Research On Recommendation Algorithm Based On Attention Mechanism

With the arrival of big data era,various kinds of information technologies is boom-ing,all walks of life have achieved informatization,and people's lives have also been greatly facilitated.With the explosive growth of data volume,finding the information that is needed in the large amount of data has become a problem that plagues users.There-fore,the recommender system as an information filtering method to effectively solve the information overload problem came into being.In recent years,recommender system technologies have been wildly studied.Among them,the graph neural network based recommendation model is a hot study topic in re-cent years,but the existing recommendation models still have some shortcomings to be improved.First of all,the existing graph neural network based recommendation models often directly pass the information of all neighbors in the process of information passing,without filtering out the more important information for the current node,which weak-ens the expressive ability of leraned embedding vectors of users and items.Secondly,in session-based recommendation,existing models usually treat an item as a whole,with-out mining the deep information of feature granularity? at the same time,the existing session-based recommendation model often use attention mechanism only in a single la-tent space,which leads to insufficient modeling of the session sequence.This thesis car-ries out in-depth research work,corresponding new recommendation models to solve the above-mentioned shortcomings are proposed,and the effectuality of the proposed recom-mendation models is verified through extensive experiments.The research work of this thesis is mainly as follows:1.The scientific research literature in related fields such as graph neural network,convolutional neural network,attention mechanism is investigated.This thesis in-depth studies existing recommendation models,analyses existing researches' problems and short-comings for improvement.2.For purpose of dealing with the insufficiency that the information passing algo-rithm of the existing graph neural network based recommendation models usually treat all the neighbors of the current user or item equally without filtering out the more important information for passing,this thesis proposes a new graph attention based message propaga-tion algorithm abbreviated as GAMP.The GAMP algorithm uses the attention mechanism to treat the neighbors of the current user or item differently,focusing on the more impor-tant information for the current task,so as to better learn the embedding representation of users and items.3.Based on graph neural network and GAMP algorithm,a new graph attention based deep neural network model named GADN is proposed.The model captures the complex interaction between users and items by explicitly modeling high-level user-item relation-ships,and updates the users' and items' high-level embedding vectors through the GAMP algorithm,thereby effectually models user interests and item features,and therefore im-proves the accuracy of recommendation.4.Aiming at the problem of insufficiency to learn deep features of a session sequence in the existing researches,the convolutional neural network is introduced into the session-based recommendation,and the horizontal and vertical convolution module is proposed.By splicing the embedding sequence of the items recently clicked by the user into a matrix,deep features between adjacent items in the session and deep features of the relationship between items at feature granularity are captured through different types of convolution kernel.And further combining the gated graph neural network and the multi-head atten-tion mechanism,a multi-head attention convolution graph neural network session-based recommendation model named MACG is proposed,which improves the performence of session-based recommendation.5.This thesis carries out extensive experiments and analysis on multiple real datasets with a variety of advanced recommendation models,verifying the effectiveness of the model proposed in this thesis.