Scientists and professionals are increasingly interested in the Internet of Things (IoT) as they help to solve numerous problems across various fields such as transportation, agriculture, healthcare, and many others. The IoT enables the connect of physical objects through short-range communication technologies like Wi-Fi and ZigBee, as well as long-range technologies such as Low Power Wide Area Network (LPWAN), including LoRa, Sigfox, and NB-IoT. These networks are characterized by low power consumption and reduced cost. Machine learning (ML), a branch of artificial intelligence (AI) that relies on algorithms capable of analyzing data for prediction and decision-making, offers solutions for optimizing and managing LPWANs. These solutions help to overcome the limitations of traditional methods. This study takes an in-depth look at the integration of machine learning (ML) techniques into LPWANs, with the aim of increasing their performance and flexibility in dynamic environments. It reviews existing research combining these two fields, identifies the main challenges facing LPWANs, and proposes ML-based solutions to overcome them. Thus, this work offers a detailed analysis of this synergy, presents a network architecture integrating ML algorithms in different levels and highlights the research questions still open, providing avenues for future contributions.