Research On Key Technologies Of Personalized Recommendation System Based On Machine Learning

In recent years,with the rapid development of information technology and Internet technology,data growth rate far exceeds human’s cognitive ability,which inevitably resulted in information overload.To solve this problem,recommendation systems actively capture the potential matching relationship between users and products utilizing artificial intelligence technology,having occupied extensive attention from academia and industry in recent years.However,most of the state of art recommendation algorithm modeling in a onesided way,resulting in limited performance,which is reflected in the following aspects: 1)based on implicit factor modeling,ignoring the auxiliary role of content features in decisionmaking process;2)modeling user in an incomplete way,such as only considering the user itself or user interest modeling;3)regarding the user as static and passive,ignoring that the recommendation results have impact on users and user interest changing over time.In order to solve the above problems,this thesis studies the key technologies of personalized recommendation system based on machine learning.First,a personalized recommendation algorithm based on multi-source and dynamic user portraits has been proposed and implemented based on deep learning.The personalized recommendation network consists of three modules including multi-source feature representation,multi-source dynamic user portrait generation and matching of users and products: In terms of features,the content embedding and the latent vector representation of users or products are fused into multi-sourse feature representation;in terms of user portraits,the user’s innate character and user interests are both modeled and expressed,and the multisource user portraits are generated dynamically according to the different candidate products;In terms of matching,to model more complex relationships,the multi-layer forward neural network is applied instead of inner product to match users and products.Experimental results show that compared with the state of art mainstream models,the model proposed is 2.21% higher than the NAIS model on Hit Ratio(HR)and 3.19% higher than the NAIS model on Normalized Discounted Cumulative Gain(NDCG).Next,this thesis studies recommendation algorithm based on deep reinforcement learning is studied,proposing a deep deterministic policy gradient online recommendation algorithm based on stratification experience replay.It is proposed to combine the Deep Deterministic Policy Gradient(DDPG)algorithm with online recommendation,making the recommended network parameters automatically adjusted according to user feedback to maximizing longterm rewards.To cope with the sparse reward situation which is common in recommended scene,a double memory(D-Memory)structure and a stratification experience replay mechanism is proposed,solving the imbalance between the quantity and quality of data in the replay memory,improving the optimization efficiency of the agent,and thus improving the long-term rewards obtained by the agent.Finally,based on the above algorithm,applying the ideas of multiple features and multiple user portraits,an adaptive marketing system is implemented.Experimental results show that the algorithm proposed can provide a fair initial value for the recommendation network,and that the reward density and long-term reward obtained by the agent are increased by more than 60% in the sparse reward scenario.This thesis studies the key technologies of personalized recommendation system based on machine learning.Related research results can be used in application scenarios such as intelligent marketing system and smart travel recommendation system.