Research On Advertising Recommendation Technology Based On Machine Learning

With the rapid development of the Internet,the scale of the online advertising market has expanded rapidly,and display advertising has become the most popular means of promotion.Besides,precise ad recommendation is the guarantee of Internet platform revenue,and accurate ad click-through rate prediction is the prerequisite for accurate recommendation.According to the requirements of update rate and real-time of advertising platform,ad recommendation technology can be divided into offline and online categories.Offline ad recommendation is mainly based on the construction of models,through which complex correlations between user and product features are mined.And that online ad recommendation is mainly based on real-time feedback from users,through which feedback information is collected and recommendation strategies are changed instantly.In this thesis,two types of ad recommendation technologies,i.e.,offline and online,are studied.The main research contributions and innovations are as follows:(1)Movie Lens Data Mining Analysis.Based on the tensor analysis method,the original highdimensional features of users and movies in the Movie Lens dataset are processed and transformed to generate a low-dimensional feature dataset suitable for online recommendation system application.On the other hand,based on the T-SNE(T-distributed Stochastic Neighbor Embedding)dimensionality reduction technique,the work on the visualization of the dataset reveals that user feedback presents complex relationships with product characteristics,and traditional online recommendation algorithms such as the generalized linear model proposed by Lin UCB are unable to fit such complex feedback interactions.(2)For offline advertisement recommendation technology,an advertisement click-through rate prediction model on the basis of attention mechanism and neural network,called CAN(CTR prediction algorithm based on Attention Machine and Deep Neural Net Works),is proposed.Additionally,to address the lack of ability of traditional models to mine higher-order feature interactions,the method uses deep neural networks to train the original features,which improves the model's ability to fit the nonlinear relationship between features.Afterward,to alleviate the problem of difficult neural network training,the method proposes to construct a low-order feature intersection layer in front of the neural network layer through an attention mechanism,which provides richer feature interaction information for the neural network layer compared to the original features.Ultimately,the method is demonstrated by simulation experiments on Movielens and Criteo datasets to outperform traditional methods such as FM(Factorization Machine)and PNN(Product-Based Neual Networks)in two metrics,Logloss and AUC(Area under curve of ROC).(3)A Gaussian process(GP)-based real-time recommendation algorithm is proposed for online advertising recommendation technology.Specifically,to address the problem of generalizability in the traditional Lin UCB online recommendation algorithm that assumes a fixed linear relationship between the item features and user feedback,a Gaussian process is utilized to map the item features to the predicted user feedback.In addition,the Gaussian process is model-free,which converts specific functions into distributional representations and can fit various complex functional relationships,thus it can solve the problem of generalizability of specific function assumptions in various environments.For the common exploration-exploitation(EE)problem of recommendation systems,two specific recommendation strategies based on upper confidence interval as well as expected improvement are put forward,which are respectively called GP-UCB(Gaussian ProcessUpper Confidence Bound),GP-EI(Gaussian Process-Expected Improvement).Moreover,a method is proposed to simulate the users' online real-time interaction with platform using offline datasets.Then,thorough experimental results demonstrated that the two recommendation strategies outperform traditional schemes such as UCB(Upper Confidence Bound)and Lin UCB in terms of cumulative regret.