Research On Click-Through Rate Prediction Methods Based On Behaviour Sequences

With the rapid development of the field of recommender systems,click-through rate prediction has been widely studied by scholars.Existing click-through prediction models lack attention to embedding techniques in data pre-processing and do not adequately consider the importance of combined feature information in feature interaction.In addition,most of the click-through rate prediction models ignore the impact of click behaviour sequence relationship on the prediction performance of the model,and cannot predict the click-through rate of users on target products more accurately.To address these problems,this paper proposes a click-through rate prediction model based on behavioural sequences.Firstly,the traditional embedding method and Auto Dis embedding method are used to pre-process different types of input features respectively,so that they can better retain the information contained in the original features;then,for historical behaviour sequences,a gated cyclic unit with an attention mechanism is used to process them,learn the dependency relationships within the sequences,simulate the dynamic changes of users’ interests with the current click target,and build a user behaviour;Finally,feature crossover is performed using a deep crossover network incorporating attention mechanisms,which effectively distinguishes the importance of combined feature information.In this paper,the model is based on three publicly available datasets,namely Movielens-1M,Amazon and Taobao,and is analysed in terms of training and click-through rate prediction results.The experimental results show that the proposed model has the best prediction results when the activation function is Relu,the optimizer is set to Adam,the embedding vector dimension is 8,and the number of gated recurrent neural networks is 128.Compared with other models,the model in this paper is more outstanding.Compared with the base model DCN,the AUC values improved by 0.9%,1.5% and 0.8% respectively,and the Logloss values decreased by 0.9%,1.7% and 0.4% respectively.The paper has 37 figures,8 tables and 54 references.