Research On User Behavior Prediction In Short Videos Based On Multi-task Learning

At present,short video has become one of the most mainstream ways of Internet content dissemination.With the gradual maturity of the market and the gradual stabilization of the competitive landscape,the growth rate of users has slowed down,and the dividend period for users has gradually faded.The focus of competition has gradually shifted from the incremental market to the stock market.How to improve the activeness of existing users and dig deeper into the value of each user is the current short video plat-form attaches great importance.Therefore,improving the performance of user behavior prediction in short videos around the multi-task learning model is of great significance for short video platforms to explore more diversified and deeper business realization models.Based on the multi-task learning model,this thesis conducts research on improving the performance of user behavior prediction in short videos.The work done includes the following aspects:First of all,this thesis proposes two adaptive optimization algorithms GW-DC(Adam with weighted gradient and dynamic bound of learning rate)and A-MSGWDC(AMSGrad with weighted gradient and dynamic bound of learning rate)based on gradient weighting,and gives proof of the convergence of the algorithms.By introducing the dynamic decay rate in the first-order moment estimation of the gradient,more memory is concen-trated on the recently gradient in the first-order moment estimation of the gradient,thereby effectively constraining the proportion of each compo-nent in the first moment estimation of the gradient.In order to verify the effectiveness of the proposed algorithms,we conducted experiments on three public data sets and five network structures of different complexity.The experimental results show that the two adaptive optimization algo-rithms based on gradient weighting proposed in this thesis not only have better generalization performance,but also have faster convergence speed.Secondly,this thesis also proposes a feature expression method based on attention mechanism.This method trains a set of weights by transform-ing the input features at the dimensional level,and does a dot product of the set of weights with the original expression of the feature to obtain a new expression of the feature.In order to verify the effectiveness of the proposed method,we conducted experiments based on the data provided by a short video platform and established two binary classification models.One is to predict whether the user will finish watching a certain video,and the other is to predict whether the user will like a certain video.The exper-imental results show that the proposed method is helpful for the improve-ment of the evaluation index in these two prediction tasks.Finally,in order to improve the performance of MMoE(Multi-gate Mixture-of-Experts)for predicting user behavior in short videos,this thesis redesigned the network structure of the MMoE model,and encapsulated the feature expression method based on the attention mechanism into a sub-layer as a module of the MMoE model.And using the gradient-weighted adaptive optimization algorithm GWDC proposed in this thesis as the op-timizer,a multi-task learning model is built on a short video data set to study user behavior in short videos.The experimental results show that the proposed method can improve the performance of MMOE in predicting users' behavior in short videos,so as to enhance the user stickiness of the short video platform,and provide strong support for the short video plat-form to tap user value and explore deep-seated business cash flow mode.