Multi-domain Data Classification Based On Multi-instance Multi-label Learning

In recent years,with the rapid development of big data and artificial intelligence,label learning has become one of its key research areas.Among them,multi-instance multi-label learning is a new learning paradigm that expands multi-instance learning and multi-label learning.MIML has a better representation of many complex and fuzzy objects in the real world,and the expression of the instances is more in line with the ambiguity of the object.Therefore,the MIML learning framework has become one of the key research topics in pattern recognition and label learning.Many scholars have proposed a variety of MIML classification algorithms,and have achieved great success in data analysis,such as text,image,audio and biological information.This paper will focus on the research of multi-domain data classification based on MIML framework.Through the in-depth study of MIML,the improvement of degradation strategy algorithm,classification algorithm and end-to-end classification algorithm in MIML classification algorithm are realized.The main research works are as follows:(1)At present,the degenerate MIML algorithm based on K-Medoids clustering treats each instance as independent of each other.Degradation process may cause more information loss,and K-Medoids clustering needs a priori knowledge of clustering cluster number K.The difference of K value has a great influence on the classification result.Aiming at this problem,an improved algorithm for multi-instance multi-label learning based on mean shift is proposed.The MIML is degraded by the mean shift algorithm,which is with weights and the non-parametric clustering,and the correlation between the instances is considered to minimize the loss of degradation process information.The effectiveness of the proposed algorithm is validated by the comparison experiment further.(2)In the traditional neural network algorithms,more network parameter settings are required,making it more likely to get a local optimal solution,instead of the global one.The extreme learning machine theory is an efficient algorithm to the single hidden layer feedforward neural network that efficiently obtains the global optimal solution by only setting the number of hidden layer nodes and randomly initializing the weights and offsets.However,randomly setting weights and biases are easily influenced by the random value and results in unstable calculations,and kernel ELM can solve this problem.Therefore,the MIML classification algorithm based on the regression kernel extreme learning machine as the base classifier is proposed to reduce the time consumption of classification while ensuring classification accuracy.Finally,the comparison experiments show the reliability of the proposed algorithm.(3)In recent years,with the rapid development of computer hardware,deep learning technology has been widely used in many fields.However,the current MIML classification algorithm based on convolutional neural networks mostly uses the Softmax function for classification.However,this function treats each class label as a mutually exclusive event and does not conform to the MIML learning framework.Therefore,it is proposed to replace the Softmax function in the CNN network with the ELM classifier to construct an end-to-end image classification algorithm.The experimental comparison between the proposed CNN-ELM-MIML model and the original CNN shows the rationality of the proposed algorithm.