Research On Shared Parameters Of Multi-task Recommendation Ranking Algorithm

With the information explosion,recommender system emerges as the times require.The core algorithm part is divided into three modules: recall module,sorting module and rearrangement module.The most important module here is sorting module.In sorting,the most frequently concerned indicators are click through rate(CTR)and conversion rate(CVR),Compared with CTR,CVR has the problems of data sparsity and sample selection bias.Data sparsity means that the scale of CVR data is much smaller than CTR data in training.Sample selection bias means that the training space of CVR is only a subspace of prediction space,and the training sample has the problem of selection bias.At present,there are two kinds of multi task learning methods in recommendation domain: task relation learning and parameter sharing.Task relational learning,such as the enter space multitask model(ESMM),can solve the problem of sample selection bias,but it can not solve the problem of data sparsity because of its lack of sufficient hidden layer sharing.Parameter sharing methods such as multi gate mixture of experts(MMOE)do not optimize the sample sparsity and sample selection bias of CVR due to their universal design characteristics.Therefore,the research contents of this paper are as follows:(1)In order to solve the problem that the existing parameter sharing model does not optimize CVR enough,this thesis proposes a parameter sharing model for CVR,which is multi-level asymmetric gate control(ML-AGC).ML-AGC adopts multi-layer sharing structure,asymmetric gate control(AGC)on task side and asymmetric customized gate control(ACGC)on feature side.The structure of AGC is a combination of hard parameter sharing and soft parameter sharing,which combines asymmetric sharing with MMOE.ACGC can be regarded as a compromise between asymmetry and consistency.Because CTR and CVR have the same feature space,ML-AGC uses multi-layer sharing to fit the nonlinearity.At the same time,because of the data sparsity of CVR to CTR,ML-AGC uses asymmetric parameter sharing to realize the asymmetry of information sharing.The asymmetry of task side and the identity of feature side make ML-AGC use different results for different levels To achieve this transition.(2)In view of the advantages and disadvantages of task relation learning and parameter sharing,this paper proposes a new CVR multi task learning model,namely,enter space asymmetry gate control(ES-AGC).The hidden layer of ES-AGC adopts ML-AGC parameter sharing,and the task relationship adopts ESMM structure,which solves the problem of data sparsity and sample selection deviation.ES-AGC also adds residual structure to CVR side to solve the problem of deep CVR network degradation caused by CVR data sparsity.ES-AGC regards each feature as an expert,and uses a feature sharing layer based on self attention mechanism to share features.