Research On Multi-instance Multi-labe Learning Based On Feature Learning

With the rapid development of Internet and multimedia technologies, there are huge amounts of data being produced every day, which contains huge amounts of data such as images, text and so on. How to effectively use and manage these kind of data is becoming a scientific research and business problem which needs to be solved. For these huge amounts of data such as images, text and so on, they present to us is not just a single data with simple content form, moreover, in our daily life, they present to us is an ambiguous data with complicated content. How to effectively deal with this kind of ambiguity data is a difficult problem for the current scientific research. When dealing with the ambiguous data with complex content, a more effective method is to use multi-instance multi-label learning approach. In this approach, the multi-instance way is employed to show the complex content of ambiguous data, the multi-label way is employed to present the multiple labels of ambiguous data, by extracting the features of data, we can carry on the model between the features and labels so that we can recognize the ambiguous data.However, when building the algorithm model, extracting which kind of feature and how to extract the feature impact the recognition rate from the source. So in this paper, based on the existing methods used to extract low feature, the existing methods used to extract high semantic feature and the deep learning technologies, we fuse these technologies to the multi-instance multi-label learning approach to propose a general multi-instance multi-label learning framework. The main research of this paper is as follows:(1)By studying the feature learning and multi-instance multi-label learning technologies, we find the deficiency of algorithms, on the basis of exiting theory and algorithms, we fuse these technologies to the multi-instance multi-label learning approach to propose a general multi-instance multi-label learning framework, this general framework is a good way to improve the deficiency of many existing multi-instace multi-label methods.(2)Based on the proposed general framework, we propose a new method called Multi-insatce Multi-label Method Based on Topic Model. In this method, the feature learning model is probabilistic latent semantic analysis. This is a shallow feature learning model, its feature learning ability is limited, so another new method called Multi-insatce Multi-label Method Based on Convolutional Neural Network is applied, the Convolutional Neural Network is a model of Deep Learning, it has a good feature learning ability.In the multi-instance multi-label learning methods, there have been proposed many algorithms such as MIMLBOOST and MIMLSVN, they are two typical algorithms in the multi-instance multi-label learning areas, so when comparing the experimental results, we compare our proposed methods with MIMLBOOST and MIMLSVN algorithms. We have done the experiments on benchmark data sets, and the experimental results show that our proposed approaches are better than MIMLBOOST and MIMLSVN algorithms.