Scholarlink

<p>Managing blood product supply and demand is challenging, especially in developing countries, where balancing shortages and wastage remains uncertain. Historical blood demand data often suffer from non-stationarity, irregular patterns, missing values, noise, and outliers, complicating accurate time series prediction. While various approaches exist for such tasks, they are limited in handling non-stationary series with sharp fluctuations. This study proposes a novel hybrid prediction framework designed for precise time series forecasting, offering a robust solution to the mentioned blood demand challenges. The proposed framework comprises three main phases. In the first phase, we decompose irregular, non-stationary signals into meaningful modes and a residual term, mitigating mode mixing and isolating transient fluctuations to yield reliable forecasts while avoiding overfitting to noise. These components are modeled via a high-order fuzzy cognitive mapping approach with robust search-based tuning, capturing multi-lag dependencies under data scarcity. In the second phase, we capture broader, slower-varying trends via a model that balances sparsity and stability and adapts to changing scales and non-stationary behavior. Parameter tuning employs regularized learning and adaptive search to ensure resilience across regimes. In the third phase, we use an ensemble approach (gradient boosting regression) to iteratively combine forecasts from the first two phases by correcting residual errors, leveraging complementary strengths to reduce bias and variance and enhance robustness to anomalies, thereby producing a more accurate final prediction. Performance evaluation of the proposed hybrid framework is conducted on three blood demand cases and eight diverse real-life datasets sourced from various applications. Furthermore, we compare our framework with sixteen state-of-the-art algorithms using the eight datasets. The experimental results demonstrate the superior accuracy and effectiveness of our proposed framework. Notably, the framework achieves an average RMSE improvement of approximately 67% across all evaluated cases, underscoring its predictive strength. This hybrid framework holds significant promise in tackling the challenges posed by non-stationary blood demand data and has broader implications for managing blood product supply and demand in developing countries.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid ensemble learning framework for blood demand forecasting under non-stationary and uncertain conditions

  • 期刊论文

  • 发表日期：

    2025年8月11日

查看全文
  • Ali Nikseresht,
  • Mostafa Zandieh,
  • Davood Golmohammadi

摘要

Managing blood product supply and demand is challenging, especially in developing countries, where balancing shortages and wastage remains uncertain. Historical blood demand data often suffer from non-stationarity, irregular patterns, missing values, noise, and outliers, complicating accurate time series prediction. While various approaches exist for such tasks, they are limited in handling non-stationary series with sharp fluctuations. This study proposes a novel hybrid prediction framework designed for precise time series forecasting, offering a robust solution to the mentioned blood demand challenges. The proposed framework comprises three main phases. In the first phase, we decompose irregular, non-stationary signals into meaningful modes and a residual term, mitigating mode mixing and isolating transient fluctuations to yield reliable forecasts while avoiding overfitting to noise. These components are modeled via a high-order fuzzy cognitive mapping approach with robust search-based tuning, capturing multi-lag dependencies under data scarcity. In the second phase, we capture broader, slower-varying trends via a model that balances sparsity and stability and adapts to changing scales and non-stationary behavior. Parameter tuning employs regularized learning and adaptive search to ensure resilience across regimes. In the third phase, we use an ensemble approach (gradient boosting regression) to iteratively combine forecasts from the first two phases by correcting residual errors, leveraging complementary strengths to reduce bias and variance and enhance robustness to anomalies, thereby producing a more accurate final prediction. Performance evaluation of the proposed hybrid framework is conducted on three blood demand cases and eight diverse real-life datasets sourced from various applications. Furthermore, we compare our framework with sixteen state-of-the-art algorithms using the eight datasets. The experimental results demonstrate the superior accuracy and effectiveness of our proposed framework. Notably, the framework achieves an average RMSE improvement of approximately 67% across all evaluated cases, underscoring its predictive strength. This hybrid framework holds significant promise in tackling the challenges posed by non-stationary blood demand data and has broader implications for managing blood product supply and demand in developing countries.