Research On Multi-Objective Optimization Recommendation Algorithm Based On Deep Learning

With the rapid development of the Internet and the advent of the era of digital intelligence,personalized recommendation has been widely used in various recommendation scenarios such as commodities,videos,music,movies,and short videos,and has played a crucial role.However,how to further improve the accuracy of recommendations and bring a higher sense of user experience has been the focus of relevant research.The prediction model commonly used in existing recommendation algorithms based on deep learning is targeted at users’ clicks,which does not take into account other behaviors generated by users’ clicks,thereby reducing the interest network and reducing user satisfaction.Moreover,there are problems such as sparse samples,unbalanced use of explicit and implicit feedback,"seesaw" phenomena between targets,and target gradient conflicts.In response to the shortcomings of the above in-depth recommendation algorithms,this paper proposes a solution by studying explicit and implicit feedback data,high-level feature interaction,multi-expert integration,and dependency relationships among multiple objectives,combined with a spatiotemporal hybrid multi-objective optimization method of joint focus loss,and studies improving the prediction performance of user behavior in recommendation scenarios.The work done includes the following aspects:(1)Aiming at the shortcomings of existing single target recommendation models such as low accuracy,sparse samples,sample selection bias,and "seesaw" phenomena,a multiobjective joint optimization algorithm integrating explicit and implicit feedback and multiple expert integration is proposed.The prediction of user probabilities in various behaviors is taken as the target of the model to learn,and multiple objectives are integrated into a single model for learning.Firstly,research is conducted based on the use of explicit and implicit feedback balancing,using the embedding layer to change the feature vectors,so that the underlying layers of the model share the same feature embedding,and using factor decomposition machines and multitasking learning to construct high-level feature interactions;Then,a multi expert ensemble is constructed,using a gated network constructed by fully connected neural networks and multi-level expert networks to learn the complex characteristic relationships of user behavior;Finally,a dependency relationship between targets is established to improve the accuracy of the model.Experiments on UCI census data sets have demonstrated the effectiveness of the algorithm.Experiments on real 9 million video data sets have demonstrated the effectiveness of the algorithm’s optimization.The results show that the playback target,viewing duration target,and average AUC are optimized by 0.18%,7.27%,and 3.57%compared to the single target model,respectively,and the overall AUC is improved by 1.57%compared to the commonly used multi-objective baseline model,Shared Bottom.(2)Aiming at the problem of sample imbalance and gradient conflict in multi-objective algorithm optimization,a spatiotemporal hybrid multi-objective recommendation algorithm with joint focus loss is proposed,which is improved from two aspects: loss function and gradient conflict.In the optimization of the loss function,the weight of negative sample loss is reduced by using focus loss to balance the data contribution of positive and negative samples.In terms of gradient optimization,an algorithm structure of space domain plus time domain is constructed.First,the gradient is changed by projecting each gradient onto the normal plane of another to prevent the interference component of the gradient from being applied to the network.Then,both first and second order momentum are considered in the time domain,and gradient updates are performed together in the spatial domain.In order to verify the effectiveness of the proposed optimization algorithm improvement,experiments were conducted on 10 million real recommendation system datasets,and the algorithm designed in this article was compared with previous algorithms.The experimental results showed that the proposed improved algorithm improved the model results by up to 7.28%.