As video traffic increases with plentiful multimedia services and the proliferation of mobile devices such as smartphones, stream mining to extract valuable information out of multimedia big data is garnering attention. By applying cloud computing to stream mining, resource-scarce mobile devices can offload the workloads of heavy applications to a remote cloud. However, resource provisioning for task scheduling is an inherent challenge of stream mining in cloud computing. In this paper we consider problem of resource provisioning and bit rate scaling for multimedia big data processing. We aim to minimize the virtual machine (VAI) leasing cost and the classification error cost while satisfying the deadline constraints of workloads which is formulated as a mixed integer nonlinear programming. Deadline based task scheduling and hit rate scaling are developed to find near optimal solution of the NP-hard problem. The upper and lower bounds of the required number of VMs are obtained for infeasible and feasible schedules respectively. Scaling down the highest bit rate first in the bit rate set of a workload is suggested to guarantee the minimum increase of error cost. Our simulation results show the efficiency of bit rate scaling in task scheduling. 5-10% cost reduction is achieved by bit rate scaling in a cloud computing environment.