Rapid development in IT and revolutionary monitoring/measuring technologies have made it possible to collect large scale, sequential data in many practical fields. It has been a challenge to efficiently analyze such data to discover hidden patterns, association/correlation, and trend of changes.
In machine learning and data mining, cost-sensitive learning and sparse online learning are two important research areas. Many algorithms have been proposed for these learnings separately. Generally, an algorithm performs well in one field may be less good in another field. Very few work has been published combining these two fields together.
By formulating a new convex optimization problem, the framework intends to balance misclassification cost and sparsity, two mutual restraint factors. We will present the theoretical analysis on the bounds of the regret of actions and cost, and the comparison to those of the existing methods. Evaluated on eight real-life streaming, high-dimensional, severely-skewed datasets, the proposed method outperforms other traditional ones.