Aquaculture presents itself as one of the most rapidly developing means of sustainable production of animal protein to feed ever-growing populations. Recirculating aquaculture systems offer higher control and fewer inconveniences than traditional systems, making them an attractive option for fish production. Although the sector’s digitalization is in its early stages, its application should increase its rentability while conserving the environment. This paper aims to promote the sector’s evolution by assessing parameter importance in mortality with tree-based machine learning models, verifying the method’s natural robustness and how it compares to a specially devised one, and at the same time evaluating the concept’s relevance in predicting categorical mortality values. In particular, to better understand the aquaculture production process through a systematic data evaluation, an exploration based on real-time data acquisition is fully needed. Moreover, algorithm robustness is a key ingredient in this application since measurements are greatly affected by errors. This invalidates the application of traditional machine learning methods, where models are sensitive to production data variations and sensor noise. The study found the parameters that play relevant roles in the production phases, such as pH and nitrate concentration. While the obtained predictive metrics are still sub-optimal, further enhancements could be achieved through rigorous analysis of feature engineering, fine-tuning model hyperparameters, and exploring more advanced algorithms. Additionally, incorporating larger and more diverse datasets, refining data pre-processing techniques, and iteratively optimizing the model architecture may contribute to significant improvements in predictive performance. Despite that, the impact costs of using adjusted machine learning metrics are clear, as are the importance of data rounding in pre-processing and directions for improvement regarding data acquisition and transformation.

Decision Trees (DTs) remain one of the most important algorithms in machine learning for their simplicity, interpretability, and often satisfactory performance. Furthermore, they are critical foundational components for more performant models such as Random Forests (RFs) and Gradient Boosted Trees. Central to DTs is the splitting process, where data is partitioned according to criteria traditionally based on information-theoretic measures such as Shannon entropy or Gini index. In this article, we propose a novel parametric entropy-based information gain criterion designed to generalize and extend classical entropic measures to improve classification performance in DTs and RFs. We introduce a five-parameter entropy formulation capable of replicating and extending known entropy measures. This new criterion was incorporated into DT and RF classifiers and evaluated on a collection of 18 benchmarking datasets, including both synthetic and real-world data retrieved from publicly available repositories. Performance was assessed using 5-fold cross-validation and optimized via Bayesian hyperparameter search, with weighted F1-score as the primary metric. Compared to splitting criteria based on existing entropy/purity measures ( e.g ., Gini, Shannon, Rényi, and Tsallis), our method yielded statistically significant improvements in classification performance across most datasets. On multiclass and imbalanced datasets, such as the Wine Quality dataset, F1-score improvements exceeded 40% using RF algorithms. Bayesian signed-rank tests confirmed the robustness of our method, which never underperformed relative to standard approaches. The proposed entropy-based splitting criterion offers a flexible and effective alternative to classical information-theoretic measures, delivering improvements in classification performance.

Aquaculture is the world’s fastest-growing food production sector, yet it lags behind other industries in adopting upcoming digital technologies. Challenges, such as integrating multimodal data and maintaining reliable network connectivity, have hindered the development of digital twins for monitoring aquaculture systems. This paper addresses these challenges through two main contributions: (i) a novel edge-based architecture for digital twinning that enables distributed, localized monitoring and actuation, reducing dependence on centralized systems and robust networks; and (ii) a three-stage algorithmic approach for mortality monitoring tailored to edge computing environments. This approach enables early detection of rising mortality rates using data fused from diverse sources, including directly monitored environmental parameters (e.g. pH and temperature), and novel optical biosensors that make use of lightweight computer vision and machine learning techniques for the estimation of bacterial concentrations within edge devices. The algorithmic strategy was tested in a real-world recirculating aquaculture system for Solea senegalensis , where bacterial concentration was estimated with an F1-score of 0.83 across five concentration levels using biosensor imagery. Moreover, a multimodal drift detection algorithm successfully identified abnormal data trends aligned with significant changes in input distributions, with preemptive drift signals preceding critical 7-day mortality spikes.

