Research On Classification Algorithms For Multi-task Learning

Compared with single-task learning(STL),multi-task learning(MTL)obtains a better classifier by sharing information between tasks in a multi-task model.In multi-task learning,the target task uses the relevant experience information possessed by the training signals of multiple non-target tasks to improve the generalization effect of the model.In the training process,the data information carried by each task is information about a certain field,but there are differences between the data of each task.Most of the existing multi-task learning methods only focus on the target task data during the training process,and ignore the nontarget task data that may be included in the target task.In addition,in real life,many problematic data sets have multiple labels,most of the data used for multi-task learning model training only contains positive and negative labels,and those data that do not belong to any positive or negative labels are called Universum data.In order to further improve the performance of the multi-task model and data utilization,we add Universum data as prior knowledge to the training of the multi-task learning classifier.The main research work of this paper includes:(1)First,conduct systematic theoretical and application research on the multi-task learning algorithm,understand and master the relevant theoretical background and application process,and summarize its advantages and disadvantages.Then,understand the theoretical background of Universum learning,and understand how the prior knowledge encoded by Universum learning has a positive effect on the model.(2)In order to build a priori knowledge about the data distribution in multi-task classification,we incorporate Universum learning into multi-task learning and propose a multi-task learning method for Universum data.The proposed model uses parameters to link each task together.In order to make full use of the Universum data(used to encode the prior knowledge information of the training set),each task has a corresponding Universum data.Then,we construct a hyperplane based on the original data and Univerusm data,and obtain an accurate classifier for the Universum data located near the hyperplane.(3)Experiment with the proposed multi-task learning support vector machine for Universum data on 20 newgroups,reuters-21578,web-kb,and Landmine data sets.In the course of the experiment,text extraction methods and multiple Universum data construction methods were used to construct the data needed for the experiment.