Research On Personalized Recommendation Algorithms With Auxiliary Information

The rapid development of Internet technology has led to an explosive growth of information,which makes people hardly capture the information they need timely and accurately thus causing the “information overload”.Recommender system(RS)can effectively cope with the "information overload" problem by helping people discover the information they need or are interested in.Most of the early work in RS only incorporated the relationship between the users and the items,which may suffer from low accuracy.Recently,work uses the context information(such as time,location,etc.)or side information of users and items(such as categories,brands,etc.)as the auxiliary information to improve the performance of the RS.We find most of the existing work has the following problems: 1)Rely on the manual work of experts to construct features and selecting features;2)Difficult to learn the non-linear and complex structure in real-world data and;3)Methods based on deep neural network suffer from high complexity and lack of interpretability.To tackle with these problems,we proposes a representation learning framework(RLF)in this paper for RS,which can integrate auxiliary information into recommendation effectively.1)RLF is an end-to-end framework that automatically learns the effective hybrid representation of users and items without manually work.2)We design an N-way concatenation pooling strategy in RLF to learn non-linear and complex structures in real-world data.3)RLF could adopt a succinct implementation to meet the need of actual needs in different scenario,which enhances the interpretability of the model.We believe that different types of auxiliary information have different characteristics,which should be taken into account to improve recommendation performance and enhance the interpretability of the model.Based on types of auxiliary information,we propose two different implementation model of RLF: a hierarchical hybrid representation model(HHRM)for context information and a behavioral inference representation model(BIRM)for side information.For the HHRM,we believe that a user's behavior heavily depends on the user's general preference and contextual information.We implement the RLF into a hierarchical hybrid representation model(HHRM)in context scenario.The BIRM adopts a two-layer RLF to learn the dynamic preference of user w.r.t the context: The first layer forms contextual representations of diverse contextual features,and the second layer of the BIRM combines them with users' general taste into a hybrid representation that models the dynamic preference of user w.r.t the context.For the BIRM,we believe that the attribution(side information)of users and items helps to model the users' preferences and items' characteristics.However,users and items may “lie” or “keep silent” about their attribution,which leads to wrong or deficient modeling of users and items.To this end,we use behavior-related meta-paths in the heterogeneous information network(HIN)to infer implicit profile information of users and items,which helps to model users and items effectively.We implement the RLF into a behavior inference representation model(BIRM)in side information scenario.BIRM firstly digs out the implicit profile features of users and item in the HIN,and then models these features into uniform representations by an inference model.Finally,BIRM inputs these representations into the RLF to learn the effective hybrid representation of users and items.Extensive experiments show the proposed models consistently outperform state-of-theart methods on real-world data sets,which indicates that: 1)The proposed RLF in this paper can effectively explore the non-linear and complex structure in the real-world data,and does not require massive workload of experts.2)The RLF can use different types of auxiliary information for the recommendation,and it can adopt different implementation w.r.t type of auxiliary information,so that the model has strong interpretability.