By default, dynamic approaches dont offer much advantage over Static Leiden
When we restrict the refinement phase to affected communities, the speedup is good
On large graphs:
By default, DF Leiden has "ok" speedup
When we restrict the refinement phase to affected communities, the speedup goes into the sink
It appears that restricting the refinement phase actually ends up increasing the aggregation time for DF Leiden. To my understanding, this is due to poor load balancing in the aggregation phase. Our original assumption for the refinement phase was that the communities to be converted to super-vertices are roughly identical. This is not the case when the refimenement phase is restricted!
wolfram77
commented
3 months ago
The observations are as follows.
On temporal graphs:
On large graphs:
It appears that restricting the refinement phase actually ends up increasing the aggregation time for DF Leiden. To my understanding, this is due to poor load balancing in the aggregation phase. Our original assumption for the refinement phase was that the communities to be converted to super-vertices are roughly identical. This is not the case when the refimenement phase is restricted!