Deep User Interest Evolution Model Based On Graph Data Augmentation

The increasingly huge data resources have promoted the prosperity of the digital economy,and the precious information in the data can effectively improve the operating efficiency of various industries.But it is very difficult to mine out the real and effective information from the huge data.Click-through rate prediction is the task of mining information from massive user data.It is necessary to analyze the probability of the user’s next click.This requires accurate matching of users and targets,the most accurate recommendation,and the elimination of information overload influence.And then it has become a hot spot in the field of user behavior analysis.At present,many mainstream click-through rate models assume that the user’s click behavior is determined by the underlying interest behind it.They want to extract user interest through the user’s click sequence data,and then predict the user’s next click from the extracted interest.Behavior.However,due to the nature of the click prediction task data,the data volume of a single user’s click sequence data is very small,and it will be inaccurate to deduce the interest latent variables behind it based on the user’s click behavior from such data.In addition,in the current mainstream click prediction tasks,the embedding layer of the model is often initialized with a random value or a single value,so that the final trained embedding will lose a lot of data information in order to balance the overall click prediction model.In response to the above two problems,this paper proposes a new graph sequence data augmentation method(GSDA),and an improved negative sampling embedding algorithm(NegMC,Negative Markov Chain),and combines these two methods to the user In interest evolution network,a new paradigm of click prediction is proposed.The final experiment proves that the GSDA method and NegMC method proposed in this paper can improve the final AUC accuracy of most mainstream click-through rate prediction models.Based on this result,this paper compares the proposed method with the latest model,and the method proposed in this paper can perform better than the latest large-scale parameter click prediction model on a smaller-scale data set.In addition,the GSDA method can also increase the time for the model accuracy to reach the inflection point to a certain extent and prevent the model from entering the local optimal value.Small and medium-sized data volumes have the ability to surpass complex models.