In today's business marketplace, many high-tech Internet enterprises constantly explore innovative ways to provide optimal online user experiences for gaining competitive advantages. The great needs of developing intelligent interactive recommendation systems are indicated, which could sequentially suggest users the most proper items by accurately predicting their preferences, while receiving the up-to-date feedback to refine the recommendation results, continuously. Multi-armed bandit algorithms, which have been widely applied into various online systems, are quite capable of delivering such efficient recommendation services. However, few existing bandit models are able to adapt to new changes introduced by the modern recommender systems.