Industrial Internet of Things (IIoT) revolutionizes the future manufacturing facilities by integrating the Internet of Things technologies into industrial settings. With the deployment of massive IIoT devices, it is difficult for the wireless network to support the ubiquitous connections with diverse quality-of-service (QoS) requirements. Although machine learning is regarded as a powerful data-driven tool to optimize wireless network, how to apply machine learning to deal with the massive IIoT problems with unique characteristics remains unsolved. In this paper, we first summarize the QoS requirements of the typical massive non-critical and critical IIoT use cases. We then identify unique characteristics in the massive IIoT scenario, and the corresponding machine learning solutions with its limitations and potential research directions. We further present the existing machine learning solutions for individual layer and cross-layer problems in massive IIoT. Last but not the least, we present a case study of massive access problem based on deep neural network and deep reinforcement learning techniques, respectively, to validate the effectiveness of machine learning in massive IIoT scenario.