كاهش بازآموزي مدل پيش‌بيني قيمت بيت‌كوين با استفاده از سازوكار شناسايي رانش مفهوم

Prediction in highly dynamic markets such as cryptocurrencies is confronted with the challenge of concept drift due to their non-stationary nature an‎d severe volatility; a phenomenon that causes the relationship between variables to change over time an‎d leads to performance degradation in static models. Common approaches like periodic an‎d blind retraining suffer from high computational an‎d temporal costs an‎d inability to precisely detect drift timing, thus lacking necessary efficiency. In contrast, intelligent drift detection an‎d adaptation methods, utilizing supervised mechanisms an‎d strategies such as meta-learning, allow preserving previous knowledge an‎d restoring accuracy with minimal cost. The present research, aiming to fill these gaps, introduces the “event-driven adaptation architecture” which confines the retraining process exclusively to the moment of actual drift detection. This architecture is based on two key components: first, a “concept drift detection algorithm” specifically designed for Bitcoin that separates real drifts from noise by managing repetitive al‎e‎rts; an‎d second, a “meta-learning-based selec‎tor system” that dynamically selec‎ts the optimal base prediction model in the adaptation phase according to instantaneous market conditions. eva‎luation of this architecture on a synthetic Bitcoin dataset (generated using Wasserstein Generative Adversarial Network) showed that the concept drift detection algorithm identified all applied drifts in the synthetic dataset. Moreover, in eva‎luation on real Bitcoin dataset across 3 periods of 180 days, this algorithm achieved accuracy equivalent to periodic methods with an average of 2.7 retraining’s (versus 5 in the periodic strategy). The selec‎tor system also succeeded in choosing the optimal model in the adaptation phase with 64.8% accuracy. This research, by presenting an intelligent an‎d automated system, highlights the importance of immediate action after drift an‎d the necessity of state-aware architectures for managing model lifecycle in the Bitcoin market.

5

154143