Research Of Recommendation System Model Based On Deep Learning

With the rapidly development of science and technology today,the amount of data in the internet is growing exponentially,and the phenomenon of information overload makes it difficult to achieve efficient access to resources.With the continuous development of science and technology,the channels for obtaining information are constantly updated.From letters to newspapers in the past,to the internet,more and more people regard the Internet as the primary way to obtain information.Recommendation system plays a vital role in the process of collecting data on the internet.It can effectively filter massive information and recommend resources that meet users’ needs.In recommender systems,click-through rate prediction(advertising recommendation scene)and TOP-K recommendation(product recommendation scene)are the most important parts of ranking stage.The click-through rate prediction and TOP-K recommendation models in previous studies have problems such as a large amount of calculation,high memory consumption,long feature screening time,and insufficient feature interaction,which can no longer meet the needs of recommendation tasks in current life scenes.Under the above background,this paper studies the recommendation algorithm based on deep learning,and proposes a recommendation system model based on separated embedding interaction network and a recommendation model based on self-attention.The main work of this thesis is as follows:(1)Aim at those problems that the feature interaction method in the existing deep learning recommendation model cannot fully utilize the information of embedding vectors and the lack of accuracy,a deep learning recommendation model based on separated embedding interaction networks(SEIN)is proposed.First,the model transforms the sparse feature vectors into dense embedding vectors by using embedding neural network layers;then,separates the features matrix of different dimensions for feature interaction,and explicitly controls the order of feature interaction by separated embedding interaction network layers;finally,sum pooling the obtained matrix of each hidden layer,and the final output is obtained by the prediction layer.The AUC(Area Under the Curve),accuracy and recall of the recommendation results are used as evaluation metrics on three publicly available datasets.The results show that the based on SEIN recommendation model better performance than the existing six typical clickthrough prediction models and five based on graph neural network recommendation models.(2)Existing deep learning recommendation algorithms can effectively solve problems of data sparsity and high-order feature interaction,but there is still the problem of insufficient utilization of global feature vector information.To solve this problem,this paper proposes an improved recommendation model CIN-SA that combines self-attention and compressed interaction network(CIN).The model uses the compressed interaction network for explicit high-order feature interaction,and trains the original feature vector through the attention module,making full use of the global feature vector to enhance the representation ability of the model.Experimental verification is carried out on the Criteo and Auto ML datasets with AUC and Logloss of recommendation results as evaluation indicators.Experimental results show that the proposed model has better recommendation performance than popular recommendation models.Based on the research of deep learning recommendation system model,this paper propose recommendation models based on separated embedding interaction network and self-attention.The proposal of two models aims to solve the problem of insufficient information utilization and not utilizing the global relationship between feature vectors in the feature interaction stage of the recommendation model.Through experimental comparison in three datasets and two scenes,the experimental results show that recommendation system models proposed have better performance than current popular models.