Research On Recommendation Algorithm Based On Deep Walk

With the wide application of recommendation algorithm in the scenes of video push and e-commerce,more and more scholars have paid attention to the research of recommendation algorithm.Meanwhile,the recommendation algorithm faced many problems during its development,including the scalability problem caused by the large dimension of the user-item rating matrix as well as the sparsity problem caused by the lack of user's rating information.To address the above problems,a recommendation algorithm based on Deep Walk is proposed in this paper.Specifically,the research details of this paper are as follows:This paper introduces the theoretical background of random walk.Besides,the measurement methods of network's node similarity based on random walk have been analyzed from their definition and application.This paper points out that these methods can measure the similarity between nodes using the neighbor structure information of each node.This paper builds a user undirected graph with the rating information of every user.The truncated random walk algorithm is used to capture the local structure information in the user undirected graph,which effectively improves the scalability of the algorithm.The truncated random walk algorithm is applied to generate the random walk path on the user undirected graph.The node sequence on the path can capture the neighbor structure information of the node more accurately.At the same time,the truncated random walk is easy to parallelize in different processes,which is conducive to be applied to large graphs.Skip-gram model and Glove model are applied to calculate the low-dimensional implicit representation of every user in the path of truncated random walk,which can effectively solve the sparsity problem of the user-item rating matrix.These two models can accurately map the structure information of user undirected graph to a low-dimensional space.The relative position of users in euclidean space can be measured by the users' vectors calculated by the model.In this paper,we use cosine distance to calculate the interest similarity between users.Moreover,the rating information is used to calculate the recommendation list for every user.This paper applies the proposed method to movie recommendation scene.The performance of the algorithm is verified on commonly used evaluation.The results show that the proposed method outperforms the collaborative filtering recommendation algorithm in terms of accuracy,recall,coverage and popularity of movie recommendation task.The ability of the algorithm about alleviating the cold start problem is also verified by applying it in different proportions of the training set.