kanscalibratie

Kanscalibratie, often translated as probability calibration, is a technique used in machine learning and statistics to ensure that the predicted probabilities from a model accurately reflect the true likelihood of an event. In many classification models, particularly those based on decision trees or neural networks, the raw output probabilities can be overly confident or under-confident. For example, a model might predict a class with 95% probability, but in reality, instances classified with this confidence only occur about 80% of the time. This discrepancy can be problematic when the probabilities themselves are used for decision-making, risk assessment, or are fed into downstream systems that rely on accurate probability estimates.

The goal of kanscalibratie is to adjust these raw probabilities so that they are better aligned with the observed frequencies. If a model assigns a probability of 0.7 to a set of instances, then approximately 70% of those instances should actually belong to the predicted class. Several methods exist for probability calibration, broadly categorized into histogram-based methods and model-based methods. Histogram-based approaches, such as Platt scaling and isotonic regression, often involve fitting a separate model to the output scores of the original classifier. Platt scaling uses a logistic regression model, while isotonic regression uses a non-parametric, monotonic function. Model-based methods aim to integrate the calibration process directly into the training of the primary model, although this is less common than post-hoc calibration. The effectiveness of kanscalibratie is typically evaluated using metrics like the Brier score, reliability diagrams, and log-loss.