Public articles linked to the same research event.
arXiv The authors propose MAGIC, a replay-free graph few-shot class-incremental learning framework that pairs a frozen graph representation backbone with closed-form analytic continual learning, learns a topological prior from the base graph characterizing both homophilous and heterophilous relations and injects it through Potts Markov random field inference to ease novel-class overfitting, and transfers previous predictions of affected historical nodes to their updated representations via drift-aware analytic distillation; across five datasets and eight baselines, the 5-shot setting improves Mean Accuracy and Final Accuracy by 5.48 and 9.33 percentage points on average and reduces Performance Drop by 10.78 points, with substantially less training time.
The authors propose MAGIC, a replay-free graph few-shot class-incremental learning framework that pairs a frozen graph representation backbone with closed-form analytic continual learning, learns a topological prior from the base graph characterizing both homophilous and heterophilous relations and injects it through Potts Markov random field inference to ease novel-class overfitting, and transfers previous predictions of affected historical nodes to their updated representations via drift-aware analytic distillation; across five datasets and eight baselines, the 5-shot setting improves Mean Accuracy and Final Accuracy by 5.48 and 9.33 percentage points on average and reduces Performance Drop by 10.78 points, with substantially less training time.
The authors propose MAGIC, a replay-free graph few-shot class-incremental learning framework that pairs a frozen graph representation backbone with closed-form analytic continual learning, learns a topological prior from the base graph characterizing both homophilous and heterophilous relations and injects it through Potts Markov random field inference to ease novel-class overfitting, and transfers previous predictions of affected historical nodes to their updated representations via drift-aware analytic distillation; across five datasets and eight baselines, the 5-shot setting improves Mean Accuracy and Final Accuracy by 5.48 and 9.33 percentage points on average and reduces Performance Drop by 10.78 points, with substantially less training time.
The authors propose MAGIC, a replay-free graph few-shot class-incremental learning framework that pairs a frozen graph representation backbone with closed-form analytic continual learning, learns a topological prior from the base graph characterizing both homophilous and heterophilous relations and injects it through Potts Markov random field inference to ease novel-class overfitting, and transfers previous predictions of affected historical nodes to their updated representations via drift-aware analytic distillation; across five datasets and eight baselines, the 5-shot setting improves Mean Accuracy and Final Accuracy by 5.48 and 9.33 percentage points on average and reduces Performance Drop by 10.78 points, with substantially less training time.