Multi-Instance Multi-Label Hashing

Recent years, with the rapid development of the multimedia, the datasets of images and texts are growing explosively. How to manage and retrieve this information accurately and quickly becomes a big challenge. People proposed lots of solutions to settle the above problem, including the traditional supervised learning, multi-instance learning, multi-label learning, multi-instance multi-label learning, and so on. Since these models can describe the date in the real world better, those models have got more reasonable and excellent performance in many applications.Multi-instance multi-label learning (MIML) is a powerful framework, which deals with the problem that each example is represented as multiple instances and associated with multiple class labels. Previous works mostly focus on accuracy, while scalability for large scale datasets has been rarely addressed. In this paper, we present a novel framework --Multi-instance Multi-label Hashing (MIMLH) to tackle both accuracy and scalability issues of MIML tasks, which means that it can’t only get good accuracy, but also fast learning speed. MIMLH leverages hashing technique. Specifically, it exploits the hashing approach in two perspectives-bag-level hashing and instance-level hashing, which replaces the dot-product kernel operator in the previous methods and effectively maps the entire samples into hamming space, speeding up the process of learning tremendously. Moreover, we also take the label information into account to enhance our framework.We evaluate our approach on two popular data sets of MIML task, which were derived from two real world applications scene classification and text categorization. The experimental results show that the proposed framework performs better than previous works on accuracy and efficiency in a balanced way.