Today, as more and more industries are involved in the artificial intelligence era, all business enterprises constantly explore innovative ways to expand their outreach and fulfill the high requirements from customers, with the purpose of gaining a competitive advantage in the marketplace. However, the success of a business highly relies on its IT service. Value-creating activities of a business cannot be accomplished without solid and continuous delivery of IT services especially in the increasingly intricate and specialized world. Driven by both the growing complexity of IT environments and rapidly changing business needs, service providers are urgently seeking intelligent data mining and machine learning techniques to build a cognitive ``brain" in IT service management, capable of automatically understanding, reasoning and learning from operational data collected from human engineers and virtual engineers during the IT service maintenance. The ultimate goal of IT service management optimization is to maximize the automation of IT routine procedures such as problem detection, determination, and resolution. However, to fully automate the entire IT routine procedure is still a challenging task without any human intervention. In the real IT system, both the step-wise resolution descriptions and scripted resolutions are often logged with their corresponding problematic incidents, which typically contain abundant valuable human domain knowledge. Hence, modeling, gathering and utilizing the domain knowledge from IT system maintenance logs act as an extremely crucial role in IT service management optimization. To optimize the IT service management from the perspective of intelligent data mining techniques, three research directions are identified and considered to be greatly helpful for automatic service management: (1) efficiently extract and organize the domain knowledge from IT system maintenance logs; (2) online collect and update the existing domain knowledge by interactively recommending the possible resolutions; (3) automatically discover the latent relation among scripted resolutions and intelligently suggest proper scripted resolutions for IT problems. My dissertation addresses these challenges mentioned above by designing and implementing a set of intelligent data-driven solutions including (1) constructing the domain knowledge base for problem resolution inference; (2) online recommending resolution in light of the explicit hierarchical resolution categories provided by domain experts; and (3) interactively recommending resolution with the latent resolution relations learned through a collaborative filtering model.