Research And Implementation Of CTR Prediction Model Based On Deep Interest Network

CTR prediction is a key link in the recommendation system and online advertising platform.As a key technology,CTR prediction which directly affects the company’s revenue and improves user experience has become the focus of industrial and academic research.Recently,artificial intelligence technology has made breakthrough in many fields,such as search engines,computer vision,and natural language processing.Various research institutions and Internet enterprises have applied related technology to CTR prediction and achieved many outstanding results.With the analysis and research on CTR prediction models published recently,we found that there are two main challenges in CTR prediction: First,the input features of CTR prediction models are characterized by large scale and high dimension.Therefore,how to capture effective information from these feature interactions becomes the key to improve the accuracy of CTR estimation.Second,user behavior sequence contains implicit and explicit feedback information.How to accurately mine users’ interests from the sequence is also the focus of many researchers.Based on the feature interaction and user behavior sequence modeling,this paper improves the existing model to improve the accuracy of CTR estimation.The main contribution and achievements of this paper are as follows:（1）Based on the classical Deep Interest Network（DIN）model,we proposed Deep Interest Factorization Machine（DIFM）model.DIFM has two improvements on the basis of DIN.First,Transformer structure is used in DIFM to model user behavior sequence.Compared with the activation unit with target attention mechanism in the DIN model.Transformer structure can better express user interest through multi-head self-attention mechanism.Second,factorization machine（FM）module is added on the basis of DIN model to capture low-order feature interaction information through FM,so that the model can simultaneously take into account low-order and high-order feature interaction information.Experiments were carried out on Movie Lens and Amazon data sets,and the performance superiority of DIFM model proposed in this paper was proved through comparative analysis,and the effectiveness of Transformer and FM modules was proved through ablation analysis.（2）This paper proposes a Deep Multi-Interest Network（DMIN）model,DMIN has two improvements on the basis of DIN: First,a two-layer Gated Recurrent Unit（GRU）is used to model the user behavior sequence.The first layer of GRU helps express preliminary interest of users,and the second layer of GRU combines the target attention mechanism to express the user’s final interest.Through the GRU,the dependencies of each behavior behind the behavior sequence can be considered,making the model expression ability stronger.Second,the target attention mechanism is used to process contextual information,user portraits and other features,and then the fine-grained feature information is used to estimate click rate.Experiments are carried out on Movie Lens,Amazon and Userbehavior data sets.Comparative analysis proves that DMIN model has better performance than DIN model,and ablation analysis proves that each module in DMIN model has different degrees of influence on model performance.