Research On The Algorithm Of Multi-instance Learning Based On Logistic Regression Model

As the technology of electronic information and computer is consistently progressing, machine learning has been applied to many fields in real life. People utilize the technique of machine learning to enable the computer to analyze and research automatically and intelligently. As a new kind of learning framework in machine learning, multi-instance learning has received tremendous attentions since its fundamental theory has been built. By transforming the problems to the ones which can be solved in multi-instance learning framework, many problems in real life could be solved well. Nowadays, the technique of multi-instance learning has many applications, such as object detecting, image retrieval, web mining, text categorization and computer aided diagnosis.Logistic regression model is a kind of mathematical model which is commonly used in machine learning. In recent years, researchers applied it to multi-instance learning, and developed multi-instance learning algorithms based on the logistic regression model. Among these existed algorithms, however, there are generally two kinds of problems. On one hand, when the algorithm uses logistic regression model, it often doesn't accord with standard multi-instance learning assumption, which is that the bag would be positive if there is at least one instance of the bag being labeled positive; on the other hand, the objective function of the algorithm is often non-convex, which makes the optimization problem of the algorithm be limited by the local optimum.Aiming at those two problems mentioned above and by viewing the multi-instance learning research works and the standard learning assumption, a new approach of multi-instance learning based on logistic regression model and aggregate function is proposed in this paper. Firstly, a linear classifier is built in the algorithm to classify the examples and based on the logistic regression model, a new likelihood function is used to capture the relationship between the bag's label and its instances'latent ones, which accords with the standard multi-instance learning assumption well. Then, the aggregate function is used to transform the nondifferentiable objective function to a smooth concave function. As a result, the parameters of the classifier can be trained by solving a general unconstrained optimization problem, which avoids the problem of local optimum well. The proposed algorithm is tested on the standard datasets including the MUSK datasets, three image retrieval datasets, and on a text categorization problem. The experimental results show that the proposed algorithm achieves superior performance to the published multi-instance learning algorithms in the aspect of classification's accuracy.