Mittejuhitud

Mittejuhitud is an Estonian term used in machine learning to describe unsupervised learning—a paradigm in which models learn patterns from unlabeled data without explicit supervisory signals. It stands in contrast to juhitud learning (supervised learning), where models learn mappings from inputs to known outputs. In mittejuhitud learning, the objective is to uncover the intrinsic structure, regularities, or representations present in the data rather than to predict a predefined target.

Typical tasks in mittejuhitud learning include clustering, dimensionality reduction, density estimation, and anomaly detection. Clustering algorithms such as k-means and DBSCAN group similar data points; dimensionality reduction methods like PCA, t-SNE, and UMAP reveal compact representations of high-dimensional data. Generative models and autoencoders learn latent representations or data distributions without labeled outcomes, enabling tasks such as compression or feature extraction.

Evaluation in mittejuhitud learning is often indirect because there are no ground-truth labels. Metrics may include silhouette scores or Davies-Bouldin indices for clustering, reconstruction error for autoencoders, or intrinsic measures of data efficiency and separability. The lack of explicit targets makes model selection and hyperparameter tuning more challenging and sometimes subjective, requiring task-specific considerations.

Applications span market research, image and text representation learning, anomaly detection, and pretraining features for downstream, supervised tasks. Mittejuhitud learning is frequently contrasted with semi-supervised learning, which uses small amounts of labeled data to guide the learning process, and with supervised learning, which relies on labeled outcomes.