Decision trees are a cornerstone of interpretable machine learning and are widely used for their simplicity and effectiveness in classification tasks. To address the growing need for models that can operate on continuous, unbounded data, decision trees have been reinvented for the data stream setting, where they must learn incrementally under constraints such as limited memory, evolving distributions, and delayed supervision. A critical component of these tree-based models, particularly those based on the Hoeffding Trees, is the split criterion, which determines how the input space is partitioned. This study introduces a new split criterion for stream-based Hoeffding trees, based on a unified five-parameter entropic formulation that generalizes several well-known measures, including Shannon, Gini, Tsallis, and Rényi entropies. While such formulations have been explored in batch learning, their application to streaming scenarios has not been made. By incorporating this criterion into a variety of established streaming classifiers and evaluating performance on standard benchmark datasets, we demonstrate consistent and statistically significant improvements over existing methods, including those implemented in the River library. Notably, we report gains of up to 40% in immediate evaluation metrics, along with consistent wins and some draws on the prequential Macro-F1, with no observed losses against baseline criteria. The generality of the approach introduces additional computational overhead and also enables greater expressiveness and adaptability in handling uncertainty and nonstationary data. This work advances the integration of information-theoretic principles into online learning and highlights the importance of efficient hyperparameter tuning and adaptive entropy selection in streaming environments.

Decision trees are a cornerstone of interpretable machine learning and are widely used for their simplicity and effectiveness in classification tasks. To address the growing need for models that can operate on continuous, unbounded data, decision trees have been reinvented for the data stream setting, where they must learn incrementally under constraints such as limited memory, evolving distributions, and delayed supervision. A critical component of these tree-based models, particularly those based on the Hoeffding Trees, is the split criterion, which determines how the input space is partitioned. This study introduces a new split criterion for stream-based Hoeffding trees, based on a unified five-parameter entropic formulation that generalizes several well-known measures, including Shannon, Gini, Tsallis, and Rényi entropies. While such formulations have been explored in batch learning, their application to streaming scenarios has not been made. By incorporating this criterion into a variety of established streaming classifiers and evaluating performance on standard benchmark datasets, we demonstrate consistent and statistically significant improvements over existing methods, including those implemented in the River library. Notably, we report gains of up to 40% in immediate evaluation metrics, along with consistent wins and some draws on the prequential Macro-F1, with no observed losses against baseline criteria. The generality of the approach introduces additional computational overhead and also enables greater expressiveness and adaptability in handling uncertainty and nonstationary data. This work advances the integration of information-theoretic principles into online learning and highlights the importance of efficient hyperparameter tuning and adaptive entropy selection in streaming environments.

Conformal Prediction (CP) has emerged as a robust paradigm for trustworthy machine learning, offering mathematically guaranteed marginal error bounds. However, its practical utility is often constricted by predictive efficiency. Standard non-conformity functions typically treat data instances uniformly, neglecting the localized complexity of decision boundaries and resulting in prediction regions that fail to adapt to underlying data heteroscedasticity. To address this validity-efficiency trade-off, this study introduces the integration of the Multiscale Impurity Analysis (MIA) algorithm into the Inductive Conformal Prediction framework. This novel, model-agnostic approach dynamically scales non-conformity scores by weighing prediction uncertainty against local data separability. For classification tasks, MIA utilizes generic entropic measures to adjust prediction sets, while for regression, it employs localized variance analysis to adapt interval widths. Additionally, a Unified Conformal Evaluation Metric (UCEM) is proposed to holistically evaluate classification efficiency by balancing set inclusion, exclusivity, and empty-set abstentions. Extensive experimental validation across 26 standard benchmarking datasets (12 classification and 14 regression) demonstrates that the MIA-enhanced CP method significantly improves predictive efficiency, yielding smaller prediction sets and tighter intervals, while maintaining valid marginal coverage.