Meta-Learning Based Click-through Rate Prediction Model

With the booming development of network technology,the Internet has profoundly influenced people's production and life,and it has become especially important to find valuable information from the massive data.Internet advertising is growing rapidly and has accounted for about 80%of the overall advertising market.Ad bidding needs to predict the expected value in advance,which involves click-through rate prediction model.The goal of the click-through prediction task is to predict the likelihood that a user will click on a certain product and thus improve the success rate of recommendations.On the one hand,it can achieve economic effect and generate revenue for the company enterprise,on the other hand,it can improve user experience,improve user stickiness and sustainable development.Existing research work mainly starts from the user's perspective,predicting the user's preference for various types of advertisements,and thus anticipating the revenue after the advertisement is placed.Such a perspective has many shortcomings in the actual operation of real industrial environments:only the association between users and advertisements is considered,ignoring the factor of scenarios,and both users and advertisements in different scenarios will show different tendencies,thus leading to different results;the association between scenarios is difficult to be rated straightforwardly,and they may show positive,negative or irrelevant trends,which cannot be simply determined in the form of coefficients and so on;the new scenarios or The problem of small scenes,the sample size of new scenes or small scenes cannot support the network to train a reliable model for them alone,and in the actual production environment,due to the restricted resource conditions,separate training requires both offline and online resource services,and the adjustment of parameters requires a lot of manpower to support,which also leads to the dilemma brought by rising costs.To address the above challenges,the research in this paper revolves around the task of advertising scenario click-through rate prediction.For the problems brought by scene factors,this paper further deals with the solutions with the help of meta-learning methods.Through meta-learning,we propose a framework suitable for the click-through rate prediction task,equipped with a variety of deep learning networks,and further improve the model effect for new scenes or small sample scenes with the help of some large-capacity mature scene data.The main work of this paper includes:1.Research and analysis of advertising click-through rate tasks.With the help of multi-scene data sets of ad click-through rate prediction tasks from the largest ad monitoring platform in China,we comprehensively analyze the impact of ad scenarios on click-through rate prediction tasks.2.Propose a click-through rate prediction model based on meta-learning.Traditional models are trained to learn features by training data of a scene individually,and the models trained by this method require a high amount of scene data and consume large resources in the context of multiple scenes.In this paper,we propose a meta-learning-based approach to equip the existing model with a metalearning framework to reduce the training cost and overhead of model adaptation to scenes and achieve accurate performance in new scenes.Its effectiveness is demonstrated through multiple experiments on two datasets.3.Propose a click-through rate prediction model for learning vector representation.To address the problem of learning ID vector representations that require a large number of samples,we propose to combine product features and historical user behavior to accelerate learning embedding vector representations using new integrated representations.Experiments are conducted to demonstrate its feasibility with the help of advertising scenario dataset.