Research On Top-N Recommendation Based On Deep Neural Network

Today is the information explosion era of the development of Internet.Various forms of information are waiting to play their value.People are in the contradiction between "information overload" and "information cocoon room".The recommendation algorithm will analyze the known historical data,achieve personalized data filtering,and provide users with a way to obtain effective information.As an increasingly popular recommendation form,top-N sequence recommendation improves users' sense of participation and acquisition.At the same time,with the development of deep neural network,top-N recommendation based on deep neural network has become an important research direction in the field of recommendation.Taking the top-N recommendation of deep neural network as the background,this paper starts from the perspective of users and projects,pays attention to the way of feature extraction and combination,integrates it into the deep network and predicts the potential preferences of users,and mainly does the following two aspects:1.By analyzing the depth neural network recommendation algorithm,it is found that the existing algorithm ignores the influence of the correlation between features on the depth network and is limited by the expression of features,which makes it difficult for the model to achieve good recommendation effect.Aiming at this deficiency,this paper proposes factorization machine deep network(FMN).Based on the embedded features and the idea of factorization machine,the algorithm interacts between the embedded feature elements to obtain the decomposition machine vector,which can represent the second-order cross features,and then learns the nonlinear implicit features through the multi-layer perceptron.Experiments were conducted on real movie datasets and store recommendation datasets.Compared with the baseline algorithm,FMN improved in the evaluation indicators of accuracy and ranking.2 In the face of the data set with too large difference between the number of users and the number of items,in order to obtain the embedded vector with appropriate length and avoid the problem of underrepresentation and over representation in the traditional matrix decomposition algorithm,this paper proposes an adaptive matrix factorization(AMF).The algorithm obtains the embedding vector suitable for the data set in the embedding layer.There is no information loss in feature extraction,and the number of useless parameters is reduced.In order to obtain the adaptive vector with the same length,the decomposition machine vector is supplemented in the adaptive combination layer,and the vector has the correlation feature at the same time.In the comparative experiment,the image recommendation data set with the ratio of the number of users to the number of items of 5.56 is selected.Through the analysis of the experimental results,it is concluded that AFM algorithm has better recommendation performance than other algorithms.