Controlled Dynamics Attractor Transformer for Graph Anomaly Detection

Transformer architectures have significantly advanced representation learning through self-attention. Concurrently, associative memory frameworks offer interpretable retrieval by mapping representations onto energy landscapes. However, the continuous-time inference dynamics of these frameworks often lack biological plausibility compared to classical Continuous Attractor Neural Networks (CANNs). To bridge this gap, a new model called Controlled Dynamics Attractor Transformer (CDAT) has been proposed. CDAT combines a mixture von Mises-Fisher (Mo-vMF) attention energy with a Hopfield refinement energy. It further enhances energy descent with an excitation-inhibition modulation inspired by CANNs. This design results in a topology-constrained dynamical system where couplings encode relational structures among tokens, linking attractor-style dynamics to modern energy-based attention. The robust and structured dynamics of CDAT have led to state-of-the-art performance in benchmarks for graph anomaly detection and graph classification.