Recommendation System Based On User Sequence And Social Network

The number of Internet users is enlarging due to the success and rapid growth of mobile Internet.The massive data generated by users has a huge impact on users and service platforms.The challenge for service platforms is finding out how to use these data to generate value for both customers and themselves,and the most efficient way to handle above problems is recommendation algorithms.The recommendation system predicts users’ preference by modeling their portrait,behavior data,the feature of the item and the relevant context data,then make recommendation for the user.In academic research,several researchers have devoted to the study of recommendation system and have developed numerous recommendation models.The industry has also made an in-depth exploration of the recommendation system so that the recommendation method cannot only improve user satisfaction,but also bring much revenue to the platform.The most conventional recommendation model is based on user rating and they utilize traditional methods such as collaborative filtering,matrix factorization and factorization machines to make perdition.But these methods have some limitations such as cold start problem and data sparsity,which will lead to inefficient performance.In recent years,deep learning models have attracted more attention from various fields with their excellent capabilities of data processing and model generalization,and there are many novel recommendation models based on deep learning technology in the research field of recommendation systems.In this paper,we also study recommendation algorithms based on deep learning models.Our main work and contributions can be summarized as follows:1)We propose a sequential recommendation framework based on Hierarchical Pairwise Gating Model(HPGM).We observed the following phenomenon in sequential recommendation scenario: First,each item in user’s historical behavior sequence has different influence on the user’s next action,because not all items are equally important for modeling user preferences.Secondly,since it is often a feature or aspect of an item that attracts user,user will assign various attention to each feature of an item.In addition,many methods do not effectively model the complex relationship between items in user historical behaviors.This paper proposes a new recommendation model based on gating mechanism to solve above problems.Specifically,we map users and items to a latent vector space and embed them.Then we introduce a pairwise encoding layer to construct a 3-way tensor,which is utilized to store the relations between items in user interaction history.We also construct two gating layers to filter item features,so that the model can effectively model transition relationships and capture user’s short-term preferences.In addition,the user embedding is also used to model user’s long-term preferences.Finally,we make recommendation by combining the user’s long-term and short-term preferences;2)We propose a sequential recommendation algorithm based on social network(SASE).Cold start problem and sparse dataset are inevitable problems in recommendation system.The way to solve these problems is not to predict them in isolation,but to leverage additional signals of user’s historical behavior.Therefore,we can improve the performance of recommendation model in the cold start scenario by using sequential data of user actions and social network.The core idea of the proposed model is to find user’s intimate friends based on social network and determine which items are related to the user’s next behavior in item sequences,and then make prediction with these features.Specifically,we utilize convolutional filters with different sizes to capture sequential patterns in user interaction sequence,and learn users’ social features from social networks through graph embedding model,so that SASE can be better at dealing with the cold start issue in the recommendation system by fusing two features;3)In order to verify the effectiveness of HPGM,we compare the performance of HPGM and seven baselines on Amazon and other real-world datasets,and conclude that the proposed model can efficiently mine user preferences;Secondly,we construct the corresponding cold start datasets on the basis of three recommendation datasets for proving the feasibility of SASE in the cold start scenario.By comparing the AUC of SASE and other baselines on three cold start datasets,we can conclude that our method can effectively work in the cold start scenario.Finally,we also discuss the hypeparameters of the proposed model.To sum up,this paper focuses on the sequential recommendation system and introduce deep learning technology into the recommendation system for achieving better performance.At the same time,we conduct a series of comparative experiments and the metrics indicate the validity and feasibility of the proposed model.