The proliferation of mobile devices, such as smartphones and Internet of Things gadgets, has resulted in the recent mobile big data era. Collecting mobile big data is unprofitable unless suitable analytics and learning methods are utilized to extract meaningful information and hidden patterns from data. Mobile devices have matured as a reliable and cheap platform for collecting data in pervasive and ubiquitous sensing systems. Specifically, mobile devices are: Sold in mass market chains, Connected to daily human activities, Supported with embedded communication and sensing modules. The Collecting mobile big data is unprofitable unless suitable analytics and learning methods are utilized to extract meaningful information and hidden patterns from data. In this paper the deep learning with Multi-instance learning (MIL) algorithm has been implemented to improve the prediction accuracy. In order to deal with large scale big data MIL based on the vector of locally aggregated descriptors representation (miVLAD) algorithm has been used. This Entire module works under the Spark-based framework with data frame techniques and MIL, this speeds up the learning of with deep models that consisting of many hidden layers and millions of parameters. We use a context-aware activity recognition application with a real-world dataset containing millions of samples to validate our framework and assess its speedup effectiveness.