Mobile crowdsensing (MCS), as a practical paradigm for large-scale urban sensing systems, has met a bottleneck as the limited budget isn’t able to cover the whole urban sensing area. To reduce the sensing cost, Sparse MCS, emerged as a variant of traditional MCS, only senses the data in a few subareas and then infers the data of the unsensed subareas using the spatiotemporal relationship of the sensed data. Previous works make a linear assumption and present some methods based on the compressive sensing or matrix completion, which are the top trends in Sparse MCS. However, the spatiotemporal relationship of the sensed data is actually collected in a nonlinear system, leading to the low data inference accuracy seen in prior attempts. To better understand these challenges and solutions, this chapter proposes a deep learning-enabled data inference scheme. Moreover, we continue to explore several real-world scenarios such as the “outlier data” caused by special events and the “fractured data” related to the unpredictability of a spatiotemporal fracture. This chapter summarizes some efficient methods to tackle these issues. Finally, this chapter provides a comprehensive outlook of the future research on Sparse MCS.