Research On Personalized Recommendation Algorithm Based On Deep Learning

With the advent of the Internet era,the frequency and duration of people ’s use of smart phones has increased,and the amount of user information they generate has exploded,leading to more serious information overload problems.How to solve the problem of information overload and maximize the value of data is one of the challenges facing the era of big data.How to help users get the products they are interested in in the shortest time,and help companies sell more products to achieve the purpose of increasing revenue is another problem in the era of big data.In order to solve the above problems,the research on personalized recommendation algorithms has been favored by experts in this field.In order to solve the above problems and improve the accuracy of the recommendation system,this paper constructs a new personalized recommendation algorithm.It first comprehensively discussed the Logistic regression algorithm,GBDT algorithm,FM model,FFM model and the latest deep learning-based neural networks and other algorithms,and adopted several methods for fusion,followed by the addition of two different DIN and CENet Attention mechanism,at the same time,construct the user’s historical data.Finally,a hybrid model is designed,which has better performance than the mainstream model.It is based on user characteristics and advertisement characteristics,adding historical click data of the user to the personalized recommendation algorithm,so as to predict whether the user clicks on an advertisement to determine Advertising strategy for users.There are four improvements in the newly constructed recommendation algorithm in this paper.One is to accurately describe the user’s interest in different advertisements,adding an attention mechanism,which can make the user characteristics and advertisement characteristics into a mapping that depends on historical click data,thereby improving the robustness of the model.This article introduces two different attention mechanisms,from the simple interaction of user features and advertising ID features to the interaction of user features and advertising features.The second is to explicitly learn the interaction relationship of different features.In this paper,we use the FFM method to explicitly learn the second-order interaction relationship of different features,and embed the Multi-Hot encoded features into a low-dimensional dense feature vector for processing.Conducive to efficient operation of model learning.The third is because deep neural networks can only learn high-level interactions of different features invisibly,so we use PNN-based methods to artificially use second-order feature interactions as input to deep neural networks,and have also tested the first-order and second-order interactions.The vectors of the first-order interaction relationship are replaced with different embedding vectors to improve the generalization ability of the model.The fourth is to expand the original data to improve the performance of the model,that is,the user ’s actual click data is modified as the user ’s historical data,and the process of modifying the click data is to use the advertising ID and advertising characteristics as new inputs to the modified model.Thus,the performance of the model is improved.In general,the personalized recommendation algorithm designed in this paper has improved in accuracy and effectiveness.It can more efficiently and accurately advertise users,and guide users to consume quickly and accurately.While reducing costs,you can get more benefits.