Understanding how patterns of interaction between neural populations reorganize with behavior is central to systems neuroscience and to applications that decode or modulate brain activity. These interactions are often framed under the umbrella of brain connectivity, distinguished as structural (anatomical pathways), functional (statistical dependencies), and effective (directed influence)--complementary lenses on how distributed circuits communicate and adapt to task demands [1, 2]. Reviews emphasize both the promise and the challenges of connectivity-centric approaches, particularly the need to capture short-timescale dynamics while maintaining interpretability [3, 4]. Graph-based formulations make this agenda concrete: representing brain regions as nodes and their relationships as edges enables quantitative analysis of modularity, hubs, and task-dependent reconfiguration [5, 6]. Such network perspectives motivate methods that go beyond static, correlation-only descriptions toward dynamic, directed estimates that better align with mechanistic questions. We focus on intracranial local field potentials (LFPs), a high-temporal-resolution measure of population activity recorded simultaneously from multiple deep-brain regions [7, 8, 9, 10]. Multi-region LFP imposes unique modeling requirements: (i) many micro-contacts per region must be consolidated into coherent region-level trajectories; (ii) neural activity reconfigures across well-defined behavioral phases; and (iii) the spatial layout of contacts is high-dimensional and non-Euclidean [11]. Existing pipelines often average across behavior with a single correlation-based graph [2] or impose hard-coded anatomical connectivity [12, 13], limiting their ability to capture phase-specific, directed interactions. Recent machine learning advances in graph learning from neural time series highlight the opportunity but do not yet fill this gap.

Class-Incremental Learning (CIL) is a practical and challenging problem for achieving general artificial intelligence. Recently, Pre-Trained Models (PTMs) have led to breakthroughs in both visual and natural language processing tasks. Despite recent studies showing PTMs' potential ability to learn sequentially, a plethora of work indicates the necessity of alleviating the catastrophic forgetting of PTMs. Through a pilot study and a causal analysis of CIL, we reveal that the crux lies in the imbalanced causal effects between new and old data. Specifically, the new data encourage models to adapt to new classes while hindering the adaptation of old classes. Similarly, the old data encourages models to adapt to old classes while hindering the adaptation of new classes. In other words, the adaptation process between new and old classes conflicts from the causal perspective. To alleviate this problem, we propose Balancing the Causal Effects (BaCE) in CIL. Concretely, BaCE proposes two objectives for building causal paths from both new and old data to the prediction of new and classes, respectively. In this way, the model is encouraged to adapt to all classes with causal effects from both new and old data and thus alleviates the causal imbalance problem. We conduct extensive experiments on continual image classification, continual text classification, and continual named entity recognition. Empirical results show that BaCE outperforms a series of CIL methods on different tasks and settings.