Graph Domain Incremental Learning (GDIL) aims to acquire knowledge from a continuous stream of graph domains while mitigating catastrophic forgetting. While parameter-isolation methods leveraging graph parameter-efficient adaptation show promise, prompt-based techniques struggle to adapt to GDIL, and low-rank adaptation methods based on a shared classification layer lead to knowledge confusion.Our empirical observations reveal that transferable knowledge is primarily concentrated in the representation layer. Further, we argue that domain-agnostic representations that are not tied to the classification characteristics are needed to assist the new model in capturing more discriminative features for graph domain incremental learning.Motivated by these insights, we propose COllabOrative Knowledge Extraction and integRation (COOKER) method for GDIL to mine inter-domain relationships and uncover the potential of domain-agnostic representations. Specifically, COOKER employs domain-specific LoRA modules and classifiers to capture specific knowledge. A domain-agnostic LoRA module is instantiated to extract transferable knowledge through contrastive acquisition and topology alignment. We introduce collaborative dynamic integration of dual representations to enable adaptive integration, guided by a complementarity loss to eliminate information redundancy. Extensive experiments demonstrate that COOKER significantly outperforms existing baselines, achieving up to a 4.7% improvement in average performance.
Jialu Li, Yu Wang, Wanyu Lin et al.· Proceedings of the 32nd ACM...· 0 citations
Class-incremental learning is commonly instantiated as a single-model paradigm, where a unified model sequentially adapts to an unbounded stream of sessions. While effective under mild distributional shifts, this formulation becomes strained when successive sessions induce incompatible optimization directions, leading to destructive interference and catastrophic forgetting. We argue that such forgetting reflects a structural limitation of enforcing heterogeneous learning dynamics within a single parameter space. Motivated by social solidarity theory, we propose Socialized Division and Collaboration (SDC) as a reformulation of continual learning that decomposes session learning across specialized models in response to optimization conflicts, while enabling coordinated collaboration. To support this formulation with a principled allocation mechanism, we introduce an energy-based session-model compatibility criterion grounded in Helmholtz free energy, which guides adaptive session allocation and model evolution under conflicting objectives. This framework integrates session assignment, model evolution, and collaborative inference into a unified pipeline, offering an alternative to monolithic continual learning formulations and highlighting a broader design principle for learning under persistent optimization conflicts.
Xinjie Yao, Zhihe Fan, Yunqi Zhu et al.· 0 citations
Bridging Literature, Agents, and Zero-gap Experimentation (BLAZE), a paradigm of socialized scientific intelligence, makes discovery more traceable, reproducible, and cumulative while preserving human creativity, judgment, and responsibility.
Xinjie Yao, Xingxin Xu, Xiyuan Gao et al.· 0 citations
