Wireless networks are more complex than ever with an increasing configurability. Understanding all aspects of complex wireless systems and providing solutions becomes infeasible, even for domain experts. As a consequence, in wireless systems there is an ongoing trend to apply machine learning, a technique were computers learn from data to avoid the need of knowledge from experts. Machine learning is currently used in wireless networks for tasks such as performance prediction, recognition of technologies such as Wi-Fi, LTE, Bluetooth, etc. and network optimisation (e.g. localization system with heterogeneous settings). However, state of the art approaches require big datasets with annotations regarding performance and technologies of samples. This approach requires companies to spend a huge amount of time on (manual) dataset generation and annotation before any value is created, making big data learning approach too expensive for many problems encountered by Flemish networking companies. Therefore, the goal of this project is to develop techniques that enable "small data"-based wireless machine learning, either by intelligently collecting less but more relevant information or by reusing insights from past optimisations. We propose novel machine learning approaches to cope with partially annotated datasets, methods for transferring knowledge towards new technologies or environments and finally methods for efficiently identifying which labels are required for fast learning.