The rapid growth of blockchain platforms like Ethereum facilitates digital asset circulation but also enables fraud due to inherent anonymity and complex transactions. Effective transaction analysis and anomalous account detection are thus critical. However, conventional anomaly detection methods applying graph mining or machine learning to Ethereum data exhibit significant limitations. Specifically, prevalent graph embedding algorithms (e.g., Node2Vec, DeepWalk) inadequately integrate crucial transaction attributes, particularly amount and timestamp, during network construction. This oversight results in imprecise node representations, constraining detection accuracy. To address these issues, this paper introduces a novel framework for detecting anomalous Ethereum accounts. Our core innovation lies in a biased random walk strategy incorporating an adjustable parameter. This design systematically enhances the influence of transaction value and temporal dynamics on the structure learned during walks. Furthermore, we implement a dynamic weighting mechanism within the node embedding process. This mechanism adaptively balances the contributions of transaction monetary value and recency when generating the final node vectors. This combined approach significantly improves the model’s capacity to discern anomalous transaction patterns. Extensive experimental validation demonstrates the framework’s effectiveness, achieving a high fraud detection precision of 92.7% and enabling the accurate identification of fraudulent accounts on the Ethereum network.