Personalized Within-basket Item Recommendation Based On Deep Learning

Personalized recommendation systems recommend content that users are most likely to be interested in based on their past behaviors and different attributes.Within-basket recommendation,which predicts the next related item to be added to the basket from the item corpus,is prevalent in grocery shopping and e-commerce.In this scenario,besides useritem collaborative filtering information,the retail platform also needs to combine the items that the user currently owns to make a recommendation.Within-basket recommendation takes into account more complete information about the user and shopping intent,which benefits both the user and the platform.Research on traditional recommender systems has focused on improving the precision and response time of individual recommended items.In fact,users purchase items aims to satisfy a set of demands,which usually requires the platform to provide multiple associated items.Most of the previous works solve the task by rule mining or incorporating various types of associations.However,the representation of the basket and the high-order feature interaction is hardly investigated previously.In this work,we propose a deep learning-based model named DBFM(Deep BasketSensitive Factorization Machine)to address these problems,as follows:(1)To address the problem that using ID to encode baskets may cause feature missing in the feature learning process,this paper makes a personalized representation for a basket based on its constituent items by latent factor learning.Also,the influence of different vector aggregation methods on basket’s representation is explored.(2)To address the problem of underutilizing the complementarity or compatibility relationships among the three entities {user,basket,target item},this paper models the associations by triples sampled from the basket instead of the regular two-tuples used in traditional recommender systems,where the association between the user and the target item is modeled to capture the compatibility relationship,the association among the items in the basket is modeled to capture the complementarity relationship,and the association between the items in the basket and the target item is modeled to capture the consistency relationship.Meanwhile,multi-association modeling can effectively alleviate the data sparsity problem.(3)To address the problem of high computational complexity of the second-degree factorization machine,after obtaining the personalized representation of the basket,a deep network is incorporated to combine both low-order and high-order feature patterns among different entities to capture the sophisticated structures from inputs,and finally the affine transformation is used to integrate the output results of different patterns.Experiments and statistical significance tests on three real-world datasets demonstrate higher performance of our model over state-of-the-art methods.