problems,this paper proposes a query suggestion model combining autoencoder and reinforcement learning(Latent Variable Hierarchical Recurrent Encoder-Decoder with Time Information of Query and Reinforcement Learning,VHREDT-RL).VHREDT-RL introduces a reinforcement learning joint training generator and discriminator,thereby enhancing the context awareness of generating query suggestion,using latent variable hierarchical recursive autoencoders that integrate query time information as a generator,and making query suggestion better time-sensitive and diversity.The experimental results on the AOL data set show that the VHREDT-RL model proposed in this paper achieves better accuracy,robustness and stability than the benchmark method.