Graph Adversarial Domain Adaptation For Non-shared-and-Imbalanced Transfer Learning Via Hierarchy Graph Reasoning

Unsupervised domain adaptation technology(UDA)is widely used to(1)use the rich label data from the relevant source domain to deal with the scarcity of the label data of the target domain;(2)reduce the cost of training downstream models,so that the model can be used on other networks Reuse.Existing deep UDA solutions often ignore or automatically delete non-shared big data to achieve the optimal migration effect between small data distributions under the same shared category,and cannot meet the needs of fast and effective learning and massive complex data in the context of the big data era.This subject researches the key technology of small application task migration learning for non-shared big data.It uses the food big data restaurant to identify small task migration scenarios as the entry point to model the process of big data and small task migration learning,and is committed to achieving big data at the same time.Automatic cleaning,non-sharing and unbalanced transfer learning.The main work involved in this article can be summarized as follows(1)Aiming at the problem of high-noise automatic cleaning and denoising,this article uses multi-label classification and convolutional neural networks to build an Auto-Clean model to establish a big data automatic cleaning mechanism;to ensure that the large data automatic cleaning module,multiple The task Auto-Clean classification model and the highconcurrency web deployment are designed for the architecture;the functional requirements analysis,the use of the Caffe deep framework training model,the use of Tornado to build the Restful API for the food image classification service are described in detail;after many tests,the multi-task Auto-The top 1 accuracy rate of the clean classification model reached96.3%,the top 5 accuracy rate reached 99.5%,and the QPS in the system pressure test could reach 43.6,meeting high concurrency requirements;finally,the log was analyzed and a high concurrency improvement plan was proposed.(2)Aiming at the problem of migration from the laboratory to the real environment and the problem of unbalanced food data,formalize the definition of non-shared and unbalanced migration learning tasks;innovatively propose a general graph adversarial domain adaptation framework based on hierarchical graph reasoning GADA,the goal is to use prior-level knowledge to enhance domain confrontational alignment feature representation with graph reasoning.Hierarchical graph reasoning(HGR)is added to the adversarial domain adaptive method,and direct semantic patterns are learned through hierarchical attention,nonlinear mapping,and graph normalization.The HGR layer aggregates local features into hierarchical graph nodes through node prediction.And through the hierarchical graph reasoning of all source classes,the domain adversarial alignment feature is enhanced.As a general graph adversarial domain adaptation model,GADA can be embedded in all adversarial-based domain adaptation algorithms,and it is easy to extend most UDA schemes and greatly improve the migration effect.Under the dual-domain unbalanced setting,there are 8 transfer tasks in the Meal-300 data set,and 6 transfer tasks in the Office-Home data set.The GADA(HGR)proposed in this paper is based on the adversarial domain adaptation algorithm(such as CDAN,MDD and GVB-GD)have been greatly improved,which proves the efficiency of the proposed algorithm model and the powerful reasoning ability of the hierarchical graph.