The Disadvantages of Crowdsourcing Network Data
In a recent article, we discussed how our sensor provides visibility into networks at the cell site level — resulting in an accurate, impartial way of determining the size and density of wireless networks. We found all sorts of market-analysis and development implications in the data.
But what about crowdsourced (application-based) data, which captures information from devices operating within the network, such as active channels and device performance?
Are these two types of data collection compatible? Which is better than the other?
Aurora Insight vs. crowdsourced
To explore fusing these data sets, we compared our dataset of cell locations in Cheyenne, Wyoming with data from a crowdsourced application. Here’s what we learned:
- Aurora Insight’s network independent assessment creates a significantly different snapshot of the spectrum landscape.
- Crowdsourced datasets are comprised of individual measurements over several months, while our dataset can be compiled within a few hours.
- In the crowdsourced data, Sprint, Verizon, and Union appear to have no identified cells, or very little. However, we know this to be incorrect.
- The crowdsourced data favors LTE, compared to any other standards because users are most likely to have LTE phones, whereas our method captures all mobile standards with equal accuracy.
Pictured Above: Number of cells located per network operator
This is seen in the following chart, which shows the number of cells located per technology:
Pictured Above: Number of cells located per technology
When using crowdsourced data, it is important to remember that the operators’ network is determining what frequency band(s) to use so users will not get the full picture of the network’s true capabilities and thus has narrowed implications on the data collected by crowdsourcing.
Isolating cell tower locations is challenging. In the plot below, we compared the accuracy of cell tower location predictions from the two data sets. The crowdsourced predicted locations were less accurate, with a typical error radius of about 4.4 kilometers. Our data has a typical accuracy of 50-100 meters:
Pictured Above: Accuracy of cell tower location predictions
What does all this mean?
From these analyses we can see a significant bias in crowd-source datasets, imparted by the users, devices, and the operator’s network. Still, these are useful perspectives for understanding network behavior under real-world conditions.
For analyzing infrastructure existence, investment trends, and capacity, however, our network-independent approach shows a more accurate and holistic view of existing wireless infrastructure and the complete RF environment.