In this talk, we will explore the power and flexibility of ant colony decision forests, a family of hybrid, nature-inspired learning methods designed to tackle complex and dynamic recognition tasks. The collective intelligence paradigm of ant colonies can be used to allow these decision forests to adapt to evolving data distributions while maintaining high levels of accuracy and robustness. We will discuss the various adaptation mechanisms such as pheromone-driven search and incremental model updating that enable ant colony decision forests to overcome challenges common in real-world data, such as class imbalance, noise, and nonstationary environments. One of the key aspects of this approach is the use of preference-driven measures to guide both the construction and evaluation of the models. By incorporating expert domain knowledge and user-defined preferences, these metrics enable more interpretable results and facilitate fine-tuning performance criteria beyond traditional accuracy-based measures. This perspective is particularly valuable in applications where costs, risks, and contextual constraints play a significant role in decision-making, ranging from anomaly detection in sensor networks to personalised recommendation systems. Drawing on a series of studies and practical implementations from my previous work, I will illustrate how the combination of biologically inspired algorithms with preference-driven evaluation provides an effective framework for building resilient and adaptive computer recognition systems. Real-world examples will demonstrate the effectiveness of these methodologies in coping with diverse data sources and varying operational conditions.

A predictive model is developed to assist medical professionals in making clinical decisions in Intensive Care Units (ICUs). This model is designed to (a) improve the early prediction of mortality, (b) support more efficient decision-making for high-risk patients, and (c) assess the effectiveness of new treatments or identify shifts in clinical practices. The model is a hierarchical machine learning framework based on Bayesian classifiers, constructed using data from a real-world ICU cohort. It evaluates mortality risk and predicts either the patient's discharge destination if they survive or the cause of death if they do not. The model, called Ensemble Weighted Average (EWA), combines five base Bayesian classifiers into a weighted ensemble using the average ensemble criterion. EWA's performance is benchmarked against other advanced machine learning predictive models. The results demonstrate that EWA surpasses its competitors and offers an additional advantage over majority-vote-based ensembles by providing confidence levels for its predictions. Furthermore, we demonstrate the benefit of locally recalibrating the standard APACHE II-based mortality risk model with new data, even though it proves to be less predictive than the other models.