TrAdaBoost Transfer Learning Based On Particle Swarm Optimization

With the development of technology, it has become a hotspot to obtain information from the mass of data in machine learning. In practical applications, human labels on the huge amounts of data will consume a large amount of manpower and resources.Thus, limiting the number of labeled samples and the development of machine learning. Therefore, transfer learning using a few labeled samples and large quantity labeled samples in related domain to study can overcome the shortcomings of traditional machine learning, and it has become a hot research in machine learning.The differences between classifiers are an important factor affecting the integration of classifiers. Therefore, differences between classifiers have also impacted on the performance of transfer learning algorithm. PSO is used to optimize the feature weight of samples. Based on the optimal feature weight, random feature subspace is constructed, which is applied into the training process of transfer learning. Thus, the differences between classifiers are improved and the performance of transfer learning is advanced.To solve the problem that the convergence of source weights is too quick in Tr Ada Boost, an approach called Adaptive Tr Ada Boost Based on Particle Swarm Optimization is proposed. Using the source domain samples help target domain samples training classifier in random feature subspace, by means of adding the correction factor to update the weight of source domain samples. There may lead to negative transfer if the distribution between source domain and target domain is too large in transfer learning. So, the paper proposes a Weighted Multi-source Tr Ada Boost Based on Particle Swarm Optimization. Using PSO to select the random feature subspace of different source domain and target domain. In the random feature subspace, the method is to assign a weight to each source domain according to its classification accuracy on auxiliary training, then, different source domain have different influence on the target task, this method can also weaken the negative impact of unrelated sample. The task learning performance can be effectively improved using labeled samples of multiple sources areas. Comparison experiment on datasets to verify the feasibility and validity of the proposed methods.