基于张量分解的排序学习在个性化标签推荐中的研究

Study on Learning to Rank Based on Tensor Decomposition in Personalized Tag Recommendation

标签的使用给系统提供了一个划分并管理用户和物品的途径,而个性化的标签推荐则不仅方便用户输入标签,而且有助于提高系统标签的质量。进而,系统可以获得更多关于用户和物品的信息,提升后续推荐的精度,改善用户体验,因此在淘宝、滴滴等类似的业务场景中具有重要的作用。然而,现有的大多数标签推荐都没有关注推荐列表中的排序问题,列表中过于靠后的标签极易丧失让用户使用的机会,造成用户和物品信息的缺失,阻碍后续的精准推荐。针对上述问题,提出了一种基于张量Tucker分解和列表级排序学习的个性化标签推荐算法,采用优化MAP的方式进行训练,并在Last.fm数据集上进行了仿真实验,不仅验证了算法的有效性,而且充分探讨了学习率、核张量维度等参数对算法的影响。实验结果表明,该算法能较好地优化推荐列表的排序问题,且随列表长度的增加,其性能呈线性下降,算法的实现有利于更好地根据用户喜好来推荐服务。

The use of tags provides a way for the system to divide and manage users and items,while personalized tag recommendations not only facilitate users input,but also help to improve the quality of system tags.In turn,the system can obtain more information about users and items,improve the accuracy of subsequent recommendations,improve the user experience.Therefore,it plays an important role in similar business scenarios such as Taobao and Didi.However,most existing tag recommendations do not pay attention to the ranking issues in the recommendation list.The tag that is too late in the list is easy to lose the opportunity for user use,resulting in the lack of information about users and items,and hindering the subsequent accurate recommendation.Aiming at the above problems,a personalized tag recommendation algorithm based on tensor Tucker decomposition and list-wise learning to rank is proposed.The algorithm is trained by optimizing MAP,and the simulation experiment is carried out on Last.fm dataset,which not only verified the effectiveness of the algorithm,but also fully explored the influence of learning rate,the dimension of core tensor and other parameters on the algorithm.Experimental results show that the algorithm can optimize the ranking problem of the recommendation list greatly,and its performance decreases linearly with the increase of the length of the list.The implementation of the algorithm is conducive to better recommendation services according to the user preferences.