They take the number of people from the same IP into account. IPs are broken down into public IPs vs private IPs based on traffic/timing of usage etc. There are research papers on this sort of feature contruction using only IPs. Cross device especially uses it extensively to be able to probabilistically ascertain if the person from your house who is checking their phone is the same person who checked smth on the Desktop computer last night based on your online timing, IPs, behaviour over the day. They can figure out, for instance, your office vs home browsing timing, interests etc with the same methods.