Advertisement Recommendation Based On Sequence Feature Extraction

With the continuous progress of science and technology and the rapid development of information technology,the Internet has become an indispensable part of people's material and cultural life,and has long penetrated into every aspect of people's lives.People can use the Internet to do almost everything they want to do such as entertainment,shopping and learning.With the arrival of the 5G era,new kinds of APPs such as various short video APPs and news APPs are gradually becoming tools for people's ordinary leisure and entertainment.There are rich business opportunities here,many businesses put advertisements inside the APP,live commerce,attracting this audience to browse,click and buy various goods.At the same time,people's browsing,personal preferences and other information will be recorded,Internet companies will use this information to make accurate recommendations,the accuracy of the recommendation will directly affect the income of Internet companies.How to realize the accurate recommendation of advertisement has attracted a lot of attention from academia and industry.CTR is an important metric for evaluating the accuracy of advertisement recommendation.The LR model is used to estimate CTR in the earliest stage of advertisement recommendation.The features used are all original features or manual features.In order to better capture the intersecting features,the FM model was used to add a pair of feature combinations.With the development of deep neural networks,the use of deep networks to extract abstract features for recommendation has gradually become the trend of advertising recommendation,such as FNN and SNN networks.Later,in order to preserve the memory of the network,low-level features plus highlevel features gradually became the mainstream of prediction,such as Wide & Deep network and Deep & Cross network.In order to better extract the features and better combine the features,this thesis establishes sequence and relationship diagrams for the products purchased and clicked advertisements by the user,and uses the Deep Walk and Word2 Vec algorithms to extract the embedding vectors from the sequence and the relationship diagram respectively.The obtained feature vector can better represent the characteristics of the advertisement and the product,and then use the self-attention mechanism to extract the feature vector of the sequence,and merge it with the network used for feature extraction of the original advertising information and user information to form a complete prediction network.In summary,this thesis studies the advertising data,uses Deep Walk,Word2 Vec and other algorithms to extract and predict features,and verifies the improvement of the effect of model through comparative experiments,which can prove that the extraction of sequence features through the attention mechanism can be effective improving the results of the forecast is of instructive significance for the industry to place advertisements.