Research On Several Machine Learning Methods And Their Applications In Fingerprint Segmentation And Cross Selling

With the development of information technology, especially the development of data acquisition and data storage techniques, people have the ability to acquire and store volumes of data at anytime and anywhere. However, the value of data is very limited, if we cannot abstract knowledge from data. And we will feel confused faced with volumes of data without the help of an effective analytical and mining tool. Luckily, machine learning and data mining techniques provide us the tool. Machine learning is defined as systems promote itself via experience. In the past 30 years, machine learning has achieved great development and successful application in several domains that has not been seen before. The thesis conducts a deep research on several machine learning methods with fingerprint segmentation in biometrics and cross selling in financial domains as backgrounds.Fingerprint segmentation is an important preprocessing step in fingerprint identification. The aim of fingerprint segmentation is to separate fingerprint foreground with texture features from background of acquired image. Generally speaking, traditional fingerprint segmentation methods are supervised trained with manually segmented fingerprints or designed by experts, so they are labor consuming. With the application scenarios of biometrics as authentication techniques becoming more and more, Fingerprint identification meets sensor interoperability. Different from traditional fingerprint segmentation methods which make use of common information of lot of fingerprints when training, the thesis proposed personalized fingerprint segmentation based on semi-supervised learning which is only trained on input fingerprint image itself. The new fingerprint segmentation model has better sensor interoperability, is free of manual fingerprint segmentation, and improves the automatic degree of automatic fingerprint identification system.Besides, the thesis conducts a systematic research on class imbalance problem and its application in cross-selling of financial domain. The research recognized several domains where class imbalance exists, reviewed academic activities on class imbalance learning, illustrated performance evaluation metrics for class imbalance learning methods, and showed an introduction to class imbalance learning methods from four levels. Then, it took PAKDD2007 data mining competition as a case study, analyzed several challenges in the task, and proposed an ensemble framework EnSVM to solve the cross-selling problem, which can predict potential cross-selling customers, and provide support to make decision for managers.