Graph neural networks are powerful on relational data, but when classes arrive one after another, they tend to overwrite earlier knowledge. The result is a recency bias: predictions skew toward the most recently seen classes, and performance on older classes collapses.

To address this, the paper introduces MAEM, a replay-free, capacity-controlled dynamic mixture of experts. Instead of storing old examples, MAEM allocates expert capacity dynamically as new classes appear, and combines expert outputs to keep past and present knowledge in balance. The design explicitly targets the recency bias that dominates class-incremental graph learning.

Because the method avoids replay, it may be more practical in settings where storing past data is costly or restricted. The authors position MAEM as a way to make incremental learning on graphs more stable without extra memory overhead.