The Internet of Things empowers citizens to interconnect their devices, such as smart phones, into large-scale participatory decentralized networks, which they can use to make real-time collective measurements as public good, for instance, crowd-sourcing the monitoring of traffic in a city. This approach is an alternative to big data analytics systems that are often expensive to access, privacy-intrusive and allow discriminatory and profiling actions over citizens' data. On the contrary, large-scale decentralized networks are complex to manage and collective measurements, i.e. computations of aggregation functions, need to encounter several dynamics such as continuously changing input data streams and highly varying temporal demand for access to the collective measurements. This paper proposes a highly reactive self-adaptation model to tackle the challenge of dynamic computational demand in large-scale decentralized in-network aggregation. The self-adaptation process makes nodes self-aware about other nodes that join and leave the network and therefore it makes them capable of self-orchestrating the communication to improve accuracy and minimize communication cost. The model is simple, yet agile. This is shown when applied in DIAS, the Dynamic Intelligent Aggregation Service without introducing architectural changes. Evaluation using data from a real-world smart grid pilot project as well as extreme demand profiles that scale up and down the demand 50% on average confirm the cost-effectiveness of in-network aggregation empowered by self-adaptation. The findings are confirmed both in simulation and a large-scale live deployment in a cluster infrastructure with 3000 independent Java virtual machines each running a DIAS node. Overall, the results encourage new promising pathways towards the broader adoption of self-adaptive participatory data analytics in large-scale decentralized networks.