基于度量学习的多空间推荐系统

隐式反馈具有数据获取成本小、形式广泛的特点,因此在现代推荐系统中被广泛使用.由于用户的隐式反馈通常是稀疏,不平衡,且含义不明确的.因此,想要准确学习用户和物品之间的复杂交互具有挑战性.传统的基于矩阵分解的推荐方法只能建模用户-物品之间的相似性.同时,矩阵分解使用点积运算作为相似度评估方式,而点积运算不满足三角不等式,即不能将用户-物品相似性传递到用户-用户以及物品-物品的相似性建模.因此,矩阵分解不足以在隐式反馈中充分建模用户和物品的关系.尽管现在有基于隐式反馈使用欧式距离来度量用户-物品相似度的度量学习方法,使得对应的推荐方法能够满足三角不等式.但是,现有的度量方法通常会将每个用户或者物品表示为度量空间中的单个点,进而在单个空间内通过用户-物品之间的距离来表征用户-物品之间的相似性.由于在不同的环境下,用户对于同一种类型的物品的偏好也可能存在差异.基于单个空间的用户、物品嵌入向量有可能无法满足用户具有的多种偏好和物品具有的多种属性,进而限制了推荐系统的性能.为了充分刻画用户和物品,我们尝试从多个侧面对于用户和物品进行表示,并提出了一个基于多空间的度量学习(MML)框架.通过设计整合多个空间相似性的度量方式,我们将用户和物品投影到多个空间中进行细粒度的表示.另外,我们设计了一种经过校准的优化策略,包括经过校准的最大间隔损失函数和经过校准的采样方法.在保持多空间度量学习表示能力的同时,确保框架的有效性.最后,模型通过训练好的用户、物品向量,对于稀疏的用户-物品交互矩阵进行填补.在动态更新空间权重的同时,可以赋予模型新的训练视角,最终实现端到端的训练.通过四个真实世界推荐数据集上进行的大量实验表明,MML可以在Recall和nDCG衡量指标上将目前最优的对比算法提高40%以上.

Multi-Space Recommender Systems via Metric Learning

Compared with explicit feedback,implicit feedback is more abundant and easier to obtain,which is widely explored by modern recommender systems.However,the implicit feedback is often sparse,imbalanced,and has ambiguous meaning,which poses great challenges to the learning of complex interactions among users and items.Based on the interaction matrix,one of the most popular methods is matrix factorization,which has been widely studied and applied to model user preferences and item properties.However,the performances of these methods are restricted due to the implicit feedback,since most of them can only treat the missing entries as negative feedback.Moreover,the matrix factorization methods cannot satisfy the triangle inequality.In other words,it can only model the similarity between users and items,which cannot be transferred to model the similarity of user-user and item-item relationship.To overcome this limitation,the recommendation models based on metric learning aim to capture some relationships beyond user-item interactions from implicit feedback,which have been popular and achieved great performance in many fields such as computer vision.However,the existing methods based on metric learning for recommendation represent users and items as single points in the metric space,and then the preference is modeled by the distance between the user and the item.In this case,the model ignores the fact that users can have multiple preferences,and items can have multiple properties,which cannot be modeled by single vectors and limit the model’s performance in the recommendation.To fully capture and exploit the multiple facets of user preferences and item properties,we propose a novel framework of Multi-space Metric Learning(MML)for the recommendation.Firstly,by designing a cross-space similarity measurement,we project users and items into multiple spaces for fine-grained representation.The different spaces aim to capture different facets of users and items,which alleviate the mix-up problem of both single user s and item s embedding.Secondly,based on the increasing dimension of the user and item embedding,it is important to alleviate the learning cost of the model.Considering the effectiveness and efficiency of the multi-space modeling,we design a calibrated optimization strategy,including a calibrated margin loss function and a calibrated sampling method.Finally,to fully integrate the learned user-item,user-user,and item-item relationships for further mining the multi-facet of user and item,we design an iterative loop mechanism for the End-to-end training.By imputing the missingness of the user-item interaction matrix,we can update the multiple spaces projection of users and items,which can provide a new perspective for the proposed MML model to mine the relationship between users and items.Extensive experiments on four real-world recommendation datasets show that MML can achieve up to 40%improvements over the state-of-the-art baselines regarding both Recall and nDCG metrics.