Adaptive learning from data flow and imbalanced data.

摘要

This dissertation presents studies on adaptive machine learning algorithms deliberately designed to address the issues of learning from data flow and data set of imbalanced class distribution. It is of particular importance to the machine learning community due to the prevalence of these problems in many fields of modern business world, such as medical diagnosis, online fraud detection, etc.;Unlike conventional machine learning situations, data flow becomes continuously available over time. Under such scenario, the challenge is how to transform the vast amount of stream raw data into information and knowledge representations, and accumulate the experience over time to support future decision-making processes. In light of this, this dissertation introduces an ADAptive INcremental learning (ADAIN) framework. Starting with description of system level architecture, this research explores the design of mapping function for knowledge transfer, theoretically studies the error bound, and concludes the discussion by presenting simulation results over a video clip as well as other real-world data sets.;Traditional machine learning techniques generally fail to achieve satisfying performance for imbalanced data due to the assumption of balanced class distribution. Based on the adaptive over-sampling technique and boosting algorithm, this dissertation presents the study of RAnked Minority Over-Sampling in Boosting (RAMOBoost) algorithm to alleviate the difficulty brought by such a problem. Integrated with AdaBoost learning framework, RAMOBoost adaptively ranks minority class examples by proportioning their chance of being sampled for creation of synthetic instance to the ratio of majority class cases in their k-nearest neighbors at each iteration. In this way, the decision boundary could be progressively shifted towards difficult-to-learn minority class and majority class examples simultaneously. The dissertation covers algorithmic description of RAMOBoost and extensive simulations for validation of its competitiveness over existing methods.;This dissertation also explores a Recursive Ensemble Approach (REA) for learning from nonstationary data flow with imbalanced class distribution. REA pushes a selected part of previous minority class examples into the current data chunk to balance its class distribution, upon which a classifier is built and added into an ensemble for future prediction. Theoretical and empirical studies are both implemented to compare REA with other methods.