A PSO-based CNN Algorithm For Keywords Selection On Google Ads

Google Ads is a paid online advertisement platform which allows users to select diversified types of ads including text,graph and video ads by entering different keywords or searching for what the users really want on Google.com.From this point of view,many advertisers use Google Ads and its functional advantages to serve ads.However,in the face of the wide variety of functions and algorithms of Google Ads,advertisers cannot quickly choose the optimal algorithm for themselves.But it was found in the experiment that if an advertiser can pick out the most appropriate information in the advertisement(the keyword of the advertisement),Google Ads can be used to reduce the cost of the advertisement and improve the effectiveness of the advertisement,thereby increasing the revenue of the advertisement.However,it seems much more difficult than we thought when conducting the practical keyword selection tasks due to the following possible problems.Firstly,we may face the problem of natural language processing.Since keywords,in many cases,tend to be a mixture of different languages.For instance,we are used to entering keywords containing both Chinese characters and English words.Consequently,it will give rise to the issue of imbalance of data quality owing to quite few features of that type of keywords.In addition,it is not easy for classifiers in the deep learning field to be well matched with our case in this thesis.More importantly,for the models in deep learning,they usually can hardly fulfill the requirements of fast classification speed and simultaneously satisfying accuracy,as there are many models' parameters waiting to be adjusted,which may be the major challenge for us to consider.To tackle the issues mentioned above,first of all,it requires us to study how to extract text information and improve the performance of classifiers.To be specific,this thesis intends to solve the problem of mixed language words by choosing corpus;address the issue of few text features by the means of embedded words;deal with the problem of imbalance of data quality.The experimental results prove the feasibility and reliability of these proposed solutions.Then convolutional neural network(CNN)is deployed as a classifier to select qualified keywords.However,to better apply in our case,we aim to use the corresponding behavioral information such as clicks,rankings,and transition rates in the keywords as features,and the association features are embedded in the last layer of the convolutional neural network.In this fashion,the improved convolutional neural network will be more suitable for this scenario.Afterwards,due to the fact of an enormous number of parameters in the convolutional neural network,we also did some research about how to choose the most suitable neural network structure from many parameters.Through experimental and theoretical research,it shows that by combining evolutionary computing with deep learning,taking precision as the target,and using a single-objective particle swarm optimization algorithm to optimize the structure of the convolutional neural network,it can successfully introduce an appropriate neural network structure without much manual parameters adjustment.Finally,the study shows that merely obtaining a network structure with high accuracy cannot meet the needs for our case.As in many industrial applications,they tend to be required to not only care about the accuracy but care more about their time cost.Therefore,after further research,it is found that the multi-objective particle swarm optimization algorithm can be used to optimize the network structure with accuracy and speed being as two goals,and can meet the needs of these two goals at the same time.The experimental results demonstrate that the suggested algorithm is an effective one in terms of well dealing with our case and possibly being used in more practical cases.