A Recommendation Method Combining Network Embedding And Collaborative Filtering

With the development of the times,the rapid development of Internet technology,the problem of Internet and big data is the low utilization of information due to information overload.In this case,the research of recommendation system gradually attracts people's attention.The recommendation system can recommend to users what they may be interested in.By further exploring the user's behavior,understanding the personalized needs of different users,and recommending a large number of long tail products to users who may be interested in it,it not only saves the user's time cost,but also brings more benefits to the website itself using the recommendation system,and even can recommend to users some products they are interested in but it's hard to find them.Since the development of recommendation system,there are three kinds of technical methods,from the initial content-based model to the collaborative filtering model,and then to the hybrid model of the first two methods.The collaborative filtering algorithm is the most classical and widely used type of recommendation algorithm,which is the most mainstream recommendation technology at present.Its main advantages are simple,easy to understand,easy to implement and good recommendation effect.What the collaborative filtering algorithm needs is the interaction behavior data between users and products,in which the potential similarity between users and projects can be found,and then complete the recommendation based on the similarity of this group.Collaborative filtering algorithm not only has a lot of research in academia,but also has a lot of extensive applications in industry.There are two main collaborative filtering methods,one is based on neighborhood,including userbased method and project-based method,the other is based on model,including matrix decomposition,Latent Factor Model and some collaborative filtering algorithms based on deep learning in recent years.In the development of recommendation system,there are also many improved algorithms for the above methods,mainly through better use of some external information,such as user attributes,tags,etc.,to enhance the effect of recommendation or to solve the cold start problem,the acquisition of external information often needs additional costs.The research of this paper only focuses on the interaction behavior data between users and products,so the problem of cold start is not considered in this paper.In the absence of external information,it is difficult to mine the characteristic attributes of users and products themselves.The information mainly comes from the behaviors such as purchase between users and products.This characteristic is similar to the complex network in which the amount of information of nodes is small and the amount of information is concentrated on the association between nodes.The traditional collaborative filtering method is too direct and simple to establish the indirect and deep relationship between users and products.In this paper,after building a complex network,the data is processed by network embedding method,and the information between users and products is mined to improve the traditional collaborative filtering method.Through the interaction behavior data between users and products,i.e.the historical behavior of users,the network with users as nodes and the network with products as nodes are constructed respectively.Then,the network embedding method node2 vec,which is considered as the most efficient and accurate way to encode network nodes so far,is used to encode nodes.In the neighborhood based collaborative filtering method,the original similarity calculation method can be replaced by the vector representation of nodes(users or products);in the LFM,the vector representation of nodes(users or products)is explicitly added to the matrix decomposition.The empirical research of this paper selects Movielens and Tencent Weibo users' attention data set.Among them,movielens is an explicit feedback data set,which contains the user's specific ratings for the movie.We use it to compare the differences between node2 vec combined with collaborative filtering method and the traditional method in the rating prediction task.Tencent Weibo user focus data set is an implicit feedback data set,which only contains the user's attention behavior.We use it to compare the methods in the Top N recommendation task.Through the experimental research,the improved method in this paper is slightly worse than the traditional method in the explicit feedback,the reason may be that the construction of network edge weight fails to make full use of the information;in the implicit feedback,it has a certain improvement compared with the traditional method.Considering that there is not much research on the parameter adjustment of node2 vec in this paper,it can be considered that the collaborative filtering method combined with node2 vec is effective.The empirical results of this paper show that it is feasible to improve the traditional collaborative filtering method by constructing a complex network and using the network embedding method to obtain the vector representation of nodes.