There is a system with $S$ transmitters/service points, $U$ subscribers/users. The link quality/gain between any transmitter indexed by $s$ and a user indexed with $u$ is denoted by $h_{u,s}$. Any transmitter can either serve the user or act an an interferer if it is not serving.
Now, we have a predefined number of groups/clusters, defined as $G$.
For a given cluster, all the transmitters in the cluster serve all the users in that cluster, but act an interferer to the users in other clusters.
What will be an optimal or near optimal/heauristic clustering solution for this problem?
The objective is to optimise the summation of service quality of all the users. The service quality of a given user is defined by the ratio of serving signal to the interfering signal. Serving signal is the summation of signals from the transmitters in user's own group and interfering signal is the summation of signals from the transmitters in other groups.
An MI(N)LP formulation should be avoided as this will be very difficult to solve for large system. And I have a large system.
I support Rob Pratt's linearized MIP formulation (his answer to the question posted on OR SE). That said, the problem is definitely amenable to heuristic solutions. Since there are no limits on the number of users in a group, and since the objective function is additive (meaning the quality of service for one user is unaffected by the assignment of other users), for a given clustering of transmitters you can just compute the service quality for each individual user and each group and assign the user to the group that maximizes service quality. That puts the focus on clustering the transmitters.
Now suppose you can cook up a measure of the interference between transmitters in different groups. If the issue here is radio interference, that should be pretty straightforward, and presumably would be based on some combination of distance between transmitters, signal power and maybe frequency assignment. Let's call that measure $\delta_{s,s'}$ for transmitters $s < s'$. You can cluster transmitters so as to minimize the sum of the $\delta$ values for pairs of transmitters in different clusters, or to maximize the sum of the $\delta$ values for pairs in the same cluster (under the assumption that being in the same cluster will mitigate the interference), or some other measure. There are a variety of clustering algorithms in the literature, or you can do it using a MIP model (which will be smaller than Rob's model since it does not include the users). Once you have a clustering of the transmitters, assign each user to that user's best cluster.
Update: I put this to the test with a small test problem (100 users, 20 servers, 4 clusters), using a random key genetic algorithm. The GA does not find the optimal solution. Whether the solution is good enough, and whether it scales acceptably, are left to the reader as an exercise. :-) I coded it in R. The full description and code are in an R notebook.